{ "metadata": { "venue": "ICLR.cc/2027/Conference", "active_venue_id": "ICLR.cc/2027/Conference/Submission", "source_url": "https://openreview.net/group?id=ICLR.cc/2027/Conference#tab-active-submissions", "api_url": "https://api2.openreview.net", "checked_at": "2026-10-09T05:47:09Z", "fetched_at": "2026-10-09T05:19:57.757Z", "status": "complete", "total_active_submissions": 42368, "candidate_count": 994, "automated_candidate_count": 1133, "scope_excluded_count": 139, "reviewed_count": 6, "abstract_reviewed_count": 12, "source_provenance": { "origin": "official_openreview_browser_session", "api_url": "https://api2.openreview.net", "transport": "official_Webfield2_api_get", "started_at": "2026-10-09T05:15:40.614Z", "page_count": 43, "public_fields_only": true, "final_count_verified": true }, "source_sha256": "5f1a2de3f9e06fccad0078949f756c116915103d433f31f39333f3c1a1a7f023", "method": "经验证的完整公开投稿 JSON 导出；标题、摘要和关键词检索；按官网摘要排除纯表示、tokenizer、codec 等范围外记录；分类为初筛标签。", "message": "已完整获取本次可公开访问的 active submissions；候选需逐篇核验。", "scope": "quantization", "official_pdf_receipt": "iclr-2027-official-pdf-receipt.json", "fulltext_reviewed_at": "2026-10-09T05:47:34Z" }, "categories": [ { "id": "ptq", "label": "训练后量化", "description": "校准、舍入、旋转与重建等训练后量化方法。" }, { "id": "qat", "label": "量化感知训练", "description": "将量化误差引入训练或微调过程。" }, { "id": "training", "label": "低精度训练", "description": "低比特数值格式与训练稳定性。" }, { "id": "kv-cache", "label": "KV Cache", "description": "注意力缓存与长上下文推理中的量化。" }, { "id": "hardware", "label": "硬件与系统", "description": "内核、部署和真实运行效率。" }, { "id": "reasoning", "label": "推理能力", "description": "量化与推理、数学和代码能力的关系。" }, { "id": "diffusion", "label": "扩散模型", "description": "图像、视频与其他扩散模型的量化。" }, { "id": "multimodal", "label": "多模态", "description": "视觉语言及其他多模态模型的量化。" }, { "id": "theory", "label": "理论与分析", "description": "量化误差、界、鲁棒性和机理分析。" } ], "papers": [ { "id": "Oou4LAMTNI", "number": 23, "title": "HiLoRe: What to Store, Compress, or Recompute for Efficient GRPO Training", "abstract": "Group-relative policy optimization (GRPO) makes learner-side activations a major memory-computation bottleneck: gradient checkpointing reduces activation memory through recomputation, but fixed schedules can leave roughly 18 GB unused on a 48-GB GPU despite substantial recomputation overhead. Existing activation-management methods set state fidelity from execution cost, tensor properties, or generic compression sensitivity, without explicitly incorporating GRPO's analytic update structure into state-fidelity allocation. We formalize this dependence as policy-update exposure, linking the current GRPO loss coefficients to state-level approximation sensitivity. These coefficients are available before backward without an additional backward pass. We introduce **HiLoRe**, which allocates graph-attributed recovery units among $\\underline{\\textbf{hi}}$gh-precision storage, $\\underline{\\textbf{lo}}$w-precision compression, and deterministic $\\underline{\\textbf{re}}$computation using measured recovery utility and update-conditioned approximation risk. It combines high-precision storage and deterministic recomputation with low-precision recovery under a calibrated risk budget. Across five model-task settings with 2K responses and memory $\\le1.10\\times$ GC's per-GPU actor-update peak, HiLoRe's actor-update throughput gains reach 13.5\\% over GC and 7.9\\% over the fastest evaluated baseline, with paired mean downstream-score differences below 0.6 percentage points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "GRPO", "Efficient Training" ], "forum_url": "https://openreview.net/forum?id=Oou4LAMTNI", "pdf_url": "https://openreview.net/pdf/f2a15cf5d927de7b6e3ee04dc92616ca06e7efa4.pdf", "updated_at": "2026-10-05T16:44:46.320Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nSsaEp75xm", "number": 40, "title": "EmbTracker: Traceable Black-box Watermarking for Federated Language Models", "abstract": "Federated language model (FedLM) training enables collaborative learning without sharing raw data. However, every client receives valuable model instances that may be redistributed without authorization, raising important intellectual property (IP) protection challenges.\nMost existing federated watermarking methods require white-box access or client-side cooperation, and provide only group-level ownership evidence rather than client-level attribution. Moreover, many methods are designed for image classification tasks and have limited utility for language models. We propose EmbTracker, a server-side framework for traceable black-box watermarking tailored for FedLMs. EmbTracker embeds a backdoor-based watermark that can be verified through API queries and assigns each client a distinct registered trigger through token-embedding replacement, avoiding per-client watermark retraining. Experiments on language and vision-language models across classification, question-answering and visual question-answering tasks show verification rates near 100\\% in most settings, primary-task changes typically within 1--2 percentage points, and resilience to fine-tuning, pruning, quantization, noise, and watermark overwriting under the evaluated conditions.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Watermark", "Federated Learning", "Language Model", "Black-box", "Traceable" ], "forum_url": "https://openreview.net/forum?id=nSsaEp75xm", "pdf_url": "https://openreview.net/pdf/6c5c354e420cf0ded9de901f54615d1c23eabb99.pdf", "updated_at": "2026-10-05T16:44:49.058Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mhr6MQbs2E", "number": 48, "title": "CheeSe : Channel Permutation for Semi-structured KV Cache Pruning", "abstract": "The KV cache is the memory-bandwidth bottleneck in long-context LLM inference.\nElement-wise unstructured pruning preserves quality best but obtains no compute acceleration, and its storage overhead keeps the effective compression ratio below the nominal one; structured methods accelerate computation but lose accuracy quickly as compression grows.\n2:4 pruning offers both acceleration on sparse tensor cores and a high effective compression ratio, yet prior work rejected it for the KV cache because of its quality loss, and subsequent work accelerated it with dedicated kernels while leaving that loss in place.\nWe propose CheeSe, which recovers this loss with two techniques.\nFirst, we show that the loss arises almost entirely on the key cache and stems from how channels are grouped rather than from which elements are removed: removing the same number of key elements costs $-0.25$ under unstructured pruning but $-2.14$ under 2:4, because important channels cluster within the same group.\nA sorting-based channel permutation, computed once during calibration and transferable across tasks, recovers most of this loss, whereas more elaborate permutation searches from the weight pruning literature cost up to four orders of magnitude more compute and bring no further gain.\nSecond, output-aware selection, which is harmful for 2:4 alone, becomes beneficial once the permutation is in place.\nAcross five models, CheeSe stays within 0.21 points of unstructured pruning and recovers more of the 2:4 loss than keeping sink tokens and the recent context dense, at no cost in compression.\nIt improves the effective compression ratio by 9.2 percentage points and accelerates decode attention by $1.80\\times$ with an existing 2:4 kernel at a runtime overhead of 2.5%, and it composes with token eviction and quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "2:4 sparsity", "semi-structured sparsity", "N:M sparsity", "channel permutation", "efficient LLM inference" ], "forum_url": "https://openreview.net/forum?id=mhr6MQbs2E", "pdf_url": "https://openreview.net/pdf/2b4305e032923a4bf0e72d3b5314af16825d878b.pdf", "updated_at": "2026-10-05T16:44:51.513Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CzIctGOUNV", "number": 49, "title": "PICO: Lossless Compression Stacked on Pruning and Quantization for Efficient LLM Inference", "abstract": "Large Language Models (LLMs) are memory-bound during inference, where memory access dominates latency and energy consumption, making the combination of pruning and quantization essential for efficient deployment. Unstructured pruning stays near dense accuracy up to about 50% sparsity without retraining, whereas structured pruning degrades sharply at such sparsity without fine-tuning, but storing\nthe irregular positions of the remaining weights limits its memory savings. Previous XOR-based compression regularizes such sparse weights into fixed-size blocks, yet at this sparsity it generates excessive correction patches that offset its gains,\nand shift-register-based variants reduce patches at the cost of substantial hardware and encoding overhead. We propose PICO (Patch-minimized Inverted XOR COmpression), a lossless, training-free scheme that stacks on existing pruning and quantization pipelines. PICO combines a hardware- and encoding-efficient XOR map with grouped inverted-XOR encoding, which selects the best map per group\nand skips patch metadata for mismatch-free groups. PICO applies regardless of unstructured pruning criterion or quantization format while preserving the regular, fixed-size access patterns required for parallel hardware decoding. Across LLaMA2/3.1, Phi-4, and Qwen2.5/3, PICO delivers 11.9% additional memory reduction at 50% sparsity with the smallest XOR decoder, 31% less area and 30% less power\nthan the strongest prior scheme, while encoding 5.9× faster, and its sequential variant PICO-S doubles the reduction of that scheme.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Compression", "Lossless", "Unstructured Pruning" ], "forum_url": "https://openreview.net/forum?id=CzIctGOUNV", "pdf_url": "https://openreview.net/pdf/19b5e057c4f37a54d2024c4b3f23e452e85b59d6.pdf", "updated_at": "2026-10-05T16:44:51.627Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PTL1u5Zn05", "number": 279, "title": "Fine-Tuning Leaves Clues: Calibration-Free Quantization via Task-Aware Basis", "abstract": "Fine-tuning foundation models has become the standard way to create domain-specific expert models. \nWhile calibration-free quantization offers an efficient deployment solution without requiring downstream data, existing methods often suffer significant performance drop at low bit-widths. \nThese methods typically neglect the task-specific information encoded in the task vector, which is defined as the parameter difference between the fine-tuned and pre-trained weights.\nIn this work, we propose **T**ask-aware bas**I**s to **G**uid**E** calibration-f**R**ee **Q**uantization (TIGER-Q). \nTIGER-Q extracts a task-aware basis from the task vector to serve as a data-free proxy for the Hessian matrix. This basis establishes a task-sensitive subspace that guides the quantization process to minimize error along dimensions critical to the target task. \nTo efficiently solve this objective, we introduce a decoupled continuous-discrete refinement pipeline.\nIt computes closed-form analytical solutions for continuous quantization scales, while employing a local search for discrete weights, coupled with a row-wise acceptance mechanism to maintain monotonic error reduction.\nExtensive experiments show that TIGER-Q achieves superior performance over state-of-the-art calibration-free methods on both language and vision models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "calibration-free quantization", "PTQ" ], "forum_url": "https://openreview.net/forum?id=PTL1u5Zn05", "pdf_url": "https://openreview.net/pdf/1616d385018eeccd1aa675fc88368c9399b280ba.pdf", "updated_at": "2026-10-05T16:54:45.216Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PCAODQ61Qw", "number": 326, "title": "BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank Allocation", "abstract": "Long-context inference makes the key-value (KV) cache a major memory bottleneck for large language models. We propose BELT, a low-rank KV cache compression framework. Its key-side objective targets attention distributions rather than raw keys or pre-softmax logits, yielding a quadratic surrogate with a closed-form low-rank solution. A learned global rank allocator uses differentiable soft selection during training and hard selection at deployment to meet a user-specified component budget. Experiments demonstrate improved model quality under aggressive compression and increased decoding throughput at comparable accuracy. Evaluated across three LLMs and multiple benchmarks, BELT achieves up to 70\\% low-rank KV cache compression while outperforming the strongest baseline by up to 23.3 points on LongBench, and up to $17.8\\times$ overall KV cache reduction when combined with quantization. With Triton-based GPU kernels and a quantized cache, BELT delivers up to $2.7\\times$ speedup for the attention module and $3.3\\times$ end-to-end decoding throughput.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-Cache", "Low-Rank Compression", "Large Language Model" ], "forum_url": "https://openreview.net/forum?id=PCAODQ61Qw", "pdf_url": "https://openreview.net/pdf/02259c117b980a90009875c7b4490599bd57e38c.pdf", "updated_at": "2026-10-05T16:54:52.887Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "x9qK9ehqNi", "number": 352, "title": "Coordinated Mixed-Precision KV Cache Quantization for Efficient LLM-based Multi-agent Systems", "abstract": "LLM-based multi-agent systems solve complex tasks through problem decomposition, specialized tool use, and iterative verification. However, collaboration produces multiple growing key-value (KV) caches that can coexist in memory, making cache storage a bottleneck. KV-cache quantization reduces this footprint, but independently applying single-model policies ignores differences in agents' precision needs. This can over-compress sensitive computations while spending precision on more tolerant caches, potentially degrading collective reasoning. In this paper, we introduce **Co**ordinated **Mix**ed-Precision **KV** Cache Quantization (**CoMixKV**), a novel method that integrates online sensitivity estimation with joint bit-width allocation under a shared memory budget. To our knowledge, **CoMixKV** is the first mixed-precision KV-cache quantization framework to jointly allocate precision across simultaneously resident agent caches under such a budget. An online estimator derives each agent's sensitivity from a local attention-error bound using recent queries, keys, and values, without calibration data, trial quantization, or additional model passes. A shared controller combines these scores with remaining quantizable capacity and byte-level storage costs to minimize total estimated error. Allocations are updated as agents start or finish and fresh observations become available. Selected widths apply only to newly quantized blocks; recent states retain native precision and previously quantized blocks remain unchanged. Across four benchmarks and three LLM backbones at average 2-bit, 3-bit, and 4-bit KV-cache settings, **CoMixKV** achieves strong accuracy compared with the evaluated quantization baselines. At the average 3-bit setting on Gemma-4-31B-it, **CoMixKV** achieves 68.35\\% average accuracy, outperforming the strongest quantized baseline by 0.59\\%. On Qwen3-32B, **CoMixKV** further reduces peak KV-cache memory by 32.08\\% and achieves a 1.38$\\times$ inference speedup over FullKV.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM-based Multi-agent Systems", "KV Cache Quantization" ], "forum_url": "https://openreview.net/forum?id=x9qK9ehqNi", "pdf_url": "https://openreview.net/pdf/052a827dced059360905162c2a620234a931c1a2.pdf", "updated_at": "2026-10-05T16:54:59.748Z", "categories": [ "ptq", "kv-cache", "reasoning", "theory" ], "matched_terms": [ "2-bit", "3-bit", "4-bit", "bit-width", "mixed-precision", "quantizable", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LxwIDZer16", "number": 353, "title": "When Cleaner Caches Break: Training Recipes Shape MLA Cache Quantizability", "abstract": "Quantizing the latent cache of multi-head latent attention (MLA) changes both\nreconstruction error and final predictive quality. We study how these outcomes\ndepend on the trained model and on which cache coordinates share a quantization\nscale. Our evaluation uses three production models and six matched AdamW/Muon\ncheckpoint pairs across three sizes. A streaming temporal policy preserves causality\nby quantizing groups only when they close. Under streaming INT3, Muon has\ngreater loss damage in every pair despite cleaner marginal statistics, while retaining\nbetter final quality. Under tokenwise INT3, Muon has worse final quality at 125M\nand 350M, but better quality at 600M. At INT4 Muon keeps better final quality\nin every pair under both policies, so the reversal is specific to the seven-level\nsymmetric INT3 reference. The activation measurements motivate a geometry-\nspecific interpretation: channel outliers can enlarge neighboring channels’ steps\nunder tokenwise grouping, but not under per-channel temporal grouping. This\ngrouping argument does not order the recipes: Muon has the lower outlier ratios\nyet the larger tokenwise damage at 125M and 350M. Controls that hold the closure\nschedule fixed while regrouping channels, and that tile the latent width evenly,\npreserve key recipe-dependent differences; an asymmetric INT3 quantizer, which\nstores one extra FP16 value per group, reduces Muon’s streaming damage by\n51–77% across the six checkpoints. Achieved fit remains a possible mediator of the\nsensitivity differences. Token-local noise probes, full-prompt MMLU, FP8 controls,\nand independent engine tests through 32K complement the causal comparisons.\nWe report all matched seeds and distinguish compression damage from final quality\nat the tested training budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multi-Head Latent Attention", "MLA", "KV Cache", "Latent Cache Quantization", "KV Cache Quantization", "Low-Bit Quantization", "Post-Training Quantization", "LLM Inference", "Cache Compression", "Muon", "AdamW", "Training Recipe", "Quantization Robustness", "Representation Sensitivity" ], "forum_url": "https://openreview.net/forum?id=LxwIDZer16", "pdf_url": "https://openreview.net/pdf/c27d97726623a59d0d414ece79d7a7ee7dc3daa5.pdf", "updated_at": "2026-10-05T16:55:00.367Z", "categories": [ "ptq", "training", "kv-cache", "theory" ], "matched_terms": [ "fp8", "int3", "int4", "low-bit", "quantizability", "quantization", "quantizer", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ps7rUvLZGb", "number": 430, "title": "MetricKV: Decoded-Space Geometry for Visual KV Compression in Streaming Video Understanding", "abstract": "Streaming video models observe visual evidence before future queries are known, making compact storage of historical visual key--value (KV) states important for long-running inference. We introduce METRICKV, a query-agnostic visual-KV codec derived from the decoded-KV reconstruction objective. For any positive-definite quadratic decoded-space distortion and full-column-rank decoder, decoded error decomposes exactly into a continuous projection term and an additional finite-rate latent error under a decoder-induced pullback metric. This metric determines the same ordering over finite-rate candidates as the decoded-KV objective; absorbing it into the latent coordinates yields a canonical Euclidean coding space in which standard nearest-neighbor quantization is exactly decoded-error aligned. Within this canonical family, minimizing the coder-independent continuous projection term reduces to PCA in the distortion-transformed KV space. We instantiate the framework with Layer-RMS reconstruction geometry and progressive residual vector quantization (RVQ). The experimental results demonstrate that a single progressive RVQ stream achieves $64\\times$ dynamic-KV compression while retaining more than 98% of FP16 performance on both OVO-Bench and StreamingBench. Code will be released soon.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Streaming Video Understanding; KV Compression; Decoded-Space Geometry" ], "forum_url": "https://openreview.net/forum?id=Ps7rUvLZGb", "pdf_url": "https://openreview.net/pdf/b583fcacd8d97fa97f3383921f976d4ea6f36184.pdf", "updated_at": "2026-10-05T16:55:08.984Z", "categories": [ "kv-cache", "multimodal", "theory" ], "matched_terms": [ "quantization" ], "review_status": "abstract", "summary": "MetricKV 在未来查询尚未确定时压缩历史视觉 KV，以解码后的 KV 重建误差作为目标。它把 decoder-induced pullback metric 吸收到潜坐标，使最近邻量化对齐解码误差，再在变换后的 KV 空间做 PCA，并实例化为 Layer-RMS 几何与渐进残差 VQ。", "contribution": "摘要报告一条渐进 RVQ 流实现 64 倍 dynamic-KV 压缩，并在 OVO-Bench 和 StreamingBench 都保留超过 98% 的 FP16 性能。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。64 倍针对 dynamic-KV 而非整模型显存；理论分解要求正定二次失真和满列秩 decoder，摘要未报告端到端速度。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "流式视频理解中的历史视觉 KV 状态及 progressive RVQ", "bit_width": "摘要未注明", "evidence": [ { "location": "官方摘要", "claim": "单一渐进 RVQ 流实现 64 倍 dynamic-KV 压缩。" }, { "location": "官方摘要", "claim": "OVO-Bench、StreamingBench 均保留超过 98% FP16 性能。" } ] }, { "id": "6HZCBDqoCa", "number": 583, "title": "DisCoMesh: Efficient and Flexible Artist-Mesh Generation via Sparse Computation powered Discrete Diffusion", "abstract": "Meshes used in games and interactive 3D applications require not only accurate geometry, but also clean topology, fine details, and editable semantic components. However, existing generative methods struggle to produce such assets within a practical computational budget. To address this challenge, we introduce DisCoMesh, a framework based on the view that a multi-part artist mesh is better represented as a set of parts coupled by sparse global relationships rather than as a single global sequence or an ordered sequence of parts. Each semantic part is canonicalized and quantized independently using the full coordinate vocabulary, preserving local geometric detail across parts of different scales. A shared discrete diffusion model then denoises all parts simultaneously, removing the need for an artificial part-generation order. To coordinate simultaneous part generation, we introduce a global interaction module that integrates part placement and global geometric context with sparse token-level communication across parts, preserving coherent connections and spatial relationships. This part-wise formulation enables flexible control over parts, supporting selective part regeneration while preserving the rest of the mesh. Experiments demonstrate that DisCoMesh outperforms existing methods in generation quality, maintains competitive inference speed, and supports diverse downstream applications.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "3D Generation", "Artist Mesh Generation" ], "forum_url": "https://openreview.net/forum?id=6HZCBDqoCa", "pdf_url": "https://openreview.net/pdf/4d005002bfc19e78c930e5162379d328493bc394.pdf", "updated_at": "2026-10-05T16:55:38.237Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zYUaM9Xm0l", "number": 798, "title": "Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers", "abstract": "Reinforcement Learning with Verifiable Rewards (RLVR) replaces costly human labeling with automated verifiers. To reduce verifier hacking, many RLVR systems binarize rewards to {0,1}, but imperfect verifiers inevitably introduce false negatives (FN, rejecting correct answers) and false positives (FP, accepting incorrect ones). We formalize verifier unreliability as a stochastic reward channel with asymmetric noise rates $\\rho_0$ and $\\rho_1$---the FP rate and the FN rate, respectively. From this abstraction we derive two lightweight corrections: (i) a backward correction that yields an unbiased surrogate reward and thus an unbiased policy-gradient estimator in expectation, and (ii) a forward correction that reweights score-function terms so the expected update aligns with the clean gradient direction and requires only the FN rate. We implement both as lightweight hooks in a group relative policy optimization pipeline, both corrections improve RLVR reasoning under synthetic and real verifier noise, with the forward variant being more stable under heavier noise. We further use a lightweight appeals mechanism to estimate FN rates online and show that the correction can complement stronger verifier-side baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Reinforcement Learning", "Reasoning" ], "forum_url": "https://openreview.net/forum?id=zYUaM9Xm0l", "pdf_url": "https://openreview.net/pdf/d9ec61659c7f2275451a0ec9542f384ea7708005.pdf", "updated_at": "2026-10-05T16:56:08.586Z", "categories": [ "training", "reasoning" ], "matched_terms": [ "binarize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vSqsA99XH4", "number": 801, "title": "Quantization-Aware Pre-Training with Constrained Empirical Weight Distribution", "abstract": "Quantization-Aware Pre-Training (QAPT) can increase the inference efficiency of DNNs, but a problematic behaviour known as rounding boundary weight oscillation can introduce detrimental noise into the training process and significantly reduce convergence speed. While existing methods can reduce this detrimental noise, they either introduce additional hyperparameters or memory overhead, or cannot consistently improve model accuracy. In this work, we propose optimization with **C**onstrained **E**mpirical **W**eight dis**T**ribution (CEWT), the first hyperparameter-free memory-overhead-free oscillation suppression method that consistently improves QAPT performance: an optimizer post-update step that projects weights to the nearest point in weight space whose empirical distribution (histogram) matches a zero-mean Gaussian. Our key insight is many quantizers are designed with the implicit assumption that the to-be-quantized data are permutations of samples from a zero-mean Gaussian, and this assumption is not true during QAPT. By enforcing the zero-mean Gaussian prior as a hard constraint, CEWT can suppress this detrimental noise. Empirical results on various combinations of SOTA quantizers and hypersphere optimizers suggest, that with a geomean increase of 4\\% in training time, CEWT can consistently reduce the pre-training perplexity (by an average of 2.5 and up to 21 points) of low-precision (down to 1-bit activations and weights and up to 610M parameters) LLaMA/GPT models without introducing any hyperparameters or storage overhead.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "QAT", "Efficient ML", "Edge Inference", "QAPT" ], "forum_url": "https://openreview.net/forum?id=vSqsA99XH4", "pdf_url": "https://openreview.net/pdf/04d5f921359928febabbbd921b61ba83341d1d9f.pdf", "updated_at": "2026-10-05T16:56:09.076Z", "categories": [ "ptq", "qat", "training", "theory" ], "matched_terms": [ "1-bit", "low-precision", "qat", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MdyFYfABEK", "number": 960, "title": "How Evaluation Choices Affect MAUVE Scores and Generator Rankings", "abstract": "Since its introduction, \\mauve{} has been widely adopted for evaluating open-ended text generation, serving both to compare language models and to design and select decoding strategies. Because it compares generated and human-written text in a learned feature space, its scores depend on evaluation choices such as the representation model, the evaluated text length, and the quantization settings. These choices are known to shift scores, but less is known about whether they also change which generation model is preferred. We study this question across seven generation models, three datasets, and 180 evaluation settings, separating variation due to evaluation choices from variation due to decoding configurations. For identical generations, evaluation choices produce a wider score spread than decoding changes, and they alter model rankings even after aggregating over decoding strategies: two settings that differ only in the representation model select the same top-ranked model in 30\\% of comparisons. Among the choices examined, the representation model has the largest effect on ranking stability, and it also affects how closely \\mauve{} rankings agree with human preference judgments. These findings suggest that interpreting a MAUVE comparison requires considering its evaluation configuration and the sensitivity of model rankings to that configuration, and show that human evaluation can identify representation models whose rankings align more closely with human preferences, providing practical guidance for choosing among them.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "evaluation; open-ended text generation" ], "forum_url": "https://openreview.net/forum?id=MdyFYfABEK", "pdf_url": "https://openreview.net/pdf/99c9705425d2e3685409fbc45fe7693dcd6414a6.pdf", "updated_at": "2026-10-05T16:56:35.385Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "k9j0YWUVI2", "number": 1013, "title": "AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM Generation", "abstract": "Fine-grained mixed-precision quantization promises efficient large language model inference, but local precision choices can conflict with regular GPU storage and computation units. This precision-boundary mismatch limits the translation of compression into practical acceleration. We introduce AlignQuant, a post-training quantization method that uses GPU-compatible two-dimensional weight tiles as the common unit of precision allocation, compact storage, and execution. This shared partition lets precision follow sensitivity within output channels. Joint prefill/decode calibration scores precision reductions using projection-output perturbations weighted by language-model loss gradients under quantized activations. Phase-normalized scores prioritize higher precision for tiles important to either phase under a model-wide weight-storage budget. Each tile stores one selected representation, while phase-specialized kernels reuse the packed model and expand lower-bit weights for INT8 computation with 8-bit activations. Across four LLMs spanning 3B to 14B parameters, AlignQuant achieves up to $2.50\\times$ generation speedup over BF16 while preserving model quality. Evaluations further cover three GPUs and contexts up to 64K tokens. These results show that local precision flexibility and regular GPU execution can coexist through a shared tile unit. The implementation is available at https://anonymous.4open.science/r/AlignQuant-0B64.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixed-precision quantization", "post-training quantization", "LLM inference", "GPU acceleration", "hardware-aware quantization" ], "forum_url": "https://openreview.net/forum?id=k9j0YWUVI2", "pdf_url": "https://openreview.net/pdf/0452d204fac5ea80756e39f35f91ae31e7c188a7.pdf", "updated_at": "2026-10-05T16:56:46.556Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "8-bit", "int8", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JNo1goQAdN", "number": 1122, "title": "GRIK: Grouped Relevance-Integrated Key Pruning for Grouped-Query Attention LLMs", "abstract": "The key--value (KV) cache is a major memory bottleneck in long-context inference, and existing K-channel pruning does not shrink it under grouped-query attention (GQA): in the evaluated implementations, ThinK materializes K and V per query head ($3.20\\times$ total KV) and SparK reconstructs a full-width cache ($1.00\\times$).\nWe identify two structural conditions, a mask shared across the query heads of each KV group and across all cached tokens, under which channel pruning yields a dense shared-key cache whose width equals the retained channel count.\nWe introduce GRIK, a training- and calibration-free method that satisfies both conditions by construction and derives a prompt-adaptive mask per KV head from output-projection influence, key-projection norms, and attention energy.\nAt 40\\% K-channel pruning, GRIK stores K at $0.60\\times$ and total K+V at $0.80\\times$ the bf16 baseline.\nWithout token eviction, its LongBench average is within $0.23$ points of Full KV on three GQA backbones; with H2O eviction it is the best non-Full-KV method in 19 of 24 (model, budget, ratio) cells, and at approximately matched total-KV memory it improves over H2O alone by $1.55$--$1.62$ points.\nThe reduction composes with other axes: adding GRIK to SnapKV, FP8, or int4 KV changes LongBench by at most $0.21$ points while cutting total KV to as low as $0.20\\times$.\nOn RULER, its eight exact-retrieval tasks average at least $99.0\\%$ through 32K, although the 13-task average trails Full KV by $6.90$--$8.59$ points.\nIn a fixed-memory H100 serving test at 50\\% pruning, the smaller cache admits $1.33\\times$ more concurrent sequences and raises aggregate request throughput by $1.10\\times$, with per-user decode latency $1.35\\times$ that of the baseline at the higher concurrency.\nCode and configurations are available at \\url{https://anonymous.4open.science/r/GRIK/}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV Cache Compression" ], "forum_url": "https://openreview.net/forum?id=JNo1goQAdN", "pdf_url": "https://openreview.net/pdf/693ef514979e782de5ec4cfe68f4dfa7ab2ebb4b.pdf", "updated_at": "2026-10-05T16:57:10.429Z", "categories": [ "ptq", "training", "kv-cache", "hardware" ], "matched_terms": [ "fp8", "int4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1ihfTCOu91", "number": 1550, "title": "Adamas: Hadamard Sparse Attention for Efficient Long-context Inference", "abstract": "Large language models (LLMs) now support context windows of hundreds of thousands to millions of tokens, enabling applications such as long-document summarization, large-scale code synthesis, multi-document question answering and persistent multi-turn dialogue. However, such extended contexts exacerbate the quadratic cost of self-attention, leading to severe latency in autoregressive decoding. Existing sparse attention methods reduce these costs by attending to only a subset of key-value (KV) tokens, but fixed or coarse-grained selection strategies can miss important tokens or retain substantial redundancy under constrained token budgets. We introduce **Adamas**, an accurate and efficient token-level sparse attention mechanism for long-context inference. Adamas applies the Hadamard transform, bucketization and 2-bit compression to produce compact representations, and uses lightweight distance-based estimation for efficient top-$k$ candidate selection. The selected positions are then used to gather the corresponding original KV states for sparse attention. Comprehensive evaluations show that Adamas maintains near-lossless performance within the effective context length of Llama-3.1-8B-Instruct. In efficiency evaluations, Adamas achieves up to $6.5\\times$ self-attention speedup and $1.7\\times$ end-to-end decoding speedup at a 32K context length. Adamas also maintains performance comparable to full attention on challenging long-form reasoning benchmarks, including AIME 2024 and AIME 2025. Code is publicly available at [Anonymous Adamas Repo](https://anonymous.4open.science/r/Adamas-36EA).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Sparse Attention", "Long-context Inference" ], "forum_url": "https://openreview.net/forum?id=1ihfTCOu91", "pdf_url": "https://openreview.net/pdf/ea4294bec8217db915cb00744429b0c4d6ab9e82.pdf", "updated_at": "2026-10-05T16:58:24.247Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "2-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Nh6KWK7YfF", "number": 1579, "title": "Event Generative Compression", "abstract": "Event cameras offer high temporal resolution and dynamic range. However, in dynamic scenes, they can generate event streams with data rates more than two orders of magnitude higher than conventional video, making transmission a major bottleneck for remote perception. While existing approaches primarily aim to preserve the event stream itself, we argue that when the downstream consumer only needs visual content, compression should instead preserve information useful for visual reconstruction and perception, rather than events themselves. We propose Event Generative Compression (EGC), which encodes each temporal window into a compact set of discrete symbols and uses a pretrained video diffusion model to recover the missing visual information. EGC learns an event-to-generation interface through latent alignment, discrete quantization, entropy coding, and generative adaptation, while reusing generated history for continuous reconstruction without transmitting reference frames. On DSEC and BS-ERGB, EGC achieves compression ratios of approximately 5,000--8,000 relative to losslessly coded event streams while retaining information for both temporally coherent video reconstruction and downstream tasks. The same low-rate messages support semantic segmentation, optical flow, and depth estimation, demonstrating the effectiveness of EGC in preserving useful visual information under ultra-low bitrate compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Event Camera", "Video Compression" ], "forum_url": "https://openreview.net/forum?id=Nh6KWK7YfF", "pdf_url": "https://openreview.net/pdf/d233ca8ba8ca57e463e69a99cce78a0e177efa91.pdf", "updated_at": "2026-10-05T16:58:33.062Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9MUpbkeP2Q", "number": 1642, "title": "PRISM: Post-hoc Recognition of Image Source Models", "abstract": "Image generative models (GMs) now produce photorealistic content that is virtually indistinguishable from natural images. This creates a critical need for reliable provenance to distinguish between synthetic vs authentic images and attribute generated content to its source. While diffusion models (DMs) have been the dominant paradigm for visual content generation, recent image autoregressive models (IARs) match DMs' generation quality and substantially improve the sampling speed by adapting the next-token prediction paradigm from large language models. We propose a unified post-hoc attribution method that identifies the source GM of a given image by exploiting model-specific artifacts (fingerprints) that arise in both DMs and IARs. Our key insight is to combine the fingerprints from the two main constituent parts of GMs: (1) the autoencoder that maps between pixel and latent spaces and (2) the latent generator that defines the image distribution in DMs and produces visual tokens in IARs. In the first stage, we identify the autoencoder family using reconstruction and quantization errors. In the second stage, we attribute the image to the latent generator within the identified model family using conditional log-likelihoods for DMs and token log-probabilities for IARs. Across diverse model families and versions, our two-stage framework enables accurate model attribution, providing a unified approach to the provenance of synthetic images.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model attribution", "image autoregressive model", "diffusion model" ], "forum_url": "https://openreview.net/forum?id=9MUpbkeP2Q", "pdf_url": "https://openreview.net/pdf/56d7bea10fc58397e7b124b3140859d1bbacc21e.pdf", "updated_at": "2026-10-05T16:58:46.693Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IF3P9Idbdf", "number": 1847, "title": "DASH-KV: Accelerating Long-Context LLM Inference via Asymmetric KV Cache Hashing", "abstract": "The increasing memory footprint of the KV cache and the high computational cost of query-key similarity computation \nconstitute two fundamental bottlenecks for large language models in \nlong-context inference. While existing KV cache compression methods primarily alleviate the former, \nthey often sacrifice generation quality and leave the latter largely unaddressed. This paper introduces \nDASH-KV, an innovative acceleration framework that reformulates attention as approximate nearest neighbor search via \nasymmetric deep hashing. Under this paradigm, we design an asymmetric encoding architecture that differentially maps \nqueries and keys to account for their distinctions in precision and reuse characteristics. To balance efficiency\nand accuracy, we further introduce a dynamic mixed-precision mechanism that\nadaptively retains full-precision computation for critical tokens. Experiments on\nsix LongBench tasks across three backbones show that DASH-KV achieves \nperformance comparable to full attention. At a 128K context length, DASH-KV\nachieves a 2.732x speedup over pre-expanded SDPA for the complete single-layer\ndecoding attention path on Qwen2-7B-Instruct.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Efficient Inference", "KV Cache Compression", "Large Language Models", "Sparse Attention" ], "forum_url": "https://openreview.net/forum?id=IF3P9Idbdf", "pdf_url": "https://openreview.net/pdf/202759fa4a89382b73ef4bd00e2ef96d33ba4868.pdf", "updated_at": "2026-10-05T16:59:20.721Z", "categories": [ "kv-cache" ], "matched_terms": [ "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zWaWggiWWP", "number": 1873, "title": "GWT: Scalable Optimizer State Compression for Large Language Model Training", "abstract": "Training large language models (LLMs) requires substantial memory, a significant fraction of which is consumed by the moment states maintained by adaptive optimizers such as Adam. Existing memory-efficient approaches commonly compress parameters or gradients through low-rank projections, which may discard information outside the selected subspace or incur additional costs from subspace construction and singular value decomposition (SVD). We introduce \\textbf{Gradient Wavelet Transform (GWT)}, an SVD-free framework that applies a multilevel Haar transform to matrix gradients. GWT maintains first- and second-moment states only for the compact approximation coefficients, while retaining all detail coefficients transiently in each update. We show that the orthogonal transform preserves the complete current gradient and establish gradient alignment and a stationary-point guarantee for smooth objectives under bounded preconditioning and a directly verifiable coarse-momentum coherence condition. Experiments on language-model pre-training and downstream fine-tuning show that GWT reduces estimated model-and-optimizer-state memory by up to 52\\% while maintaining competitive or improved model quality. On LLaMA-3B pre-training, GWT achieves $1.9\\times$ the throughput of 8-bit Adam and is compatible with multiple optimizer families, including Adam, Adam-mini, and Muon.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Wavelet transform", "large language model", "memory efficient" ], "forum_url": "https://openreview.net/forum?id=zWaWggiWWP", "pdf_url": "https://openreview.net/pdf/438e28aad6cb8ec341ed0e97e647ad10ca20d8b2.pdf", "updated_at": "2026-10-05T16:59:24.541Z", "categories": [ "training", "hardware" ], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EAPLudodCg", "number": 1958, "title": "DESA: Delivered-Energy Spectral Allocation for Merging Reinforced Experts", "abstract": "Model merging provides a practical way to combine multiple experts into a single model without retraining or increasing inference cost. However, existing merging methods are largely developed for supervised fine-tuning and do not account for the distinctive structure of reinforcement learning with verifiable rewards (RLVR) updates. \nWe show that RLVR task vectors exhibit two properties that are especially important for merging: (i) apparent sparsity under low-precision representation, and (ii) substantial heterogeneity in update energy across experts. These properties make sparsification-based heuristics unreliable and cause existing spectral methods to impose an uneven, supply-determined allocation of delivered energy, which strongly affects axis-wise recovery after merging. \nMotivated by this observation, we propose DESA (Delivered-Energy Spectral Allocation), a training-free spectral merging method that explicitly reallocates expert-level delivered energy while preserving a shared spectral frame. DESA introduces controlled budget equalization across experts together with retention floors that prevent under-allocation to experts whose capabilities are sensitive to reduced budget. \nAcross both language and vision-language RLVR expert pools, DESA consistently outperforms strong baselines in average accuracy and recovery. Our results identify explicit delivered-energy allocation as a key ingredient for merging RL-trained experts.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model merge", "RLVR", "LLM" ], "forum_url": "https://openreview.net/forum?id=EAPLudodCg", "pdf_url": "https://openreview.net/pdf/b6fa7bfae4a7e79f0182aaf90b50ff6da13522e5.pdf", "updated_at": "2026-10-05T16:59:36.579Z", "categories": [ "multimodal" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ITI0eC0Jwe", "number": 2235, "title": "Prune-Quantize-Prune: Towards Better Joint Model Compression by Getting Free Sparsity from Symmetric Quantization", "abstract": "The deployment of large language models (LLMs) at scale imposes substantial demands on memory bandwidth and computational throughput, making model compression essential for efficient inference. Among post-training compression techniques, 2:4 semi-structured sparsity and low-bit quantization are particularly attractive, yet effectively combining them remains challenging. Existing approaches typically apply pruning and quantization sequentially, either in a prune-then-quantize or quantize-then-prune manner, without explicitly exploiting the additional sparsity naturally induced by quantization. To address this limitation, we propose a three-stage **prune-quantize-prune** framework tailored to the combination of semi-structured sparsity and symmetric quantization. In the first stage, we introduce a *Quantization-aware Mask Selection* strategy that avoids prematurely enforcing the exact 2:4 constraint and instead reserves room for quantization-induced free sparsity. After applying low-bit symmetric quantization, the resulting additional zeros are explicitly exploited. In the final stage, we re-prune the quantized weights to strictly satisfy the target 2:4 sparsity pattern. To mitigate the reconstruction error introduced by this post-quantization pruning, we further propose a *Grid-constrained Compensation* method, which performs second-order weight compensation directly in the quantized space while preserving the discrete quantization grid. Extensive experimental results on popular LLMs demonstrate that this method outperforms existing state-of-the-art methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Compression", "Prune", "Quantization" ], "forum_url": "https://openreview.net/forum?id=ITI0eC0Jwe", "pdf_url": "https://openreview.net/pdf/f3006ba5b5983c03a9d287cd93ff8a2852ae0902.pdf", "updated_at": "2026-10-05T17:00:34.165Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BZAhEGxrVx", "number": 2236, "title": "Improve MXFP4 Quantization via Inter-block Permutation and Micro-block Givens Rotation", "abstract": "Large language models (LLMs) have achieved remarkable performance across diverse tasks, but their growing computational and memory costs pose substantial challenges for efficient deployment. Post-training quantization (PTQ) has emerged as a practical solution, while recent hardware support for microscaling formats such as MXFP4 further enables efficient low-bit inference. However, MXFP4 remains highly sensitive to massive activation outliers due to its power-of-two E8M0 shared scales and low-precision E2M1 values. Existing block-wise Hadamard transformations mitigate outliers by uniformly spreading their energy across an entire quantization block, but such magnitude reduction is not necessarily optimal for MXFP4. We observe that MXFP4 contains a set of favorable, exactly or nearly representable magnitudes, termed **lossless quantization points**, and quantization error is strongly determined by the distance of transformed values to these points. Based on this observation, we propose a novel transform framework for MXFP4. First, we propose *Inter-block Permutation* to group massive outliers with low-importance channels, reducing their interference with important features under block-wise shared scaling. Second, we propose *Micro-block Givens Rotation* to adaptively control the number of participating channels and redistribute outlier energy toward favorable quantization points, instead of uniformly dispersing it across the entire block. Experiments on the LLaMA2/3 and Qwen3 model families demonstrate that our method better adapts to the quantization characteristics of MXFP4 than the Hadamard transform and achieves superior quantization performance across multiple LLM benchmarks, providing new insights into the design of quantization methods for next-generation microscaling low-precision formats.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Microscaling FP4", "Transform", "LLM Quantization" ], "forum_url": "https://openreview.net/forum?id=BZAhEGxrVx", "pdf_url": "https://openreview.net/pdf/8c3dc442eec76f918c82b92df77a502f5c59df7f.pdf", "updated_at": "2026-10-05T17:00:33.254Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "low-bit", "low-precision", "mxfp4", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "j8fLzLZ27h", "number": 2369, "title": "MirrorQuant: Learnable Householder Rotations for Low-Bit LLM Quantization", "abstract": "Orthogonal rotations have become an effective technique for low-bit LLM quantization by expressing hidden representations in more quantization-friendly bases. However, existing approaches predominantly parameterize learned rotations as dense matrices, incurring quadratic learned state and costly optimization. Their deployment efficiency largely depends on foldability: at non-foldable interfaces, dense rotations must remain on the inference path, with both quadratic storage and computation of the rotation dimension. To overcome these limitations, we propose MirrorQuant, a unified framework for efficient and learnable rotations across Transformer interfaces. MirrorQuant parameterizes learned rotations using short chains of Householder reflections, a construction that preserves exact orthogonality, reduces the learned state to linear in the rotation dimension for fixed chain depth, and enables learnable log-linear online execution without materializing dense rotation matrices. We further introduce projection-guided learning, which complements representation reconstruction by penalizing quantization distortion after the immediate downstream linear projection. Experiments across 23 matched settings spanning model families, scales, and low-bit configurations demonstrate accuracy and efficiency gains. MirrorQuant improves 0-shot$_9$ accuracy over DartQuant in every setting, with gains up to 1.81%, while its online rotation reduces prefill latency by 29.6% relative to a dense implementation. Source code can be found in the supplementary materials.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Householder Rotations", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=j8fLzLZ27h", "pdf_url": "https://openreview.net/pdf/1f8895aceea8e1a2a3eebbc4552af80882d2a1fd.pdf", "updated_at": "2026-10-05T17:00:59.037Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hYEPYy0e8d", "number": 2496, "title": "DepthQuant: Depth Error Propagation for Mixed-Precision Bit Allocation", "abstract": "Mixed-precision quantization assigns a bit width to each unit of a large language model under a fixed bit budget, but most current methods decide the allocation from single-layer information and fall well short of full precision. This may be partly because single-layer methods miss information from deeper layers: similar quantization errors can produce different responses at the model output. To address this, we propose DepthQuant, a mixed-precision quantization method that uses depth error propagation as the signal that drives the bit allocation. By measuring the response each precision change induces one layer deeper and how much it changes the model output, DepthQuant spends the bit budget where more precision reduces that measured change the most. Built on this principle, DepthQuant introduces a\ncohesive multi-scale design: short-depth response scoring ranks precision changes by the response they induce one layer deeper, full-depth allocation then adjusts the resulting allocation by the error's effect on the model output, and a final step sets each row's range from information in its own layer. Experiments show that DepthQuant improves on the compared baselines in most evaluated settings and in every reported average-2-bit setting, reducing perplexity by up to 40.6\\% at 2.5-bit average precision relative to the best-performing baseline and raising zero-shot accuracy by up to 11.6\\% at 2-bit average precision, where these methods degrade the most.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixed-precision quantization", "low-bit quantization", "post-training quantization" ], "forum_url": "https://openreview.net/forum?id=hYEPYy0e8d", "pdf_url": "https://openreview.net/pdf/6f1beda88fb970b426e8ca58a5923419f8a8e94d.pdf", "updated_at": "2026-10-05T17:01:20.252Z", "categories": [ "ptq" ], "matched_terms": [ "2-bit", "5-bit", "bit width", "low-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7KG7PAwqGD", "number": 2510, "title": "AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive Coding", "abstract": "Long-running agents retain KV caches across turns and sessions for hours or days. Compressing these persistent caches requires reducing storage while controlling reconstruction error. We present AlignPred, a codec that combines local alignment with predictive coding under a configurable KV reconstruction-error budget. Local orthogonal transforms align related heads or layers into a common coordinate frame, allowing AlignPred to combine token-wise prediction with same-token prediction across heads or layers. Residuals are quantized on an error-bounded lattice and entropy-coded, without learned predictors, encoder-side mode search, or large joint transforms. Across three attention mechanisms, AlignPred requires less storage than the evaluated baselines at comparable measured reconstruction error. Against the evaluated CPU codec implementations, it decodes at least 1.6× faster while matching or improving encode time. AlignPred achieves near-native task scores at substantially compressed operating points on LongBench-Pro at 128K context. AlignPred's SGLang HiCache integration reduces persistent target-KV storage by 42–72% relative to native HiCache, with comparable latency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression; persistent KV storage; predictive coding; error-controlled compression; LLM serving" ], "forum_url": "https://openreview.net/forum?id=7KG7PAwqGD", "pdf_url": "https://openreview.net/pdf/7fdb9a206962af5fbfd8f1de178604010f420e36.pdf", "updated_at": "2026-10-05T17:01:22.650Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YEXEpbOENl", "number": 2581, "title": "JetDiffusion: Pushing the Speed Limits of Video Generation with NVFP4 Sparse Attention", "abstract": "Attention over long sequences is a major computational bottleneck in video generation, which can be alleviated through sparsity and low-bit quantization. We propose JetDiffusion, a video generation system built on trainable Video Sparse Quantized Attention (VSQA), with custom inference kernels that combine native NVFP4 computation and kernel fusion. At 90% sparsity, VSQA accelerates attention by 17-29x compared with FlashAttention on different hardwares. To further accelerate video generation, we combine VSQA with NVFP4 linear quantization and step distillation. Although each technique can individually preserve generation quality, jointly adapting to all three remains challenging, particularly under aggressive NVFP4 quantization. Through empirical studies, we develop an effective staged training recipe that enables stable adaptation to quantization, sparsity, and few-step denoising. On Wan2.1-14B, the resulting model achieves 282x end-to-end video generation speedup on the RTX 5090 while maintaining comparable video quality.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Generation", "Diffusion Transformers", "Few-Step Distillation", "Sparse Attention", "Quantization-Aware Training" ], "forum_url": "https://openreview.net/forum?id=YEXEpbOENl", "pdf_url": "https://openreview.net/pdf/10bcb74dc874f96a3909921583de14335030ac3a.pdf", "updated_at": "2026-10-05T17:01:35.586Z", "categories": [ "qat", "hardware", "diffusion" ], "matched_terms": [ "low-bit", "nvfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "owpMcF7vNT", "number": 2599, "title": "AuditGym: Can Black-Box Audits Verify the Model Behind an LLM API?", "abstract": "Large language models (LLMs) are increasingly accessed through APIs and embedded in agent systems, yet users have limited evidence that the API actually serves the model it claims. Existing black-box behavioral audits approach this problem through equality testing, model identification, and fingerprinting. However, their outcomes are difficult to interpret and compare: a detected discrepancy may reflect either a different model or a change in how the same model is deployed. We introduce AuditGym, a benchmark for systematically evaluating LLM identity verification under controlled deployment changes. AuditGym separates model specification, including checkpoints, learned adapters, and quantization, from generation policy and the interaction layer, and tests whether auditors reject changes to the former while accepting changes to the latter. We evaluate four auditing approaches across seven change families and 19 checkpoints, including models running inside real agent harnesses. All evaluated methods achieve only limited reliability for the identity verification: all four detect checkpoint substitution but miss finer model changes such as adapters and quantization. More seriously, every method rejects unchanged models under certain generation, watermarking, system-prompt changes, or agent harnesses. With these results, we reveal that current auditors confuse model changes with deployment changes, and AuditGym provides a controlled foundation towards closing this gap. Our anonymous codebase is available at https://anonymous.4open.science/r/auditgym_iclr-BC1E/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Equality Test", "Model Provenance" ], "forum_url": "https://openreview.net/forum?id=owpMcF7vNT", "pdf_url": "https://openreview.net/pdf/884189463c6e48509744e32a9b58d4ad5a659a83.pdf", "updated_at": "2026-10-05T17:01:37.494Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DtdxkedRiX", "number": 2601, "title": "Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning", "abstract": "Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that places teacher supervision where the quantized model actually goes. Starting from a QAD checkpoint, the student generates through the quantized forward path used at deployment and receives feedback from a frozen full-precision teacher on its own prefixes, combining dense token-level guidance with task-verifier rewards. Across four models at 2.79 and 1.88 effective bits, OPD raises average BF16 performance retention from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval while preserving short-form performance, with reasoning gains substantially exceeding those of continued teacher-forced QAD in matched-budget comparisons. By coupling QAD's stable low-bit initialization with OPD's on-policy reasoning recovery, our framework provides a comprehensive sub-3-bit solution that preserves broad capabilities while restoring long-form reasoning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "on-policy-distillation", "reasoning" ], "forum_url": "https://openreview.net/forum?id=DtdxkedRiX", "pdf_url": "https://openreview.net/pdf/5c2f53d9e0ddc556dca80965f7c3b742ce1331fc.pdf", "updated_at": "2026-10-05T17:01:38.278Z", "categories": [ "qat", "hardware", "reasoning" ], "matched_terms": [ "3-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6ge2NdeE7W", "number": 2945, "title": "SealKV: Accurate Low-Bit KV Compression for Faster LLM Inference", "abstract": "Long-context reasoning and agentic workloads have made the key–value (KV) cache the dominant cost of LLM serving: it is re-read at every decoding step and grows with every generated token, bounding both decoding speed and concurrency.\nCompressing it pays off only if it combines high quality, low memory use, and fast serving, yet existing quantization methods sacrifice at least one: they require offline calibration, lose accuracy at low bit widths, or save memory without a measured speedup.\nWe introduce SealKV, the \\emph{first} KV-cache compression method that achieves all three, by bringing a trellis-coding quantizer to a cache that grows continuously.\nOur key insight is that quantization can take place in sealed blocks, each coded only once all its tokens have arrived; sealing decouples trellis encoding from the newest token and exposes a subtractable per-channel offset, shared by neighboring tokens, that rotation alone cannot remove; subtracting it also lifts three existing quantizers by up to 32 points on GPT-OSS-120B retrieval.\nWe also observe that a shorter trellis is also affordable online at little loss: the code's quality depends far less on its window size than its encoding cost does.\nCo-designed fused dequantization--attention kernels read the codes directly and reconstruct values in registers without writing them back to memory.\nWithout calibration, SealKV is effectively lossless at three bits and above and state-of-the-art at two, where it shrinks KV cache memory by $6.6\\times$ while losing as little as 0.37 points on Qwen3-14B and outscoring all evaluated baselines.\nIntegrated end to end into vLLM, SealKV raises peak decode throughput at 64K context to $2.81\\times$ that of BF16 at 4 bits and $3.44\\times$ at 2 bits, ahead of all evaluated baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "calibration-free quantization", "KV cache compression", "efficient reasoning", "efficient agentic serving" ], "forum_url": "https://openreview.net/forum?id=6ge2NdeE7W", "pdf_url": "https://openreview.net/pdf/0f8ec34f3bfa63299382d4427b17b1d66e99a938.pdf", "updated_at": "2026-10-05T17:02:41.969Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "low bit", "low-bit", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WV17ZUKkZX", "number": 2970, "title": "Not All Tokens Are Equal: Rethinking Quantization for Diffusion Language Models", "abstract": "Quantization is a standard approach for large language models' deployment. Most existing quantization methods for large language models are designed for autoregressive models and the quantization objective is based on layer-wise mean square reconstruction error. In this paper, we show that such objective is mismatched to diffusion language models. Unlike autoregressive models, diffusion language models distinguish between visible tokens, which provide conditioning context, and masked tokens, which are the actual prediction targets. This structural asymmetry induces a fundamentally different quantization objective. Our analysis shows that minimizing the standard layer-wise mean square reconstruction error can yield a loose bound for quantization error in diffusion language models. Motivated by the above analysis, we explicitly disentangle visible and masked tokens as two groups and propose MV-Quant, which minimizes the worst-group reconstruction error. We further prove that, for an n-layer Transformer, MV-Quant yields a tighter quantization error bound than the standard MSE objective. Extensive experiments across multiple diffusion language models and multiple benchmarks validate the effectiveness of our proposed MV-Quant.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language model" ], "forum_url": "https://openreview.net/forum?id=WV17ZUKkZX", "pdf_url": "https://openreview.net/pdf/a19196c3802f461eaf9ee4ac65f9899b48a096f4.pdf", "updated_at": "2026-10-05T17:02:45.948Z", "categories": [ "ptq", "hardware", "diffusion", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SGpeYjd8rc", "number": 2982, "title": "Quantization-Aware Kalman Estimation for Diffusion Sampling", "abstract": "Quantization offers a practical path to deploying diffusion models with reduced memory and computation, but aggressive compression can cause quantized outputs to deviate substantially from their full-precision counterparts. Sampling-stage correction methods seek to compensate for such deviations during sampling, but existing approaches rely primarily on local information and underexploit trajectory history, limiting their ability to correct errors that propagate across timesteps. In this work, we formulate sampling with a quantized denoiser as an online estimation problem, using the history of quantized denoiser outputs to recover the underlying full-precision outputs required by the sampler. We propose QuAKE, a Quantization-Aware Kalman Estimator that combines a smooth trajectory prior with a conditional Gaussian observation model. At each sampling step, QuAKE recursively updates the posterior over the output window in closed form and feeds its posterior mean to the sampler. QuAKE is a lightweight plug-and-play corrector that requires no modification to the quantized network and naturally supports arbitrary high-order multistep ODE samplers. Experiments across W4A4-quantized text-to-image diffusion models show that QuAKE consistently outperforms existing methods in reducing the distributional discrepancy from full-precision sampling.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "Model Quantization", "Online Estimation" ], "forum_url": "https://openreview.net/forum?id=SGpeYjd8rc", "pdf_url": "https://openreview.net/pdf/308f7ea20d9719588ba9d9999f579518f450830e.pdf", "updated_at": "2026-10-05T17:02:47.706Z", "categories": [ "qat", "diffusion" ], "matched_terms": [ "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sPAEy9VbRn", "number": 3039, "title": "AutoQS: Automated Quantization Strategy Design with Large Language Models", "abstract": "Quantization reduces LLM deployment costs, but effective strategies must balance model quality and storage. LLM-driven search can automate strategy design through repeated proposal and evaluation. This process must identify useful strategies within a limited evaluation budget. The key challenge is that _actual quantization is expensive, while inexpensive proxies may not reflect quantized-model performance._ To address this challenge, we present AutoQS, an **Auto**mated **Q**uantization **S**trategy Design with LLMs. AutoQS keeps weight precision fixed and searches group-size allocations across layers and operators under deployment constraints. Evaluation history guides coordinated LLM proposals, while deterministic validation ensures compatible allocations. An inexpensive quality proxy evaluates candidates, and an candidate repository retains alternative allocations to inform subsequent proposals. Periodic actual quantization complements these proxy evaluations with measured quality and checkpoint storage, providing feedback for further search within the evaluation budget. Experiments on Qwen2.5-7B-Instruct show that AutoQS improves seven-task average accuracy over uniform-group GPTQ and AWQ by 0.70 and 1.12 percentage points, respectively, at four-bit weight precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization", "Audo Quantization", "LLM-Driven Algorithm" ], "forum_url": "https://openreview.net/forum?id=sPAEy9VbRn", "pdf_url": "https://openreview.net/pdf/6babcd016774b57d2fe0877e55c4b78a1830619b.pdf", "updated_at": "2026-10-05T17:02:55.042Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "uXnqTBzXlz", "number": 3080, "title": "BitsMoE: Cost-Aware Bit Allocation in Spectral Space for MoE LLM Quantization", "abstract": "Mixture-of-Experts (MoE) large language models incur substantial memory costs due to their large expert parameter counts. Mixed-precision quantization reduces these costs by allocating different bit-widths to experts or linear blocks according to their importance. However, assigning a single precision within each expert or linear block overlooks its internal structural heterogeneity. This limitation motivates two key questions: (1) how to define a fine-grained unit for quantization within a linear transformation; and (2) how to characterize the quantization cost of each unit under actual activation patterns and different bit-widths. To address these two questions, we propose BitsMoE, a cost-aware mixed-precision quantization framework built on two complementary techniques: (1) Shared-basis Spectral Decomposition (SSD) separates expert weights into a shared basis and expert-specific spectral components, defining structural quantization units while exploiting cross-expert redundancy. (2) Factorized Quantization Cost Modeling (FQCM) estimates component-wise costs from output reconstruction loss by combining intrinsic spectral importance, activation-dependent importance, and bit-width-dependent distortion. Using these component-wise costs, we formulate bit allocation as an integer linear program (ILP) that minimizes total modeled quantization cost under a fixed memory budget. On Qwen3-30B-A3B at 2-bit, BitsMoE achieves 64.29\\% average accuracy over seven downstream tasks, outperforming the evaluated MoE-specific methods, including those using ILP-based bit allocation, and exceeding GEMQ by 2.80 percentage points. Under the same setting, it achieves a $16.47\\times$ end-to-end offline quantization speedup over GEMQ. It also achieves up to $6.46\\times$ the decode throughput of GPTQ.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts; Post-Training Quantization; Mixed-Precision Quantization; Spectral Decomposition; Efficient LLM Inference" ], "forum_url": "https://openreview.net/forum?id=uXnqTBzXlz", "pdf_url": "https://openreview.net/pdf/470e8f500747f34491f36583fec4f30c73c77599.pdf", "updated_at": "2026-10-05T17:03:01.006Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "bit-width", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6t1egKiB0v", "number": 3089, "title": "Quantization-Induced Decision Changes Have Regime-Dependent Intervention Spectra", "abstract": "Quantization can preserve aggregate task performance while replacing individual predictions. However, neither score changes nor disagreement rates reveal how the associated behavioral effects are organized across representation directions. To address this gap, we introduce intervention-spectrum analysis, which tracks paired-decision transfer along ordered shared directions and normalizes aggregate prefix transfer by the complete task-supported transfer rate. Using directions fitted to other folds' full-precision states, we project each held-out paired-state difference onto successive prefixes. On disagreeing inputs, we add these projections at their native magnitude to the quantized residual state, testing agreement with the original full-precision prediction regardless of correctness. Across five checkpoints and three quantizers, final-site task-supported interventions outperform dimension- and norm-matched ambient controls in fixed-choice answer selection and teacher-forced full-vocabulary prediction. Despite this shared directional advantage, quantizer-level summaries show that the first 16 directions attain 91.8-92.1% of the complete task-supported one-way transfer rate in fixed choice, versus 12.6-15.1% in full vocabulary. This contrast persists under plain pooled-PCA, while candidate restriction within the full-vocabulary computation does not reproduce compact accumulation. These findings distinguish directional advantage from concentration: effective transfer can remain broadly distributed along a shared representation ordering. An anonymous code repository is available at https://anonymous.4open.science/r/quantization-behavioral-drift-CD71/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "LLMs", "model compression", "representation intervention" ], "forum_url": "https://openreview.net/forum?id=6t1egKiB0v", "pdf_url": "https://openreview.net/pdf/2842f07118fc1ccb259b61645c910e173e6bca99.pdf", "updated_at": "2026-10-05T17:03:02.589Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantized", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZuNamQK6PF", "number": 3106, "title": "Flow-Q: Flow Consistency for Post-Training Quantization of Vision Transformers", "abstract": "We present Flow-Q, a general post-training quantization framework for Vision Transformers. Our analysis shows that the existing sensitivity-weighted reconstruction paradigm induces feature distortion during quantization, degrading the robustness of quantized feature representations. To overcome this challenge, we first propose an evolution-aware learning (EAL) strategy that captures feature evolution information to construct a manifold representation with global topological structure. Moreover, a geometric constraint rectification (GCR) module is designed to utilize the manifold representation to rectify quantized features from both direction and magnitude perspectives, thereby alleviating feature distortion. EAL provides the geometric structural foundation for GCR, enabling the quantized model to inherit the robustness of visual representations from the full-precision network. Extensive experiments on nine tasks across twelve benchmarks demonstrate the strong generalizability of Flow-Q, which achieves new state-of-the-art quantization performance across DINOv3, conventional ViTs, and mobile-efficient models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Vision Transformers" ], "forum_url": "https://openreview.net/forum?id=ZuNamQK6PF", "pdf_url": "https://openreview.net/pdf/a8d96d052a295ee98e66cb002eca8dcf0292568d.pdf", "updated_at": "2026-10-05T17:03:03.978Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CwB8eaqsMQ", "number": 3138, "title": "MaskTac: Mask-guided Learned Lossy Compression for Sparse Tactile Glove Data", "abstract": "Tactile sensing provides direct feedback about physical contact for a wide range of perception and interaction applications. However, high-quality force-based tactile datasets remain limited. The widespread use of tactile data also creates a growing need of efficient compression for storage and transmission. To mitagate these gaps, we firstly introduce TacGlove, a tactile glove dataset collected using a customized wearable tactile glove. It contains 4,660 sequences and 781,969 frames, covering 32 object classes and 8 grasp types. The glove records three-axis contact forces from 1,207 taxels distributed across the hand. Such force-based tactile glove data exhibit pronounced spatiotemporal sparsity and strong correlations within contact regions. Based on these characteristics, we then propose MaskTac, a learned lossy codec for sparse tactile glove data. By combining sparsity-aware quantization with a mask-guided contextual entropy model, MaskTac exploits data spatiotemporal sparsity and reduces spatiotemporal redundancy within contact regions for efficient compression. We further establish a benchmark for tactile compression by adapting traditional and learned visual codecs to TacGlove. The benchmark evaluates both reconstruction fidelity and object and grasp classification accuracy. Experiments show that MaskTac outperforms the benchmarked codecs, achieving BD-rate improvement relative to VTM of 34.04\\% under PSNR, 32.57\\% under MS-SSIM, and 58.83\\% under Contact F1. Ablation studies further validate the effectiveness of the proposed design.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Tactile", "Tactile Dataset", "Lossy Compression", "Data Compression" ], "forum_url": "https://openreview.net/forum?id=CwB8eaqsMQ", "pdf_url": "https://openreview.net/pdf/f97e197b9ec8fe47e394aa22b165263853c81e12.pdf", "updated_at": "2026-10-05T17:03:07.702Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OToCbu4d9u", "number": 3255, "title": "Pushing the KV Cache Beyond the 2-Bit Cliff via Spectral Transform Coding", "abstract": "Large language models (LLMs) are increasingly applied to very long inputs, such as whole documents,\n codebases, and conversations that run for many turns. The key-value (KV) cache stores the attention\n keys and values of all past tokens. Because the cache grows with the context length and is re-read\n in full at every generated token, it dominates the inference memory at long context.\n Most existing methods reduce this cost by quantizing the cache: storing every value at the\n same low precision. These methods reach about two bits per value, but they rarely go below two bits,\n because accuracy drops sharply at that point. The reason is that two bits allow only four levels per\n value, which are too few for a cache with outlier channels, which is known as the 2-bit cliff. \n In our study, we find that once we project the cache into a latent space that removes the\n correlations between channels, the contributions of the different channels of the KV cache to the\n attention output are highly uneven. And we find that spending the bit budget on those few important channels is far more accurate than spreading the bit budget evenly. Guided by these observations, we develop SPECTRA, a training-free compression method for the KV cache\n that requires no change to the model. SPECTRA transforms the KV cache into a latent space in which the channels are ordered by their\n contribution to the attention output, and then allocates the bit budget across these ordered\n channels. On long-context benchmarks and three instruction-tuned models, every uniform quantization methods that we\n compare against has fallen below the 2-bit cliff at 8$\\times$ compression rate, and none of these\n methods has a configuration beyond 9$\\times$. SPECTRA matches the uncompressed model at\n 8$\\times$ on Llama-3.1-8B, stays within 1 point of that accuracy up to 11$\\times$, and recovers every\n planted needle in needle-in-a-haystack at 12$\\times$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "KVCache", "Quantization", "Pruning", "Efficiency", "Compression" ], "forum_url": "https://openreview.net/forum?id=OToCbu4d9u", "pdf_url": "https://openreview.net/pdf/35c6f8f970f1f73626ab5fcaf31c676b321042e6.pdf", "updated_at": "2026-10-05T17:03:24.690Z", "categories": [ "kv-cache" ], "matched_terms": [ "2-bit", "low precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xL7T2h9rE4", "number": 3256, "title": "BellowsKV: One Latent Cache for Many Layers, Expanded on Demand", "abstract": "Large language models (LLMs) are now widely utilized as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To serve such a context without recomputing the past, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache and re-reads it at every generation step, so KV-cache memory can become the main bottleneck for GPU serving. Most compression methods operate within a single layer, by quantizing the cache entries, reducing the channel dimension, or evicting less important tokens from the cache. All three lose substantial accuracy at high compression ratios. A more recent line of work shares one compressed cache across adjacent layers to go beyond this limit. However, existing cross-layer methods also degrade sharply at high compression ratios. We find that the KV cache representations of adjacent layers are quite different, whereas the input hidden states are highly similar. We also find that, under the same memory budget, reducing the dimension of this shared latent loses far more accuracy than quantizing it. Based on these findings, we propose BellowsKV, a training-free method that maps the hidden states of n adjacent layers to one shared latent per token. At inference time, each layer reads its own keys and values from this latent through a layer-specific projection. We keep the latent at full dimension and compress it in the two other ways instead: BellowsKV-Quant stores the latent at low precision, which is nearly lossless because the latent has no dominant outlier channels; BellowsKV-Evict makes one eviction decision per block, which the layers can share because they rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant retains about 99% of full-cache accuracy at 3.6$\\times$ compression and 97–98% at 13–14$\\times$, where the strongest existing cross-layer method retains at most 84% at 16$\\times$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "KVCache", "Quantization", "Pruning", "Efficiency", "Eviction", "KV Cache Sharing" ], "forum_url": "https://openreview.net/forum?id=xL7T2h9rE4", "pdf_url": "https://openreview.net/pdf/73ebff579f2e424b690edce171461152c6d17375.pdf", "updated_at": "2026-10-05T17:03:24.679Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "low precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pOPF5E5PhW", "number": 3295, "title": "Don’t Compress Away the Evidence: Secure Provenance Auditing for LLM KV Caches", "abstract": "Efficient KV cache compression is essential for deploying long-context large language models, but it introduces an unexamined conflict with model ownership auditing. Existing fingerprinting methods embed ownership evidence through trigger patterns that are deliberately positioned away from routine task semantics—precisely the kind of patterns that cache compression discards as ``unimportant\". Therefore, we propose Fingerprint-Preserving Causal KV Cache Compression (FinCa), a framework that treats ownership evidence and task utility as joint constraints. Our key insight is to replace static trigger patterns with a low-rank perturbation in a KV projection that generates an input-conditioned structured signal in the runtime cache. Complementing this generation mechanism, we reserve fingerprint-bearing coordinates during compression and apply quantization-index modulation to protect them, ensuring that the ownership signal survives cache reconstruction. As a result, FinCa embeds a robust fingerprint signal from only 20 documents and achieves $100\\%$ signal recovery in most evaluated settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Provenance; Ownership Auditing; Large Language Models; KV-Cache Compression; Causal Rate-Distortion Coding" ], "forum_url": "https://openreview.net/forum?id=pOPF5E5PhW", "pdf_url": "https://openreview.net/pdf/e2837aecb511226de261c87bdd622631160fc876.pdf", "updated_at": "2026-10-05T17:03:35.575Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8qOKRk4Hvt", "number": 3413, "title": "Why does Post-Training Quantization Work?", "abstract": "Post-training quantization (PTQ) compresses large language models (LLMs) by reducing weight precision and can largely preserve downstream task performance, \neven though they are never exposed to quantization error during pretraining.\nA common intuition is that low-bit formats such as NVFP4 induce only small errors in each weight matrix, so errors accumulate slowly through the model. \nHowever, we show that this explanation alone does not fully explain PTQ's empirical effectiveness.\nSpecifically, when quantizing randomly initialized and pretrained models to the same precision, \nwe observe that their weight quantization errors are nearly identical, yet randomly initialized models exhibit substantially larger hidden-state error than pretrained models.\nWe attribute this discrepancy to a counteracting mechanism present in pretrained networks: the error newly introduced by a layer tends to oppose the error propagated from the preceding layer.\nThese two components partially cancel, substantially slowing error accumulation across layers. \nMoreover, at the output layer, the probabilities of high-ranked tokens are more robust to the propagated errors. \nWe quantify these mechanisms and validate our findings across multiple model families and quantization settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Post-Training Quantization", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=8qOKRk4Hvt", "pdf_url": "https://openreview.net/pdf/7ac1c9474eae0d265d334fc841e9a135bc1a478d.pdf", "updated_at": "2026-10-05T17:03:56.346Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "nvfp4", "ptq", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xP0AsDPofp", "number": 3423, "title": "COMPRESS FOR RECOVERABILITY: STREAMED COUPLED CORRECTIONS FOR MOE EXPERTS", "abstract": "Mixture-of-Experts (MoE) models activate few experts per token, but their expert banks consume substantial GPU memory. We study recoverability as a third axis for compression, alongside resident memory and uncorrected accuracy: how much lost computation can be restored under a limited correction budget? A compressed model with poor standalone accuracy can still provide an effective base if its errors admit sparse, input-dependent correction. This perspective makes expert merging a viable compression strategy when combined with selective recovery. We introduce Delta-MoE, which keeps compressed experts on the GPU and streams selected residual channels from host memory. Each correction couples the gate, up, and down projections of a SwiGLU channel and recomputes its nonlinear contribution. Our analysis characterizes recovery through residual error spectra and explains why improving an uncorrected merge need not improve its corrected counterpart. On Qwen3-30B-A3B, the fused implementation achieves 0.810 GSM8K accuracy at 16× resident expert compression, with MATH-500 and IFEval point estimates similar to an 8× quantized base, and Mixtral-8x7B also exhibits recovery from merging.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "model compression", "expert merging", "quantization", "error recoverability", "residual correction", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=xP0AsDPofp", "pdf_url": "https://openreview.net/pdf/e267f2c1a63ad183d58be12777e005e8bb4c7520.pdf", "updated_at": "2026-10-05T17:03:57.840Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rHJLwXYerC", "number": 3515, "title": "HyperQuant: A Near-Entropy-Optimal Lattice Quantization Pipeline for Large Language and Diffusion Models", "abstract": "Data-free post-training quantization has converged on rotation plus a fixed-rate codebook: whiten a weight tile toward Gaussianity, then round it to an MSE-optimal grid. That last step spends $\\log_2 N$ bits per index however unevenly the indices are used.\n\nWe introduce HyperQuant (Hadamard, optimallY Packing, Entropy Rice-coding), which replaces that codebook with a lattice and a variable-length code in one calibration-free pipeline for both the linear layers and the KV cache of large language and diffusion transformers. A per-tile Walsh--Hadamard rotation whitens each tile, an optimal low-dimensional lattice ($E_8$, $D_4$, $A_2$, or $\\mathbb{Z}$) quantizes it, a lossless strip discards the bits lattice membership already fixes, and a Rice code carries the remainder to within $\\sim0.1$ bps of that lattice's rate--distortion ideal. Decoding is algebraic, so dimension costs no codebook memory. For the KV cache we add subtractive dither, which under an idealized sampler leaves every cached inner product unbiased per vector, and measurably so in practice, where rotation alone is unbiased only on average.\n\nTable1 collects operating points. On Llama-3.1-8B weights the default $E_8$ leads HIGGS at $3$--$5$ bps, by $0.93\\pm0.07$ PPL at $b{=}3$ over $4$ seeds. Neither OCTOPUS nor TurboQuant released usable code, so we reproduce both, validate each against published numbers, and run all three codecs in one harness under one itemized rate accounting over $8$ seeds. \\method leads both at equal or lower effective rate, apart from a tie with both at $4$ bits and a loss to TurboQuant at $2$ bits on C4 under QJL; the leads hold at $p\\le0.016$ over shared seeds. The same pipeline quantizes the 19B-parameter LTX-2 video DiT with no visible artifacts. On one H100 at batch $1$ it cuts resident weight memory $\\sim2.8\\times$ and the KV-cache encoding rate $\\sim3.8\\times$ (computed rather than measured on a filled cache), though the decode kernel stays slower than bare bf16 GEMV.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "LLM", "DiT" ], "forum_url": "https://openreview.net/forum?id=rHJLwXYerC", "pdf_url": "https://openreview.net/pdf/659efe175e335b1f3c9079a425635c706cb462c4.pdf", "updated_at": "2026-10-05T17:04:10.369Z", "categories": [ "ptq", "kv-cache", "hardware", "diffusion" ], "matched_terms": [ "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1ZHmTZk1D2", "number": 3543, "title": "TameProbe: An Empirical Tameness Spectrum for Neural Decision Boundaries", "abstract": "Model-theoretic o-minimality (``tameness'') has recently been proposed as a lens on neural network geometry, but it is a binary structural property that is not directly measurable on a trained model. We turn tameness into an operational, continuous diagnostic spectrum: two cheap, model-agnostic statistics of a network's decision boundary---the oscillation complexity ($\\mathrm{OC}$, sign flips and local extrema of the decision margin along random 1D probes) and the decomposition complexity ($\\mathrm{DC}$, number of distinct linear regions crossed). Across a controlled matrix of MLP architectures, seeds and post-training operators (an alignment-style Jacobian/weight regularizer, knowledge distillation, and 8-bit quantization; $14{,}400$ measured slices), we find that (i) tameness is not scalar: the alignment proxy lowers both $\\mathrm{OC}$ ($-35.4\\%$) and $\\mathrm{DC}$ ($-30.3\\%$), whereas distillation lowers $\\mathrm{OC}$ ($-17.9\\%$) while raising $\\mathrm{DC}$ ($+1.0\\%$), so the two axes decouple; (ii) 8-bit quantization does not move the spectrum ($p>0.18$), a negative result that validates the metrics as capturing function geometry rather than numerical noise; and (iii) on an external adversarial-robustness anchor ($n=120$), $\\mathrm{OC}$ strongly predicts PGD robustness (Spearman $\\rho=-0.68$, $p<10^{-17}$; $\\sim3\\times$ the effect of $\\mathrm{DC}$), connecting our spectrum to the curvature--robustness literature. We further show, on a balanced compositional truth concept across five same-family base/instruct LLM pairs, that both base and post-trained LLMs already sit at the tame end (a nonlinear RBF kernel never beats a linear probe), with post-training leaving this position essentially unchanged. TameProbe gives a reproducible, quantitative vocabulary for how far along the tameness spectrum a model sits and how training operators move it.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "neural decision boundaries", "tame geometry", "o-minimality", "model diagnostics", "adversarial robustness" ], "forum_url": "https://openreview.net/forum?id=1ZHmTZk1D2", "pdf_url": "https://openreview.net/pdf/6cdac6321a0b6fa1731eeb500d044a53d2883652.pdf", "updated_at": "2026-10-05T17:04:14.941Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "8-bit", "binary", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Cn4HeZm7Ky", "number": 3648, "title": "Mix-Quant: Quantized Prefilling, Precise Decoding for Agentic LLMs", "abstract": "LLM agents have recently emerged as a powerful paradigm for solving complex tasks through planning, tool use, memory retrieval, and multi-step interaction. However, these agentic workflows often introduce substantial input-side overhead, making the compute-intensive prefilling stage a key bottleneck in long-context, multi-turn inference. \nIn this work, we propose Mix-Quant, a simple and effective phase-aware quantization framework for fast agentic inference. We first investigate FP4 quantization in agentic LLM workflows and observe that, despite being the dominant source of computation, prefilling exhibits substantially higher quantization redundancy than autoregressive decoding. Based on this insight, we apply high-throughput NVFP4 quantization to the prefilling phase while preserving BF16 precision for decoding. By decoupling prefilling acceleration from decoding quality, Mix-Quant combines phase-aware algorithmic quantization with hardware-efficient NVFP4 execution to alleviate the inference bottleneck in LLM agents. \nExtensive experiments across long-context and agentic benchmarks demonstrate that Mix-Quant largely preserves task performance while delivering significant improving efficiency, achieving up to a 3.74$\\times$ speedup during prefilling.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Agents", "Quantization" ], "forum_url": "https://openreview.net/forum?id=Cn4HeZm7Ky", "pdf_url": "https://openreview.net/pdf/eb39a316bfb9a2cd5dd27f00abdf4d2dc7823c88.pdf", "updated_at": "2026-10-05T17:04:33.990Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "nvfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qwVJuNWqkO", "number": 3715, "title": "Quantization-conditioned Normalizing Flows for Multi-granularity Graph Anomaly Detection", "abstract": "Multi-granularity Graph Anomaly Detection aims to design a unified framework for detecting abnormal patterns at node, edge, and subgraph granularities, occupying an important position in academia and industry. Existing methods usually focus on exploiting a unified representation space for different granularities, while overlooking the distinct characteristics and anomaly patterns across granularities, leading to poor identification ability. To address these limitations, a quantization-conditioned normalizing flow framework for multi-granularity GAD (QNMGAD) is proposed to tackle this challenge. On the one hand, QNMGAD introduces prototype quantization to capture discriminative semantics across granularities and utilizes residual patterns quantization to extract latent semantics within a specific granularity. On the other hand, QNMGAD introduces conditioned normalizing flows based on granularity prototypes to model the distribution of normal patterns and calculate density estimation to identify anomalies. Extensive experiments compared with state-of-the-art baselines show the superiority of QNMGAD.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "multi-granularity graph anomaly detection", "unsupervised learning", "normalizing flow" ], "forum_url": "https://openreview.net/forum?id=qwVJuNWqkO", "pdf_url": "https://openreview.net/pdf/66e96cd15088b08e2ccbe1a3d1a12c25fb06d0d2.pdf", "updated_at": "2026-10-05T17:04:44.232Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BhYm577QTB", "number": 3851, "title": "AquaPIC: Boosting On-device LLM Quantization via Prefetching Input-wise Compensation", "abstract": "Post-training quantization (PTQ) enables efficient local LLM inference, but aggressive quantization can substantially degrade accuracy. Recovering this lost accuracy typically requires additional correction data, but the gains are constrained by limited device memory budget. Prefetching correction data from second-tier memory can overcome this capacity limit, yet limited transfer bandwidth constrains their size, making compact low-rank compensation particularly suitable for this setting. We introduce input-partition-wise compensation (IPC), which exploits additional second-tier capacity by storing specialized low-rank compensators for different input regions without increasing the amount of data fetched per input. Building on IPC, AquaPIC combines output-error-aligned clustering, shared block-level partitions, and a lightweight residual-based partition estimator to select and prefetch compensators while the preceding block executes. Across two LLMs and two quantization methods, AquaPIC improves zero-shot accuracy under 2-bit quantization by 6.8 percentage points on average over the strongest same-rank baselines, while incurring only 2.27% average latency overhead across discrete GPU and unified memory setups.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization", "Post Training Quantization", "Local LLM Inference", "Low-rank Compensation", "Efficiency", "ML System" ], "forum_url": "https://openreview.net/forum?id=BhYm577QTB", "pdf_url": "https://openreview.net/pdf/b9d50f54cd03eabdd467c5d61e5061813d0cc396.pdf", "updated_at": "2026-10-05T17:05:15.608Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a70W2l4Lxx", "number": 3959, "title": "Latent Diffusion Model Watermarks can be Removed by VAR Regeneration", "abstract": "Diffusion regeneration has emerged as a universal removal attack against invisible image watermarking, but it is not able to remove latent-domain watermarks, while the underlying reason remains unaddressed. In this work, we reveal that the reason lies in the natural preservation of generation--inversion symmetry in diffusion regeneration. Motivated by this, we propose visual autoregressive regeneration (VARR), a simple yet powerful universal watermark removal attack that breaks this symmetry. Specifically, during the VAR model’s multi-scale quantization, it keeps low-scale semantic tokens while replacing high-scale ones with their predictions using VAR's autoregressive transformer. Subsequently, the partially modified token sets are used to reconstruct the image with the watermark removed. Across eight representative latent-domain watermarking methods, VARR achieves an average removal success rate of 92.8\\% while maintaining high visual fidelity with an average FID of 28.27, significantly lower than state-of-the-art regeneration methods (FID across 40 to 75). Furthermore, we show that the VAR domain is more effective for watermark removal compared with the pixel and frequency domains, empirically supporting the effectiveness of VARR. Our findings expose a fundamental vulnerability of latent-domain watermarks and establish autoregressive regeneration as a new paradigm for evaluating watermark robustness.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Watermark Removal", "Model Watermarking", "Autoregressive Models" ], "forum_url": "https://openreview.net/forum?id=a70W2l4Lxx", "pdf_url": "https://openreview.net/pdf/f1bb380e38a1d504e9d6b022ca47bc40dbd331ba.pdf", "updated_at": "2026-10-05T17:05:43.744Z", "categories": [ "diffusion", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4Feb9juVnH", "number": 3988, "title": "FlashLoop: Fast and Memory-Efficient Looped Transformers via Lazy Updates", "abstract": "Looped Transformers have attracted substantial attention as a parameter-efficient approach to increasing computational depth through repeated application of shared Transformer blocks. However, their practical advantages over conventional Transformers remain under debate: each additional loop incurs another Transformer pass and requires caching another set of KV states, causing inference FLOPs and KV-cache memory to grow continuously with loop depth. This overhead becomes particularly severe at large loop counts and long context lengths, preventing the parameter efficiency of Looped Transformers from translating into practical inference efficiency.\nIn this paper, we find that much of the additional computation and storage introduced by looping is redundant. As recurrence proceeds, state changes become increasingly concentrated on a small subset of tokens; attention-output differences are dominated by a sparse and stable subset of key columns; and KV residuals between adjacent loops become progressively more amenable to low-bit quantization. Building on these observations, we introduce FlashLoop, a training-free inference framework that reduces cross-loop redundancy through token-sparse updates, sparse attention, and KV-residual quantization. Across five Ouro and Huginn variants, FlashLoop largely preserves model accuracy while achieving $1.52$--$1.64\\times$ end-to-end speedup and $5.18$--$6.06\\times$ KV-cache memory reduction. At a 32K context length, FlashLoop reduces KV-cache memory by $82.8\\%$ while delivering approximately $1.6\\times$ speedup, substantially improving the practicality of scaling Looped Transformers to greater computational depths.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Looped Transformer;Sparse Attention;KV Cache Compression" ], "forum_url": "https://openreview.net/forum?id=4Feb9juVnH", "pdf_url": "https://openreview.net/pdf/522fa146c9a1e2ca9bb83518bc2a062613712524.pdf", "updated_at": "2026-10-05T17:06:03.193Z", "categories": [ "kv-cache" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xvloOO0IWo", "number": 4097, "title": "GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives", "abstract": "Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control. Yet vision-language-action (VLA) and world-action models (WAMs) take perception from general-purpose vision-language and video-generation backbones, which still fail in these settings. We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI establishes a new state of the art, averaging 73.68\\%, above the larger GPT-6 Astra (71.54\\%). As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. Across RoboTwin 2.0 and RoboCasa-GR1, it outperforms all evaluated backbones in all four out-of-distribution settings, with relative gains of up to 24.8\\% over the strongest baseline. On RoboCasa-GR1, GroundingPI trained with 50\\% of the demonstrations outperforms those baselines trained with 75\\%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m. We analyze GroundingPI's pretraining in scale and data composition. Grounding quality and downstream manipulation improve with pretraining scale, with generalization tracking perceptual gains most closely. Analyzing the data recipe across these 11 perceptual capabilities shows dense grounding's substantial benefits for both, and OCR's potential as a catalyst for perceptual learning. These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Foundation Models", "Vision-Language Models", "Vision-Language-Action Models" ], "forum_url": "https://openreview.net/forum?id=xvloOO0IWo", "pdf_url": "https://openreview.net/pdf/c949d7d4cc2bc12f01bb64c219225d916de52051.pdf", "updated_at": "2026-10-05T17:06:58.007Z", "categories": [ "ptq", "multimodal", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7zDzdnMh1m", "number": 4177, "title": "TraceGate: Routing-Action History for Block Reuse in Video Diffusion", "abstract": "Video diffusion transformers repeatedly evaluate costly block stacks and attention routes. Block caching skips stack evaluations; sparse attention reduces work within executed stacks. TraceGate connects these decisions: a planner reuses, repairs, or refreshes query and key clusters on executed calls, and its historical actions inform later residual reuse at the block level. The outer gate combines this delayed feedback with conditioning drift and residual magnitude history; skipped calls do not compute their own attention. A product quantization shortlist reduces planner candidate scoring. Settings focused on efficiency achieve 3.36×, 2.45×, and 3.02× overall speedups on Wan 2.1, Wan 2.2, and HunyuanVideo, respectively, with tradeoffs in task quality. On the Wan 2.1 calibration cohort of 108 videos, the setting with feedback has 0.13 dB higher mean PSNR and 0.003 higher mean SSIM than a retuned setting without feedback; both speedups round to 3.30× and LPIPS is equal at reported precision. This comparison is exploratory because the cohort also informed calibration. Substituting a stale map on 72 selected events tests propagation of inner routing error, not the outer execution versus reuse action.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "video generation", "training-free", "diffusion transformers" ], "forum_url": "https://openreview.net/forum?id=7zDzdnMh1m", "pdf_url": "https://openreview.net/pdf/4015e59749b03e406c0b6cc7785b56429c9c43e4.pdf", "updated_at": "2026-10-05T17:07:44.388Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MJE81EQgGk", "number": 4200, "title": "OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers", "abstract": "Diffusion transformers (DiTs) achieve state-of-the-art image and video generation, but their multi-step sampling and growing parameter count make inference expensive. Post-training quantization (PTQ) is the natural remedy, yet DiT activations shift across timesteps, prompts, and guidance branches, forcing prior methods to re-fit calibration data for every new checkpoint or modality. We present OrbitQuant, a data-agnostic weight-activation quantizer that bypasses range estimation by quantizing in a normalized, rotated basis. In this basis, a randomized permuted block-Hadamard (RPBH) rotation balances coordinate variances and maps activations toward a stable, known marginal across inputs, so a single Lloyd-Max codebook serves all timesteps, prompts, and layers of a given input dimension. We extend the same quantizer to weight rows offline, absorbing the rotation into the weights so that it cancels inside each linear layer and only a forward rotation on the activations remains at runtime. The same recipe transfers from image to video without per-modality tuning. Across FLUX.1, Z-Image-Turbo, Wan 2.1, and CogVideoX, it sets the state of the art for PTQ at several low-bit settings. It also pushes PTQ of image DiTs to W2A4 with usable generation quality.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Transformer" ], "forum_url": "https://openreview.net/forum?id=MJE81EQgGk", "pdf_url": "https://openreview.net/pdf/91985b896339387e64a0273a750d8c1a0bd3b2c8.pdf", "updated_at": "2026-10-05T17:08:05.396Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantizer", "quantizing", "w2a4" ], "review_status": "abstract", "summary": "OrbitQuant 避免逐 checkpoint、逐模态重新估计激活范围，在归一化并旋转的坐标中量化权重和激活。RPBH 旋转平衡坐标方差，使同一输入维度可共用 Lloyd-Max 码本；离线权重旋转吸收到线性层权重，运行时只保留激活侧的前向旋转。", "contribution": "摘要报告同一配方无需逐模态调参即可从图像迁移到视频，在 FLUX.1、Z-Image-Turbo、Wan 2.1、CogVideoX 的若干低位 PTQ 设置取得领先质量，并将图像 DiT 推至具有可用生成质量的 W2A4。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。摘要未给出 FID 等质量数值或实测速度；W2A4 的可用质量陈述明确针对图像 DiT，不能扩展为全部视频模型同位宽结果。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "图像/视频 DiT 权重与激活", "bit_width": "明确报告图像 DiT 的 W2A4；其他低位设置摘要未列出", "evidence": [ { "location": "官方摘要", "claim": "报告图像 DiT 的 W2A4 可用生成质量。" }, { "location": "官方摘要", "claim": "配方从图像迁移到视频无需逐模态调参；具体质量数值未在摘要列出。" } ] }, { "id": "QPdhwnOh9a", "number": 4203, "title": "SoloQ: Calibration-Free Quantization for Diffusion Language Models", "abstract": "Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models through bidirectional diffusion-based token generation. However, their growing model sizes and high inference costs make efficient deployment challenging: full-sequence denoising repeatedly invokes compute-intensive forward passes, while block-diffusion models additionally introduce a memory-intensive KV-cache. Low-bit weight-activation quantization is therefore attractive, yet existing dLLM post-training quantization methods rely on calibration data despite activation distributions shifting across masking states and denoising steps. We present SoloQ, a calibration-free quantization framework that maps weights and activations into a normalized rotated basis with a predictable marginal distribution, enabling data-independent quantization. SoloQ combines a structured K-RPBH rotation with a lightweight rescaling correction for calibration-free quantization. Its predictable post-rotation distribution supports both distribution-matched codebooks and hardware-native NVFP4. For block-diffusion models, SoloQ further applies commit-time KV-cache quantization to compress persistent states without perturbing the actively denoised block. Across full-sequence dLLMs (LLaDA and Dream) and block-diffusion dLLMs\n(Fast-dLLM v2 and Nemotron-Labs-Diffusion), SoloQ retains accuracy under 4-bit quantization and outperforms calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, SoloQ reduces peak memory by up to 2.61$\\times$ and accelerates end-to-end inference by up to 2.24$\\times$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Language Model" ], "forum_url": "https://openreview.net/forum?id=QPdhwnOh9a", "pdf_url": "https://openreview.net/pdf/7251c7c3cbdc03075cda0e7bfbd166ae0fa93139.pdf", "updated_at": "2026-10-05T17:08:04.190Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning", "diffusion" ], "matched_terms": [ "4-bit", "low-bit", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CQDc2CcjFm", "number": 4235, "title": "$\\mu$-Zero: Zero-Anchored Log Companding for KV Cache Quantization", "abstract": "Long-context LLM inference is increasingly limited by KV cache memory traffic, making low-bit cache quantization essential for efficient deployment. Most calibration-free methods use per-group affine min--max quantization, uniformly allocating levels across the observed range. \n We show that this rule is poorly aligned with attention: after standard key/value grouping, online cache groups exhibit dense zero-centered mass and sparse, approximately log-tailed extremes, causing uniform grids to waste codes in low-density regions. \n Moreover, query-key logits are often residuals of cancelling signed products, so small perturbations near zero can induce large relative errors or sign flips despite low entrywise reconstruction error.\n These observations motivate protecting the zero neighborhood rather than only the range endpoints. \n We propose $\\mu$-Zero, a zero-anchored piecewise log-companded quantizer that applies a monotone $\\mu$-law transform to the positive and negative sides around an exact zero level. \n $\\mu$-Zero concentrates resolution near zero, preserves asymmetric side extents, and uses coarser logarithmic spacing in the tails.\n A lightweight prefill-time rule selects the compander strength per layer and cache type from scale-free statistics and reuses it during decoding. \n Across reasoning, code generation, and long-context understanding, $\\mu$-Zero achieves the best quantized results across all evaluated 2-4-bit settings against representative scalar, transform/sketch-domain, and mixed-precision baselines. \n Further evaluations support long-output reasoning and applicability to models up to 70B and selected MQA/MLA architectures. \n System-level evaluations show up to 2.38 $\\times$ higher throughput and 69% lower peak memory usage than BF16, while extending feasible batch size by at least 3.3$\\times$ and context length by 4.7$\\times$.\n Code available at: https://github.com/muZeroQuant/muZero.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "kv-cache", "compression", "LLM" ], "forum_url": "https://openreview.net/forum?id=CQDc2CcjFm", "pdf_url": "https://openreview.net/pdf/99ef2797025961085f638f9a71efbe8e3abd77b6.pdf", "updated_at": "2026-10-05T17:08:08.554Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "4-bit", "low-bit", "mixed-precision", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XFg3wEzJXQ", "number": 4345, "title": "ASQ: An Improved Approach to Unifying Sparsification and Quantization of LLMs", "abstract": "Sparsification and quantization are widely adopted for the efficient deployment of large language models (LLMs).\nHowever, they are often optimized independently, overlooking their intricate interactions and yielding suboptimal compression trade-offs. \nTo address this, we propose ASQ, an ADMM-based framework that jointly optimizes sparsification and quantization under a unified formulation, supported by theoretical convergence guarantees. \nAt its core, ASQ admits an exact, closed-form joint projection that concurrently resolves sparse support and quantized values by accounting for both their importance to the objective criterion and their representability on a quantization grid. \nAcross multiple model architectures and scales, ASQ substantially outperforms existing joint compression methods under aggressive compression, with the performance gap widening significantly under extreme sparsity and low-bit regimes.\nMoreover, under hardware-compatible 2:4 sparsity with INT4 quantization, ASQ achieves a 4.78$\\times$ memory reduction and up to 3.20$\\times$ end-to-end decode speedup on Qwen3-8B-Base relative to dense BF16 inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sparsity", "quantization", "model compression", "LLMs" ], "forum_url": "https://openreview.net/forum?id=XFg3wEzJXQ", "pdf_url": "https://openreview.net/pdf/6d6a820b404c4e5f06a7cafd6828c41db519d630.pdf", "updated_at": "2026-10-05T17:08:58.721Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "int4", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Tb8ZKaLFR9", "number": 4403, "title": "Corruption Robustness in Post-Training Quantization of Vision Transformers: A Calibration Perspective", "abstract": "Post-training quantization (PTQ) enables efficient deployment of pretrained models using only a small calibration set, but its robustness under distribution shifts remains underexplored.\nWe observe that quantized models retain substantially less of their full-precision performance on corrupted inputs than on clean data, particularly under low-bit quantization.\nWe study this problem from the perspective of calibration data and show that different transformations of the same fixed-budget calibration set lead to markedly different corruption robustness.\nOur analysis reveals that effective transformations tend to reduce activation energy, exhibiting a characteristic energy shift similar to that observed under common corruptions.\nWe associate this reduction with two changes in self-attention: lower attention concentration and weaker directional alignment among value vectors associated with different tokens.\nMotivated by these observations, we introduce *spatial entropy of difference energy*, a model-independent score that measures how broadly a transformation distributes its changes across an image.\nThis score strongly correlates with both activation-energy reduction and post-quantization corruption robustness, enabling transformation selection without model inference.\nBased on this proxy, we propose **Entropy-$K$**, which selects multiple high-scoring transformations while preserving the original calibration budget.\nEntropy-$K$ requires neither corrupted target-domain samples nor modification of the underlying PTQ algorithm.\nExperiments across multiple Vision Transformer architectures and low-bit PTQ settings on ImageNet and ImageNet-C demonstrate consistent improvements in corruption robustness while maintaining competitive clean accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Robustness" ], "forum_url": "https://openreview.net/forum?id=Tb8ZKaLFR9", "pdf_url": "https://openreview.net/pdf/49aa283c46db3b4e1380340fb6a53f7a909efedf.pdf", "updated_at": "2026-10-05T17:09:13.715Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tyokGv9qrV", "number": 4712, "title": "LoadMatch: Balancing Memory and Compute Loads in Quantized Models using Speculative Decoding", "abstract": "Quantization has become a standard practice for efficient large language model (LLM) inference. However, the exact mechanism behind this acceleration is often misunderstood. A common misunderstanding is that quantized weights directly speed up arithmetic operations: in reality, low-bit GEMM still incurs higher-precision dequantization, rescaling, and accumulation, so its compute-side speedup is much smaller than its memory-side gain. Consequently, the computation time itself remains largely unchanged. However, the perceived speedup stems from the efficiency gained by reducing the transfer time of quantized weights from the GPU memory to its cache. Because of this, quantized LLMs achieve significant speedups only in memory-bound scenarios, whereas their improvements in compute-bound regimes are marginal. To truly maximize the decoding throughput of quantized LLMs, a system must achieve a load-matched state where the memory transfer duration and computation time are equal so that neither the memory bus nor the compute units remain idle. To reach this state, we propose LoadMatch, which balances memory and computation costs using speculative decoding. We leverage speculative decoding, in which a draft model predicts multiple future tokens, called candidate tokens, and the target model verifies them in parallel. By precisely controlling the candidate-token count processed at once, our method deliberately matches the memory and computation workload. Empirically, LoadMatch compounds the speedups from quantization and speculative decoding, delivering a measured end-to-end speedup of $8.8 \\times$ over FP16 autoregressive greedy decoding on standard benchmarks. We will release the project upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Dequantization Overhead", "Memory-Compute Load Match", "Speculative Decoding", "Speculative Parallelism Throttling" ], "forum_url": "https://openreview.net/forum?id=tyokGv9qrV", "pdf_url": "https://openreview.net/pdf/178dd1ddb1876e42434cc3e326ddd2d14c8acdba.pdf", "updated_at": "2026-10-05T17:12:03.181Z", "categories": [ "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "G7btG9LR60", "number": 4826, "title": "PrefLUT: Reusable and Refinable Personalized Color Editing from Pairwise Preferences", "abstract": "Photographic color editing is inherently personal: the same image can appear too warm, too muted, or already satisfactory to different users. Most lookup table (LUT) and reference-guided methods target a specified appearance rather than model persistent preferences from repeated user choices. To address this gap, we introduce PrefLUT, a reusable and refinable user-preference modeling framework for deployable 3D LUTs, encoding ordered preferred/non-preferred image pairs into a lightweight Reusable User Profile that is reused across queries and refined using additional user preference pairs, without per-user optimization. A Query-Conditioned LUT Predictor combines this profile with each image to predict a LUT latent vector and edit strength. An Identity-Residual LUT Decoder and Edit-Strength Controller then produce an exportable 3D LUT. Experiments on three datasets demonstrate effective personalized editing and general-purpose enhancement. Each quantized profile requires only 260 bytes, and editing takes 1.365 ms/image on an RTX 5090 GPU. We also introduce the Preference-Conditioning Verification Protocol (PCVP), an evaluation protocol to verify whether personalized image edits depend on user preferences and the query image through controlled changes to user profiles, preference orders, pair correspondences, and query images.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Color Grading", "Color Editing", "Image Enhancement", "User Preference Modeling" ], "forum_url": "https://openreview.net/forum?id=G7btG9LR60", "pdf_url": "https://openreview.net/pdf/48e3b59c53c9c4ad86fbbd1d75ba1d88fd5baa1c.pdf", "updated_at": "2026-10-05T17:12:53.989Z", "categories": [ "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "K6mOMIOghl", "number": 4854, "title": "SAGE: Pushing 2-Bit Vector-Quantized LLMs toward Real-World Usability", "abstract": "Extreme low-bit weight-only quantization is an effective approach to reducing the deployment costs of large language models.\nBelow 3 bits, scalar quantization often suffers severe accuracy loss, while vector quantization (VQ) better preserves accuracy through more expressive joint encoding.\nHowever, existing VQ methods still exhibit substantial degradation on real-world tasks, and two design challenges remain underexplored: (1) protecting sensitive weight channels within jointly encoded vectors; and (2) allocating a fixed bit budget among representation components, such as direction and radius, to suit different weight matrices.\nTo address these challenges, we propose $\\textbf{SAGE}$: $\\textbf{S}$plit $\\textbf{A}$daptation and $\\textbf{G}$rouping by $\\textbf{E}$nergy for weight-only vector quantization.\nFirst, Energy-Stratified Grouping (ESG) couples channel grouping with activation-weighted fitting, grouping channels from different energy tiers within each subvector to prioritize sensitive channels.\nSecond, Adaptive Bit Split (ABS) balances direction-codebook capacity and radius precision per matrix, selecting the split with the lowest activation-weighted reconstruction error under a fixed budget.\nAcross benchmarks covering knowledge, mathematics, code generation, instruction following, and tool use, SAGE achieves state-of-the-art performance in the 2-bit quantization setting.\nOn Qwen3-30B-A3B, it retains 95.3\\% of the BF16 model's average benchmark score.\nWith our custom-designed CUDA kernel, SAGE delivers up to a 2.55$\\times$ non-streaming end-to-end speedup over FP16. We will release our code, models, and kernels.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Extreme Low-Bit Quantization", "Weight-Only Quantization", "Vector Quantization", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=K6mOMIOghl", "pdf_url": "https://openreview.net/pdf/c1d4a87cdcab5bc8bcc8db04ca4d5ed1d8575b3c.pdf", "updated_at": "2026-10-05T17:12:56.487Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "09gtv3rua8", "number": 4873, "title": "Calibrate the Decisions That Change the Future: On-Policy Post-Training Quantization for Multimodal Large Language Models", "abstract": "Post-training quantization (PTQ) lowers deployment cost for multimodal large\nlanguage models, but calibration typically reconstructs fixed sequences with\nlocal objectives. This overlooks autoregressive feedback: a quantization-induced\ntoken change redirects the prefix and changes future states. Yet on-policy\ncoverage alone is insufficient because many decision mismatches barely affect\nfuture generation. We propose OnPTQ, an on-policy framework that calibrates\non trajectories visited by the current quantized policy. On shared prefixes,\nOnPTQ identifies quantization-eroded boundaries, evaluates competing tokens\nthrough short counterfactual rollouts, and combines current discrepancy with\nbranch consequence into a Decision--Consequence risk. The risk prioritizes\ncritical states, while context anchoring and trajectory refresh preserve\nmultimodal behavior and keep calibration aligned with the updated policy. We\nfurther derive a Decision--Consequence bound linking behavioral deviation to\ncurrent policy discrepancy and action-conditioned future-value span. Across\nvision--language and omni-modal Qwen models under multiple low-bit settings,\nOnPTQ improves downstream performance and yields fewer correctness flips\nagainst the corresponding Dense/FP16 references, without changing the deployed\ninference graph.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "On-Policy Learning", "Vision-Language Models", "Behavioral Robustness", "Counterfactual Reasoning" ], "forum_url": "https://openreview.net/forum?id=09gtv3rua8", "pdf_url": "https://openreview.net/pdf/66481e03cf7755fa46ef913ff797abc1dcf5673c.pdf", "updated_at": "2026-10-05T17:12:59.049Z", "categories": [ "ptq", "hardware", "reasoning", "multimodal", "theory" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9RG3pdm8Zw", "number": 4949, "title": "Rescue Vector: Bringing Quantized Reasoning Back on Track", "abstract": "Low-bit quantization substantially reduces the storage and inference costs of large language models, but can cause pronounced degradation on reasoning tasks. Existing studies have mainly characterized this degradation through task-level performance or sought to mitigate it through improved quantization methods, leaving how quantization changes the reasoning process and internal representations less understood. In this work, we compare paired responses generated by the same model in its unquantized and quantized forms. We find that the two models often follow comparable reasoning initially, but the quantized model can deviate at critical intermediate steps and consequently produce incorrect answers. We further examine their hidden representations and find that the differences between the unquantized and quantized models exhibit a consistent directional pattern across failure cases. Motivated by this finding, we propose **Rescue Vector**, which aggregates these hidden-state differences to correct the intermediate representations of the quantized model.\nRescue Vector can be applied directly to a quantized model, without retraining, modifying its quantized weights, or changing the original quantization procedure. Experiments across diverse language models and reasoning benchmarks show that Rescue Vector improves average accuracy under FP4 quantization from 78.76\\% to 81.16\\%, recovering 77.2\\% of the performance lost after quantization. On Llama-3.2-3B-Instruct, the recovery reaches 93.2\\%. Further analysis shows that, in rescued cases, the corrected hidden states move closer to those of the unquantized model. Together, these findings provide a new perspective on reasoning degradation under quantization from the perspective of internal representations and suggest a practical way to mitigate it.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "low-bit quantization", "reasoning", "representation analysis", "representation intervention" ], "forum_url": "https://openreview.net/forum?id=9RG3pdm8Zw", "pdf_url": "https://openreview.net/pdf/ad04bf99d73af65f82719dbf57b15f98da929d95.pdf", "updated_at": "2026-10-05T17:13:09.212Z", "categories": [ "training", "reasoning" ], "matched_terms": [ "fp4", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "57WmEFt3Qx", "number": 5029, "title": "High-Degree Node Energy Smearing for Graph Neural Networks quantization", "abstract": "Graph Neural Networks (GNNs) achieve state-of-the-art accuracy on irregularly structured data, yet efficient deployment on resource-constrained platforms remains challenging. This challenge is growing in importance as GNNs are increasingly incorporated into real-world products. In recent years, graphs underlying real-world applications have increased dramatically in size, making it necessary to operate under strict energy, latency, and memory constraints. Quantization has emerged as a promising approach to address these challenges. However, its accuracy degradation in message-passing GNNs is often severe. We show that this degradation is fundamentally topological: activation energy grows with node degree, causing hub nodes to dominate the dynamic range and reducing quantization resolution. Existing methods treat quantization error as a calibration or smoothness problem and fail to address this deterministic range explosion. We propose High-Degree Node Energy Smearing (HD-NES), a topology-aware method that applies lightweight, reversible rotation transformations to diffuse the feature energy of hub nodes. By reducing activation variance while preserving information via reversibility, HD-NES mitigates hub-induced outliers and enables accurate low-bit quantization. HD-NES integrates into existing GNN quantization pipelines without architectural changes and consistently outperforms prior approaches.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Graph Neural Networks", "Quantization" ], "forum_url": "https://openreview.net/forum?id=57WmEFt3Qx", "pdf_url": "https://openreview.net/pdf/f7850f6c44e472206e68e46f00af45260c6ec567.pdf", "updated_at": "2026-10-05T17:13:20.273Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "InkbogQZ19", "number": 5232, "title": "Deep Interest Mining for Intent-Enriched Semantic IDs in Multimodal Generative Recommendation", "abstract": "Semantic IDs (SIDs) provide the discrete item vocabulary used by generative recommendation, but their quality depends on\nwhat item evidence is preserved before quantization. In product recommendation, surface metadata often misses latent usage\nintent, visual evidence may be only weakly reflected in text, and downstream policy learning provides sparse feedback about\nwhether a generated SID corresponds to a semantically useful item. We introduce DeepInterestGR, an intent-enriched SID\nframework for generative recommendation. Before SID quantization, CMSA enriches item representations through two\ncomplementary evidence paths: recommendation-oriented VLM captions and projected image embeddings. DCIM then uses an LLM to\nmine item-side intent descriptors, which capture latent usage motivations implied by product content rather than\npersonalized user states. During policy training over the constructed SIDs, QARM adds a target-gated descriptor-quality\nbonus on top of standard SID rewards, applying the bonus only when the generated SID decodes to the target item. Thus,\ndescriptor quality cannot reward a fluent but irrelevant item prediction. Experiments on three primary Amazon Product\nReview categories show that DeepInterestGR improves over competitive generative and RL-based baselines, with relative gains\nof up to 15.1% in NDCG@5 and 13.9% in NDCG@10 over the strongest per-metric baseline; additional category-level checks on\ntwo more categories provide a compact robustness check without expanding the claim to full-catalog generalization.\nComponent ablations, CMSA branch analyses, reward variants, and SID-level case studies support a bounded claim: enriching\npre-quantization item evidence with visual cues and item-side intent descriptors, together with target-gated\ndescriptor-quality weighting, improves SID-based generative recommendation under the evaluated settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Generative Recommendation", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=InkbogQZ19", "pdf_url": "https://openreview.net/pdf/3aa5f84558a2fcb88210325761fe6aa2433316b4.pdf", "updated_at": "2026-10-05T17:15:03.287Z", "categories": [ "multimodal", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vp7YLGcPO4", "number": 5242, "title": "Think Global, Act Local: Efficient KV Cache Quantization via Global Prefill Calibration of Early Decoding", "abstract": "Efficient mixed-precision KV cache quantization requires identifying which cached tokens will remain important during future decoding, before aggressively compressing them. To address this challenge, we propose \\textbf{ForeKV}, an efficient KV cache quantization framework that estimates future token importance from early decoding queries and calibrates this local estimate with global query information from prefill. Our key insight is that the first few decoding queries provide the earliest direct evidence of the model's emerging attention behavior, but such evidence is limited by the short observation window. Therefore, we delay compression for a few decoding steps to collect real decoding queries, estimate token importance from these local observations, and calibrate the estimate using the mean query over the global prefill sequence. The resulting query directly scores cached keys to guide mixed-precision allocation without materializing the full attention matrix. We further design a paged attention kernel for K caches with token-wise K4/K2 allocation and channel-wise quantization, enabling efficient decoding directly over the compressed cache. Experiments across multiple LLM backbones on LongBench show that ForeKV remains close to BF16 at an effective compressed-cache storage ratio of approximately $5.0\\times$, with the average score dropping by no more than approximately 1\\%. On Qwen3-8B, our end-to-end ForeKV implementation achieves up to $3.1\\times$ the generation throughput of the BF16 baseline under the same 80\\,GB GPU memory budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "long-context inference", "mixed-precision quantization", "efficient LLM inference" ], "forum_url": "https://openreview.net/forum?id=vp7YLGcPO4", "pdf_url": "https://openreview.net/pdf/6f3a99422c8dd6d296c2dcab14ed1f3be616c80c.pdf", "updated_at": "2026-10-05T17:15:11.184Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8MZLDC9kv1", "number": 5288, "title": "SOLM: SELF-ORGANIZING LATTICE MEMORY FOR LLM AGENTS", "abstract": "Long-term memory for LLM agents requires imposing structure on an evergrowing\nhistory, but several recent memory architectures pay for that structure\nwith real, per-message LLM calls: A-Mem generates a structured note and\nre-evaluates evolution for a set of neighboring memories on every insertion,\nMemTree re-aggregates content across an average of over three ancestor nodes\nper insertion, and Zep extracts entities, relations, and deduplicates edges via multiple\nLLM calls per episode. This construction cost scales with memory volume,\ntying a deployment’s throughput directly to model inference capacity.\nSOLM’s structure-building (bottom-up placement, quantization-error-driven splitting,\ncross-branch weak links) needs zero LLM calls, computed from embeddings\nalone. Its only LLM usage: one fixed call per query for generation, and one optional,\ncheap small-model call per message for importance scoring; this is the\nsole cost scaling with volume, far lighter than A-Mem, MemTree, or Zep’s perinsertion\ncosts. Weak links directly address what we call aggregation dilution: a\ncluster’s centroid can fail to resemble a query even when a member buried inside\nit would match well, causing that whole branch to be wrongly discarded. (A nongenerative\nreranking step narrows and reorders an already-retrieved pool at query\ntime and is not counted here; see Section 3.5.)\nOn LongMemEval-S, a controlled ablation (mean over 3 independent runs) shows\nweak links yield a relative accuracy improvement of 7.7% (68.6% to 74.0%) at\nmatched settings; at nearly identical retrieved-context size, SOLM outperforms a\nflat similarity-retrieval baseline by a relative 13.8% (mean over 3 runs: 76.0% vs.\n66.7%); and accuracy scales predictably with the retrieved-context budget: 76.0%,\n83.0%, and 86.0% when the top-5, top-10, and top-15 retrieved memories, respectively,\nare used as answer context. These results show that competitive long-term\nmemory accuracy does not require scaling LLM usage in proportion to memory\nvolume.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Long-term Memory", "Memory", "Gen AI", "Agentic AI", "Agentic Memory", "RAG" ], "forum_url": "https://openreview.net/forum?id=8MZLDC9kv1", "pdf_url": "https://openreview.net/pdf/68702110dd1c9785e4c5ba472e741ac05b251a16.pdf", "updated_at": "2026-10-05T17:15:33.239Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1ivz5AasHb", "number": 5322, "title": "Beyond Suppression: Measuring Concept Erasure Durability in Diffusion Models", "abstract": "This paper first investigates a vulnerability that quantization can reverse concept erasure in diffusion models. \nTo this end, we first quantize diffusion models edited by different concept-erasure methods, and the results demonstrate that the target concept remains suppressed after quantization. However, the quantized erased model exhibits a novel vulnerability: restoring just 5\\% of the edited weights to their original values can markedly increase the generation rate of the erased concept.\nIn this paper, we refer to the ability to keep suppression under such restoration as erasure durability and evaluate it by tracking target generation across restoration budgets. This yields a recovery curve showing how target generation changes as the restoration budget increases. We summarize this curve with the half-recovery budget, defined as the minimum restoration needed to raise the target-generation rate to the midpoint between its erased-model baseline and fully restored level.\nBased on this, we construct DuraBench, a durability-based benchmark that distinguishes erasure methods with similar initial suppression but different resistance to weight restoration. We further propose SPARC, a post-training method that helps erased models maintain target suppression when some edited weights are restored.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Concept Erasure", "Robustness", "Quantization", "Diffusion Models" ], "forum_url": "https://openreview.net/forum?id=1ivz5AasHb", "pdf_url": "https://openreview.net/pdf/499236139a7e36141561447ae75b328de56460e2.pdf", "updated_at": "2026-10-05T17:16:01.003Z", "categories": [ "ptq", "diffusion", "theory" ], "matched_terms": [ "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mHLERvJE8P", "number": 5354, "title": "Gamma-Aware Block Rotations for MXFP4 LLM Quantization", "abstract": "Rotation-based post-training quantization reduces outliers by inserting orthogonal transforms and fusing them into adjacent weights. For the residual-stream rotation R1, however, the learned scaling parameter $\\gamma$ of RMSNorm does not commute with R1 in general, so existing offline methods first fold it into the weights of subsequent linear layers. This operation\ncan amplify magnitude disparities within quantization blocks and degrade the quantized model. We propose Gamma-aware Block Rotation (GaBR). GaBR first permutes channels to group similar $|\\gamma|$ values within quantization blocks. It then decomposes the scaling parameters into a block-constant $G$ retained in RMSNorm and a residual $D$ fused into weights, and alternately\noptimizes $G$ and a block-diagonal R1. With a $32\\times32$ block-diagonal rotation, retaining one scale per 32 channels in RMSNorm remains exactly equivalent in full precision. Extensive experiments on seven Llama and Qwen models under MXFP4 W4A4 quantization show that GaBR consistently improves upon existing rotation-based PTQ methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "MXFP4", "rotation-based quantization", "block-scaled quantization" ], "forum_url": "https://openreview.net/forum?id=mHLERvJE8P", "pdf_url": "https://openreview.net/pdf/8fc629734d7143418896a7e89b5136bff2fc8402.pdf", "updated_at": "2026-10-05T17:16:11.754Z", "categories": [ "ptq" ], "matched_terms": [ "mxfp4", "ptq", "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "geAguvzv84", "number": 5376, "title": "SQGen: Structured Quantum Image Generation with Latent-Modulated Quantized Tensor Trains", "abstract": "Generating images directly from quantum systems is an attractive but unresolved goal on NISQ hardware. Existing quantum generators face several coupled obstacles: barren plateaus that block trainability, expensive quantum circuit preparation, and hardware noise that erodes quantum information with depth. A further difficulty is generating image-scale outputs without a classical decoder. We propose SQGen, a decoder-free quantum generator whose circuit follows a quantized tensor train (QTT) skeleton with latent modulation. Specifically, SQGen promotes the QTT bond index of the target pixel distribution to ancilla bond qubits, so that each circuit site acts on a bond register, the two physical qubits that carry the row- and column-bit of one image scale, and a coarse-to-fine link to the previous scale. The latent enters through latent modulation: an exact angle-level factorization of each re-uploading rotation into a trainable main path plus an additive latent term, reducing to the main path when the latent term is disabled. During training, we create a differentiable model in the classical system under gate-compatibility constraints, with a torus prior as the latent distribution. Every trained operator maps one-to-one to a native quantum gate, yielding a compact, deployable quantum circuit with no classical decoder in the inference path. Extensive experiments on image datasets and synthetic data demonstrate that SQGen trains stably, generates images end-to-end from a shallow circuit with no classical decoder, and shows promising feasibility on real quantum hardware.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantum machine learning", "Quantum image generation" ], "forum_url": "https://openreview.net/forum?id=geAguvzv84", "pdf_url": "https://openreview.net/pdf/15d7211ed9d3a136dfc148b995b2d7864abe5175.pdf", "updated_at": "2026-10-05T17:16:19.079Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "k3BcJBjXro", "number": 5627, "title": "IDENTITY RECOVERY FROM REPEATED CAMPRO OUTPUTS UNDER SYNTHETIC ACQUISITION NOISE", "abstract": "Privacy-oriented imaging aims to suppress facial recognition, but repeated\nreleases may support identity recovery. We examine identity recovery from\nrepeated outputs of the CamPro image signal processor (ISP), using independent\nsynthetic pre-ISP noise and fixed, aligned face images. Comparisons include\nseparately selected single-output attacks, reconstruction averaging, joint\ninversion and matched pixel-mean inversion, alongside color-channel and\nfixed-weight detection analyses. Eight observations improve FaceScrub\nclosed-gallery Rank-1 by 14.6--20.3 percentage points over the selected\nsingle-output attack; NTU gains are less consistent. The FaceScrub gain persists\nafter 8-bit lossless export, while strict open-set acceptance remains limited\nunder the tested calibration. A 500-identity synthetic study reports mean gains\nof 5.9--9.2 points across recognizers over three independently trained and\nselected fits; averaging recovers much of this gain. Joint inversion outperforms\nmatched pixel-mean inversion on FaceScrub. Eight of twelve constructed color\npairs with identical noiseless outputs have distinguishable noisy output\ndistributions. Added noise lowers native person AP from 0.518 to 0.505; task and\nidentity inputs use different spatial preprocessing. These results motivate\nrepeated-release privacy evaluation and show that deterministic ambiguity need\nnot survive random acquisition.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Privacy-preserving imaging", "Identity recovery", "Repeated observations", "Image signal processing", "Set inversion", "Synthetic acquisition noise" ], "forum_url": "https://openreview.net/forum?id=k3BcJBjXro", "pdf_url": "https://openreview.net/pdf/32a60359ffca02296e1ec3eec5bbc4e27a8e1463.pdf", "updated_at": "2026-10-05T17:19:36.100Z", "categories": [ "ptq" ], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gXM4AMGLZF", "number": 5685, "title": "RATIO: Reasoning Analysis and Token-level Inference Optimization for Quantized Reasoning Models", "abstract": "Post-training quantization (PTQ) has become a widely adopted technique for reducing the memory footprint and inference cost of large language models (LLMs). However, recent studies reveal that when applied to reasoning models, PTQ not only degrades reasoning performance but also exacerbates overthinking, leading to longer reasoning trajectories. These issues may offset the efficiency gains expected from lower-precision inference. Existing approaches to mitigating these problems mainly rely on complex optimization procedures. More recent lightweight inference strategies instead use predefined overthinking markers, limiting their adaptability across quantized models. To address these issues, we propose **R**easoning **A**nalysis and **T**oken-level **I**nference **O**ptimization (RATIO), a framework that identifies model-specific overthinking tokens and assigns each a tailored penalty. RATIO first introduces Quantization-aware Reasoning Behavior Analysis (QRBA) to identify overthinking tokens by analyzing discrepancies between full-precision and quantized models. It then adopts Token-Specific Penalty Determination (TSPD), which leverages full-precision guidance to derive token-specific penalties without additional training. Extensive experiments show that RATIO achieves a better accuracy-efficiency trade-off than existing token-level interventions. Specifically, RATIO achieves up to 9.8 points accuracy improvement and reduces chain-of-thought (CoT) length by up to 51.3\\% compared with quantized baselines. We will release all the code and implementation of RATIO.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Efficiency" ], "forum_url": "https://openreview.net/forum?id=gXM4AMGLZF", "pdf_url": "https://openreview.net/pdf/540fdcc6139c7a5f3f5d834912f4b3215729ee23.pdf", "updated_at": "2026-10-05T17:20:09.618Z", "categories": [ "ptq", "qat", "reasoning" ], "matched_terms": [ "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TMJfwImQGZ", "number": 5697, "title": "Binary Semantic Indexing for Generative Retrieval", "abstract": "Retrieval systems must organize large collections of knowledge while allowing queries to reach relevant information efficiently. As retrieval increasingly relies on learned representations and generative models, the representation itself becomes central to both semantic modeling and access. This raises a broader question: can different retrieval mechanisms share a common semantic foundation? We introduce **Binary Semantic Indexing (BSI)**, a framework that connects sparse indexing and generative retrieval through a **shared binary semantic space**. A jointly trained semantic encoder and binary spherical quantizer compress each document into a compact code that preserves retrieval information. The code's local components form addressable units in an inverted index. Complete-code and local-address generators provide complementary query interfaces, while direct query encoding accesses the same representation. By **separating document organization from query modeling**, BSI allows these mechanisms to operate over a common document space. Experiments show effective retrieval across query interfaces and domains, while controlled analyses characterize the tradeoffs between code capacity, index access, and retrieval quality.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Generative retrieval", "Semantic representations", "Binary spherical quantization", "Sparse retrieval", "Semantic indexing" ], "forum_url": "https://openreview.net/forum?id=TMJfwImQGZ", "pdf_url": "https://openreview.net/pdf/1994ee8a2ff22b773cfd77b594fb354f28f6a536.pdf", "updated_at": "2026-10-05T17:20:24.532Z", "categories": [], "matched_terms": [ "binary", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pKRT8nTFj6", "number": 5723, "title": "Gefen: Optimized Stochastic Optimizer", "abstract": "AdamW is a default optimizer for modern deep learning, but its first and second moment states add roughly two parameter-sized buffers to training memory, increasing the already substantial cost of large-scale pretraining. We propose Gefen, a memory-efficient optimizer that automatically shares second-moment estimates across parameter blocks and quantizes the first moment using a learned codebook, thereby reducing AdamW’s memory footprint by up to 8× while maintaining the same performance, corresponding to a reduction of 6.5 GiB per billion parameters. Prior work reduces optimizer memory by sharing second moments across parameters grouped along the Hessian’s block-diagonal structure, but relies on hand-specified architectural rules and leaves unexplained why such grouping works. We prove that large mixed Hessian entries constrain the ratio of squared gradients toward one. This supplies the missing explanation, since a shared second moment is accurate when the squared gradients it pools are similar. It also shows the Hessian need not be computed: a block-structured Hessian forces the squared gradients to inherit that structure, so the blocks can be found in the squared gradients themselves. Gefen therefore infers block structure from the initial squared gradients, requiring no architecture-specific metadata or user-tuned hyperparameters beyond AdamW defaults. Gefen learns an exact histogram-based dynamic-programming quantization codebook and reuses the same blocks for first-moment scaling. Across diverse pretraining experiments, Gefen achieves the lowest peak optimizer memory among the compared methods that maintain AdamW-level performance. In single-machine and distributed training, the reduced footprint enables larger microbatches and substantially improves throughput over AdamW, making Gefen a drop-in replacement that can train larger models or use larger global batch sizes. We provide the complete Python implementation, including fused CUDA kernels.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "optimizer", "memory-efficient optimization", "adaptive gradient methods", "optimizer-state quantization", "deep learning optimizers", "large-scale model training" ], "forum_url": "https://openreview.net/forum?id=pKRT8nTFj6", "pdf_url": "https://openreview.net/pdf/83e7acab0dcb0055ded1236a2ad41b27dad37a1a.pdf", "updated_at": "2026-10-05T17:20:38.445Z", "categories": [ "training", "hardware" ], "matched_terms": [ "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DH1IOY829U", "number": 5743, "title": "Rank Trajectories for Adversarial Robustness via Quantized Tensor Train", "abstract": "Many defenses against adversarial examples rely on an explicit or implicit model of the natural-image distribution and are evaluated primarily by reconstruction quality. We pursue a different question: whether the multiscale tensor structure of an image provides an interpretable and computable basis for analyzing adversarial robustness. We study this question through quantized tensor-train (QTT) representations, whose modes encode spatial scales and whose bond ranks measure their interactions. Tensors satisfying nontrivial QTT-rank constraints form lower-dimensional algebraic sets, and natural images empirically concentrate near this low-rank structure. An adversarial perturbation, in contrast, comes from a search constrained only by a norm budget and a misclassification objective, and neither constraint confines it to these sets. To connect this geometry to model behavior, we introduce rank trajectories, which track predictions along progressively higher-rank reconstructions from a single decomposition sweep. Clean examples typically stabilize at moderate ranks, whereas adversarial examples often emerge later. The same geometry also has a constructive use inside randomized smoothing. A training-free filter suppresses the weak singular directions that carry most of the Gaussian smoothing noise. Experiments show that it consistently improves certified accuracy over plain smoothing in the high-noise regime. Together, these results establish QTT rank as a multiscale coordinate for analyzing adversarial robustness, and improves certified robustness under strong smoothing noise.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Robustness", "Tensor Network" ], "forum_url": "https://openreview.net/forum?id=DH1IOY829U", "pdf_url": "https://openreview.net/pdf/db2151aba69348a9fd0750fa867ce2aa99b5406b.pdf", "updated_at": "2026-10-05T17:20:51.331Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "uG51pMvR1g", "number": 5937, "title": "Distill Globally, Train Locally: Error-Localized Recovery for 4-bit LLMs", "abstract": "Low-bit quantization enables efficient large language model (LLM) inference, but maintaining accuracy with 4-bit weights and activations (W4A4) remains challenging. Local calibration in post-training quantization (PTQ) lacks direct supervision of final predictions, limiting further accuracy recovery. Full-model quantization-aware distillation (QAD) can recover near-BF16 accuracy through global supervision, but requires costly model-wide optimization. To preserve the accuracy of full-model QAD while minimizing training cost, we propose LoDiQ, a lightweight error-localized distillation framework with three components: (1) propagated-error localization identifies sensitive input channels; (2) quantized residual compensation strengthens recovery within expanded low-bit GEMMs, with Gradient-Guided Residual Compensation Allocation (GRCA) supporting lower-budget configurations; and (3) localized distillation trains only the selected corrective weights using compact teacher targets, avoiding full-vocabulary teacher-logit storage. Experiments across the Qwen and Llama model families span language modeling, zero-shot commonsense reasoning, broad-domain knowledge, and mathematical reasoning. LoDiQ surpasses state-of-the-art PTQ baselines in downstream accuracy under both MXFP4 and NVFP4 W4A4 quantization. On a matched reasoning evaluation, LoDiQ outperforms full-model QAD using only 1/100 of the distillation tokens and 4.67% of its trainable parameters, substantially reducing optimization workload and trainable-state memory. Inference benchmarks on RTX 5090 show up to 3.20× and 3.14× time-to-first-token (TTFT) speedups over BF16 in MXFP4 and NVFP4, respectively.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization", "Quantization-Aware Distillation", "Parameter-Efficient Training", "Efficient Inference", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=uG51pMvR1g", "pdf_url": "https://openreview.net/pdf/484ea322566334051b2f104d53ddcb192ad475e8.pdf", "updated_at": "2026-10-05T17:23:17.518Z", "categories": [ "ptq", "qat", "training", "reasoning" ], "matched_terms": [ "4-bit", "low-bit", "mxfp4", "nvfp4", "ptq", "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3xpzqlgCgE", "number": 5943, "title": "JustQuant: You Don’t Need Smoothing, SVD, or Rotation for 4-Bit Activation Quantization", "abstract": "Recent generative models have become increasingly powerful, but their inference\ncost continues to grow. Model quantization is a promising path for compressing\nthese models and accelerating inference. However, when quantization is pushed to\n4 bits, activation quantization becomes substantially more difficult than weight\nquantization. Recent PTQ (post-training quantization) and QAT\n(quantization-aware training) methods have made progress on 4-bit activation\nquantization by introducing smoothing, SVD branches, rotations, mixed precision,\nor advanced formats such as NVFP. Yet these operators and data types impose\ndemanding requirements on inference engines and hardware, limiting the broad\nadoption of low-precision models. **Can we quantize using only plain\nlow-bit operators?** To answer this question, we propose **JustQuant**, a simple yet effective\nframework that moves the complexity required for low-bit quantization from\ndeployment-side operators into an acceptable training process. First, we revisit\nmodel quantization from the perspective of knowledge distillation, and show that\na key reason existing PTQ and QAT methods fail is that they typically exploit\nsupervision at only a single level. We then introduce Theseus QAD\n(quantization-aware distillation), a quantization-aware distillation method\nthat progressively applies multi-level supervision, analogous to the gradual\nreplacement process in the Ship of Theseus. Extensive experiments on DiT and\ndLLM show two regimes. For smaller models, \\method{} can serve as a light\nwarm-up stage that substantially improves QAT with plain operators, while naive\nQAD can collapse in the same setting. For larger models, Theseus QAD provides a\nstronger distillation training path than ordinary QAD. Across these regimes,\nTheseus QAD improves plain quantization while avoiding the complex \noperators required by many PTQ methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Model", "Efficiency" ], "forum_url": "https://openreview.net/forum?id=3xpzqlgCgE", "pdf_url": "https://openreview.net/pdf/fb92d6de045df8daba7bbffe829bd2984b27e226.pdf", "updated_at": "2026-10-05T17:23:23.032Z", "categories": [ "qat", "training", "diffusion", "reasoning", "hardware" ], "matched_terms": [ "4-bit", "low-bit", "low-precision", "mixed precision", "ptq", "qat", "quantization", "quantize" ], "review_status": "reviewed", "summary": "JustQuant 把 plain A4 量化的难点转移到训练路径：Theseus QAD 从短段蒸馏逐步合并到长段，以 teacher activation 作为每段输入、继承上一粒度的 student 参数。固定合并宽度如 (1,2,4,8,16,28)，γ 分配各粒度训练预算而不改变宽度；这不是把 teacher/student 混合路径逐层替换。小模型之后做任务 QAT，FLUX 与 LLaDA 在预算和 batch 限制下只做 Theseus QAD。", "contribution": "Table 3 中 DiT-XL/2 的 W4A4、50K/250-step FID 从直接 QAT 的 5.15 降至 3.32，W1.58A4 从 5.57 降至 3.30。官方 Figure 3 另有实际部署：A6000、batch 16 下 JustQuant 延迟 49.5 ms，对 FP 的 245.4 ms 为 4.96 倍；模型大小 148 MB，对 FP 的 2575 MB 为 17.4 倍压缩。质量恢复不等于所有设置都最优，算子增强 RobuQ 在 W4A4/50K 的 FID 仍更低。", "limitations": "“Plain”仍需量化、packing 与低位矩阵乘内核；Figure 3 的速度限于 A6000/batch 16 和该优化预算，图未明确其延迟覆盖整个生成 pipeline。FLUX W4A4 仅用于 transformer-block linears，text encoders、VAE、边界模块保留高精度；LLaDA 实验使用 NVFP4，不能作为摆脱专用数值格式的证明。未评测自回归 LLM 或充足资源下的大模型长期端到端训练，且除 ELF 多 seed 外许多结果未量化运行间波动。附录 Table 10 的 ELF-B quantization-function 行与 W1.58A4/W4A4 列标题不一致，具体实现配置仍需澄清，未自行纠正。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "abdda216034974fe79a6c029678e92798ffb6713b95ec4ff8c88437e4358c62c", "reviewed_pdf_url": "https://openreview.net/pdf/fb92d6de045df8daba7bbffe829bd2984b27e226.pdf", "reviewed_source_updated_at": "2026-10-05T17:23:23.032Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "DiT、FLUX 与 diffusion language models 的 transformer 权重/激活；FLUX 限定 block linears", "bit_width": "DiT/ELF-B：W4A4、W1.58A4；W1.58A32 为中间/参考设置。FLUX：INT4 W4A4 group64；LLaDA：NVFP4 W4A4", "evidence": [ { "location": "PDF p.5–6，Sections 3.2–3.3 / Figure 7", "claim": "teacher-fed student 段由短到长合并，γ 仅分配预算；反向次序令 W4A4 FID 从 6.28 升至 8.70。" }, { "location": "PDF p.7，Table 3", "claim": "W4A4 的 FID-50K 为 3.32，对直接 QAT 5.15；W1.58A4 为 3.30，对直接 QAT 5.57。" }, { "location": "PDF p.7，Table 3", "claim": "W4A4/50K 下带额外算子的 RobuQ FID=2.71，低于 plain JustQuant 的 3.32。" }, { "location": "PDF p.7，Table 4 / PDF p.8，Section 4.2", "claim": "FLUX-schnell FID/GenEval=17.62/0.7149；FLUX-dev FID=20.21，但 QAD 有更高 CLIP-IQA。" }, { "location": "PDF p.8，Table 5", "claim": "ELF-B W4A4 的 Qwen3 judge gen-PPL 为 14.854±0.344，FP 为 14.729±0.152；四 seed。" }, { "location": "PDF p.8，Table 6 / Section 4.3", "claim": "LLaDA 使用 NVFP4；GSM8K greedy 为 0.6543，FP 为 0.6748；部分 transfer 指标仍略低于 QAD。" }, { "location": "PDF p.3，Section 1 / Figure 3", "claim": "A6000、batch 16 实测图：JustQuant 49.5 ms/148 MB，FP 245.4 ms/2575 MB；图示 4.96 倍速度、17.4 倍模型大小压缩。" }, { "location": "PDF p.9，Table 7 / Section 5", "claim": "仅 1% 数据供两阶段时 FID=15.12；本文未报告自回归 LLM 实验或充分资源的大模型长期训练。" }, { "location": "PDF p.16，Appendix C Q6–Q9", "claim": "大模型只做 Theseus QAD；plain 仍需低位内核，FLUX 外围保持高精度；非重复运行不估计 run-to-run variation。" }, { "location": "PDF p.15，Appendix B / Table 10", "claim": "ELF-B 的 W1.58A4 列写 INT4 weights、W4A4 列写 ternary weights，与标题相反；按原图保留为配置不一致。" } ], "results": [ "PDF p.7，Table 3：DiT W4A4、10K/50-step 的 FID/IS 为 6.48/238.31，直接 QAT 为 10.72/196.53，直接 QAD 为 188.26/3.89。", "PDF p.7，Table 3：DiT W4A4、50K/250-step 的 FID/IS 为 3.32/256.56，QAT 为 5.15/220.53；FP FID 为 2.27，带 Norm/SVD/Rotation 的 RobuQ 为 2.71。", "PDF p.7，Table 3：DiT W1.58A4、50K/250-step 的 FID/IS 为 3.30/241.95，直接 QAT 为 5.57/166.06；JustQuant 110k 训练步骤、QAT 100k。", "PDF p.7，Table 4：FLUX-schnell W4A4 的 MJHQ FID=17.62、GenEval overall=0.7149；naive INT4=18.37/0.6753。FLUX-dev 的 FID=20.21、ImageReward=0.956，FP 为 20.26/0.958；并非所有指标优于 QAD。", "PDF p.8，Table 5：ELF-B W4A4 在 Qwen3-8B judge 下 gen-PPL=14.854±0.344，FP=14.729±0.152、direct QAT=19.625±0.327；四 seed 结果。W1.58A4 的 14.626±0.482 略低于 FP 14.823±0.208，差异在所报波动内。", "PDF p.8，Table 6：LLaDA-8B NVFP4 W4A4 的 GSM8K greedy=0.6543，QAD=0.6475、PTQ=0.6035、FP=0.6748；HumanEval transfer=0.4451，略低于 QAD/FP 的 0.4512。", "PDF p.3，Figure 3：A6000、batch 16 的 FP/RobuQ/JustQuant 延迟为 245.4/82.7/49.5 ms，模型大小为 2575/195/148 MB；图以 24 小时 kernel 优化预算为背景。", "PDF p.9，Table 7：两阶段共享 1% 数据子集时 FID=15.12；Stage 1 用 5% 且 Stage 2 用全数据时报告 FID=6.59，说明 warm-up 省数据不代表最终任务阶段同样无需足够数据。" ], "reading_basis": "实际逐页阅读本次已核验的 16 页官方原稿，包含方法正文、Tables 3–7、附录 A/B/C；另查看官方 PDF p.3、7、8、15 的渲染图以核对实际部署、旋转表格与配置列。数字仅来自此次官方稿，未运行训练/内核或复现实验。" }, { "id": "Uilsbq8dfC", "number": 5949, "title": "LumaFlux: Structure-Aware SDR-to-HDR Conversion with Monotone Luminance Flows", "abstract": "The rapid adoption of HDR-capable displays has created a growing need to con-\nvert legacy and user-generated 8-bit Standard Dynamic Range (SDR) content\ninto perceptually faithful and physically plausible 10-bit High Dynamic Range\n(HDR). Existing inverse tone-mapping methods often rely on fixed tone map-\npings or narrowly defined training degradations, which can lead to clipped high-\nlights, chromatic shifts, and temporally unstable tone reproduction. We intro-\nduce LUMAFLUX, an Diffusion Transformer (DiT) conditioned directly on phys-\nical measurements of the SDR input, while training only 0.57% of total param-\neters. LUMAFLUX combines Physically-Guided Adaptation (PGA), Perceptual\nCross-Modulation (PCM), an HDR Residual Coupler, and a monotone Rational-\nQuadratic Spline (RQS) tone-field decoder to recover HDR appearance while\npreserving the luminance ordering of the input. HDR latents are reconstructed\nfrom an SDR-conditioned endpoint through direct rectified-flow transport in only\n8 steps. The same image model extends to video inverse tone mapping without\ntemporal modules or video-specific training, shared bridge noise and temporally\nsmoothed RQS parameters stabilize deterministic framewise transport. We further\ncurate a 314,396-pair UGC and PGC training corpus in PQ BT.2020 and introduce\nLumaEval, a unified evaluation protocol spanning eight SDR degradation types.\nOn LumaEval, LUMAFLUX achieves 24.23 dB PSNR and 56.54 ∆EITP, improv-\ning over the strongest baseline by 0.49 dB and 5.06, respectively, while reducing\nFR-HIDROVQA from 0.703 for HDRTVNet++ to 0.631. For video, temporal\nstabilization reduces signed excess flicker by 53.3%, while pooled raw flicker re-\nmains within 3.3% of the HDR reference. Code will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Generation", "HDR", "Inverse Problem", "Diffusion Transformer", "Inverse tone mapping", "HDR reconstruction", "rectified flow", "parameter-efficient adaptation", "video temporal consistency" ], "forum_url": "https://openreview.net/forum?id=Uilsbq8dfC", "pdf_url": "https://openreview.net/pdf/d4686e1b9a1a8fc932ee1cef2da16e75dd83ee6d.pdf", "updated_at": "2026-10-05T17:23:29.312Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UkH0vfqnDF", "number": 6026, "title": "OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond", "abstract": "The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. Extreme low-bit quantization has emerged as a fundamental imperative to reclaim memory efficiency and sustain high-throughput inference. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0× speedup in decoding, reduces memory footprint by 5.3×, and increases throughput by 4.1×. Code for result reproduction is provided in the anonymous supplementary materials.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "low-bit quantization", "efficient inference", "large language models", "multimodal models", "token norm imbalance" ], "forum_url": "https://openreview.net/forum?id=UkH0vfqnDF", "pdf_url": "https://openreview.net/pdf/9d0076c693a2f273d840699aed98a950aad74c6f.pdf", "updated_at": "2026-10-05T17:24:19.334Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning", "multimodal", "theory" ], "matched_terms": [ "int2", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nhL3D1XE2R", "number": 6027, "title": "QuantMLA: Function-Aligned Dual-Path Quantization for Low-Bit MLA KV Caching", "abstract": "Multi-Head Latent Attention (MLA) enables expressive multi-head attention with compact caches for its content and decoupled RoPE paths, yet cache memory still scales linearly with context length and batch size. Existing methods primarily quantize the content cache to FP8 while retaining the RoPE key cache at high precision. Low-bit RoPE quantization remains poorly understood, leaving joint low-bit compression of the content and RoPE caches largely unexplored. In this work, we establish a systematic model of MLA's dual-path quantization errors, characterizing their distinct effects on attention-output distortion and explaining the pronounced amplification of RoPE-path errors. Guided by this analysis, we introduce QuantMLA, a function-aligned framework for low-bit dual-path quantization. We derive path-specific transformation spaces that preserve full-precision computation while remaining fully fusible into model parameters offline, eliminating online transformation overhead. Within these spaces, QuantMLA learns path-specific transformations with function-aligned objectives: attention-output reconstruction captures the content path's coupled matching and aggregation errors, while positional QK reconstruction preserves the RoPE-induced component of the attention logits and admits a theoretical bound on output distortion. Across four MLA model families, QuantMLA enables, to our knowledge, the first reported joint INT4 caching of the content and RoPE caches with minimal accuracy degradation. Further compressing the content cache to INT2 while retaining the RoPE key cache at INT4 maintains competitive performance on challenging reasoning and code benchmarks. We develop a native low-bit MLA attention kernel that integrates unpacking and dequantization directly into attention computation. The physical cache layout provides 3.59× compression at 128K context, while a cache-pressure serving workload achieves 5.168× higher whole-job output throughput than BF16. Our code is provided in the supplementary material.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multi-Head Latent Attention", "KV Cache Quantization", "Low-Bit Quantization", "Model Compression", "Efficient LLM Inference" ], "forum_url": "https://openreview.net/forum?id=nhL3D1XE2R", "pdf_url": "https://openreview.net/pdf/edb4d9d1e065c40550e876a415ddd702950a98f5.pdf", "updated_at": "2026-10-05T17:24:14.635Z", "categories": [ "kv-cache", "theory", "hardware" ], "matched_terms": [ "fp8", "int2", "int4", "low-bit", "quantization", "quantize" ], "review_status": "reviewed", "summary": "相同缓存重建误差在 MLA 的内容与 RoPE 路径会造成不同输出损失。QuantMLA 分别学习可离线融合的旋转和缩放：内容路径重构注意力输出，RoPE 路径重构位置 QK，并把后者限制在与频率对旋转兼容的空间内。原生 FlashMLA 后端在注意力 tile 内解包和反量化，保留部分 BF16 token，同时维持全历史的统一 softmax。", "contribution": "将双路径功能误差分析、等价变换和实际打包缓存执行组合起来。预训练参数冻结，每路径用 128 条 WikiText-2 序列优化 300 步，以 32 条独立序列选 checkpoint，推理无需新增在线变换。原稿声明代码放在补充材料中，本次未运行该代码。", "limitations": "服务比较是一次 C4R4 运行对历史 BF16 基线，无独立重复置信区间；负载的总 token 需求略超 BF16 容量，不能把 5.168× 推广到没有缓存压力的服务。GPU 型号未说明，原生正确性测试使用独立页，未验证共享前缀所有权。SmoothQuant/QuaRot 是作者的 MLA 适配，精度表的比较不能视为完整原算法重现；3.59× 也不是全模型峰值显存下降。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "fcc289b71edf2e0e4cc08623381eaee38e36548b5a18cb1c3d2be9dd944ec48b", "reviewed_pdf_url": "https://openreview.net/pdf/edb4d9d1e065c40550e876a415ddd702950a98f5.pdf", "reviewed_source_updated_at": "2026-10-05T17:24:14.635Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "MLA 的共享内容 latent 缓存与解耦 RoPE key 缓存；系统测试压缩 KV 缓存，模型权重仍为 FP8、激活为 BF16。", "bit_width": "主设置 C4R4（内容 INT4、RoPE INT4），另测 C2R4；per-token 非对称 affine、group 64。原生 C4R4 部署保留前 4 个 sink 和最近 128 个 token 的双路径 BF16 缓存。", "evidence": [ { "location": "PDF p.3–4，§2.1–2.3，Eq.1–5", "claim": "内容误差同时影响匹配与聚合，RoPE 误差影响位置 logits；相同缓存 NMSE 不能代表相同功能误差。" }, { "location": "PDF p.5–6，§3.1–3.3", "claim": "内容旋转/缩放及 RoPE 频率对兼容变换可离线融合；内容用输出重构，RoPE 用位置 QK 重构。" }, { "location": "PDF p.7，Table 1 / §4.1；p.21，Table 7", "claim": "DeepSeek-V2-Lite 的十任务均值 BF16 66.80%、C4R4 66.54%；每路径 300 步，128/32 条训练/验证序列，group 64。" }, { "location": "PDF p.6，§3.3；p.22–23，Appendix C.5", "claim": "每个 token 有 packed 副本；前 4 和最近 128 个 token 另保留预量化 BF16；不淘汰历史 token，保护与量化项共享全局 softmax。" }, { "location": "PDF p.23，Appendix C.5，Eq.40", "claim": "持久 KV 分配压缩在 128K 为 3.59×，计入保护缓冲和映射，但排除模型权重等全局及临时开销。" }, { "location": "PDF p.25–27，Appendix F.1 / Table 9", "claim": "注意力计时为 attention-main 至 split-result combination，包含解包/反量化，排除 writer/metadata；64K 起相对 BF16 的延迟绝对差小于 0.5%。" }, { "location": "PDF p.26–27，Appendix F.2 / Table 10", "claim": "八 GPU、五个 128K 输入/1K 输出请求：吞吐 2.614→13.512 tokens/s，抢占 32→0；单次低比特运行对历史 BF16，无独立重复 CI，5.168× 不是注意力 kernel speedup。" }, { "location": "PDF p.24–25，Appendix D.2；p.1，Abstract", "claim": "比较方法是作者的 MLA-specific SmoothQuant/QuaRot 适配；摘要声明代码在 supplementary material 中提供。" } ], "results": [ "Table 1：DeepSeek-V2-Lite 十任务等权均值，BF16 66.80%，C4R4 QuantMLA 66.54%；同精度 RTN 61.02%、作者适配的 SmoothQuant 63.44%、QuaRot 64.67%。C2R4 QuantMLA 为 64.42%。", "§4.4 / Appendix C.5：sink4/local128 的持久 KV 分配在 4K/128K 上下文分别缩小 3.22×/3.59×；计入 BF16 缓冲、冗余 packed 副本和映射，排除模型权重、共享页表、分配器开销及临时 workspace。", "Table 9：batch 8/32、1K–1M context 的原生注意力延迟相对 BF16 为 −7.42% 至 +4.29%，64K 起绝对差小于 0.5%。每项来自两次独立运行、120 个样本；计时包括解包/反量化，排除 cache writer 与 metadata construction。", "Table 10：八 GPU 上 DeepSeek-R1-0528（两组均 FP8 权重、BF16 激活），五个同时到达、各 131072 输入及 1024 输出的请求，整批输出吞吐由 2.614 增至 13.512 tokens/s（5.168×）；抢占 32→0，重算 tokens 3996255→0。这是避免缓存容量不足重算的收益，不是 kernel 快五倍。" ], "reading_basis": "实际阅读本次取得的 27 页官方投稿 PDF：正文 §2–4、Tables 1–3、Appendix B 的误差模型与干预、C 的变换融合/量化/缓存生命周期、D/E 的评估与基线、F 的内核和服务协议。另直接查看 PDF p.27 图像核对 Tables 9–10；本笔记仅依据该官方原稿，数字为作者报告，未独立复现。" }, { "id": "STszbpLa5G", "number": 6081, "title": "QuoVLA: Quotient Space for Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models adapt pretrained visual-language features for robot control through action supervision. Action supervision can guide task-specific feature learning and the suppression of action-irrelevant variation. To limit this variation, we introduce QuoVLA, which combines joint feature learning with a quantized prefix adapter. The adapter processes visual features and language embeddings before a shared language backbone and action expert, using flow matching and relative temporal-complexity regularization. This regularization uses a stop-gradient raw-prefix reference to discourage excess temporal variation in the flow predictions of the learned policy. We interpret this policy through an action quotient that groups representations with the same chosen optimal action law. For this fixed action law, a conditional characterization defines an ideal compression target without assuming sufficient pretrained features or guaranteeing that training attains the target. The target motivates paired representation diagnostics alongside task success evaluation. Task success improves over the primary baseline on four simulation benchmarks and four real-robot tasks, while the diagnostics support selective compression under the evaluated perturbations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision language action" ], "forum_url": "https://openreview.net/forum?id=STszbpLa5G", "pdf_url": "https://openreview.net/pdf/6ca9cc8a74847c6541cb564658ddc16d6e83ef53.pdf", "updated_at": "2026-10-05T17:24:44.052Z", "categories": [ "training", "diffusion", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UvUSTi1qjy", "number": 6186, "title": "REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent", "abstract": "Post-training quantization (PTQ) is essential for deploying large language models (LLMs) under strict resource constraints. State-of-the-art PTQ methods quantize each layer with a single closed-form second-order solver: to remain analytically tractable, they heavily approximate the global loss (dropping cross-channel coupling, pooling output rows into groups), and they then freeze the resulting Hessian across the entire layer, with no way to refresh it as the loss landscape shifts column by column---a phenomenon we call \\textit{information misalignment}. We propose \\textbf{REAL-Q} (\\textbf{R}eal-time \\textbf{E}2E-loss \\textbf{A}ligned \\textbf{L}LM \\textbf{Q}uantization), a novel PTQ paradigm that breaks this compromise: instead of diluting the objective for the sake of analytic tractability, REAL-Q targets an end-to-end-aligned surrogate of the global loss and refines it via fine-grained, dynamic Block-wise Gradient Descent applied after every column block (128 columns). By coupling this fine-grained correction with a sliding window mechanism for smooth cross-layer transitions, REAL-Q effectively mitigates error propagation across the network. Experiments on LLaMA-3.1-8B-Instruct and Qwen3-4B/8B/32B cover twelve weight-only and weight--activation settings, with REAL-Q achieving the lowest KL loss and ranking first on downstream tasks in most settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models; Post-Training Quantization; Model Compression; Fisher Information; Gradient-Based Optimization" ], "forum_url": "https://openreview.net/forum?id=UvUSTi1qjy", "pdf_url": "https://openreview.net/pdf/53eb16b8f96fec666d3e5c458d98076f489dbd97.pdf", "updated_at": "2026-10-05T17:25:37.475Z", "categories": [ "ptq", "training" ], "matched_terms": [ "ptq", "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ojnmlXUtnV", "number": 6243, "title": "miniTE: Diagnosing and Bridging the Execution Gap in Native NVFP4 Reinforcement Learning", "abstract": "Native NVFP4 promises to accelerate reinforcement learning (RL) for large language models (LLM), but achieving stable learning across separate training and sampling engines remains challenging. We show that matching model weights and numerical formats is insufficient to ensure consistent policy evaluation, highlighting the importance of cross-engine numerical execution. We introduce miniTE, a lightweight framework that combines execution diagnostics, cross-engine numerical alignment, and efficient native W4A4 (4-bit weights and activations) GEMM computation for LLM RL. It improves numerical execution consistency without modifying the RL objective, while reducing implementation overhead to realize the benefits of native low-bit computation. Long-horizon RL experiments on Qwen3-30B-A3B demonstrate stable training and benchmark performance near to BF16, outperforming naive Transformer Engine (TE) NVFP4 in evaluation quality. On single NVIDIA B300 GPU node, the optimized learner reaches 33,540 tokens/s, yielding 1.39$\\times$ and 1.47$\\times$ the throughput of BF16 and TE NVFP4, respectively, for complete learner updates. We released the framework, kernels, diagnostics, and training configurations to make native low-bit RL easier. Our framework, kernels, diagnostics, and training environments are available at https://anonymous.4open.science/r/miniTE.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Reinforcement Learning", "LLMs", "High-performance Training" ], "forum_url": "https://openreview.net/forum?id=ojnmlXUtnV", "pdf_url": "https://openreview.net/pdf/d2815dd854910165dfb451983f210d14e1d8f677.pdf", "updated_at": "2026-10-05T17:25:58.754Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "low-bit", "nvfp4", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3TkUHSFp3e", "number": 6244, "title": "TRIAGE: Direction-Aware Mismatch Stabilization for Native NVFP4 Reinforcement Learning", "abstract": "Low-precision execution can substantially accelerate reinforcement learning (RL) for large language models, but discrepancies between learner and sampler execution can destabilize policy optimization. In this paper, we characterize the interaction between mismatch and the policy-gradient direction, distinguishing locally amplifying from contracting update contributions that mismatch magnitude alone\ncannot identify. In native NVFP4 runs, we observe an early imbalance between the two amplifying regions, favoring negative-advantage, negative-gap updates. Their tail tokens become concentrated in a small fraction of response segments before mismatch spreads globally.\nMotivated by these findings, we introduce TRIAGE, a direction-aware stabilization method that uses segment-level diagnosis to selectively rebalance policy-gradient updates and applies bounded repair to residual severe mismatch. TRIAGE modifies the optimization objective while retaining native NVFP4 weight-and-activation 4-bit (W4A4) forward execution on both\nthe sampler and learner. Experiments on Qwen3-4B and Qwen3-30B-A3B show stable optimization throughout the evaluated training horizon and achieve full precision level performance across five mathematical reasoning benchmarks, \nwhile native NVFP4 with TRIAGE provides up to 2.3$\\times$ higher rollout throughput than BF16. \nOur code is available at https://anonymous.4open.science/r/TRIAGE.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Reinforcement Learning", "LLMs" ], "forum_url": "https://openreview.net/forum?id=3TkUHSFp3e", "pdf_url": "https://openreview.net/pdf/8cebee8d9beb0bd414d06f16a399cd0390f6db11.pdf", "updated_at": "2026-10-05T17:25:59.535Z", "categories": [ "training", "hardware", "reasoning" ], "matched_terms": [ "4-bit", "low-precision", "nvfp4", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CAAyqk9d58", "number": 6272, "title": "SPARCQuant: Sparse Adaptive Repair of Critical Channels for MLLMs Quantization", "abstract": "Post-training quantization (PTQ) for multimodal large language models is challenging because heterogeneous modalities induce distinct activation distributions in the shared decoder. Smoothing-based quantization has proven effective and is widely adopted for large language model quantization, yet directly extending it to these models can cause smoothing factors estimated from mixed calibration data to be dominated by high-magnitude modalities, thereby over-smoothing other modalities and increasing activation quantization error. Against this backdrop, the analysis in this paper reveals a previously overlooked structural property: **high sensitivity to modality-specific smoothing mismatch is concentrated in a small subset of channels**. Motivated by this finding, this paper proposes **SPARCQuant**, a sensitivity-guided framework for **sparse adaptive repair of critical channels** in multimodal large language model quantization. SPARCQuant selects a layer- and module-adaptive base modality to define shared smoothing and ranks channels using robust activation-range discrepancy and activation-magnitude importance. A coarse-to-fine proxy reconstruction search with online scheduling then **automatically selects the sparse channel support and its size**, adapting the repair to each layer, module, and non-base modality **without manual channel selection or per-layer budget tuning**. Selected-channel ridge compensation corrects residual errors while preserving a single shared quantized weight. Experiments on vision-language and omni-modal models demonstrate **improved robustness under low-bit quantization**. A CUDA implementation with custom kernels enables efficient deployment, achieving **2.66×–2.87× end-to-end inference speedup over BF16**.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "multimodal large language models; low-bit quantization" ], "forum_url": "https://openreview.net/forum?id=CAAyqk9d58", "pdf_url": "https://openreview.net/pdf/edd9656d111ff7e5d73bf61d8e17c775a0d1c3db.pdf", "updated_at": "2026-10-05T17:26:23.924Z", "categories": [ "ptq", "hardware", "multimodal", "theory" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AYT3Kmxq8y", "number": 6395, "title": "Calibration Timing and Model Sensitivity in Rotated KV-Cache Quantization", "abstract": "Does a rotated uniform KV-cache quantizer require locally fitted scales? A comparison can conflate scale granularity with calibration timing and metadata cost. We audit these factors with non-overlapping finite windows, a dense non-evicted reference, and one-write quantize–dequantize simulation. Following development experiments, a frozen 80-condition comparison crosses model, dataset, rotation seed, calibration timing, granularity, and symmetry, with eight-bit controls. On Llama-3.1-8B-Instruct, four-bit per-write global calibration yields post-prefill perplexities of 7.508–7.557 on WikiText-2 test (reference 7.404), whereas prefill-frozen global configurations yield 125.414–39,255.244. The timing effect persists on SQuAD contexts and for two new rotations. Grouped per-write scales reduce error further but incur more metadata, preventing an equal-total-budget claim. Qwen2.5-7B-Instruct behaves differently: all tested four-bit configurations fail severely across both datasets and rotations, while eight-bit controls remain close to their dense references. Unquantized cache and rotation-roundtrip controls also remain close on a diagnostic window. These results support calibration- and model-specific conclusions rather than universal necessity of local scales. They do not establish packed-kernel speed, cross-domain generalization, or a unique mechanism for the Qwen low-bit failure.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache", "quantization", "calibration", "language models", "reproducibility" ], "forum_url": "https://openreview.net/forum?id=AYT3Kmxq8y", "pdf_url": "https://openreview.net/pdf/64b88a4e3321a8a5b32e55161d2de655bec6aff2.pdf", "updated_at": "2026-10-05T17:27:26.208Z", "categories": [ "ptq", "kv-cache", "hardware", "theory" ], "matched_terms": [ "low-bit", "quantization", "quantize", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pdQTAhD6Xq", "number": 6410, "title": "Error Self-Compensation Phase Retrieval for Binary Fringe Projection Profilometry", "abstract": "Error-diffusion-based binary-defocusing fringe projection enables high-speed 3D measurement, but systematic phase errors introduced during binarization can degrade reconstruction accuracy. To address this problem, we propose EDFv2, a unified Error Diffusion Fusion framework for arbitrary-step phase-shifting fringe projection. EDFv2 exploits the complementary phase errors generated by forward and inverse diffusion kernels and redesigns the binarization process according to the phase-shifting configuration. For three-step phase shifting, each fringe is binarized twice with complementary kernels to achieve effective error cancellation. For phase-shifting schemes with more steps, the binarization flow is reorganized to preserve the same compensation mechanism for both even- and odd-step configurations. Since the proposed strategy operates only during binary-pattern generation, it introduces no additional computational overhead during phase reconstruction. Experiments under different phase-shifting steps, fringe periods, and defocus levels demonstrate that EDFv2 consistently reduces binarization-induced phase errors while retaining the high-speed advantage of binary projection.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Binary Defocusing", "Phase-Shifting Profilometry", "Error Compensation", "Complementary Diffusion", "High-Speed 3D Measurement" ], "forum_url": "https://openreview.net/forum?id=pdQTAhD6Xq", "pdf_url": "https://openreview.net/pdf/6b438b3175aea13607531da1dd61fbdebefee9e2.pdf", "updated_at": "2026-10-05T17:27:49.494Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "binarization", "binarized", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "92qhqIKIlK", "number": 6544, "title": "HeRo-Q: A General Framework for Stable Low Bit Quantization via Hessian Conditioning", "abstract": "Post-Training Quantization (PTQ), a mainstream model compression technique, often leads to the paradoxical “low-error, high-loss” phenomenon because it focuses solely on minimizing quantization error. The root cause lies in the Hessian matrix of the LLM loss landscape: a few high-curvature directions are extremely sensitive to perturbations. To address this, we propose the Hessian Robust Quantization (HeRo-Q) algorithm, which applies a lightweight, learnable rotation-compression matrix to the weight space prior to quantization. This joint framework reshapes the loss landscape by reducing the largest Hessian eigenvalue, thereby significantly enhancing robustness to quantization noise. HeRo-Q requires no architectural modifications, incurs negligible computational overhead, and integrates seamlessly into existing PTQ pipelines. Experiments on Llama and Qwen models show that HeRo-Q consistently outperforms state-of-the-art methods—including GPTQ, AWQ, and SpinQuant—not only achieving superior performance under standard W4A8 settings, but also excelling in the highly challenging W3A16 ultra-low-bit regime, where it boosts GSM8K accuracy on Llama-3-8B to 70.15% and effectively avoids the logical collapse commonly seen in aggressive quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "PTQ.+Hessian" ], "forum_url": "https://openreview.net/forum?id=92qhqIKIlK", "pdf_url": "https://openreview.net/pdf/fc46d22601578550222e11e2af1256f0e6ebc3cf.pdf", "updated_at": "2026-10-05T17:28:53.297Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "low bit", "low-bit", "ptq", "quantization", "w3a16", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MX3GdV1CgI", "number": 6702, "title": "CoHSPAD: Continuous Human Gaussian Reconstruction from Raw Monocular 1-bit SPAD Streams", "abstract": "Monocular human reconstruction from conventional RGB video is limited by finite frame rates and exposure time. Rapid motion causes severe blur, degrading the image quality needed for reliable pose estimation and appearance recovery. SPAD cameras offer an alternative: binary photon-arrival observations at very high temporal resolution, which remain informative under low light and rapid motion. However, a single 1-bit frame is extremely sparse, and independent per-frame optimization of SMPL parameters and 3D Gaussians is easily destabilized by noisy photon gradients. We present CoHSPAD, a monocular framework for 3D Gaussian reconstruction of dynamic humans and static scenes with continuous human motion recovered directly from raw SPAD streams. Built on a differentiable SPAD imaging model, we supervise rendering directly in the 1-bit raw sensor domain. Instead of independent per-frame poses, we model motion as a continuous SMPL trajectory with sparse temporal knots that share supervision across the stream and stabilize pose gradients. We introduce a Multi-Scale Temporal Fusion Network (MTF-Net) to aggregate long and short streams per-timestamp, providing a more stable signal than a single binary frame while avoiding blur from naive temporal averaging. Experiments on our newly collected CoH-Synthesis and CoH-Real show that CoHSPAD remains robust and temporally consistent under inaccurate initial SMPL poses and rapid motion.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "SPAD Imaging", "Human Reconstruction", "3D Gaussian Splatting" ], "forum_url": "https://openreview.net/forum?id=MX3GdV1CgI", "pdf_url": "https://openreview.net/pdf/51f6ef699451ea86944707393d230a14e9ab16f0.pdf", "updated_at": "2026-10-05T17:30:32.615Z", "categories": [ "ptq", "training" ], "matched_terms": [ "1-bit", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "liUWwWlga1", "number": 6731, "title": "Latency and Accuracy Tradeoffs in Spiking Neural Networks", "abstract": "Spiking neural networks (SNNs) are attractive for low-power speech command recognition, yet their latency has received far less attention than their energy efficiency, and their multi-timestep execution is widely assumed to make them slower than quantized neural networks (QNNs). This paper challenges the assumption that more local timesteps necessarily imply higher network latency. By overlapping computation across adjacent layers at the timestep level, SNNs may complete execution in less time than a comparable bit-serial QNNs. However, this overlap relies on spikes firing on incomplete inputs, and a spike once generated cannot be withdrawn, so its error persists and reduces accuracy. Waiting for more input before firing would seem to improve accuracy at the cost of reduced overlap. Yet we find and prove that this intuition fails at some layers, where even a small increase in waiting can change spike timing and downstream computation, making the network both slower and less accurate. \nWe therefore propose a Pipeline Delay Search (PDS) method which selects each layer’s delay by balancing task-level accuracy gains against added network latency. We then adapt the selected configurations through spike-based quantization-aware training and bounded tuning of firing thresholds and initial membrane potentials. Together, these steps form Falcon, a framework for Fine-grained Analysis of Latency and Controlled firing which systematically analyzes and optimizes SNN latency under a spatial analog compute-in-memory mapping with shared digital engines.\nWe evaluate Falcon on Google Speech Commands V2 (GSC) and Spiking Speech Commands (SSC), achieving competitive accuracies of 96.31\\% and 83.02\\% at modeled network-core latencies of 119.64 and 124.00$\\mu\\mathrm{s}$, respectively. Together, our analysis and results show that SNNs can compute more yet finish faster, and wait longer yet predict worse, highlighting why Falcon matters for both latency and accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking neural networks", "Latency analysis", "Energy efficiency", "Speech recognition" ], "forum_url": "https://openreview.net/forum?id=liUWwWlga1", "pdf_url": "https://openreview.net/pdf/1d88daef34a5f3a5dd7be3de243188e0d845336b.pdf", "updated_at": "2026-10-05T17:30:46.175Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XVUzeKa5Rg", "number": 6767, "title": "HGA-Quant: Heterogeneity-Aware Quantization for AutoRegressive Visual Generation Models", "abstract": "Recently, AutoRegressive Visual Generation (ARVG) models have emerged as a competitive paradigm for high-quality image synthesis. However, their substantial computational and memory demands make them difficult to deploy on resource-constrained edge devices. Post-Training Quantization (PTQ) provides a promising solution to reduce the computational and memory costs of ARVG models, yet existing methods often suffer from substantial image quality degradation under low-bit quantization. In this paper, we propose a training-free PTQ framework for ARVG, consisting of Tail-Aware Bounded-init Grid Refinement (T-BGR) and Coarse-to-Fine Activation Precision Scheduling (CAPS). T-BGR introduces a tail-aware reconstruction objective for weight quantizer calibration, assigning greater importance to distribution-tail coefficients and preserving fine-grained information without introducing additional parameters or high-precision branches. CAPS performs coarse-to-fine activation precision scheduling by jointly considering the importance and computational cost of different generation stages, enabling efficient mixed-precision allocation under a given bit budget. Extensive experiments on multiple ARVG backbones demonstrate that our framework consistently improves robustness under aggressive low-bit quantization. Notably, on Infinity, our framework achieves near full-precision image generation quality at an effective W4A5.6 activation precision, while existing methods at W4A6 still exhibit noticeable quality degradation or inconsistent performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Autoregressive Visual Generation", "Vision Generation Models", "Low-Bit Quantization" ], "forum_url": "https://openreview.net/forum?id=XVUzeKa5Rg", "pdf_url": "https://openreview.net/pdf/e022ea6389e1654dadf5bcd8c76f31bf85415c05.pdf", "updated_at": "2026-10-05T17:31:24.300Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "low-bit", "mixed-precision", "ptq", "quantization", "quantizer", "w4a5", "w4a6" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MEaq2jZ4lI", "number": 6817, "title": "Mechanism-Aligned Observability for Training Instability of Large Language Models", "abstract": "Frontier large language model training consumes massive accelerator fleets over long wall-clock times, making stability failures extremely costly. However, global symptoms such as loss divergences often emerge only after the underlying fault has already affected the training dynamics, further increasing the sunk compute. In which internal representation can a training fault become persistently and interpretably observable before it appears in global training metrics? We propose a mechanism-aligned principle for training observability: identify the internal state into which a fault is persistently written, derive the structure induced by its mechanism, and construct monitors in representations aligned with the computation of the affected module. For FlashAttention, we show that biased backward errors accumulate into a persistent low-dimensional structure in weight updates, producing spectral concentration. For MoE routers, we characterize routing dynamics through per-token routing behavior rather than parameter similarity, which may change without altering the routing function. Experiments show that the two monitors respond selectively to changes in their corresponding modules, before degradation becomes visible in global training metrics. ur code is available at https://anonymous.4open.science/r/mindspeed-observability-F605.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Training", "Training Instability Detection", "Low-Precision Training", "Mixture-of-Experts", "Spectral Analysis" ], "forum_url": "https://openreview.net/forum?id=MEaq2jZ4lI", "pdf_url": "https://openreview.net/pdf/4d677d2f4eb434e5ce0b2216467bbeb22d706288.pdf", "updated_at": "2026-10-05T17:31:22.534Z", "categories": [ "training", "hardware" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cA6zJCAZVg", "number": 6969, "title": "When Mismatch Misleads: Stabilizing Low-Precision Reinforcement Learning via Reliable Correction", "abstract": "Reinforcement learning (RL) has become a key component of large language model post-training, but its substantial computational cost increasingly motivates end-to-end low-precision training. Despite improving training efficiency, it also introduces numerical perturbations that amplify train-inference mismatch and create a tradeoff between efficiency and stability. Existing approaches primarily stabilize low-precision RL by aligning rollout and actor policies. However, actor-rollout alignment alone does not guarantee improvement of the underlying full-precision objective, since the observed mismatch conflates genuine policy mismatch with numerical noise. We therefore formulate low-precision RL as a three-policy alignment problem, where noisy rollout and actor policies provide observable training signals while the inaccessible full-precision policy defines the learning problem. Based on this formulation, we propose **ERIS (Excess-Risk-guided Importance Sampling)**, an excess-risk-guided method that corrects high-reliability mismatch tokens while retaining identity weighting for noise-dominated tokens. Across multiple Qwen model scales and architectures, ERIS consistently improves FP8 training stability and performance. ERIS improves the average mathematical reasoning score over standard FP8 by 14.29\\% on Qwen3-8B, while achieving performance comparable to BF16. Mechanism analysis further confirms the advantage of three-policy alignment in preserving consistency with the full-precision policy. Meanwhile, ERIS preserves the efficiency benefits of FP8, achieving up to a 1.17$\\times$ end-to-end throughput improvement over BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning; Low Precision Training; Train-Inference Mismatch" ], "forum_url": "https://openreview.net/forum?id=cA6zJCAZVg", "pdf_url": "https://openreview.net/pdf/546f2c0749a3d75f854d5ca43ab4abfcf172e9fd.pdf", "updated_at": "2026-10-05T17:32:54.783Z", "categories": [ "ptq", "training", "hardware", "reasoning" ], "matched_terms": [ "fp8", "low precision", "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1x84XCyHff", "number": 7011, "title": "CurveFP: Co-Designing Numerical Representation and Product Arithmetic for Language Models", "abstract": "Low-precision formats usually optimize scalar fidelity while inheriting conventional product arithmetic. We introduce CurveFP, a block-scaled family that distributes magnitudes across interleaved logarithmic curves. Uniform curve indices make every nonzero product an exact sign and integer-index update, while a rational radix exposes the finite phase schedule required for accumulation. We instantiate the algebra as CurveFP8 E4C3/E5C2 for training and CurveFP7 E3C3 for compact inference. On four 7B-9B models, CurveFP7 beats tensorwise FP8 perplexity with one fewer element bit and stays within 1.32% of native quality. CurveFP8 lowers error in all 36 paired training-GEMM comparisons. Across three matched 3B-token pretraining triplets, it reaches mean BF16-inference perplexity 22.5366 versus 22.5407 for FP8 and has a lower format penalty in every seed. Downstream evaluation shows transfer parity and a consistent WikiText-103 gain. In a preliminary 4x4 Nangate45 spatial accelerator tile, CurveFP8 uses one fewer product register and 4.6% less area than timing-closing FP8 at 500 MHz. These results support CurveFP as a numerical and arithmetic co-design, while leaving system-level efficiency to future study.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Data Representation", "Hardware Acceleration" ], "forum_url": "https://openreview.net/forum?id=1x84XCyHff", "pdf_url": "https://openreview.net/pdf/3606440af9137eb3336049c2580c7bd8c3de28af.pdf", "updated_at": "2026-10-05T17:33:10.442Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp8", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "M7toaoVsuU", "number": 7073, "title": "SCORE: A Model Equality Test for Auditing Black-Box LLM APIs", "abstract": "Users increasingly rely on black-box inference APIs as their primary interface to large language models (LLMs), for both closed- and open-weight models (e.g., Llama models are now widely accessed through Amazon Bedrock and Azure AI Studio), with little insight into which model is actually deployed. To reduce costs or maliciously alter model behavior, API providers may silently deploy quantized models or substitute entirely different ones, potentially degrading model performance and compromising safety. Detecting such substitutions is challenging because users cannot inspect model weights and, in most cases, lack access to internal model signals such as output logits, leaving generated text as the only observable. To address this problem, we propose \\textbf{S}emantic \\textbf{C}ontrast for \\textbf{O}utput-only \\textbf{R}ank-based Model \\textbf{E}quality Testing (SCORE), an output-only auditing method that tests whether a target LLM API and its official reference API induce the same response distribution, thereby assessing whether the black-box API behaves consistently with the model it claims to serve. SCORE measures the change in semantic recovery induced by a control prefix, converts each target-model score into a randomized rank calibrated against reference-model scores, and tests the resulting ranks for uniformity at significance level \\(\\alpha\\) using only generated text. We evaluate SCORE across multiple threat scenarios, including model quantization, harmful fine-tuning, and full model replacement, as well as on commercial inference API endpoints. Across these settings, SCORE consistently achieves higher statistical power than existing methods under stringent black-box constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model substitution detection", "output-only black-box testing", "Auditing LLM APIs" ], "forum_url": "https://openreview.net/forum?id=M7toaoVsuU", "pdf_url": "https://openreview.net/pdf/6c0d9c33d6a63d99e6abdface17c30a8b37ba3e3.pdf", "updated_at": "2026-10-05T17:33:39.711Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SORR42klNg", "number": 7244, "title": "Bridging KV-Cache Quantization and Linear Attention: From Theory to Pretrained Weight Migration", "abstract": "KV-cache quantization and linear attention are two representative approaches to tackling the storage and computational costs of Transformers.\nKV-cache quantization compresses individual KV entries into discrete codes but retains all entries, whereas linear attention recurrently aggregates multiple historical KV contributions into a fixed-size continuous state but can introduce interference.\nThis contrast raises the question of whether per-KV compression and multi-KV aggregation can be bridged within a single mechanism for efficient attention.\nWe identify RAM-Net as such a bridge through soft assignments over a discrete address space.\nThese assignments determine recurrent updates to the continuous slot state associated with each address.\nUnder a restricted RAM-Net construction, we prove that soft address assignments extend hard quantized matching to a separable read-write overlap that locally approximates full-attention similarity and supports recurrent aggregation.\nThese connections further enable Transformer-to-RAM-Net weight migration through a new path based on a soft-quantized intermediate construction.\nAcross nine pretrained Transformer models from 0.3B to 7B parameters, RAM-Net recovers an average of 87.1\\% of the teachers' accuracy gains over random guessing across six commonsense and knowledge tasks using only a 500M-token budget per model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Efficient Attention", "Linear Attention", "KV-cache Quantization", "Pretrained Weight Transfer", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=SORR42klNg", "pdf_url": "https://openreview.net/pdf/ed0bda98d4bb07595c20d006d945484adadddf79.pdf", "updated_at": "2026-10-05T17:35:31.206Z", "categories": [ "kv-cache" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZnHqRTCr2K", "number": 7246, "title": "Preserving Updates in Low-Bit Recurrent State Quantization", "abstract": "Linear attention and recurrent models compress past tokens into a fixed-size\nstate that is updated at every step.\nHowever, storing these states in high precision incurs substantial memory and\nbandwidth costs at large batch sizes, motivating low-bit state quantization.\nUnlike weights and key--value caches, a recurrent state is requantized after\nevery update, so its quantization errors accumulate across steps.\nOur analysis shows that these errors are highly coherent across steps, so\nreducing per-step errors alone is insufficient.\nThis coherence arises mainly from small updates that are lost in rounding.\nTo preserve these updates, we intervene both before and at rounding.\nBefore rounding, we propose Low-Rank Bypass (LRB), which accumulates small\nupdates in a compact low-rank buffer and periodically merges its sufficiently\nlarge elements into the state.\nAt rounding, we observe that a state element can even move opposite to its\nupdate, and such reversals prove particularly harmful.\nTherefore, we propose Bounded Directional Rounding (BDR), which rounds reversed\nelements toward their updates without additional storage.\nAlthough BDR enlarges the rounding error, it improves the model, suggesting that\nthe direction of an update matters more than its rounding error at a single\nstep.\nAcross Gated DeltaNet, GLA, Mamba2, and RetNet, our training-free method\nsubstantially outperforms INT4, while reducing state memory by up to\n5.3$\\times$ and speeding up decoding by up to 1.9$\\times$ relative to FP32\nstates.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization; Linear Attention" ], "forum_url": "https://openreview.net/forum?id=ZnHqRTCr2K", "pdf_url": "https://openreview.net/pdf/740eee8e0d619b8901b18ad283f2b2ab389c17b2.pdf", "updated_at": "2026-10-05T17:35:33.299Z", "categories": [ "ptq" ], "matched_terms": [ "int4", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CJf168TUPG", "number": 7305, "title": "Marrying Capacity and Precision for Efficient 2D Gaussian Image Representation", "abstract": "2D Gaussian representations provide an explicit and efficient way to represent images, but their rate-distortion performance under limited budgets depends critically on how Gaussian capacity and parameter precision are allocated. Existing methods typically optimize these two resources separately, overlooking content-dependent capacity demands and heterogeneous quantization sensitivity. We present CP-GS, a capacity-precision coupled Gaussian splatting framework for budget-aware 2D Gaussian image representation that jointly optimizes capacity allocation and attribute precision under a shared rate constraint. CP-GS first constructs a content-adaptive Gaussian representation by dynamically reallocating a fixed Gaussian budget from redundant regions to areas with higher representation demand. It then introduces hierarchical distortion-aware representation quantization, using efficient rendering-based calibration to allocate precision across Gaussian attributes and further refine it within heterogeneous attribute groups. Finally, it jointly optimizes Gaussian capacity and attribute precision by constructing candidate representations with different Gaussian counts and precision configurations and selecting operating points according to decoded distortion and actual serialized size. Experiments on DIV2K and Kodak show that CP-GS achieves improved rate-distortion performance over representative 2D Gaussian and implicit neural representations across a range of budgets, while maintaining a fixed overall representation budget and low optimization cost. These results demonstrate the benefit of jointly optimizing spatial capacity and parameter precision for compact Gaussian image representation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "2D Gaussian Representation; Image Representation; Adaptive Quantization; Budget-Aware Optimization" ], "forum_url": "https://openreview.net/forum?id=CJf168TUPG", "pdf_url": "https://openreview.net/pdf/6e928d5a523b9b9599557ae3ef75cf69ed6da731.pdf", "updated_at": "2026-10-05T17:36:20.125Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HfjKCjgg09", "number": 7394, "title": "TrajQuant: Trajectory-Aware Quantization of Interactive Video World Models", "abstract": "Interactive video world models generate future visual states conditioned on actions and reuse those states as history for subsequent interaction. Their repeated denoising and long autoregressive rollouts make low-bit inference attractive, but conventional local reconstruction does not explicitly account for how quantization errors propagate along action-conditioned trajectories. We study this mismatch as a **control-to-trajectory error** problem: a local perturbation introduced during one chunk alters the inputs encountered by later chunks, while layer-wise tensor reconstruction may fail to capture errors in the complete state transition. We propose **TrajQuant**, a trajectory-aware post-training quantization framework for action-conditioned autoregressive video world models. Its **MLGR** module applies structured rotations whose angles are refined against complete quantized Transformer-block outputs, improving reconstruction at the transition level. Its **FGCC** module (Fisher-Guided Chunk Calibration) estimates a squared-gradient sensitivity proxy for each chunk–layer pair and uses it to weight paired full-precision and upstream-quantized reconstruction, accounting for errors propagated from preceding chunks. MLGR is fixed before FGCC performs final weight rounding in the transformed coordinates. We evaluate Matrix-Game-2 and ABot-World-0 at W6A6, W4A8, and W4A6 with WBench, LPIPS-FP16, and FVD-FP16. On Matrix-Game-2 at W4A6, TrajQuant achieves LPIPS-FP16/FVD-FP16 of 0.1294/32.71 versus SVDQuant’s 0.1709/45.26. On ABot-World-0 at W4A6, it also achieves the lowest LPIPS-FP16 and FVD-FP16 among the listed quantized methods, while the leaders at other precisions differ by metric.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "interactive video world models", "autoregressive video generation", "diffusion transformers", "low-bit inference", "trajectory-aware calibration", "action-conditioned generation" ], "forum_url": "https://openreview.net/forum?id=HfjKCjgg09", "pdf_url": "https://openreview.net/pdf/ef4ea8ba674c8f380cedf4a475c77766bf36b702.pdf", "updated_at": "2026-10-05T17:37:31.875Z", "categories": [ "ptq", "training", "diffusion" ], "matched_terms": [ "low-bit", "quantization", "quantized", "w4a6", "w4a8", "w6a6" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cmcGeC1HmI", "number": 7488, "title": "Bridging Efficiency–Accuracy Gap in Spiking Neural Networks via Operation-Constrained Quantization and Low-Spike Suppression", "abstract": "Spiking neural networks constitute a promising approach to energy-efficient intelligence by leveraging binary, event-driven computation. However, binary spike signals substantially constrain the channel capacity for information delivery, thereby limiting attainable model accuracy. To jointly improve representational capacity and inherit event-driven computation, we propose a novel multi-level spiking neuron termed QuantLIF, consisting of operation-constrained quantization and low-spike suppression. Specifically, we construct a dense quantization codebook whose values can be represented by a bounded number of signed power-of-two terms. Through shift-based re-parameterization, synaptic integration is replaced by several accumulations, without introducing excessive inference delay. To counteract the deterioration in event sparsity caused by multi-level spikes, the low-spike suppression stochastically filters low-contribution spikes during training and deterministically eliminates them during inference. We theoretically analyze QuantLIF from the perspectives of quantization and information theory, respectively. Comprehensive experiments demonstrate that QuantLIF achieves a favorable trade-off among predictive accuracy, firing sparsity, and energy consumption across both static and neuromorphic benchmarks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Multi-level Spiking Neuron", "Quantization" ], "forum_url": "https://openreview.net/forum?id=cmcGeC1HmI", "pdf_url": "https://openreview.net/pdf/6b2c070e0e565682c7402c7f7262cb3349c50687.pdf", "updated_at": "2026-10-05T17:38:46.559Z", "categories": [ "theory" ], "matched_terms": [ "binary", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NDLxJ6yqt1", "number": 7514, "title": "Watt Does a Benchmark Point Cost? Measuring the Energy Efficiency of LLM Capability under Production Serving", "abstract": "Benchmark scores tell us what an LLM can do, but not how much energy is required to deliver that capability.\nWe directly measure this trade-off for 86 open-weight model checkpoints, evaluated on seven benchmarks and six hardware platforms ranging from a consumer GPU to an eight-GPU server. We find large differences in energy efficiency even among models with similar capability. Most configurations lie away from the score--energy Pareto frontier: they use a median of 2.6--4.8$\\times$ more energy than another configuration that scores at least as well, with differences reaching 27$\\times$. Reasoning can improve capability, but enabling thinking increases energy by about 3.7$\\times$ on average, and the benefit depends strongly on the model family and output budget. Generated-token count is also an imperfect proxy for energy, misordering 18--24\\% of model pairs. In contrast, FP8 reduces energy by roughly a quarter with little change in score, while changing hardware largely rescales energy without changing the relative ranking of models. These results show that capability alone gives an incomplete picture of efficiency. We argue that benchmark scores should be reported together with measured energy consumption under a specified serving regime and output budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "energy efficiency", "benchmarking", "evaluation", "inference serving", "green AI" ], "forum_url": "https://openreview.net/forum?id=NDLxJ6yqt1", "pdf_url": "https://openreview.net/pdf/61dd4f61dcdd4034be41988f90de3c882ab2c44a.pdf", "updated_at": "2026-10-05T17:39:02.784Z", "categories": [ "training", "hardware", "reasoning" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "M33a4XnrZE", "number": 7556, "title": "Quantize, Evict, or Both? When Mixed KV-Cache Allocation Pays, and Why its Distortion Proxy Misleads", "abstract": "A key-cache budget can be spent by quantizing every token, by evicting tokens,\nor by mixing zero and finite precisions across tokens; values stay exact. Hybrid\nsystems exist. We ask when this mixed interior beats both endpoints, and whether\nan answer measured on attention-output distortion carries over to end tasks. If\neviction is treated as a zero-rate tier, a finite-bit tier is dominated when its logit\u0002noise cost exceeds the cost of zero rate. The share of heads whose 2-bit tier is\ndominated orders the interior’s exactly recomputed output-error advantage across\nsix LLMs and 25 model–context cells (Spearman ρ = −0.978; −0.94 controlling\nfor model). Where the interior stops paying, however, depends on the architecture,\nand longer contexts move a model toward eviction. Over 4,096 decode tokens a\nhead’s best policy family is stable, but its token allocation goes stale within a few\nsteps. On retrieval tasks we compare against three dense quantizers (TurboQuant,\nKIVI, KVQuant) and six eviction methods (H2O, SnapKV, DropKV, Ada-KV,\nLaProx, OBCache). Routed per KV head, the allocator that is optimal under a local\nsurrogate of output distortion beats every eviction method (mean score 0.77 vs.\n0.60), partly by sending 1–25% of heads to dense quantization; its mixed interior\nalone scores 0.47. Yet dense quantization at the same key code bits does better on\nall three end-task models, and the best quantizer depends on the model. TurboQuant\nwins on Llama-3.1-8B (0.99 vs. 0.65 for the allocator) and Mistral-7B (0.95 vs.\n0.70). On Qwen3-8B TurboQuant loses to the allocator (8k–32k: 0.87 vs. 0.96),\nand per-channel KIVI or KVQuant win (0.99). Output distortion ranks methods\nonly weakly and overstates the cost of rotated quantization noise. The allocator\nmainly fails by truncating multi-token answers; smoothing its token scores over\nneighboring positions largely removes this failure (Llama, 8k–32k: about 40% to\n10% of answers). At both levels, the choice of baselines and of when the cache is\ncompressed moves verdicts more than prompt sampling does.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression; llm inferences; Quantization; eviction policies" ], "forum_url": "https://openreview.net/forum?id=M33a4XnrZE", "pdf_url": "https://openreview.net/pdf/a77b178ef3b4eaf5fce4da3ff67c5afde676e0d5.pdf", "updated_at": "2026-10-05T17:39:25.271Z", "categories": [ "kv-cache" ], "matched_terms": [ "2-bit", "quantization", "quantize", "quantizer", "quantizers", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mBjfVGFp8m", "number": 7760, "title": "SchurQuant: Groupwise Discrete Optimization for Layer-Wise LLM Quantization", "abstract": "Weight-only post-training quantization (PTQ) enables the deployment of large language models under tight memory budgets, but accuracy often collapses at 2--3 bits. \nExisting backpropagation-free PTQ optimizers have two limitations: group decisions ignore the correction that the remaining continuous suffix can absorb, and discrete refinements typically keep the affine quantization grid fixed. \nWe introduce SchurOpt, which analytically eliminates the suffix's optimal continuous response, yielding an exact groupwise quadratic with Schur-complement curvature.\nIt then alternates closed-form row-wise scale/zero-point refitting with coordinate descent over integer codes. With the GPTQ objective fixed, SchurOpt improves mean zero-shot accuracy on 2-bit Qwen3-4B by 11.88 percentage points (pp). \nAt higher precision, however, tighter reconstruction does not consistently improve end-model metrics.\nSchurQuant therefore combines SchurOpt with quantized-prefix teacher reconstruction, reference-weight regularization, residual-add targets, and teacher-decision token weighting. \nAcross eight Llama and Qwen models, SchurQuant achieves the highest mean zero-shot accuracy among the evaluated backpropagation-free PTQ baselines, outperforming the strongest baseline by 9.65 pp at 2 bits.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "quantization" ], "forum_url": "https://openreview.net/forum?id=mBjfVGFp8m", "pdf_url": "https://openreview.net/pdf/ace6835ff1ba830d2b2718f947fd9929aa502b25.pdf", "updated_at": "2026-10-05T17:41:29.923Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "2-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YmLA7r8Hne", "number": 7921, "title": "Beyond Straightness: Non-Crossing Flow Matching via Quantile AlignTree Coupling", "abstract": "The performance of Flow Matching largely depends on the quality of the coupling between the source and target distributions. However, independent coupling often leads to path crossings and local velocity ambiguity, while OT-based couplings typically incur high construction costs. To address this challenge, we propose Quantile AlignTree Flow Matching (QAT-FM), an efficient structured coupling strategy that constructs a hierarchical coupling between a Gaussian prior and the target data distribution via a quantile-aligned tree structure. QAT-FM constructs the coupling in $\\mathcal{O}(Nd\\log N)$ time and supports per-pair source sampling with $\\mathcal{O}(d)$ complexity, enabling scalable training for large-scale high-dimensional generative tasks. Theoretically, we prove that the QAT coupling satisfies marginal consistency, induces non-crossing linear interpolation paths, and consistently improves path separation at intermediate times compared with independent coupling, thereby alleviating local velocity ambiguity. QAT-FM further extends naturally to conditional generation, enabling structured conditional coupling while preserving global Gaussian alignment. Experiments across diverse benchmark datasets demonstrate that QAT-FM achieves competitive generative performance while substantially reducing coupling construction cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Flow Matching", "Generative Models", "Data Alignment", "Binary Trees" ], "forum_url": "https://openreview.net/forum?id=YmLA7r8Hne", "pdf_url": "https://openreview.net/pdf/bc3e46e7fa35147b221477abe235bc3938ed82af.pdf", "updated_at": "2026-10-05T17:43:07.518Z", "categories": [ "qat", "diffusion", "theory" ], "matched_terms": [ "binary", "qat" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EutVjE5fM2", "number": 7932, "title": "Byte-Constrained Layer-Wise Quantization with Cost-Normalized Sensitivity", "abstract": "Mixed-precision post-training quantization allocates storage across tensors of different sizes and sensitivities, but nominal bit width omits format metadata. We instantiate a byte-constrained multi-choice knapsack formulation with measured sensitivity and runtime-supported formats, then allocate precision using quality gain per byte. Across five models from three families, the resulting GGUF artifacts lower perplexity relative to six production llama.cpp presets at 25 of 30 matchedsize points. Paired block-bootstrap intervals resolve 23 improvements and one regression. Controlled ablations separate storage accounting from ranking quality. Uncorrected nominal budgets overshoot file-size targets by 10.6% to 17.9%, but nominal and byte pricing produce similar perplexity after size matching. In contrast, raw-sensitivity ranking is worse at 29 of 30 matched-size points, by up to 6.36 perplexity. Across six downstream tasks, Benjamini–Hochberg correction at 5% identifies 21 gains and three regressions among 180 comparisons, with 19 gains on models of at most 3B parameters. A same-model untying control attributes most shared-embedding sensitivity to the language-model head. These results distinguish accurate artifact sizing from effective precision allocation, while exposing the limits of a separable sensitivity objective.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "mixed-precision quantization", "model compression", "sensitivity-aware bit allocation" ], "forum_url": "https://openreview.net/forum?id=EutVjE5fM2", "pdf_url": "https://openreview.net/pdf/85558fda51fec828f108a1b8bbcb9e5fc362f96a.pdf", "updated_at": "2026-10-05T17:43:09.580Z", "categories": [ "ptq" ], "matched_terms": [ "bit width", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W1nwTn1Nsu", "number": 7949, "title": "PRECISION ENGINEERING AN ACTIVATION FUNCTION", "abstract": "Activation functions are typically designed as mathematical abstractions and evaluated independently of their underlying hardware representations. In this work, we approach the activation function as an engineering problem constrained by three requirements: (1) sublinear growth that mirrors the diminishing marginal impact of large-magnitude features while remaining unbounded above, (2) execution via standard Arithmetic Logic Unit (ALU) instructions rather than Special Function Units (SFUs), and (3) continuity and differentiability almost everywhere, sufficient for stable gradient-based optimization. To satisfy these constraints simultaneously, we introduce ReLog, an \\textit{engineered} activation, constructively defined by the IEEE 754 floating-point bit layout. By isolating the exponent and mantissa, ReLog replaces the linear positive domain of ReLU with a piecewise-linear behavior.\n\nThis hardware-driven construction yields several testable geometric and optimization properties, which we verify empirically rather than assert. Topologically, ReLog possesses a finite number of piecewise ``kinks,'' bounded by $2^{N_E-1}$ for exponent field width $N_E$ (128 for FP32/BF16, 16 for FP16/FP8-E5M2, 8 for FP8-E4M3), with strictly linear sub-regions between them. Its gradient evaluates to exactly $2^{-(E-B)}$ for exponent field $E$ and bias $B$ -- a magnitude-dependent decay we show acts as an intrinsic per-node damping mechanism: across a spectral-radius sweep up to $\\rho=5.0$ (240 trained configurations), ReLog shows zero observed divergence (Wilson 95\\% CI $[0\\%, 1.6\\%]$), while standard clipping norms fail to reliably rescue an unclipped baseline at the same instability levels. Because ReLog is defined by the bit pattern, numerical precision governs its own stability: we find a clean, monotonic training-instability dose-response throughout the precision formats. We do not position ReLog as a universal replacement for all architectures, but rather test and report which domains benefit from it and which suffer.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "activation functions", "hardware-aware neural network design", "IEEE 754 floating-point", "low-precision training", "quantization", "piecewise-linear networks", "training stability", "gradient clipping", "recurrent neural networks", "diffusion transformers" ], "forum_url": "https://openreview.net/forum?id=W1nwTn1Nsu", "pdf_url": "https://openreview.net/pdf/743f3be83021c5941838677f90b619fcfc063071.pdf", "updated_at": "2026-10-05T17:43:17.528Z", "categories": [ "training", "hardware", "reasoning", "diffusion" ], "matched_terms": [ "fp8", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FYmK6zNgGt", "number": 8031, "title": "Quality Recovery for Quantized KV Caches via Low-Rank Attention Adaptation", "abstract": "Low-bit key--value (KV) caches reduce the memory required for autoregressive\ndecoding, but the resulting quality loss depends on the model and quantizer.\nWe keep the quantizer fixed and distill the floating-cache model's behavior into\nlow-rank Q/K/V projection updates while the student executes a physically\npacked incremental cache. Across three seeds, 4-bit affine-cache adapters recover\n$54.24\\%\\pm2.47\\%$ of the held-out perplexity gap on TinyLlama-1.1B and\n$75.96\\%\\pm4.04\\%$ on Gemma-4-12B. On the same frozen NF4\nLlama-3.1-8B base, one validation-selected run per quantizer recovers $60.42\\%$\nunder KIVI\nK2V2 and $37.61\\%$ under KVarN K4V2, while preserving 180-case associative\nretrieval. Gemma's score on an official 4K/8K RULER subset rises from 42.80\nwith the unadapted 4-bit cache to 48.33 after adaptation (46.15 floating), with\nsubstantial task heterogeneity. Finally, a 2-bit rank--token sweep reduces\nTinyLlama's 2-bit PPL from 576.10 to\n$11.4000\\pm0.0059$ across three seeds, versus 10.3988 floating, but restores\nonly 11--12 of 180 retrieval cases. These results show that low-rank projection\nadaptation can recover held-out quality across fixed cache formats, while\nperplexity recovery need not restore long-context retrieval.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "LoRA", "error correction", "knowledge distillation", "cache compression" ], "forum_url": "https://openreview.net/forum?id=FYmK6zNgGt", "pdf_url": "https://openreview.net/pdf/65ce4b411236b046f2e68959c3fe50d50c06d0bc.pdf", "updated_at": "2026-10-05T17:43:32.146Z", "categories": [ "kv-cache" ], "matched_terms": [ "2-bit", "4-bit", "low-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1HBkg8Exll", "number": 8221, "title": "MosaicQuant: Inlier–Outlier Disaggregation for Unified 4-Bit LLM Quantization", "abstract": "4-bit quantization significantly reduces the memory footprint and accelerates the inference of large language models (LLMs). \nHowever, its limited bit-width representation struggles to faithfully capture both dense common values (*inliers*) and rare large-magnitude values (*outliers*), causing substantial accuracy degradation. Existing mixed-precision methods mitigate this by retaining outliers in high precision, but at the cost of breaking the uniformity of low-bit execution, introducing precision conversion and extra data movement that undermine practical speedup. We propose **MosaicQuant**, a unified 4-bit LLM quantization paradigm built on a novel principle of *inlier--outlier disaggregation*. Rather than elevating outlier precision, MosaicQuant quantizes the full weight matrix into a dense 4-bit base component, where inliers are captured faithfully while outlier are inevitably quantized. A sparse 4-bit residual component is then introduced to compensate for these quantization errors, selectively targeting the most error-critical weight blocks where output distortion is shown to be concentrated. However, a unified representation alone is insufficient, as naïvely executing the sparse residual as a separate kernel still breaks the unified low-bit inference pipeline. To bridge this gap, we introduce **ZipperEngine**, which fuses sparse block computation into the dense 4-bit GEMM kernel via an overlapped pipeline, unifying not only the representation but also the execution into a single coherent low-bit inference pipeline. Extensive experiments on LLaMA3 and Qwen3 demonstrate that MosaicQuant preserves near-FP16 accuracy while achieving up to $1.24\\times$ speedup over the W16A16 baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization", "kernel optimization" ], "forum_url": "https://openreview.net/forum?id=1HBkg8Exll", "pdf_url": "https://openreview.net/pdf/c6d25ac9ed97a7ca7dcab98aaaba01287e6bdb93.pdf", "updated_at": "2026-10-05T17:45:21.095Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "bit-width", "low-bit", "mixed-precision", "quantization", "quantized", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bub7HTzeSv", "number": 8229, "title": "Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe", "abstract": "FP4 training promises substantial reductions in memory and computation cost for LLM pretraining, yet current FP4 hardware paths and recipes, including NVIDIA Blackwell/Rubin-class systems and AMD MI350-series GPUs, remain centered on E2M1 data elements. In this study, we identify a fundamental limitation of that choice: non-uniform formats such as E2M1 inherently suffer from Shrinkage Bias, a systematic negative rounding error caused by the geometric asymmetry of their representable bins. We show that this bias accumulates multiplicatively across layers and is amplified by the Random Hadamard Transform (RHT), providing a unified explanation for the training instability observed in existing E2M1-based FP4 recipes. In contrast, uniform grids (E1M2/INT4) bypass this grid-geometry error and better convert the improved bucket utilization from RHT into higher quantization quality. Based on this finding, we propose UFP4, a uniform 4-bit training recipe that applies RHT to all three training GEMMs while restricting stochastic rounding to dY alone. On Dense 1.5B, MoE 7.9B, and MoE 124B long-run pretraining, UFP4 consistently achieves lower BF16-relative loss degradation than strong E2M1-based baselines, supported by scaling-law analysis and ablation studies. Our results suggest that future accelerators should support E1M2/INT4-style uniform 4-bit grids as first-class training primitives alongside E2M1.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "FP4", "Full Quantization Training", "LLM" ], "forum_url": "https://openreview.net/forum?id=bub7HTzeSv", "pdf_url": "https://openreview.net/pdf/0c1b3cf3e50e504eec2bddd7ede2acf636b94d88.pdf", "updated_at": "2026-10-05T17:45:26.286Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "4-bit", "fp4", "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WuldvnhGHs", "number": 8281, "title": "Quantinue: Quantization for Unlearning operator", "abstract": "Machine unlearning aims to make a trained model behave like a model retrained without designated forget data, and the resulting model is typically quantized for efficient deployment.\nLow-bit quantization, however, can undo unlearning by discarding parameter changes introduced during unlearning and recovering knowledge that was previously forgotten.\nExisting methods address this by making the unlearning stage more robust to quantization, but they still treat unlearning and quantization as two separate steps.\nWe instead replace the two stages, unlearning and quantization, with quantization itself the unlearning operator.\nQuantization maps continuous weights to discrete codes while seeking to retain the original model behavior.\nWe guide this mapping not only to retain the original model behavior on the retain data, but also to induce forgetting behavior on the forget data, directly constructing an unlearned low-bit model without a preceding continuous-space unlearning update.\nActivation statistics from forget and retain calibration data define an element-wise score that determines where and how strongly the assigned code deviates from the reconstruction-driven code.\nThe resulting process separates ordinary quantization error from code-displacement error and compensates them separately, reducing compensation along forget-sensitive directions for the code displacement error.\nBoth code assignment and error compensation use only forward calibration statistics, without gradient computation or fine-tuning.\nOn the MUSE and TOFU benchmarks, our method achieves the lowest aggregate distance to the retrained reference among the compared methods in 4 bit precision.\nThese results suggest that quantization, previously a threat to unlearning, can instead serve as an effective means of achieving it.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Unlearning", "Quantization" ], "forum_url": "https://openreview.net/forum?id=WuldvnhGHs", "pdf_url": "https://openreview.net/pdf/6cf0a10467c7d7ea5bb9f6c5be9e2e5431bb23e9.pdf", "updated_at": "2026-10-05T17:45:58.977Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "4 bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1Toh6VM9Pu", "number": 8347, "title": "Generate, Then Match: Lightweight Relevance Modeling for Generative Retrieval", "abstract": "Generative retrievers can generate relevant document identifiers yet rank the corresponding documents poorly, exposing a mismatch between generation likelihood and relevance.\nWe introduce QIMatch, a lightweight adaptation method that learns a relevance matching function over a retriever's existing query and identifier representations.\nWith the generator and identifier index fixed, two shared projections learn to match queries with complete identifiers through contrastive training followed by document ranking refinement, updating only 1.18M parameters without an additional neural relevance teacher.\nAt inference, QIMatch combines matching scores with generation likelihood to rerank generated identifiers without encoding document text.\nOn Natural Questions and MS MARCO, QIMatch achieves the highest Recall@1 and MRR@10 among the compared methods for both product-quantized and title-URL identifiers.\nAcross these four settings, it improves MRR@10 over supervised fine-tuning baselines by 3.08-12.33 percentage points.\nAnalyses show that the learned matching function transfers to identifiers excluded from its supervision in both families, while fusion can correct ranking errors that neither score resolves alone.\nThese findings demonstrate how learning to match existing representations enables effective relevance adaptation with a fixed generator.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Generative Retrieval" ], "forum_url": "https://openreview.net/forum?id=1Toh6VM9Pu", "pdf_url": "https://openreview.net/pdf/cb8e62cb3cdf4da119896f4f06074d12553d94ca.pdf", "updated_at": "2026-10-05T17:46:57.879Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sJ7KUx26PC", "number": 8571, "title": "GIFT: Geometry-Informed Low-Precision Gradient Communication for LLM Pretraining", "abstract": "Gradient communication is a major scaling bottleneck in large language model (LLM) pretraining. Low-precision formats such as FP8 can substantially reduce communication cost, but often at the expense of gradient fidelity. Existing approaches either quantize gradients directly in their original coordinates or apply data-independent transformations to suppress outliers. Our study shows that both approaches overlook a fundamental source of quantization distortion: LLM gradients are highly anisotropic, with geometry that varies substantially across layers, making low-precision error strongly direction dependent. This observation suggests that gradient fidelity depends not only on the quantizer itself, but also on the coordinates in which quantization is performed. Motivated by this insight, we design GIFT, a geometry-informed gradient communication method that transforms gradients into more isotropic coordinates before low-precision quantization. To make this transformation practical, GIFT uses a one-sided, low-rank approximation of the geometry and applies it only to layers that are most sensitive to quantization errors, while leaving the optimizer, training recipe, communication collective, and low-precision format unchanged. Across Llama-style models with approximately 600M and 2B parameters, GIFT achieves downstream task performance comparable to Muon training with FP32 gradients. On 64 NVIDIA GH200 Superchips, GIFT reduces end-to-end training step time relative to FP32 gradient communication by 10.87% for Llama-600M and 17.05% for Llama-2B.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "distributed pretraining", "gradient communication", "communication compression", "low\u0002precision communication", "geometry-aware communication", "K-FAC" ], "forum_url": "https://openreview.net/forum?id=sJ7KUx26PC", "pdf_url": "https://openreview.net/pdf/5596254f6353fb0dc52359934e662b69d60b382a.pdf", "updated_at": "2026-10-05T17:49:43.317Z", "categories": [ "training" ], "matched_terms": [ "fp8", "low-precision", "quantization", "quantize", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9UjCKdjvkl", "number": 8719, "title": "SGA-Sparse: Training Native SwiGLU Gates for Activation-Sparse and Low-Bit LLMs", "abstract": "Activation sparsity can reduce FFN weight traffic during single-user LLM decoding, but active channels must be identified early enough to avoid fetching unused weights. In pretrained SwiGLU models, training-free routing with the native gate performs substantially worse than routing with the up activation at high sparsity. We show that this ordering reverses after sparsity-aware continued pretraining. To the best of our knowledge, we present the first systematic study that converts pretrained SwiGLU LLMs into fixed-budget activation-sparse models by training the existing gate projection as the router. Our method, SGA-Sparse, updates all model parameters through continued pretraining while preserving the original SwiGLU activation and introducing no auxiliary router parameters. After approximately 30B training tokens, LLaMA models from 1B to 8B achieve uniform per-token 75--85\\% FFN channel sparsity while remaining within 0.2--2.7 pp of their original dense checkpoints in average zero-shot accuracy. We further study two complementary QAT recipes: continued-QAT largely recovers the performance at 4-bit weight quantization, while joint-QAT performs better for the more aggressive sub-4-bit target. To compare activation sparsity with dense low-bit quantization, we translate the sparsity-induced reduction in weight traffic into an equivalent bit-width. Under this common idealized metric, our 2.5-bit-equivalent models achieve average zero-shot accuracy comparable to dense 3-bit QAT models, while our 1.5-bit-equivalent models approach dense 2-bit models and\noutperform dense ternary models by 1.3 pp on average. These\nresults establish native-gate training as a practical approach for combining high FFN sparsity with low-bit weights.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "activation sparsity", "efficient LLM inference", "SwiGLU", "on-device LLM" ], "forum_url": "https://openreview.net/forum?id=9UjCKdjvkl", "pdf_url": "https://openreview.net/pdf/5b219ace6409de123ee505174348fdf8a9467b5c.pdf", "updated_at": "2026-10-05T17:51:28.508Z", "categories": [ "qat" ], "matched_terms": [ "2-bit", "3-bit", "4-bit", "5-bit", "bit-width", "low-bit", "qat", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TDcosiMqmF", "number": 8724, "title": "How Much do Language Models Regurgitate under Post-Training Quantization?", "abstract": "Quantization has become a standard practice for efficient deployment of today's large language models. However, there is still a lack of understanding of its impact on verbatim regurgitation—where models may regurgitate training data samples, which can be a privacy vulnerability. In this work, we present the first comprehensive study of how post-training quantization affects verbatim regurgitation in LLMs, uncovering multiple key insights. First, we show that reducing precision consistently lowers regurgitation rate across all model sizes, but at the cost of decreased downstream performance. Second, we find that for the same set of sequences, larger quantized models regurgitate more than smaller models, consistent across varying prefix \\& suffix lengths. We also propose a scaling law to model verbatim regurgitation under quantization. Lastly, fine-grained layer-wise and component-wise analyses reveal a consistent trend: quantizing attention modules leads to a higher decrease in regurgitation compared to feed-forward networks, and quantizing later layers has a more pronounced impact than early-middle layers, further supported by flip rate analyses. Our empirical study covers three modern post-training quantization methods and six models from the Pythia suite, providing novel insights into the impact of quantization on verbatim regurgitation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Verbatim Regurgitation", "Quantization", "Language Models" ], "forum_url": "https://openreview.net/forum?id=TDcosiMqmF", "pdf_url": "https://openreview.net/pdf/845937ae367a7b6794e7cee656fc0f0b299e822a.pdf", "updated_at": "2026-10-05T17:51:30.199Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "x5uT9IMk34", "number": 8787, "title": "CurvProbe: Damage Has Two Factors, and the Checkpoint Sets One", "abstract": "Post-training compression first has to decide where: which layers to delete, factorise or quantise. The criterion the field deploys ranks layers by displacement, how far an operation moves a layer's output, and never asks how much that move raises the loss. Expanding the loss where these operations inject their change splits the damage at a layer into that displacement times the loss's curvature along it. Second-order salience already estimates this curvature, one operation at a time. That curvature belongs to the frozen checkpoint, not to the operation, so it can be read once, before any operation is named. CurvProbe reads it as a curvature profile, one number per layer from $2L$ forward passes, with no backward pass and no Hessian. The factorisation predicts that displacement ranks layers below chance on checkpoints whose curvature is highest where displacement is lowest, as on Llama-3-8B, where the attention sublayer ranked safest to delete costs the most; the profile flags such failures before any damage is measured. One profile, never shown the operation it is judged on, repairs the damage ranking across eighteen checkpoints and eleven operations. On Llama-3-8B its layer sets keep that gain once a quarter of the layers are compressed. After low-rank approximation they hold 65.8% LAMBADA accuracy, against Block Influence's 55.8%. A checkpoint can therefore ship with its profile, and a compressor that arrives later ranks its layers with one forward pass where measuring their damage would take $L$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "layer selection", "large language models", "layer pruning", "curvature", "Hessian", "quantization" ], "forum_url": "https://openreview.net/forum?id=x5uT9IMk34", "pdf_url": "https://openreview.net/pdf/7adf5056819cabc027bf5c7bd14fc6626e09299c.pdf", "updated_at": "2026-10-05T17:51:55.209Z", "categories": [ "ptq" ], "matched_terms": [ "quantise", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tppSfpxbWH", "number": 8918, "title": "Lightweight but Global: Event-Guided Broad-Range Brightness Restoration", "abstract": "Broad-range brightness restoration maps RGB frames captured under low, normal, or high illumination to a target exposure. Events provide complementary measurements when the frames are noisy or saturated. Balancing global spatial context, temporal information, and reconstruction detail within a compact model remains challenging. To this end, we present MambaSEE, a compact spatiotemporal architecture that assigns these functions to different operators. First, fixed-width additive fusion and attentive state-space modules model each frame–event pair with global context. Next, a weight-and-activation 8-bit quantization-aware ConvGRU aggregates three first-in, first-out pairs through backward-then-forward propagation, without observations later than the prediction time. Finally, a prompt-conditioned depthwise decoder reconstructs fine image details. On the SDE and SEE-600K benchmarks, MambaSEE achieves the highest reported average PSNR with the lowest FP32-equivalent weight storage among the compared methods. On SEE-600K, its PSNR values are 2.01, 1.39, and 5.32 dB higher than the published SEENet results for low-, high-, and normal-light inputs, respectively. MambaSEE uses mixed-precision weight storage equivalent to 0.48M FP32 parameters, about one quarter of SEENet's.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Event-based Vision", "State Space Models", "Mamba", "Quantization-Aware Training", "Efficient Architecture" ], "forum_url": "https://openreview.net/forum?id=tppSfpxbWH", "pdf_url": "https://openreview.net/pdf/060552fcfef7f01325b7b5458ed19d4c5aa70049.pdf", "updated_at": "2026-10-05T17:53:01.388Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "8-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3H1szWq0K8", "number": 8968, "title": "Evolutionary-Potential-Aware Iterative Preference Learning for LLM-Based Automatic Heuristic Design", "abstract": "Automatic Heuristic Design (AHD) uses large language models (LLMs) to construct heuristics for complex optimization problems. Existing LLM adaptation methods largely learn from the immediate performance of generated heuristics, overlooking the longer-term evolutionary potential revealed by their descendants. Search trees naturally preserve parent--descendant relationships and the outcomes of subsequent exploration, providing structural information that can serve as preference supervision. We propose an evolutionary-potential-aware iterative preference learning framework for LLM-based AHD that converts this information into search-feedback signals. After each MCTS expansion, newly observed descendants reassess the relative evolutionary potential of their ancestors; changes in local rankings then yield search-feedback preference pairs. These pairs complement on-policy performance preferences from newly generated candidates to fine-tune a local LLM through direct preference optimization. The updated LLM subsequently guides MCTS exploration, closing the iterative search--learning loop. Experiments across NP-hard combinatorial optimization tasks show that our method achieves the best overall result in six of nine ACO task--scale settings and the best LLM-based AHD result in eight. Despite using a locally deployed quantized 7B model, it outperforms GPT-4o-mini-based AHD baselines across all nine ACO settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Evolutionary computation", "combinatorial optimization", "automatic heuristic design" ], "forum_url": "https://openreview.net/forum?id=3H1szWq0K8", "pdf_url": "https://openreview.net/pdf/d7b886f1e05f3a3dce43fcbaf15da30169a8200a.pdf", "updated_at": "2026-10-05T17:54:05.080Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BF2ySBHQJP", "number": 9007, "title": "When Is Component Sensitivity Compositional? A Causal Study of Neural Resource Allocation", "abstract": "Neural resource-allocation methods commonly assign each component a scalar sensitivity and rank components independently. This raises two distinct questions: whether the score tracks singleton utility, and whether ranking by singleton utility remains useful under joint allocation. We separate these two questions and test them through direct interventions under a common, logged protocol across quantization, layer skipping, structured pruning, mixture-of-experts selection, and adapter placement.\n\nThe resulting behavior forms a spectrum. Layer skipping exhibits small median pairwise interactions, yet one-shot singleton selection can still incur substantial held-out NLL loss at the allocation budget; pruning rankings depend on background sparsity; GPTQ-based group refinement and adapter placement show strong non-additivity; and expert interactions vary across the tested depths. Large additive prediction errors need not imply large selection gaps: the six-layer adapter placement realizes only 15% of its singleton-sum NLL prediction yet trails the tested last-6 heuristic by only 0.011 nats. By contrast, on Qwen2.5-14B one-shot singleton selection at a 25% layer-skip budget incurs 1.013 nats of held-out damage versus 0.540 for sequential re-measurement, and under sparse pruning backgrounds the dense ranking transfers worse than background re-measurement in 16 of 18 conditions—separating additive prediction failure from realized allocation loss. The 14B layer-skip NLL advantage does not yield a statistically resolved accuracy advantage on ARC-Challenge or HellaSwag.\n\nAs an exploratory diagnostic, residual-stream perturbation alignment is positively associated with pairwise interaction magnitude in 11 of 12 runs. We conclude that component sensitivity need not behave as an intrinsic scalar: it is a configuration-dependent marginal effect in several tested regimes, and allocation methods should validate compositionality at the deployment granularity, background, budget, and metric.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Component Sensitivity", "Compositionality", "Neural Resource Allocation", "Model Compression", "Causal Intervention", "Neural Network Interpretability" ], "forum_url": "https://openreview.net/forum?id=BF2ySBHQJP", "pdf_url": "https://openreview.net/pdf/c76e1e52bab530e2e31c15a88eaf63c935d23d6c.pdf", "updated_at": "2026-10-05T17:54:46.291Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GOPoXWnJQx", "number": 9016, "title": "Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation", "abstract": "Streaming video generation (SVG) distills a pretrained bidirectional video diffusion model into an autoregressive model equipped with sliding window attention (SWA). However, SWA inevitably loses distant history during long video generation, and its computational overhead remains a critical challenge to real-time deployment. In this work, we propose Hybrid Forcing, which jointly optimizes temporal information retention and computational efficiency through a hybrid attention design. First, we introduce lightweight linear temporal attention to preserve long-range dependencies beyond the sliding window. In particular, we maintain a compact key–value state to incrementally absorb evicted tokens, retaining temporal context with negligible memory and computational overhead. Second, we incorporate block-sparse attention into the local sliding window to reduce redundant computation within short-range modeling, reallocating computational capacity toward more critical dependencies. Finally, we introduce a decoupled distillation strategy tailored to the hybrid attention design. A few-step initial distillation is performed under dense attention, then the distillation of our proposed linear temporal and block-sparse attention is activated for streaming modeling, ensuring stable optimization. Extensive experiments on both short- and long-form video generation benchmarks demonstrate that Hybrid Forcing consistently achieves state-of-the-art performance. Notably, our model achieves real-time, long video generation at 29.5 FPS on a single NVIDIA H100 GPU without quantization or model compression. The source code and trained models will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "stream video generation", "hybrid attentioin" ], "forum_url": "https://openreview.net/forum?id=GOPoXWnJQx", "pdf_url": "https://openreview.net/pdf/06dd1504c21e7ba206d35b94ce8e28e104b23587.pdf", "updated_at": "2026-10-05T17:55:14.075Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MjSR1GFfKw", "number": 9045, "title": "OneBaseKV: Low-Bit Canonical KV Sharing with Per-Request Replacement Overlays for High-Fanout RAG", "abstract": "Long-context services increasingly process recurring content across many requests. High-fanout retrieval-augmented generation (RAG) is a representative workload in which the same document chunks are repeatedly combined with different queries, orders, and retrieval results. Autoregressive inference retains key--value (KV) states for these long inputs, so per-request cache ownership replicates document state and rapidly consumes memory even after low-bit quantization. Exact prefix sharing removes duplication only for byte-identical prefixes, while cache-fusion systems do not encode shared document state and query-conditioned repairs as jointly executable low-bit objects. We propose OneBaseKV, which stores each reusable chunk once as an immutable INT4 canonical base and represents query-sensitive differences with private INT4 replacement overlays. An architecture-normalized selector identifies repair positions, and materialization-equivalent composite attention merges the base, overlay, and dynamic KV through one global online softmax without a persistent dense request cache. We prove exact equivalence to standard attention over the corresponding materialized dequantized KV and derive a mixed-partition perturbation bound for heterogeneous precisions. Experiments show that OneBaseKV achieves the best quality among the compared efficient methods under fixed persistent-KV budgets. At a $0.52\\times$ BF16 copy-on-write (BF16-COW) persistent-KV cap, it exceeds an oracle choosing the best feasible external method for each combination of model and task by 5.48\\%. At a 12K-token context and fanout 16, OneBaseKV uses only 37.7\\% of BF16-COW persistent KV. In two fixed-capacity serving traces, it achieves $6.06\\times$--$8.57\\times$ the service-level objective (SLO) goodput of Fusion-style. On non-retrieval RULER at 32K, it exceeds every efficient external method on three of four models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "1. Retrieval-Augmented Generation（RAG） 2. KV Cache Sharing 3. KV Cache Quantization 4. Query-Aware Cache Reuse 5. Large Language Model Serving" ], "forum_url": "https://openreview.net/forum?id=MjSR1GFfKw", "pdf_url": "https://openreview.net/pdf/8d3a5418e7b1bd16df443e58bd0e02c14502c701.pdf", "updated_at": "2026-10-05T17:56:32.187Z", "categories": [ "kv-cache" ], "matched_terms": [ "int4", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Z7BQZUgzsZ", "number": 9049, "title": "Variational Inference for Dynamic Mixtures in Probabilistic Time-Series Forecasting", "abstract": "Probabilistic time-series forecasting requires modeling complex, dynamic, and often multimodal conditional distributions while maintaining computational efficiency. We propose a lightweight finite mixture framework that uses a mixture sieve model and an analytically derived optimal posterior for reconstruction, thereby reducing the variational inference problem for conditional distribution estimation to a multi-class classification task. The framework defines the model output in the fractionally differenced domain, which alleviates distribution shift in the target variable and yields a dynamic mixture representation with time-varying component locations on the original scale. The proposed objective provides a quantization-level distributional approximation guarantee and has bounded logit-level gradients. Experiments on real-world benchmarks demonstrate that our method consistently outperforms existing probabilistic forecasting approaches while generating plausible future trajectories and competitive point estimates.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Probabilistic forecasting", "Variational inference", "Mixture model", "Fractional differencing", "Multi-class classification" ], "forum_url": "https://openreview.net/forum?id=Z7BQZUgzsZ", "pdf_url": "https://openreview.net/pdf/ceff04b6a14875b06065a27db8c2e995714f03a5.pdf", "updated_at": "2026-10-05T17:56:50.010Z", "categories": [ "ptq", "training", "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wrC04D13fL", "number": 9057, "title": "Dual-QK: Sharp Queries and Flat Keys for Prunable 2-bit KV Caches", "abstract": "Long inputs and extended generation increase the storage and access costs of the key--value (KV) cache. Low-bit quantization reduces\nstorage and memory traffic, while query-channel pruning can further reduce key-cache reads. Rotation-based quantization redistributes\nthe energy of key outliers across channels. To maintain computational invariance, the same orthogonal transform must be applied to queries, preserving query--key dot products. However, this rotation can disperse query energy, weakening the separation between a few large components to retain and many small ones to prune. We introduce Dual-QK, which uses paired non-orthogonal query and key transforms to address this conflict. Using calibrated query and key statistics, Dual-QK combines partial key whitening with a query-aligned basis to balance key scales for INT2 quantization and concentrate query energy for dynamic channel pruning. Channel-0 protection and bucket-relative RoPE support low-bit accuracy over long contexts. Experiments on four models across five generative benchmarks and long-context retrieval tasks show improved accuracy over OSCAR on most tasks at 40\\% query-channel sparsity. At a 128K context, Dual-QK provides $6.8\\times$ KV-cache compression and an estimated $8.3\\times$ reduction in KV read volume relative to unpruned BF16. Under the evaluated configurations, our SGLang implementation achieves up to $3.75\\times$ the decoding throughput of unpruned BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache", "quantization", "sparse attention", "query-channel pruning", "long context LLM inference" ], "forum_url": "https://openreview.net/forum?id=wrC04D13fL", "pdf_url": "https://openreview.net/pdf/14949f0a4898156ed2676e670e2da75251b5ae30.pdf", "updated_at": "2026-10-05T17:57:47.817Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "2-bit", "int2", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "We9x9Twodv", "number": 9060, "title": "ActTune: Action-Aware Precision and GPU Operating-Point Adaptation for Energy-Efficient Vision-Language-Action Inference", "abstract": "Vision-language-action (VLA) policies repeatedly invoke inference to control robots, making graphics processing unit (GPU) energy a recurring cost of task execution. Reducing energy per inference call, however, may not reduce energy per successful task if numerical errors increase failures or slower inference prolongs execution. We therefore target GPU energy per successful task while preserving task success and limiting increases in episode duration. Our approach builds on two observations: quantization sensitivity varies across action classes, model layers, and weights versus activations; and numerical precision changes the workload, shifting favorable GPU operating points. We introduce ActTune, an action-aware framework that connects layer-wise precision allocation with workload-dependent GPU frequency selection. A lightweight recognizer selects precision before each policy call. The controller forecasts the next workload and applies GPU settings asynchronously using a lookup table calibrated under a latency budget. A shared resident quantized weight bank enables configuration switching without weight reconstruction or additional policy evaluations. On LIBERO, a benchmark for lifelong robot learning, ActTune with the evaluated selector improves mean task success by up to 2.4\\% relative to state of the art. Relative to the original BF16 implementations, it delivers up to $2.02\\times$ faster inference and, with dynamic voltage and frequency scaling (DVFS), reduces energy per successful task by up to 76.8\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action Models; Energy-Efficient Inference; Mixed-Precision Quantization。" ], "forum_url": "https://openreview.net/forum?id=We9x9Twodv", "pdf_url": "https://openreview.net/pdf/feaa34309a28a0954e0488a470449bec6f9af176.pdf", "updated_at": "2026-10-05T17:57:08.112Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MhiweC44cB", "number": 9122, "title": "Layer-wise Adaptive Gradient Compression with Generalized Error Estimation", "abstract": "Reducing gradient communication is critical for large-scale data-parallel training. While layer-wise adaptive compression schemes can outperform uniform layer-wise ones, they typically rely on $L_2$ distance in gradient space as the error measure with a fixed measurement schedule. We propose Generalized Gradient Compression (GenComp), which generalizes layer-wise adaptive compression along three axes: 1) the space in which compression error is measured, on gradients or on latent activations obtained by simulating gradient updates and forward passes; 2) the error measure; 3) the frequency of measurement. We show that gradient- and activation-based error measurements, in tandem with error measures addressing information loss and different frequencies of measurement, induce distinct layer-wise compression dynamics, and can improve task performance and compression ratio. Across sparsification, quantization, and low-rank compression schemes applied to different model architectures, GenComp achieves significantly higher compression ratios than recent layer-wise adaptive schemes while retaining similar performance, which leads to reduced end-to-end training time for larger models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Gradient compression", "distributed training", "error estimation" ], "forum_url": "https://openreview.net/forum?id=MhiweC44cB", "pdf_url": "https://openreview.net/pdf/c46ac054f5a0f5efeea5429738902d579d3be8d1.pdf", "updated_at": "2026-10-05T18:00:25.671Z", "categories": [ "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "uPstvqOoai", "number": 9174, "title": "ORQA: Fast and Memory-Efficient Hybrid Reasoning with a Quantized Reasoner", "abstract": "Reasoning improves the capability of large language models, but long reasoning traces increase latency and decoding cost. Hybrid reasoning reduces this cost by offloading intermediate reasoning to a lightweight model, which we call an efficient reasoner. Existing methods, however, require extra model parameters and KV cache, increasing memory consumption. Moreover, the capability gap between the reasoner and the target LLM often limits speedup or reduces accuracy. To this end, we propose ORQA, a hybrid reasoning framework that improves speed and accuracy over prior methods without adding memory overhead. Instead of loading an extra reasoner, ORQA uses a quantized counterpart of the target LLM and a shared KV cache pool. We propose weight infusion, which integrates parameters between the target LLM and reasoner, eliminating separate model storage while maintaining lossless inference for both. We further devise Quant-ifier, a routing signal designed for a quantized reasoner that predicts token disagreement using the quantization error distribution. Experiments show that ORQA improves\nspeed and accuracy over prior methods with no additional memory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Hybrid Reasoning", "Reasoning", "Acceleration", "Quantization" ], "forum_url": "https://openreview.net/forum?id=uPstvqOoai", "pdf_url": "https://openreview.net/pdf/827154da8c7d124af21cd41891032cb6e0336e2d.pdf", "updated_at": "2026-10-05T18:02:01.498Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XZuXPhlSJv", "number": 9223, "title": "GEAR: Response-Aware Test-Time Compute Allocation for Frozen Quantized Spiking Transformers", "abstract": "Frozen quantized spiking classifiers cannot change their weights at deployment, but they can spend extra computation on selected inputs. We study when an alternative view helps rather than merely changes a prediction. GEAR routes low-margin images from a frozen 4-bit QSD-Transformer through nested high-resolution and flipped views, then fuses their predictions. Its empirical response curve models a fixed action's net repair utility---repairs minus newly introduced errors---along the cheap-view margin rank. On ImageNet validation and five ImageNet-derived shifts, a fixed 26:8:4 policy gains 0.79--2.46 percentage points without target-set retuning. A labeled target-like probe calibrates the curve and operating width, evaluated on disjoint held-out rows; geometry interventions show why a source-optimal crop need not transfer. Tests on a second spiking family, synthetic corruptions, detection, and segmentation clarify its scope: margin routing helps when escalation corrects errors concentrated among uncertain inputs, but not when refinement mainly benefits confident ones. The response profile is conditioned on the chosen action and deployment distribution.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Quantized Spiking Transformers", "Response-Aware Inference", "Test-Time Compute Allocation", "Adaptive Test-Time Augmentation", "Distribution Shift" ], "forum_url": "https://openreview.net/forum?id=XZuXPhlSJv", "pdf_url": "https://openreview.net/pdf/524bf052ee66ad66086934fd54e44af21604460b.pdf", "updated_at": "2026-10-05T18:03:16.429Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "oV9tZGWLEy", "number": 9303, "title": "Beyond 4 Bits: Data-Centric Post-Training Quantization and Recovery", "abstract": "Post-training quantization (PTQ) aims at compressing the weights of a large trained model using a limited compute and data budget. Down to 4 bits, state-of-the-art PTQ methods are near-lossless; below this limit, the tested scalar PTQ configurations collapse to near-random accuracy and model quality becomes acutely sensitive to choices such as the quantization representation and the calibration data. At such low bitwidths, recovery fine-tuning after quantization becomes crucial, and its effectiveness depends heavily on these same choices. Yet there is little systematic evidence on how the quantization method, the quality and task alignment of calibration data, and the design of the recovery procedure interact to determine the best final 2-bit model. We present a systematic study of 2-bit quantization and recovery fine-tuning across multiple model families, scales, and data distributions, and distill our findings into a practical, architecture-dependent set of best practices for efficiently obtaining an accurate 2-bit model. This recipe, which combines a vector codebook representation, a small task-aligned calibration mixture, and a short codebook-only distillation stage, restores $63$%--$69$% average accuracy across seven question-answering tasks at 2 bits (a $71$%--$73$% bf16 ceiling), versus the ${\\sim}29$% random accuracy floor of scalar PTQ (more so on MoE than dense models), at roughly half the compute budget of the strongest Quantization Aware Training (QAT) baseline. Our contribution is empirical: a systematic, architecture-aware map of the 2-bit design space, together with released code, Hessian mixtures, and checkpoints that make the findings directly actionable and reproducible.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Foundation or frontier models", "including LLMs" ], "forum_url": "https://openreview.net/forum?id=oV9tZGWLEy", "pdf_url": "https://openreview.net/pdf/d25bb68d997829be472255eed73999b1f115a63b.pdf", "updated_at": "2026-10-05T18:05:26.389Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "2-bit", "ptq", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9anJkPeyEc", "number": 9360, "title": "PiCoKV: Dequantization-Free KV Cache Quantization via Binary Principal-Component Coding", "abstract": "The growing key--value (KV) cache limits long-context large language model (LLM) inference through its memory footprint and bandwidth demand. Quantization reduces these costs, but attention typically still requires dequantizing or unpacking the stored KV states into wider numerical operands for computation, introducing conversion overhead and limiting the computational benefits of low-bit storage.\nTo bridge this gap between compact storage and computation, we present \\textbf{PiCoKV}, a training-free, data-aware KV-cache quantization method for dequantization-free attention on packed binary codes. PiCoKV factorizes keys and values into binary codes and compact decoding matrices. By folding these matrices into attention's query and output paths, PiCoKV performs both query--key scoring and value aggregation directly on packed codes through word-level bitwise operations, processing multiple binary entries in parallel without unpacking them into wider numerical operands.\nEvaluations on long-context benchmarks for Llama-3.1-8B-Instruct and Qwen3-8B show that PiCoKV achieves accuracy comparable to full attention and outperforms low-bit quantization baselines. \nThe combination of compact storage and direct computation yields practical efficiency gains: PiCoKV achieves higher end-to-end decoding throughput than the baselines, reaching $3.45\\times$ that of FlashAttention-2 at a 256K context while compressing the KV cache by up to $6.3\\times$. Our code is available at https://anonymous.4open.science/r/PiCoKV-BFC1/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "KV Cache", "LLM Inference" ], "forum_url": "https://openreview.net/forum?id=9anJkPeyEc", "pdf_url": "https://openreview.net/pdf/57b7f04dc7b9a4fc55f01a1d59c62acc8f2fa0ef.pdf", "updated_at": "2026-10-05T18:06:19.332Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "binary", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6PlfDcC0PM", "number": 9387, "title": "Hardware-Native Joint Sparse-Quantization for Trillion-Scale Mixture-of-Experts", "abstract": "Mixture-of-Experts (MoE) architectures allow frontier language models to scale toward trillions of parameters, but their deployment is constrained by massive memory footprints and bandwidth limitations. While modern accelerators feature Sparse Tensor Cores (SpTCs) to reduce weight storage and boost throughput via low-precision semi-structured sparsity, exploiting them in MoEs is hindered by severe model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware–software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective scaled through expert-parallel compression. Systemically, we implement a custom grouped sparse GEMM kernel tailored to sparse, low-precision MoE inference on SpTCs. Across MoE scales ranging from 30 billion to frontier-scale trillion-parameter models, our framework advances state-of-the-art joint sparse quantization by up to $4.35$ percentage points in task accuracy while preserving $96.09\\%$ of the original models' performance. System-level benchmarks on NVIDIA B200 GPUs show that our kernel outpaces the vendor baseline by up to $1.65\\times$, delivering $1.18\\times$ higher serving throughput and up to a $4.03\\times$ reduction in end-to-end latency. These results establish hardware–software co-design as a practical and necessary pathway toward scalable, highly efficient MoE deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Sparsity", "Quantization", "Model Compression", "Inference Acceleration", "Mixture-of-Experts" ], "forum_url": "https://openreview.net/forum?id=6PlfDcC0PM", "pdf_url": "https://openreview.net/pdf/96618eaf097115053a4d7a9650d2df9005c73509.pdf", "updated_at": "2026-10-05T18:06:40.444Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FZBkRsBAfs", "number": 9482, "title": "QuantiBias: Benchmarking Quantization-Induced Bias in LLMs", "abstract": "Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked. We find its principal side effect is increased bias that standard safety evaluation misses. Holding the model, its training, and the prompts fixed, a quantized model still refuses harmful requests, still avoids over-refusing benign prompts, and still selects the unbiased multiple-choice answer. Yet asked an open-ended question, the same model volunteers stereotypes in all eight languages we probe, in roughly one in four open-ended answers under an independent judge (~24% to ~27% across the compression ladder): it passes every standard check and still reaches users measurably more biased. The selective gap is a robust finding; whether open-ended bias further increases with compression is less certain, sensitive to the judge that scores it. We address both with QuantiBias, a benchmark that pairs a generative, multilingual stereotype probe with the refusal and multiple-choice controls that isolate open-ended generation, contrasts each build with and without reasoning, and rates the content severity of what it generates. Across two backbone models (Qwen and Gemma), a five-family screen, and eight benchmarks, quantizers allocate their extra precision by capability data that carries no bias-prevention signal, and reasoning before answering roughly halves the effect on some families while doing nothing on others. A quantized build must be re-evaluated for open-ended bias, not only on the short-form safeguards it already passes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "bias", "fairness", "safety evaluation", "benchmark", "large language models", "multilingual", "stereotypes", "LLM-as-judge", "reasoning" ], "forum_url": "https://openreview.net/forum?id=FZBkRsBAfs", "pdf_url": "https://openreview.net/pdf/d4d98143e01fa2b61067e9dfc4c4754a7b0af38a.pdf", "updated_at": "2026-10-05T18:07:53.245Z", "categories": [ "reasoning" ], "matched_terms": [ "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "iYjRQ32Kiz", "number": 9561, "title": "RECAP: Calibrating Low-Bit Reasoning Models on the History They Read", "abstract": "Low-bit quantization reduces the cost of serving reasoning models but can substantially degrade long chains of thought. Each step reads cached keys and values spanning thousands of tokens, allowing quantization errors to accumulate along the trajectory. Standard calibration operates on short segments that omit the earlier context needed by later reasoning steps. Even when this context is retained, cache reconstruction error alone does not capture how quantization errors affect subsequent attention. We present RECAP, a post-training quantization method that calibrates reasoning models on the history they read. Dynamic-prefix calibration separates long historical context from short-block reconstruction. It regenerates prefix keys and values with the current quantizer and rotates the active block to supervise every position along the trajectory. Prefix read geometry further supervises historical cache errors according to their effects on attention outputs, weighting key and value errors asymmetrically. With both components applied only during calibration, RECAP outperforms the baseline across multiple reasoning models and quantization settings. Under W4A4KV4, RECAP improves accuracy over the baseline by 13.3 percentage points on MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and 15.8 points on AIME-120 with Qwen3-4B. These results highlight the importance of preserving historical dependencies during calibration for low-bit reasoning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "reasoning language models", "long-context calibration" ], "forum_url": "https://openreview.net/forum?id=iYjRQ32Kiz", "pdf_url": "https://openreview.net/pdf/4294c1f2bd1cc65fdf6691c8a1cda1113faf524f.pdf", "updated_at": "2026-10-05T18:08:41.515Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "low-bit", "quantization", "quantizer", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8Xk6D9Bdqn", "number": 9596, "title": "From Generation to Correction: Autoregressive Bitwise Correction for Image Super-Resolution", "abstract": "Visual autoregressive (VAR) models have achieved remarkable progress in image generation, which has boosted their application to image super-resolution (ISR).\nHowever, existing VAR-based ISR methods tend to follow a generative paradigm conditioned on the low-quality (LQ) representation directly, without explicitly disentangling reliable content from degradation-induced discrepancies. This work focuses on bitwise VAR models that operate by quantizing latent features into binary codes. Within this binary latent space, we systematically analyze the bitwise behavior between LQ and high-quality (HQ) representations. We observe that LQ and HQ representations are highly similar at the bit level, motivating us to revisit ISR through an LQ-anchored bitwise correction formulation. This formulation aims to identify and correct LQ bits that disagree with their HQ counterparts while preserving consistent ones. This selective correction enables faithful restoration by leveraging reliable LQ information and concentrating the generative prior on degradation-induced discrepancies for high-frequency detail synthesis. Building on this insight, we propose Generative-Prior-enhanced Autoregressive Bitwise Correction (GPABC), a framework that autoregressively predicts LQ-relative flip masks by leveraging corrected representations from preceding scales as context alongside a pretrained VAR prior. To balance reconstruction fidelity and perceptual quality, we further introduce a scale-adaptive asymmetric focal flip loss. This loss penalizes false flips at coarse scales to preserve reliable LQ content, while penalizing missed flips at fine scales to recover high-frequency details with the generative prior. Extensive quantitative and qualitative experiments across four benchmarks show that GPABC outperforms state-of-the-art methods in perceptual quality while maintaining competitive reconstruction fidelity, validating the effectiveness of the proposed bitwise correction formulation for ISR.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Image Super-Resolution; Visual autoregressive models; Generative models" ], "forum_url": "https://openreview.net/forum?id=8Xk6D9Bdqn", "pdf_url": "https://openreview.net/pdf/a25a8cfe5bc23c80303ac7f830050caddc25c729.pdf", "updated_at": "2026-10-05T18:09:06.496Z", "categories": [ "ptq" ], "matched_terms": [ "binary", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pUyiw6fK2I", "number": 9621, "title": "RATE-DISTORTION ADAPTIVE PRIMITIVE SELECTION FOR OMNIDIRECTIONAL GAUSSIAN SPLATTING", "abstract": "Learned image codecs (LICs) achieve high reconstruction quality, but their decoding speed is often insufficient for immersive virtual reality (VR). Gaussian splatting (GS) codecs render much faster, yet still lag in reconstruction quality and typically decide primitive allocation without considering the coding cost of each primitive. We introduce OIC-GS, an omnidirectional GS codec with a new hierarchical\nHEALPix primitive grid representation. Gaussian primitives are anchored at predefined spherical locations, eliminating explicit coordinate coding. Finer levels refine their coarser ancestors, naturally supporting coarse-to-fine reconstruction and layered transmission. The predefined grid also enables efficient viewport decoding by selecting only view-relevant primitives. We further introduce a lightweight\nentropy model for quantized primitives and optimize the codec under a spherical rate-distortion objective. Primitives with insufficient rate-distortion benefit are automatically removed when their quantized opacity becomes zero, allowing\nOIC-GS to adapt both primitive density and level of detail without a fixed primitive budget. A single bitstream supports full-sphere, viewport-dependent, and progressive decoding. The first viewport reaches final quality after decoding only\n52% of the bitstream, and is then rendered at 1,270 FPS. On a 100-image omnidirectional benchmark, OIC-GS outperforms all evaluated GS codecs, reducing WS-PSNR BD-rate by 49.6% over GaussianImage++ and 68.6% over SGI, which uses a learned entropy model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "2D Gaussian splatting", "image compression" ], "forum_url": "https://openreview.net/forum?id=pUyiw6fK2I", "pdf_url": "https://openreview.net/pdf/ffcde68c349c0598b8b44e1d9905a42f822c0e7c.pdf", "updated_at": "2026-10-05T18:09:35.109Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5WbHJ1SWCZ", "number": 9645, "title": "FidelityQ: Preserving Predictive Fidelity in Ultra-Low-Bit Language Model Quantization", "abstract": "Ultra-low-bit scalar quantization offers a regular arithmetic structure for efficient deployment of large language models, but preserving their predictive behavior at extreme precision remains challenging. With only a few reconstruction levels, accurate weight representation must translate into consistent model predictions using a small set of calibration samples. We propose FidelityQ, a scalar quantization framework for preserving predictive fidelity at weight precisions down to 2 bits. Specifically, we introduce Information-Aligned Analytic Coding (IAC) for accurate quantization. IAC shapes each weight row toward a centered symmetric source through randomized Hadamard mixing, then fits a mirrored four-level minimum-distortion quantizer to its empirical weights. Every level set in this symmetry-aligned family decomposes exactly into 2 sign-valued bases with associated scales, combining source-adaptive reconstruction with lookup-free decoding compatible with scaled integer arithmetic. Our analysis further reveals that residual predictive discrepancies are heavy-tailed and concentrate on tokens that the teacher predicts confidently, motivating Knowledge-Intensive Calibration (KIC) for consistent quantization. KIC ranks sequences using the tail of token-wise teacher-student divergence and allocates a limited calibration budget to the highest-scoring samples. It then restores predictive agreement by updating only quantization levels and scales under teacher supervision, leaving the pretrained weights frozen. Experiments on the LLaMA and Qwen families demonstrate the effectiveness of FidelityQ across weight-only and joint quantization settings. On LLaMA-2-7B, FidelityQ achieves a WikiText2 perplexity of 6.62 at W2A16 and improves the prior state-of-the-art W2A4KV4 result from 8.31 to 6.96, while reducing linear-layer weight memory by 8.0× relative to FP16. These results support accurate LLM deployment with structured ultra-low-bit quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Large Language Models", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=5WbHJ1SWCZ", "pdf_url": "https://openreview.net/pdf/03c50370ffd2a6417b9bb5664749c011b425a8c4.pdf", "updated_at": "2026-10-05T18:09:49.458Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantizer", "w2a16", "w2a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Srr981ILYv", "number": 9667, "title": "Byte Economics: Million-Token Inference of a 2.8-Trillion-Parameter MoE on a Consumer Desktop", "abstract": "Mixture-of-experts language models now exceed a trillion parameters, while the largest consumer GPU holds 32 GB. We show that serving such a model on a desktop is governed by byte economics: decode time is the bytes each token moves across the memory hierarchy divided by the bandwidth of the links, and prefill time is set by how many times the expert store is read. We build OMOE, a tiered serving system that runs the native 2.8-trillion-parameter Kimi-K3 checkpoint (1.4 TB of MXFP4 experts, unmodified) on two RTX 5090s, 192 GB of DRAM and a 3.2 TB storage tier, with outputs gated byte-exact or within $6\\times10^{-7}$ of a full-VRAM reference. A bandwidth cost model with no fitted constants guided four system generations from 4.62 to 1.65 s/token at short context and places the software ceiling for this model on this machine at 0.58 s/token. A layer-major prefill schedule reads each expert at most once per prefill, so prefill throughput falls only from 69 to 47 tokens/s between 96K and one million tokens of context, where needle retrieval passes with 6.1 h to first token and 3.53 s/token decode. The instrumented deployment measures what a production MoE asks of a memory hierarchy: an LRU hit rate of 0.34 at 3% of the store, 3.0 TB written per million-token prefill, and the bytes a model designer could remove. Concurrent consumer-scale systems report contexts up to 65K; the system, telemetry and acceptance harnesses will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixture-of-experts", "efficient inference", "LLM serving", "offloading", "long-context inference", "memory hierarchy", "consumer hardware" ], "forum_url": "https://openreview.net/forum?id=Srr981ILYv", "pdf_url": "https://openreview.net/pdf/a77fb15da3b37bf3f40c4763d59eed9270fdb8bc.pdf", "updated_at": "2026-10-05T18:09:55.587Z", "categories": [ "hardware" ], "matched_terms": [ "mxfp4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a13x7esw55", "number": 9711, "title": "Factorized AdaBoost.MH Achieves the Same Convergence Rate as AdaBoost.MH", "abstract": "AdaBoost.MH reduces multi-class classification to a collection of binary subproblems and enjoys the classical boosting-type convergence guarantee under a weak learning condition. A more structured variant, Factorized AdaBoost.MH, uses base classifiers of the form $\\mathbf{h}(x)=\\alpha \\mathbf{v} \\varphi(x)$, where a single binary classifier $\\varphi$ is shared across all classes and the label dependence is carried by a vote vector $\\mathbf{v} \\in\\\\{\\pm1\\\\}^K$. This factorization is algorithmically attractive and achieves better performance in practice, but its convergence depends on whether one can always choose a vote vector with sufficiently large induced binary weight mass. Previous work resolved this question with a lower bound $\\max\\\\{1/n,1/\\sqrt{2K}\\\\}$, which still leaves a dimension-dependent slowdown relative to the original AdaBoost.MH analysis. In this paper, we sharpen this combinatorial step and provide an exact characterization. For the minimax quantity $\\mathfrak{W}_ {n,K}$ governing the factorized edge, we prove $\\mathfrak{W}_ {n,K} = C_ {\\min\\\\{n+1,K\\\\}}$, where $C_q=1$ for $q=1$, $C_q=q/(3q-4)$ for even $q\\ge2$, and $C_q=(q+1)/(3q-1)$ for odd $q\\ge2$. Since $C_q\\downarrow 1/3$, our bounds show that $\\mathfrak{W}_{n,K}=\\Theta(1)$ uniformly over $n$ and $K$. Consequently, Factorized AdaBoost.MH achieves the same boosting-type convergence rate as AdaBoost.MH up to a universal constant factor, removing the previously suggested additional dependence on $n$ or $K$ in the number of boosting rounds.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Boosting", "Learning Theory" ], "forum_url": "https://openreview.net/forum?id=a13x7esw55", "pdf_url": "https://openreview.net/pdf/cd49c12f7b052e4b04c801a9daf44c39d028a356.pdf", "updated_at": "2026-10-05T18:10:32.314Z", "categories": [ "theory" ], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LJ3qgkXdhV", "number": 9746, "title": "Model Compression with Exact Budget Constraints via Riemannian Manifolds", "abstract": "Assigning one of K options to each of N groups under a total cost budget is a recurring problem in efficient AI, for instance in mixed-precision quantization, non-uniform pruning, and expert selection. The objective (model loss) depends on all assignments jointly and does not decompose across groups, which means that combinatorial solvers can only optimize proxy objectives. Methods such as evolutionary search evaluate the actual loss but lack gradients, while penalty-based methods enforce the budget only approximately and can require heavy hyperparameter tuning. We show that under softmax relaxation, the budget constraint defines a smooth Riemannian manifold in logit space with unusually clean geometry. Specifically, the normal vector is available in closed form, shifting logits along the cost vector changes expected cost monotonically, and the vector transport reduces to a single inner product. Building on this, we propose Riemannian Constrained Optimization (RCO), which wraps tangent projection, binary-search retraction, and momentum transport around a standard Adam step. Combined with Gumbel straight-through estimation and budget-constrained dynamic programming for discrete feasibility, RCO provides first-order optimization of the actual loss with the expected budget enforced exactly at every iterate, and no constraint-related hyperparameters. The same construction handles several simultaneous budgets, whenever an assignment must respect more than one resource limit, with no added coefficients. RCO exceeds or matches the performance of state-of-the-art methods in both synthetic problems and realistic LLM compression settings, often at considerably lower wall-clock cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Riemannian optimization;discrete assignment;constraint optimization;softmax manifold;geometric retraction" ], "forum_url": "https://openreview.net/forum?id=LJ3qgkXdhV", "pdf_url": "https://openreview.net/pdf/26c62a44e4a68b9260ef66dc123164cf2fca9082.pdf", "updated_at": "2026-10-05T18:10:54.654Z", "categories": [ "training" ], "matched_terms": [ "binary", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SLBIPbiamQ", "number": 9790, "title": "ConQuER: Modular Conditional Control of Frozen IQP Generators", "abstract": "Pretrained instantaneous quantum polynomial (IQP) generators learn data distributions but lack direct control over the generated samples when trained unconditionally. Practical applications often require generation conditioned on a specific mode or class. We study how to add this capability without updating the pretrained backbone or departing from the IQP circuit family. We propose ConQuER, a modular post-training framework that equips frozen IQP generators with a reusable conditional generation interface. Its controller adopts the same IQP structure and composes directly with the backbone to steer generation toward requested conditions. On Binary Blobs, a 784-support controller achieves mean conditional squared maximum mean discrepancy ($\\mathrm{MMD}^2$) of $0.0105$ and a mean requested-mode hit rate of $78.49\\%$. On $4\\times4$ binarized MNIST-3/5, a controller with 1,380 trainable parameters serves both classes and reduces mean conditional $\\mathrm{MMD}^2$ by $94.1\\%$ relative to the frozen unconditional backbone. Further evaluations on these tasks and Shapes20 characterize the trade-off between conditional distribution quality and parameter budget and demonstrate parameter savings through shared control across conditions.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantum Neural Network", "Quantum Machine Learning" ], "forum_url": "https://openreview.net/forum?id=SLBIPbiamQ", "pdf_url": "https://openreview.net/pdf/05f8045a60354f7ab87cad07e03c1f8beb5f906a.pdf", "updated_at": "2026-10-05T18:11:42.285Z", "categories": [ "ptq" ], "matched_terms": [ "binarized", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XMSwLKtADv", "number": 9796, "title": "Rethinking Numerical Stability in Hessian-Based PTQ: Accumulation-Order Sensitivity", "abstract": "Post-Training Quantization (PTQ) is widely used to reduce the computational and memory costs of large language models. Hessian-based PTQ methods such as GPTQ estimate empirical second-order statistics from calibration samples to guide weight quantization. Although the sum of these statistics is invariant to sample order in exact arithmetic, floating-point non-associativity can yield different empirical Hessians depending solely on the order in which the same samples are accumulated. While calibration sample selection is commonly treated as an important experimental condition, accumulation order is typically left uncontrolled. To our knowledge, we provide the first systematic evidence that reordering fixed calibration samples alone affects quantization and downstream performance in Hessian-based PTQ, with variability comparable to calibration resampling. We isolate this effect by varying only accumulation order across Qwen and Llama models, multiple Hessian-based PTQ methods, and weight bit-widths. In some settings, order-induced performance variation exceeds inter-method performance gaps and even reverses method rankings. Further analysis shows that order-dependent perturbations in the empirical Hessian are amplified through subsequent Hessian-dependent operations, ultimately altering quantization decisions. As a practical control, numerical stabilization incorporating compensated summation substantially suppresses order-induced variation. These findings highlight the need to account for accumulation order when evaluating and comparing Hessian-based PTQ methods. Our code is available at https://anonymous.4open.science/r/ReSQ-B3EF.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Hessian-based Quantization", "Numerical Stability", "Reproducibility" ], "forum_url": "https://openreview.net/forum?id=XMSwLKtADv", "pdf_url": "https://openreview.net/pdf/0c76bd65dd01dc8fcb1af29a36a531e2ea34e8b6.pdf", "updated_at": "2026-10-05T18:11:47.177Z", "categories": [ "ptq" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ecsBB7vNYv", "number": 9846, "title": "AxisQ: Depth-Adaptive Channel--Frequency Quantization for Camouflaged Object Detection", "abstract": "Camouflaged object detection (COD) depends on weak texture and boundary cues that can be destroyed by aggressive post-training quantization (PTQ). Existing PTQ methods mainly optimize quantizers or reconstruct network blocks, while using a single precision-allocation domain throughout hierarchical backbones. Instead, we introduce AxisQ, a depth-adaptive bit-allocation framework motivated by the observation that the useful allocation axis changes with feature depth. Under a fixed per-layer average activation-bit budget, AxisQ allocates precision in the channel domain for shallow features and in the local-frequency domain for deep features, using COD-aware sensitivity to better preserve subtle foreground regions and ambiguous boundaries. We also formulate a boundary-aware Hessian approximation for reconstruction by incorporating boundary-weighted gradient and error statistics. Experiments across multiple COD models and diverse benchmarks demonstrate improved accuracy over the strongest baselines at low bit widths, with relative weighted F-measure gains of 12.7% for CFRN and 33.3% for CamoFormer-P on CAMO under W2A2 quantization. These results support adapting the allocation domain to feature depth as a transferable quantization strategy for hierarchical COD models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training Quantization", "Camouflaged Object Detection", "Vision Transformers" ], "forum_url": "https://openreview.net/forum?id=ecsBB7vNYv", "pdf_url": "https://openreview.net/pdf/e1ab0769dda05ceaf13d9a0ff05f6df6873c1c7c.pdf", "updated_at": "2026-10-05T18:12:30.045Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low bit", "ptq", "quantization", "quantizers", "w2a2" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Q3dncJ4Sza", "number": 9847, "title": "TempoKV: Predictive KV Cache Quantization For Autoregressive Video Generation", "abstract": "Autoregressive (AR) video diffusion models enable streaming generation of long videos but suffer from significant memory overhead due to the rapidly growing KV cache. Existing KV cache quantization methods are primarily designed for LLMs and transfer poorly to video models, where activations exhibit different statistics and strong spatiotemporal redundancy. We observe that the video KV cache is largely predictable over time, with a clear asymmetry. Keys are highly predictable from their temporal history, while values depend primarily on the current key and need a short history of preceding values. Based on these observations, we propose \\textbf{TempoKV}, a training-free predictive KV cache quantization framework for AR video generation. TempoKV predicts keys and values according to their distinct dependencies and stores only the residuals at low precision. Both predictors are lightweight, calibrated once offline by linear regression on a small set of BF16 samples, and remain fixed across prompts and generation runs. Experiments on Self-Forcing, Causal Forcing, and LongCat-Video-13B show that TempoKV consistently achieves higher generation quality than existing KV cache quantization methods at comparable or higher compression ratios, reaching up to $6.71\\times$ KV cache compression with the best overall quality--efficiency tradeoff.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "autoregressive video generation", "video diffusion models", "long video generation", "memory-efficient inference" ], "forum_url": "https://openreview.net/forum?id=Q3dncJ4Sza", "pdf_url": "https://openreview.net/pdf/4f650d50e680531d69ccff19d2b3ce688152302a.pdf", "updated_at": "2026-10-05T18:12:33.290Z", "categories": [ "ptq", "kv-cache", "diffusion" ], "matched_terms": [ "low precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "V6RLMimId0", "number": 9954, "title": "From 201 Bins to 1 Bit: What a Binary Advantage Interface Retains in Vision-Language-Action Models", "abstract": "Vision-language-action (VLA) models are promising for general robot control, yet advantage-conditioned post-training can use a fine-grained categorical value critic even when the actor receives only a binary condition. We study what information this interface retains. Equality of the induced observation-action-label law is sufficient for equality of the supervised actor objective. For a mean-based residual, equal critic means suffice but are not necessary; we also bound objective changes under value and empirical-quantile perturbations. A covariance identity characterizes when ideal positive conditioning improves a chosen action utility. Direct binary prediction has a Bernoulli posterior as its Bayes target. In an idealized finite-alphabet model, mean-estimation sample complexity is independent of bin count, unlike full categorical recovery. Guided by this analysis, we train a scalar bootstrap score, cache recorded residuals in a Persistent Advantage Buffer, and fine-tune a prompt-conditioned flow-matching actor with classifier-free guidance. Physical evaluations on four bimanual manipulation tasks illustrate the pipeline against a shared behavioral-cloning initialization. The results separate sufficiency at the actor interface from the challenge of learning useful labels.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action Models", "Reinforcement Learning", "Embodied AI" ], "forum_url": "https://openreview.net/forum?id=V6RLMimId0", "pdf_url": "https://openreview.net/pdf/45ed176e57501332cecf353170f3f08de31ac50c.pdf", "updated_at": "2026-10-05T18:13:17.980Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "1 bit", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GViTZhaurM", "number": 10094, "title": "Coded Computation: Perceptual-Semantic Acceleration for Video Diffusion", "abstract": "Efficient video diffusion is commonly approached by making generative computation cheaper through fewer denoising steps, quantized operators, feature reuse, or selective computation. We explore a different perspective---can computation itself be coded and subsequently reconstructed? Inspired by coded acquisition in computational imaging, we introduce coded computation, a new formulation that jointly designs which generative computations are explicitly performed and how the omitted information is recovered. We instantiate this principle with a perceptual-semantic computational encoder and a lightweight code-aware generative decoder. The encoder combines bottom-up visual cues with top-down semantic relevance from the text prompt to construct a time-varying compute code, allocating richer computation to perceptually and semantically important content while encoding less critical content more efficiently. The decoder then leverages the compute code and strong spatiotemporal and diffusion priors to reconstruct the omitted generative information. Extensive experiments demonstrate that coded computation substantially accelerates video diffusion inference while preserving visual quality, temporal consistency, and prompt fidelity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Coded Computation", "Video Diffusion", "Acceleration" ], "forum_url": "https://openreview.net/forum?id=GViTZhaurM", "pdf_url": "https://openreview.net/pdf/dc38f0efc2b62a85fd34c6b4a7b7498e7767fcdf.pdf", "updated_at": "2026-10-05T18:14:40.568Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dZVIGLXkqO", "number": 10129, "title": "AMS-Quant: Adaptive Mantissa Sharing for Floating-Point Quantization", "abstract": "Large language models (LLMs) have demonstrated remarkable capabilities in various kinds of tasks, while the billion or even trillion parameters bring storage and efficiency bottlenecks for inference. Quantization, particularly floating-point quantization, is known to be capable of speeding up LLM inference by reducing memory footprint and data movement during the inference process. For the first time, we advance the floating-point quantization exploration from integer bit-widths to non-integer bit-widths, namely AMS-Quant, to further approach the quantization sweet spot. AMS-Quant incorporates two novel techniques to put it into effect: (1) it proposes Mantissa-bit Sharing, which groups k quantized weights and lets them share the least significant mantissa bit, allowing us to further approach the minimum quantization bit-width without accuracy loss. (2) It introduces Adaptive Searching, which employs an offline optimization strategy to minimize the accuracy degradation introduced by sharing. Moreover, AMS-Quant is also prototyped as efficient CUDA Linear kernels, which translates memory savings into wall-clock latency reduction by reducing memory access. Extensive experiments on large-scale datasets and models show that AMS-Quant can quantize the model to FP-5.33-e2m3 and FP4.25-e2m2, and significantly speed up the LLM decoding over FP16 inference (2.8$\\times$ and 3.2$\\times$), with negligible accuracy loss.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Quantization" ], "forum_url": "https://openreview.net/forum?id=dZVIGLXkqO", "pdf_url": "https://openreview.net/pdf/706a3aa80ca36f0e3b36b3e834852e9c26e21d32.pdf", "updated_at": "2026-10-05T18:14:52.385Z", "categories": [ "training", "hardware" ], "matched_terms": [ "bit-width", "fp4", "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Mm5qnq9ufn", "number": 10300, "title": "Rewind the Transformer: Input-Anchored Computational Replay", "abstract": "Transformer states have continuous coordinates, but the states a node can take are generated by discrete admissible sources. Reversibility is then a question about this realizable set: which source distinctions does an observation preserve? When the input's identity survives, recovering it once lets one deterministic forward pass replay every declared state without inverting any layer. For a lossy code fixed before recovery, a hybrid decoder built on SipIt and first-order token ranking makes the route executable: cheap search proposes candidates, only the observation's own rule accepts them, and ordinary exhaustion falls back once to SipIt. Behind a frozen 8-bit-per-coordinate interface on openai-community/gpt2, both a SipIt-based decoder and the hybrid recovered the 8-token input of each of 8 separately selected documents and reproduced all 14 declared nodes bitwise, whereas continuing from the dequantized state shifted the next-token distribution on every document (top prediction still agreed on all 8). In a fitted-proposal extension on 16 fixed news inputs, an affine proposer completed all 16 in both repetitions, versus 15 for non-speculative continuous recovery under a 120 s decode cap, with a median 3.5× request-time speed-up on the 15 common completions. Diagnostics at known prefixes show these distinctions coming apart for geometry, read-outs, partial views and a known edit. Discreteness alone does not guarantee distinguishability and a match is not a uniqueness certificate; these results establish a small-scale workflow, not a generally superior codec or a proof of unique recoverability over all inputs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Language Model Inversion", "Hidden Representations", "Representation Invertibility", "Activation Privacy", "Probing" ], "forum_url": "https://openreview.net/forum?id=Mm5qnq9ufn", "pdf_url": "https://openreview.net/pdf/4ceb53059cf21ddf5499a990792ad0e25bace1e3.pdf", "updated_at": "2026-10-05T18:16:39.836Z", "categories": [], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7tCLWw24Nx", "number": 10322, "title": "UniCG: A Framework for Unifying Graph Comprehension and Generation", "abstract": "The pursuit of general-purpose intelligence has motivated unified architectures that both comprehend and generate content, with notable progress in language and vision. However, prevailing graph learning paradigms still treat these capabilities separately: comprehension models map graphs to predictions, whereas generative models synthesize graphs without being evaluated for comprehension. Consequently, enabling a single model to comprehend and generate graphs through a shared representation remains a major challenge. In this paper, we propose UniCG, a novel framework for Unifying graph Comprehension and Generation that addresses both the representational incompatibility and optimization interference between these capabilities. Specifically, to establish a common graph interface, UniCG combines lossless graph serialization with residual quantization of node semantics to construct a shared discrete graph vocabulary. To support both capabilities within one model, UniCG formulates generation as next-token prediction and comprehension as a pooled read-out over a single autoregressive backbone. To alleviate the asymmetric interference exposed by joint training, UniCG routes the two objectives through task-decoupled experts while retaining the shared backbone. Extensive experiments across 20 graph domains, four comprehension tasks, and three generation regimes against 13 baselines demonstrate that UniCG jointly supports both capabilities, achieving competitive comprehension performance while generating valid and novel graphs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Unified Graph Comprehension and Generation", "Graph Representation Learning", "Graph Foundation Models", "Graph Generation" ], "forum_url": "https://openreview.net/forum?id=7tCLWw24Nx", "pdf_url": "https://openreview.net/pdf/ca6356b29e3f718d52ba564335101cde61825d11.pdf", "updated_at": "2026-10-05T18:16:49.513Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pGXWVPoQef", "number": 10339, "title": "HelixKV: Allocating Bits Where They Matter for Autoregressive Video Diffusion", "abstract": "Autoregressive video diffusion models enable streaming and long-horizon video generation, but their key--value (KV) caches grow linearly with sequence length, creating an increasingly severe memory bottleneck.\nExisting KV-cache quantization methods typically rely on uniform or coarse-grained precision allocation. This wastes memory on quantization-tolerant cache regions and leaves fewer bits for regions where precision matters most. To alleviate this problem, we introduce HelixKV, a training-free KV-cache compression framework that allocates bits where they matter. Specifically, HelixKV characterizes the impact of KV quantization through model-output distortion and allocates precision along two complementary dimensions.\nAcross attention heads, it assigns head-specific bit-widths to keys and values according to their quantization sensitivity.\nAlong the temporal axis, it further differentiates the precision of recent and historical KV states, allowing each head to retain only the precision required by different parts of its history. Combining these distortion profiles with byte-accurate storage costs, HelixKV solves a global budgeted allocation problem to jointly allocate KV precision across the head and temporal axes.\nBy compressing quantization-tolerant cache regions and reallocating the saved memory to regions where precision matters more, HelixKV improves generation quality at a fixed KV-cache budget while preserving the complete KV history.\nExperiments on Self-Forcing-1.3B and HY-WorldPlay-5B show that HelixKV achieves $6.97\\times$ and $6.61\\times$ KV-cache compression, respectively, while retaining 99.9\\% of the VBench Total score on both models, demonstrating a favorable memory--quality trade-off.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Autoregressive Video Diffusion Model" ], "forum_url": "https://openreview.net/forum?id=pGXWVPoQef", "pdf_url": "https://openreview.net/pdf/b8b613824e4b34618a5b033852914a6db5b6b69f.pdf", "updated_at": "2026-10-05T18:17:00.721Z", "categories": [ "kv-cache", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LidSUUTrJW", "number": 10382, "title": "SPARK: Spatial Policy-driven Adaptive Reinforcement learning for Knowledge distillation", "abstract": "Low-bit quantization enables deployment of image restoration (IR) networks on resource-constrained devices, but introduces rounding noise that disproportionately degrades high-frequency regions such as edges and fine textures. Existing knowledge distillation (KD) methods apply distillation signals uniformly across all spatial locations, overlooking the varying reconstruction difficulty across image regions. To address this, we propose SPARK (Spatial Policy-driven Adaptive Reinforcement learning for Knowledge distillation), a framework that adaptively allocates distillation effort using a lightweight reinforcement learning (RL) policy network. At each training step, a difficulty feature extractor computes four signals, namely Laplacian variance, pixel variance, student reconstruction error, and teacher-student knowledge gap, which are fed into a compact policy CNN that produces a stochastic spatial weight map to modulate the KD loss during quantization-aware training (QAT). SPARK is IR task-agnostic, adds no inference cost, and integrates into any existing QAT pipeline without architectural changes. Experiments on benchmark datasets demonstrate that SPARK consistently outperforms PTQ, QAT, and state-of-the-art (SOTA) KD approaches across multiple student architectures, achieving reconstruction quality closest to the full-precision teacher under significant computational constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Image restoration", "Knowledge distillation", "Reinforcement Learning", "Image enhancement" ], "forum_url": "https://openreview.net/forum?id=LidSUUTrJW", "pdf_url": "https://openreview.net/pdf/d00fad1854ba246a8fa730dc58f015884c4aaa4b.pdf", "updated_at": "2026-10-05T18:17:21.999Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "low-bit", "ptq", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VQXQTBmP5X", "number": 10404, "title": "Beyond Shadow Weights: Quantization-Aware Training as Quantized-Endpoint Descent", "abstract": "Quantization-aware training (QAT) updates a full-precision shadow weight $\\mathbf{x}$ but deploys the quantized endpoint $Q(\\mathbf{x})$. Existing explanations for QAT largely view its success through the lens of shadow weights: QAT can move $\\mathbf{x}$ toward flatter basins, gain robustness from oscillations, or balance the shadow loss $f(\\mathbf{x})$ against the quantization error $\\|\\mathbf{x}-Q(\\mathbf{x})\\|_2$. These perspectives do not directly explain the empirical observation that the deployed endpoint loss $f(Q(\\mathbf{x}))$ improves even as the shadow loss $f(\\mathbf{x})$ increases substantially. In this paper, we offer a different explanation by treating QAT as finite-grid endpoint dynamics. Motivated by the approximate normality of rescaled pretrained weights, we propose an idealized residual-phase model that yields a crossing law that determines which coordinates cross quantization boundaries after a shadow update. We also find that signal imbalance creates asymmetric oscillations that favor weaker-gradient endpoints. This motivates QAR (Quantization with Amplified Routing), an algorithmic framework that directly operates on the quantization code. We further prove feasible-gradient guarantees for QAR with a family of power amplifier. Post-training experiments on three large language models support the endpoint view and show that QAR can achieve performance comparable to QAT with up to 24.1% peak memory saving.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-aware training", "large language models", "straight-through estimator", "crossing law", "signal imbalance" ], "forum_url": "https://openreview.net/forum?id=VQXQTBmP5X", "pdf_url": "https://openreview.net/pdf/a1012eeed1fbb7c1993ea2322e1694074703b05a.pdf", "updated_at": "2026-10-05T18:17:35.230Z", "categories": [ "ptq", "qat", "training", "theory" ], "matched_terms": [ "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aznqUxfIuv", "number": 10432, "title": "Quieter, Not Safer: Quantization Inflates Oversight-Invisible Damage in LLM Agents", "abstract": "Quantization is a common step in deploying LLM agents on cheap hardware, and its cost is assumed to be a little accuracy. We show it can also change what an agent does in ways the usual checks do not see. Beyond response-level safety—refusals that fail, biased answers—an agent that acts can complete its task and, in the same run, quietly edit what it was never asked to touch. We build Tripwire, 180 file-editing tasks in three domains, each planted with registered out-of-scope temptations, so that every violation is a deterministic function of file state and is graded by the weakest oversight that would catch it. Across six base models and four 4-bit schemes at two scales we find: (1) in four of ten paired comparisons—NF4 at both scales, FP4 and AWQ at 7B, three of them surviving Bonferroni correction—quantization raises the rate of runs containing such a violation 3–5×; at 7B this happens with task success unchanged at 90/90, and at 14B NF4 raises violations inside successful runs from 3 to 22 of 90 while success falls from 85 to 78; every quantized condition adds more violating tasks than it removes, and NF4, the QLoRA default, is the most severe case at 14B and inflates at both scales; (2) task success and the agent's own report do not register the increase, and a 7B reviewer misses three quarters of the 54 violations that occur inside successful 14B runs, leaving a residual under NF4 that is five times full precision's and a strict superset of it, while a 72B reviewer reading the full diff flags them all; (3) the visible cost is small and the hidden cost large—1.6 to 2.8 MMLU points in the four inflating comparisons against 3–5× more violations—and neither bit-width, calibration, nor codebook geometry accounts for which of our configurations is affected. At 7B these agents match full precision on task success. They are not safer.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "LLM agents", "AI safety", "agentic evaluation", "oversight", "benchmark", "scope violations", "side effects", "NF4", "QLoRA" ], "forum_url": "https://openreview.net/forum?id=aznqUxfIuv", "pdf_url": "https://openreview.net/pdf/efc7651a9515a4c318142e1846240aacfb054610.pdf", "updated_at": "2026-10-05T18:17:50.333Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "4-bit", "bit-width", "fp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sF70kBTBO3", "number": 10436, "title": "PrismQ: Noise-Aware Resource Allocation for Elastic Post-Training Quantization", "abstract": "Elastic post-training quantization (PTQ) jointly calibrates multiple precisions within a single model, enabling flexible precision adaptation under different deployment budgets. However, extending reconstruction from a single precision to multiple precisions introduces two key challenges. First, quantization errors vary across linear layers and precisions, creating a trade-off between cross-precision parameter sharing and precision-specific reconstruction capacity within the reconstruction space. Second, calibration states should reflect precision-dependent error propagation, while preserving complete quantized states for every precision incurs substantial calibration memory overhead. To address these challenges, we propose PrismQ, a noise-aware elastic PTQ framework that adapts both reconstruction space and calibration states according to precision-dependent quantization errors. Specifically, PrismRank estimates the demand for reconstruction capacity at each linear layer and precision from the singular-value characteristics of clipping-induced errors, and PrismLoRA realizes the resulting capacity allocation through shared and precision-specific low-rank components. Moreover, Calibration-only Sparse Token selects quantized token states according to token-wise quantization error under a fixed cache budget and uses them to construct calibration inputs for the next block, allowing these inputs to reflect precision-dependent error propagation. Extensive experiments on language models, vision transformers, and vision-language models show that PrismQ improves quantization performance across multiple precisions, with particularly clear gains at low bit widths.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Elastic Post-Training Quantization", "Multi-Bit Quantization", "Weight-Activation Quantization", "Low-Rank Adaptation", "Noise-Aware Rank Allocation", "Calibration-State Propagation", "Large Language Model Quantization" ], "forum_url": "https://openreview.net/forum?id=sF70kBTBO3", "pdf_url": "https://openreview.net/pdf/87ac1a1a04f97591aff2608dc6d02b111e00e8d4.pdf", "updated_at": "2026-10-05T18:17:56.914Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5oSJKHhm1N", "number": 10468, "title": "The Butterfly Effect in Large Language Model Quantization", "abstract": "Quantization has been widely used to reduce the memory footprint and inference latency of large language models (LLMs). Low-bit quantizers have typically been evaluated using average metrics such as perplexity and task accuracy. Nevertheless, these average metrics do not reveal failures on individual inputs, such as correct answers becoming wrong after quantization. Focusing on tail reliability, we propose a unified source-to-decision analysis, which measures rare output failures and traces each input's quantization errors from individual modules through downstream amplification to changes in output distributions, token decisions, and autoregressive generation. Our analysis reveals the quantization butterfly effect: High-gain computation paths amplify small rounding errors on particular inputs into large output shifts and token changes. These amplified errors form an upper tail hidden by average metrics, and an early token change redirects the autoregressive continuation. Our experiments demonstrate the quantization butterfly effect on Qwen2.5-3B, where 4-bit weight-only GPTQ increases mean perplexity by only 7.3\\%, yet the 99th-percentile (p99) per-token KL divergence is 8.9 times the mean, 14.8\\% of top-1 predictions flip, and almost every greedy continuation leaves the full-precision trajectory. The same pattern recurs across model families and zero-shot, reasoning, code, and instruction-following tasks. Similar average accuracy scores do not show how often correct predictions become wrong. Our perturbation--propagation theory factorizes each module's impact into local error and downstream gain along the error's direction. We prove non-asymptotic KL bounds, derive an exact propagation-aware decision radius and a sharp multiclass KL barrier for top-1 changes, and establish guarantees for preserving greedy trajectories; finite-perturbation bounds extend Fisher-based control to every quantile and CVaR. Our theory-guided diagnostic score identifies intervention targets for causal validation. Restoring only the four highest-scoring layers to full precision while keeping all other layers at 3 bits reduces the p99 divergence gap by up to 83\\%. To improve tail reliability, we introduce two complementary mitigation methods: Propagation-Aware Precision Allocation (PAPA) protects high-impact modules at a fixed bit budget, while Butterfly-Aware Reliability Tuning (BART) learns a lightweight correction at a fixed bit-width. By updating under 2\\% of parameters, BART raises 3-bit GSM8K accuracy from 0.030 to 0.350. Tail reliability must therefore be measured and reported alongside average performance, and optimized as a quantization objective.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization", "butterfly effect", "tail reliability", "error propagation", "decision robustness", "mixed precision" ], "forum_url": "https://openreview.net/forum?id=5oSJKHhm1N", "pdf_url": "https://openreview.net/pdf/0cd0c2cfc7e508d470bab82567181ac9b4f75701.pdf", "updated_at": "2026-10-05T18:18:09.032Z", "categories": [ "ptq", "hardware", "reasoning", "theory" ], "matched_terms": [ "3-bit", "4-bit", "bit-width", "low-bit", "mixed precision", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AhG6FcJBHt", "number": 10602, "title": "LoopQuant: Loop-Aware Post-Training Quantization for Vision-Language-Action Models", "abstract": "Vision-language-action (VLA) models achieve strong manipulation performance, yet\ntheir substantial memory footprint and latency hinder real-time execution on robot\naccelerators. In practice, a VLA operates across two nested loops: an inner\nloop over denoising steps and an outer control loop that initiates each re-planning request\nfrom a new observation. Under W4A8 post-training quantization (PTQ), static four-bit weights\ncause reconstruction errors to recur at every denoising step and compound across consecutive requests\nthrough executed actions. Meanwhile, dynamic activations shift across denoising steps, and we\nfind that their channel peaks drift across observations, requiring offline calibration thresholds\nto cover both loops. We propose **LoopQuant**, a training-free W4A8 PTQ framework tailored\nto these nested loops. **Request-aware multi-observation activation extremal calibration**\n(RMAEC) records per-channel token percentiles at every denoising step, aggregates their maxima\nwithin each request and across diverse observations, and freezes the resulting scales.\nAn extreme-value model predicts that these scales grow sub-linearly with the number of observations.\nTo reduce weight error, **exact non-uniform codebook construction** aggregates repeated weight values\nto make dynamic programming tractable at layer scale, yielding globally optimal levels for fixed row scales.\nAt runtime, **fused packed execution** decodes four-bit weights in a Triton kernel for half-precision\nmatrix multiplication and executes the expert core with a single launch per denoising step.\nAveraged across four LIBERO suites, LoopQuant achieves $97.3\\%$ closed-loop success on $\\pi$0.5 and $84.3\\%$\non GR00T-N1.5, outperforming the strongest quantized baseline by $2.5$ and $4.0$ percentage points,\nrespectively. The packed $\\pi$0.5 runtime serves requests $1.88\\times$ faster than the unquantized\nBF16 model, requiring only $0.59\\times$ of its resident GPU memory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization; Model Compression; Vision-Language-Action Model" ], "forum_url": "https://openreview.net/forum?id=AhG6FcJBHt", "pdf_url": "https://openreview.net/pdf/1ef0d2cdb708cf38ee10d950cebdb56771a5cb0f.pdf", "updated_at": "2026-10-05T18:19:15.192Z", "categories": [ "ptq", "multimodal", "diffusion", "hardware" ], "matched_terms": [ "ptq", "quantization", "quantized", "w4a8" ], "review_status": "abstract", "summary": "LoopQuant 针对机器人 VLA 的两层执行循环：权重误差会随去噪步骤和连续控制请求反复传播，而激活峰值也会随观察改变。方法跨步骤和多观察记录通道极值以冻结激活尺度，用动态规划构造非均匀权重码本，并通过 Triton 打包执行减少每个去噪步骤的启动开销。", "contribution": "在四个 LIBERO 套件上，π0.5 和 GR00T-N1.5 的平均闭环成功率分别为 97.3% 和 84.3%，高于最强量化基线 2.5 和 4.0 个百分点；打包 π0.5 相比 BF16 请求执行快 1.88 倍，驻留 GPU 内存为其 0.59 倍。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。速度和内存数字针对摘要所述 π0.5 打包运行时；其矩阵乘仍为半精度，不能据 W4A8 推定为原生 INT4/INT8 运算。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "VLA 模型权重与动态激活；内层去噪循环和外层重规划请求", "bit_width": "W4A8；运行时解码权重后使用半精度矩阵乘", "evidence": [ { "location": "官方摘要", "claim": "π0.5/GR00T-N1.5 在四个 LIBERO 套件的平均成功率为 97.3%/84.3%。" }, { "location": "官方摘要", "claim": "打包 π0.5 请求执行为 BF16 的 1.88 倍速度，驻留 GPU 内存为 0.59 倍。" } ] }, { "id": "5o52SL4jFx", "number": 10606, "title": "QuIL: Quantifying Quantization-induced Information Loss in Neural Networks", "abstract": "Floating-point bit-string representations of parameters are ubiquitous building blocks of neural networks, and their quantization into lower precision formats is increasingly used to reduce their computational requirements. The degradation that quantization induces in the model is, however, not well studied, and its detection relies on task-specific heuristics. We propose QuIL: An information-theoretic approach for measuring quantization-induced information loss. By modelling the probability distribution of the underlying floating-point bit-strings of the parameters with a probabilistic circuit, we can measure information loss as the change in entropy across quantization levels. Experiments on neural-network regression, vision transformers, and GPT models show that QuIL, computed from the parameter posterior alone, identifies the bit-width at which quantization begins to degrade accuracy, calibration, and predictive density, offering a principled basis for selecting parameter precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "neural networks", "probabilistic circuits", "quantization", "entropy" ], "forum_url": "https://openreview.net/forum?id=5o52SL4jFx", "pdf_url": "https://openreview.net/pdf/91d4c096d15927c215119f8706f2a23b431ee515.pdf", "updated_at": "2026-10-05T18:19:10.614Z", "categories": [ "ptq" ], "matched_terms": [ "bit-width", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cDud0Ztrwv", "number": 10677, "title": "QuantRet: Quantization-Robust Language Model Pretraining via Geometric Retractions", "abstract": "Low-bit post-training quantization (PTQ) can fail when a small number of coordinates set the quantization range or when high-gain directions amplify local errors. We introduce QuantRet, a pretraining method for residual Transformer decoders. QuantRet retracts selected activations to fixed-radius spheres and constrains the norms of hidden weight matrices throughout pretraining. These operations fix the radial scale at selected activation sites and constrain how hidden linear maps rescale their inputs without changing the quantizer. Our analysis separates weight and activation perturbations, bounds their local amplification under the matrix constraints, and gives a conditional bound on their propagation through spherical residual updates. On Qwen3-like models, QuantRet lowers language-modeling loss under low-bit weight, activation, and key--value-cache quantization, and its strongest variant achieves higher zero-shot commonsense accuracy under the same quantized graphs. Mechanism measurements show that similar weight signal-to-noise ratios can produce different final dot-product signal-to-noise ratios and end-to-end degradation. These results support pretraining-time geometric constraints as a practical way to improve low-bit PTQ.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-robust pretraining", "post-training quantization", "manifold constraints" ], "forum_url": "https://openreview.net/forum?id=cDud0Ztrwv", "pdf_url": "https://openreview.net/pdf/eed77ad00bffa1abd8dfb670441c015530a8ec38.pdf", "updated_at": "2026-10-05T18:19:40.250Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GYZUysAiKG", "number": 10703, "title": "Understanding the Weight Averaging Mechanism in LLM Training for Post-Training Quantization", "abstract": "Large language models (LLMs) are typically pretrained in high precision but increasingly deployed with low-precision post-training quantization (PTQ). Recent studies have shown that using weight averaging during pretraining can improve PTQ performance compared with learning-rate decay, suggesting that it might provide a simple way to improve the pretraining-to-quantization transition. But the mechanism behind weight averaging remains insufficiently explained. This leads to inconsistent and fragile performance gains, thereby preventing practitioners from applying such a technique confidently. As a response, we formulate weight averaging as a trade-off between retaining training progress and improving robustness under perturbation. We further derive a continuous family of averaging kernels that unifies conventional strategies and achieves the Pareto frontier between the two competing goals. Critically, a theoretical framework for performing weight averaging under PTQ is developed. It can be shown that coarser quantization is more susceptible to perturbations, whereas finer quantization could be less affected. Thus, our results could provide unified theoretical guidance for performing weight averaging under different PTQ conditions. Experiments validate both the predicted behavior and the proposed averaging strategy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Pretraining", "PTQ", "Weight Averaging" ], "forum_url": "https://openreview.net/forum?id=GYZUysAiKG", "pdf_url": "https://openreview.net/pdf/682a4b4aa787cff85adc369c38cbcded7ad1e9a7.pdf", "updated_at": "2026-10-05T18:19:58.428Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "low-precision", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cAyqQ6lwIW", "number": 10718, "title": "SteerDistill: Quantizing LLM Agents via Online Teacher-Steered Trajectory Distillation", "abstract": "Quantization makes multi-turn LLM agents brittle: a single erroneous\ndecision alters every subsequent interaction and can derail the task. We\nobserve that quantization rarely breaks an agent's overall plan; instead, it\ncorrupts a few executable decisions, such as a tool call with the wrong\ninterface or a misgrounded argument. Existing recovery methods either learn\nfrom fixed full-precision data that misses the states a quantized agent\nreaches, or supervise the student's own rollouts without locating the error\nand rely on later rollouts to avoid it; neither corrects the turn at which a\nrollout goes wrong. We introduce SteerDistill, a quantization-aware distillation method\nbuilt on *online teacher steering*. When a quantized student's rollout\ndiverges from a successful full-precision trajectory, the teacher issues a\nsteering prompt at the divergent turn; the student regenerates that turn\nitself, the environment verifies the correction, and the rollout continues\nfrom the corrected state. The steered behavior is then distilled into the\nstudent without the prompt, so no teacher is needed at inference. Across\nfive agentic benchmarks and three models with 4-bit weights, \\nickname\noutperforms the strongest quantized baseline by 1.67--2.18 average points and\nstays within one point of BF16. On $\\tau^2$-Bench, it also exceeds\nquantization-aware reinforcement learning with 67\\% fewer training GPU hours.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Agent", "LLM Quantization", "Distillation" ], "forum_url": "https://openreview.net/forum?id=cAyqQ6lwIW", "pdf_url": "https://openreview.net/pdf/a44b89879a8a428cfbea33f27af44850c488cfa4.pdf", "updated_at": "2026-10-05T18:20:14.989Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "4-bit", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "I9sF2v1sOi", "number": 10816, "title": "SuBitKV: Sub-Bit KV Cache Compression for Autoregressive Video Generation", "abstract": "Autoregressive video generation relies on a growing key–value (KV) cache to retain information from previous frames, which raises a memory challenge as video length increases. However, retaining the full KV cache at high precision becomes impractical under limited GPU memory. Discarding past states or reducing their precision saves memory but can compromise consistency in appearance, structure, and motion over time. To address this challenge, we introduce SuBitKV, a training-free KV cache compression framework that supports storage below one bit per original KV element. Its key idea is to preserve a low-memory approximation of the retained KV cache as a persistent base, then allocate additional bits across the KV cache within the memory budget to improve reconstruction fidelity. We construct this persistent base using residual vector quantization, which jointly represents the elements of each key or value vector with a small set of compact indices rather than storing each element separately. As the cache grows, SuBitKV reduces these extra bits to stay within budget while retaining the persistent base. Experiments on three video-generation and world-model backbones show a strong memory--quality trade-off. On Wan2.2, SuBitKV achieves $21.42\\times$ compression with near-BF16 VBench scores and reaches $30.09\\times$ under tighter memory budgets.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "V cache compression", "video generation", "residual vector quantization", "low-bit quantization", "memory-efficient inference", "autoregressive video generation", "sub-bit compression" ], "forum_url": "https://openreview.net/forum?id=I9sF2v1sOi", "pdf_url": "https://openreview.net/pdf/f6d7297b70bccd9aed0d03f6270a109d8e716649.pdf", "updated_at": "2026-10-05T18:21:17.948Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rbuqRwx9F0", "number": 10959, "title": "Let Attention Flow: Energy-Dissipating Guidance for Diffusion Models", "abstract": "Diffusion models achieve remarkable generative performance by iteratively refining noisy samples through a sequence of denoising operations. An important component driving this refinement is the attention mechanism within the learned parameterization, which governs how information flows across spatial locations, semantic tokens and intermediate features. Existing training-free guidance methods often rely on score-space corrections or heuristic changes to internal activations, such as blurring attention maps. We instead guide diffusion by evolving attention through a Partial Differential Equation (PDE) inspired by Landau–de Gennes (LdG) dynamics. This energy-dissipating flow suppresses unstable attention patterns while preserving directional structure, providing a principled way to modulate the model’s internal information flow without retraining. The resulting guidance integrates cleanly with modern text-to-image samplers, requires no architectural modification, and improves generation quality both qualitatively and quantitatively. On MS-COCO, our method improves over SDXL and prior attention-based guidance methods. Under 4-bit model quantization, it improves generation quality and outperforms the unguided FP16 SDXL baseline across all reported metrics.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "PDE" ], "forum_url": "https://openreview.net/forum?id=rbuqRwx9F0", "pdf_url": "https://openreview.net/pdf/2dd624b8a97e5e0055b4d5e38ee3e48c383a45d9.pdf", "updated_at": "2026-10-05T18:22:37.687Z", "categories": [ "diffusion" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dpcP5dlK9P", "number": 11001, "title": "PyraQuant: Spatial Mixed-Precision Quantization for High-Resolution Diffusion Models", "abstract": "High-resolution diffusion generation remains expensive because denoising cost grows rapidly with image resolution. Progressive generation reduces this cost by first generating at a low resolution and then progressively increasing the resolution, while existing diffusion quantization methods mainly allocate precision across layers, channels, or denoising steps without exploiting the stage-wise and spatial structure of this process. We find that progressive refinement exhibits two forms of non-uniformity: quantization tolerance varies across resolution stages, and refinement changes differ across spatial regions. Based on these observations, we propose PyraQuant, a training-free precision allocation framework for quantized progressive diffusion models. PyraQuant keeps the base quantizer at the lowest resolution. At each subsequent resolution transition, it assigns each region a precision according to how much it changes, carrying these decisions to the next resolution. The lower-precision configuration is derived directly from the original quantized representation without additional calibration or model copies, and regions requiring limited refinement reuse cached predictions. Experiments show that on FLUX.1-schnell with the SVDQuant W4A4 base quantizer, PyraQuant reduces effective precision from 4.48 to 2.00 bits while maintaining comparable generation quality, and achieves a $1.81\\times$ end-to-end speedup on UltraHR-eval4K.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "2D", "Diffusion", "AR", "Image", "quantization", "CV" ], "forum_url": "https://openreview.net/forum?id=dpcP5dlK9P", "pdf_url": "https://openreview.net/pdf/b93aea2d94c1695e4f3ea748bc07645873196063.pdf", "updated_at": "2026-10-05T18:23:07.073Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "mixed-precision", "quantization", "quantized", "quantizer", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NVmypKWAd2", "number": 11062, "title": "MXAttention: Data-Free Optimal Scaling and Pre-Normalization Quantization for MXFP4 Attention", "abstract": "The quadratic cost of attention limits diffusion-based video generation. MXFP4 attention offers a promising path toward lower-cost inference, but direct quantization often degrades generation quality. We identify two numerical causes: power-of-two shared scaling creates a clipping–underflow trade-off within MXFP4 blocks, while quantization in the online softmax loop breaks row-wise normalization, so the attention weights no longer sum to one. We propose MXAttention, a data-free post-training quantization framework with two components. First, Universal Optimal Scaling (UOS) exploits the periodic structure of power-of-two microscaling to minimize a global MXFP4 quantization-error objective. It yields the closed-form, distribution-independent boundary Qmax = 7.25, without calibration or per-layer search. Second, Pre-Normalization Quantization (PNQ) uses the same quantized exponential tiles in the normalizer and output updates, preserving row-wise normalization. On Wan2.2 and HunyuanVideo, MXAttention substantially improves frame-level similarity to FP16 outputs. It preserves FP16-level generation quality with less than 0.01 absolute degradation on all reported VBench metrics and achieves video quality competitive with strong NVFP4-based baselines. We further evaluate LLM quantization on Qwen3-14B, observing accuracy gains for attention quantization, weight–activation quantization, and their combination. On one Ascend 950-series NPU, MXAttention delivers these accuracy gains at nearly the same attention latency as the PNQ-enabled MXFP4 baseline (0.13% difference). MXAttention retains a 2.70× attention speedup and a 1.84× total video-generation task speedup over BF16 attention.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "low-bit inference", "attention quantization", "MXFP4", "microscaling floating-point formats", "diffusion transformers", "video generation" ], "forum_url": "https://openreview.net/forum?id=NVmypKWAd2", "pdf_url": "https://openreview.net/pdf/3e496b78d10be05ba8d7e6c98d667097359e8943.pdf", "updated_at": "2026-10-05T18:23:47.639Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "low-bit", "mxfp4", "nvfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4D9gkWAVWM", "number": 11084, "title": "EvoWorld: Evolving Gaussians for 4D Occupancy Forecasting and Motion Planning", "abstract": "Future 3D semantic occupancy forecasting and motion planning are central to autonomous driving, as they require models to reason about how surrounding scenes evolve and how the ego vehicle should act. Vision-centric world models must infer this evolution from camera observations, yet many existing approaches discretize scenes into latent embeddings, volumetric features, or quantized tokens and forecast future states through fixed-step autoregressive generation. This limits temporal flexibility and makes the predicted scene evolution difficult to interpret. We propose EvoWorld, a vision-centric, non-autoregressive occupancy world model that predicts a compact set of 4D semantic Gaussian primitives from multi-view camera histories to represent the entire forecast horizon. For each primitive, EvoWorld directly predicts conditional spatial geometry, temporal support, and planar velocity, specifying its shape, motion, and temporal contribution at any queried time. Conditioning and splatting these primitives yields semantic occupancy at arbitrary timestamps, with independent temporal queries that can be decoded in parallel. Primitive velocity is further decomposed into shared ego-induced scene motion and object-level dynamic motion, and the same Gaussian representation supports open-loop motion planning. Extensive experiments on Occ3D-nuScenes and Occ3D-Waymo demonstrate improved vision-centric future occupancy forecasting and strong motion planning performance, with further analyses showing accurate primitive motion and consistent object association.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Autonomous driving", "occupancy forecasting", "motion planning", "gaussian splatting" ], "forum_url": "https://openreview.net/forum?id=4D9gkWAVWM", "pdf_url": "https://openreview.net/pdf/7d0f71fb474bd9a7d3f6ed98cb87a7e9e9c978e4.pdf", "updated_at": "2026-10-05T18:23:58.604Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VXb6iC5X71", "number": 11144, "title": "PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers", "abstract": "Smaller activation outliers do not necessarily imply better low-bit quantization: their alignment with the quantizer matters. We introduce PrismQuant, a quantizer-aware rotation framework that aligns the leading activation eigenspace with the constant group subspace of asymmetric grouped INT4. The affine offsets represent the energy in this subspace without widening the range within the group. We formulate rotation design as a Ky Fan trace maximization and derive a closed-form solution that is provably optimal for this alignment objective. Compact Householder transformations and their compact-WY representation enable gradient-free construction and efficient application at both foldable and online sites. A predictive range law further connects unaligned activation energy and group size to quantization-relevant variation. Experiments on Llama, Qwen, and Mistral span dense models up to 70B parameters and a 30B mixture-of-experts model. Under W4A4KV4, PrismQuant sets the state of the art on Llama-3.2-3B among the compared methods in both perplexity and accuracy. On Llama-3.1-70B, it attains 3.85 perplexity and 72.46% average zero-shot accuracy, only 0.22 percentage points below the full precision. In the deployment study on Llama-3.1-8B, our optimized implementation achieves 1.51× prefill and 1.22× CUDA Graph decode speedups over matched FP16 baselines, with 56.34% lower decode peak memory and only 2.35% additional Graph decode latency over Hadamard.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Model Quantization", "Low-Bit Inference", "Activation Quantization", "W4A4KV4", "Rotation-Based Quantization", "Quantizer-Aware Optimization" ], "forum_url": "https://openreview.net/forum?id=VXb6iC5X71", "pdf_url": "https://openreview.net/pdf/a843cfb6e91a78bf7390292f4b87f31967c46614.pdf", "updated_at": "2026-10-05T18:24:39.752Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "int4", "low-bit", "quantization", "quantizer", "quantizers", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RTIRFpI696", "number": 11205, "title": "OpenWan-Light: High-Quality Data Enhancing and Accelerating Video Generation", "abstract": "Recent advances in video generation have greatly improved visual quality and temporal coherence, but the growing scale of modern models makes inference increasingly expensive, limiting their use in real-time applications. While models and acceleration algorithms are becoming increasingly open, the data used to train high-quality accelerated video models remain comparatively underexplored. We introduce a 450K real-video dataset centered on rich and diverse motion while preserving fine visual detail and strong text alignment, curating from 4.55M clean, standardized candidate clips. Our data-curation pipeline combines standard filtering and agentic VLM-based review to identify and balance dynamic video content. To demonstrate its effectiveness, we further build a complete acceleration pipeline on Wan2.2 A14B, combining supervised fine-tuning and few-step distillation with a generator-matched one-step refiner for high-resolution detail recovery. Without quantization or sparse attention, our system achieves acceleration over the reference generation pipeline and generates a 5-second 480p video in approximately 4.9 seconds on single NVIDIA H100, while maintaining strong generation quality and user preference. We will publicly release the dataset, model checkpoints, and complete training and inference code to support reproducible research on efficient and dynamic video generation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video datasets; Video data curation; Few-step distillation; Real-time video generation" ], "forum_url": "https://openreview.net/forum?id=RTIRFpI696", "pdf_url": "https://openreview.net/pdf/d964665a89a75e20f2a8f5b3ac32ba2d8a6eacd1.pdf", "updated_at": "2026-10-05T18:25:12.441Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ve4LXwNy9W", "number": 11230, "title": "ZPoles: Fast Fractional Matrix Powers for Optimization via Rational Approximation", "abstract": "Fractional matrix powers underpin gradient preconditioning in Shampoo and momentum transforms in Muon, but their repeated computation can be expensive. Rational approximations use a small number of shifted matrix inverses, whose cost has limited their appeal for GPU training. We propose ZPoles, a method for computing fractional powers with a new mixed-precision linear solver. The solver keeps diagonal operations in higher precision and batches half-precision matrix multiplications across the shifted linear systems. On an RTX 5090 GPU, the solver is 3.7–4.1× faster than cuSOLVER, trading numerical accuracy for runtime. For matrices of size 1024×1024 to 12288×12288, ZPoles computes inverse fourth roots 2.6–8.1× faster than coupled Newton iteration at a fixed 10-iteration budget, with lower error when both use half-precision matrix-multiplication operands. In GPT-2 pretraining, ZPoles reduces observed Shampoo step time relative to coupled Newton, with lower or similar validation loss. Muon and fractional-power Muon provide further evidence of applicability to momentum transforms. These results support rational approximation as a practical computational primitive for GPU-based optimization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "fractional matrix power; orthogonalization; Muon; Newton-Schulz; rational approximation;" ], "forum_url": "https://openreview.net/forum?id=Ve4LXwNy9W", "pdf_url": "https://openreview.net/pdf/13366eaabcf7f8095e706239da7ed24264b81fb0.pdf", "updated_at": "2026-10-05T18:25:24.772Z", "categories": [ "training", "hardware" ], "matched_terms": [ "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "L9KtDDbUfr", "number": 11402, "title": "Beyond Full Imitation: Direction-Guided Quantization-Aware On-Policy Distillation", "abstract": "Microscaling floating-point (MXFP) formats have emerged as a promising low-precision representation for efficient large language model (LLM) training and inference, yet quantizing both weights and activations to 4 bits (also known as W4A4 quantization) can still severely degrade performance on complex reasoning tasks. While existing post-training quantization (PTQ) and quantization-aware distillation (QAD) methods struggle to fully recover BF16 performance under aggressive quantization, on-policy distillation (OPD) is naturally suited to this setting because it supervises the quantized student on its own generated trajectories and better matches the states encountered during inference. However, we find that full imitation of the BF16 teacher is unnecessary for effective recovery. Quantization can either weaken or strengthen preferences introduced by post-training, suggesting that not every deviation from the BF16 teacher should be corrected. Based on this insight, we introduce DQ-OPD, a \\textbf{D}irection-guided \\textbf{Q}uantization-aware \\textbf{O}n-\\textbf{P}olicy \\textbf{D}istillation method that selectively corrects preference-weakening deviations while preserving preference-strengthening ones. Across Qwen3-8B and Qwen3-30B-A3B under MXFP W4A4 quantization, DQ-OPD consistently outperforms all PTQ and QAD baselines. On Qwen3-30B-A3B, it narrows the average gap to BF16 to 0.71 points and even surpasses the BF16 teacher on AIME24. These results demonstrate that near-BF16 performance under W4A4 quantization can be achieved without fully imitating the BF16 teacher, highlighting the importance of deciding which quantization deviations to be corrected rather than applying teacher correction uniformly.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-aware distillation", "on-policy distillation", "MXFP4 quantization", "large language models", "low-bit quantization" ], "forum_url": "https://openreview.net/forum?id=L9KtDDbUfr", "pdf_url": "https://openreview.net/pdf/d3871734c26d3876202fb2bb9bc3779682904c8c.pdf", "updated_at": "2026-10-05T18:27:02.283Z", "categories": [ "ptq", "qat", "reasoning" ], "matched_terms": [ "low-bit", "low-precision", "mxfp4", "ptq", "quantization", "quantized", "quantizing", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "at69adBhYz", "number": 11406, "title": "TEACH ME NO REGRET: ACCESSIBILITY AND REALIZABILITY IN REGRET-BASED POST-TRAINING", "abstract": "Regret-minimization fine-tuning (RMFT) trains a decision policy on selected low-regret trajectories. Its success depends both on generating useful trajectories and on reproducing their behavior with the information available at deployment. We develop a scenario-wise analysis that separates these two requirements. For regret distributions with arbitrary atoms and ties, we derive a finite-iteration bound combining initial lower-tail accessibility with a selection-to-deployment discrepancy. A finite-sample result separates trajectory sampling, score error, and policy fitting, while margin conditions translate regret into rollout-averaged action error. We also solve repeated population imitation in a binary model with noisy observations: the policy converges geometrically to the best admissible decision rule, although an information-dependent gap to the evaluator's oracle persists. Experiments with a lightweight language model in stationary, contextual, and semisynthetic recommendation bandits show contrasting responses to direct RMFT and teacher-distilled initialization. The resulting framework distinguishes the quality of selected trajectories from the performance achieved by the deployed policy and provides a basis for interpreting initialization-dependent post-training gains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "regret minimization", "imitation learning", "LLM agents", "partial observability", "distribution shift" ], "forum_url": "https://openreview.net/forum?id=at69adBhYz", "pdf_url": "https://openreview.net/pdf/571059d880d768f315664e1579373f65119850c7.pdf", "updated_at": "2026-10-05T18:27:02.373Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ToMCufw8bq", "number": 11435, "title": "Silent Collapse: Diagnosing and Repairing Low-Precision Robot Policy Training at the Action Level", "abstract": "Robot learning is now trained largely with low precision, bf16 and fp8 alike. Low precision removes information the policy needs during training, and this can be silent: every monitor reads normal, yet the robot no longer produces the right action. We study this failure at the level of the executed action. It is not one fail\u0002ure mode but three, concentrated where the signal is weak along the activation\u0002 gradient chain. The two most dangerous ones are new: mantissa starvation and conservative collapse. Both rewrite the action output without warning, with strik\u0002ing unevenness across tasks and seeds. We build a seconds-long probe locating the weak-signal points where failures start and present CURE (Cheapest sUfficient REpair), a routing protocol assigning each failure mode its cheapest sufficient fix: a free analytic bypass where the weak signal has a closed-form gradient, exact backward fidelity from the UPH (Unified Precision Health) layer where it does not, and LockedAdamW, an optimizer locking the per-step noise-displacement budget. We test the routing across policy families from a 0.8M ACT to 7B vision-\u0002language-action models under the twin-run protocol. On the ACT collapse cell, the routed scheme brings the collapsed posterior’s KL back into the fp32 band; on the 7B CogACT, LockedAdamW breaks below the fp32 self-retraining floor on both seeds; on real training hardware, one tested Transformer Engine fp8 con\u0002figuration trains normally—consistent with the same mechanism—while unscaled variants fail in simulation. The probe answers in 3.4 seconds at 0.8M and in min\u0002utes at 7B; the repair starts free.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-precision training", "Robot learning", "Numerical stability", "Training reliability" ], "forum_url": "https://openreview.net/forum?id=ToMCufw8bq", "pdf_url": "https://openreview.net/pdf/d0e1784c75848e7d66d1c2e112c49313dd28ca3e.pdf", "updated_at": "2026-10-05T18:27:18.722Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp8", "low precision", "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xbMYkksdsK", "number": 11533, "title": "CostBit: Candidate-Cost-Informed Bit Allocation for Post-Training Quantization of LLMs", "abstract": "Post-training quantization (PTQ) is a practical route to reducing the memory and bandwidth requirements of large language models (LLMs). However, uniform bit-width quantization becomes increasingly inefficient in the aggressive sub-3-bit regime due to heterogeneous quantization sensitivity across weight blocks. Although existing mixed-precision schemes alleviate this issue, their allocation rules fail to account for the quantization error variation of each weight block across candidate bit‑widths, resulting in suboptimal accuracy–compression trade-offs. To address this limitation, we introduce CostBit, a candidate-Cost-informed blockwise Bit allocation method for PTQ of LLMs. Specifically, CostBit evaluates the Hessian-weighted cost associated with each candidate bit-width for every weight block and leverages a Lagrangian allocator to closely match the target bit budget, supporting non‑integer values. Extensive experiments on OPT, LLaMA-2, and LLaMA-3 spanning 125M to 70B parameters demonstrate that CostBit yields improved performance at comparable target bit budgets, with significant gains in the sub-3-bit regime. Further, at an average bit-width of approximately 2.5 bits, it reduces WikiText-2 perplexity from 7.69 to 6.43 and improves the average zero-shot accuracy from 60.04\\% to 63.79\\% compared with OWQ on LLaMA-2-7B. CostBit also enables fully GPU-resident inference of LLaMA-2-70B on a single RTX 4090 (24 GB) at approximately 2.5 bits, achieving 75.70\\% average zero-shot accuracy, only 1.37\\% below the FP16 baseline. Our code is available at https://anonymous.4open.science/r/CostBit.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Mixed-Precision Quantization", "Bit Allocation", "Large Language Model Compression" ], "forum_url": "https://openreview.net/forum?id=xbMYkksdsK", "pdf_url": "https://openreview.net/pdf/07cfc61c51085f68a776355627c2af74b1e6745c.pdf", "updated_at": "2026-10-05T18:28:22.651Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "3-bit", "bit-width", "mixed-precision", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yo3dumKWKk", "number": 11550, "title": "What \"Form\" of Uncertainty Does Conformal Prediction Quantify?", "abstract": "Conformal prediction (CP) is a post-hoc calibration method that returns prediction sets with distribution-free coverage guarantees, and the size of the output set is often interpreted as a notion of uncertainty. We show that the difference in set size between full and split CP is directly connected to Bayesian posterior variance, which is the epistemic uncertainty of the model. Essentially, this difference in set size is an aggregation of two opposite effects: (1) better model accuracy and therefore smaller set size achieved by training over extra calibration points, and (2) the \"formability\" of the model at test point towards every possible label. The latter effect vanishes with stronger models or more calibration data (as epistemic uncertainty is known for that behavior). It turns out that the model’s formability itself (without the quantization effect of prediction sets, and through estimating the posterior variance) is a useful tool to quantify the epistemic uncertainty. In our empirical evaluation it performs on par with the state of the art in many setups.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Uncertainty Quantification", "Conformal Prediction", "Epistemic Uncertainty" ], "forum_url": "https://openreview.net/forum?id=yo3dumKWKk", "pdf_url": "https://openreview.net/pdf/a8c71b86a34d5a0146c518f2974e230e0f6b9a95.pdf", "updated_at": "2026-10-05T18:28:29.475Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JlxR9NlrsL", "number": 11631, "title": "VFold: Symmetry-Aware Cross-Layer Value Cache Compression", "abstract": "While caching key-value (KV) states accelerates Large Language Model (LLM) decoding, this cache can dominate memory usage at long context lengths. One solution is to compress this memory by exploiting inter-layer cache similarities. However, existing techniques necessitate architectural changes to the transformer and incur substantial overhead. In this work, we propose a symmetry-aware value cache merging strategy that reduces cache memory while avoiding both harmful performance degradation and architectural overhead during decoding. Furthermore, we show that this redundancy can be exploited alongside existing cache compression techniques, seamlessly composing with high-ratio quantization and key cache pruning methods. Ultimately, our findings reveal a major source of underutilized capacity in the value cache, opening a simple yet highly effective avenue for scaling context windows under strict memory constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache", "folding", "merging", "transformer", "compression" ], "forum_url": "https://openreview.net/forum?id=JlxR9NlrsL", "pdf_url": "https://openreview.net/pdf/3c8393170a0b6f915ac32fb475038f1ef2dc46ab.pdf", "updated_at": "2026-10-05T18:29:32.947Z", "categories": [ "kv-cache" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "oUnHRKuYEZ", "number": 11672, "title": "Few Bits, One Law: Toward W2A4KV2", "abstract": "Extreme low-bit LLM compression is most challenging when weights, activations, and KV caches are quantized together: their distributions differ, and quantization errors interact throughout the network. We introduce CanonQ, a unified quantization-aware training framework that addresses these challenges by separating source canonicalization from task-aware adaptation. Fixed rotations and energy normalization map heterogeneous tensor sources to canonical coordinates, enabling frozen Gaussian-reference codebooks to be reused across layers and models. Joint training then adapts the network to the coupled errors of weight, activation, and cache quantization within a common scalar/vector interface. We bound frozen-codebook transfer error and local task loss, and derive an exact normalization-aware straight-through Jacobian that links quantization distortion to gradient bias. The strongest gains arise under joint W2A4KV2 compression: across LLaMA3-1B/3B/8B, CanonQ-Omni achieves up to 14.28× lower WikiText-2 perplexity and up to 57.9% higher mean zero-shot accuracy than prior state-of-the-art and representative quantization baselines. The benefits extend to Qwen3-1.7B, code generation, and mathematical reasoning: on instruction-tuned MobileLLM-Pro-1B at W2A16KV16, CanonQ achieves relative improvements of 41.7% in HumanEval pass@1 and 39.1% in GSM8K exact match over the strongest evaluated quantization baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Model Quantization", "Low-Bit Quantization", "Codebook Quantization", "Quantization-Aware Training" ], "forum_url": "https://openreview.net/forum?id=oUnHRKuYEZ", "pdf_url": "https://openreview.net/pdf/da4cdf0bb2474b641cfe984bc59f3f4b4eec90a1.pdf", "updated_at": "2026-10-05T18:29:58.889Z", "categories": [ "qat", "kv-cache", "reasoning", "theory" ], "matched_terms": [ "low-bit", "quantization", "quantized", "w2a16kv16", "w2a4kv2" ], "review_status": "abstract", "summary": "CanonQ 将异质张量的分布规范化与任务适配分开：固定旋转和能量归一化把权重、激活和 KV 映射到可共用高斯参考码本的坐标。联合 QAT 随后适配三类误差的耦合，同时推导 normalization-aware STE Jacobian 和码本迁移误差界。", "contribution": "在 LLaMA3-1B/3B/8B 的 W2A4KV2 比较中，摘要报告相对所比较量化基线最高 14.28 倍的 WikiText-2 困惑度改善和 57.9% 的平均零样本准确率相对提升；MobileLLM-Pro-1B 的 W2A16KV16 下 HumanEval、GSM8K 相对提升为 41.7%、39.1%。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。准确率增益是相对提升，不能写成百分点；W2A4KV2 与 W2A16KV16 的结果来自不同模型设置，摘要未提供真实部署效率。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "LLM 权重、激活与 KV 缓存的联合量化", "bit_width": "重点 W2A4KV2；另报告 W2A16KV16", "evidence": [ { "location": "官方摘要", "claim": "W2A4KV2 的 LLaMA3 比较报告最高 14.28 倍困惑度改善、57.9% 平均零样本准确率相对提升。" }, { "location": "官方摘要", "claim": "MobileLLM-Pro-1B 在 W2A16KV16 下 HumanEval/GSM8K 相对提升 41.7%/39.1%。" } ] }, { "id": "HQAdNxxrcj", "number": 11674, "title": "JacQuant: STE‑Free Quantization‑Aware Training via Learned Jacobian Surrogates", "abstract": "Quantization-aware training (QAT) usually propagates gradients through hard quantizers with a fixed straight-through estimator (STE). At ultra-low precision, this backward rule ignores how quantized weights respond to finite perturbations. We introduce JacQuant, a QAT framework that learns a lightweight surrogate of quantizer sensitivity while retaining the hard forward quantizer. A periodic Gaussian probe fits one scalar gain per weight group; the gains are averaged over time and reused between refreshes. We identify the population regression target through Gaussian smoothing and separate it from the clipped estimator used in practice. For a local linearized surrogate, we establish non-convex stationarity bounds and linear convergence to an error neighborhood under a Polyak–Łojasiewicz condition, with explicit bias and variance terms. Across 1B–8B LLaMA and Qwen models with 1–2-bit weights, JacQuant improves both perplexity and mean zero-shot accuracy over the ParetoQ and WinQ baselines. On LLaMA3-1B W2A16, it lowers ParetoQ perplexity by 8.9% and recovers roughly one third of the gap to FP16 in both metrics. Gains persist at all three 8B checkpoints through 90K steps. In its default configuration, JacQuant incurs only a 1.6–3.5% reduction in training throughput on 1B models and 0.76% on 8B relative to the corresponding STE baselines. It requires no additional computation at inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Aware Training; Large Language Models; Ultra-Low-Bit Quantization; Straight-Through Estimator; Learned Jacobian Surrogates; Gradient Estimation; STE-Free Quantization; Model Compression" ], "forum_url": "https://openreview.net/forum?id=HQAdNxxrcj", "pdf_url": "https://openreview.net/pdf/13f9f6b6e7208090b6c6c850616ab92a76b11f2c.pdf", "updated_at": "2026-10-05T18:30:01.026Z", "categories": [ "qat", "training", "hardware", "theory" ], "matched_terms": [ "2-bit", "low precision", "low-bit", "qat", "quantization", "quantized", "quantizer", "quantizers", "w2a16" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9lEvG6Tfku", "number": 11736, "title": "JARQ: Joint Alternating Refinement for Quantization", "abstract": "Group-wise post-training quantizers for large language models round weights onto a grid that is not refit to the resulting integer codes. We show that this leaves accuracy on the table: the best grid depends on the codes, input correlations couple the errors of different groups, and useful code changes often involve many codes at once. We propose JARQ (Joint Alternating Refinement for Quantization), a plug-in refinement that starts from any group-wise quantizer and alternates a joint least-squares fit of all group scales with bounded Babai proposals that move many codes of a group together on the current grid. The problem is a bilinear box-constrained mixed-integer least-squares problem; the solver is backpropagation-free, does not increase the layer-wise objective under exact scale solves, and keeps the host's bit width, groups, zero points, and inference cost. Across Llama-2, Llama-3, and Qwen models with RTN, GPTQ, OmniQuant, and AWQ hosts, JARQ lowers perplexity in 90 of 96 comparisons, cuts three-bit RTN perplexity by up to 36%, raises mean multiple-choice accuracy in 23 of 24 configurations, and improves QEP, QuaRot, and OJBKQ outputs, at under a minute per 7B block.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Optimization", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=9lEvG6Tfku", "pdf_url": "https://openreview.net/pdf/3323101a0aaf566e69a09bd220a524d13e2f6449.pdf", "updated_at": "2026-10-05T18:30:42.298Z", "categories": [ "ptq" ], "matched_terms": [ "bit width", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5n4xeaG7bM", "number": 11798, "title": "CurveTQ: Rotation-Free Trellis Quantization of LLM Weights via Curvature-Weighted Search", "abstract": "The best two-bit weight quantizers for large language models, such as QTIP and Proteus, share a recipe. Each weight matrix is first rotated by a random orthogonal transform so that its entries look Gaussian, and this rotation has to be undone at every decoding step. The rotated weights are then encoded with a trellis or lattice code under a Euclidean search, and the layer Hessian is used only in the error-feedback step between coding blocks. We show that this recipe leaves part of the Hessian unused. Error feedback turns the loss into a weighted sum of per-coordinate rounding errors, and the weights, the diagonal of the Hessian's LDL factorization, are computed by existing quantizers but never read. We put these weights into the search's distance, the Viterbi branch metric, so that the search follows the curvature within each coding block. This also explains what the rotation is for. The rotation removes this within-block variation, so after it the weight has little left to exploit. Weighting in the native basis and rotating are therefore substitutes. We confirm this on three models: the weighted native search matches the baselines' rotation, a full-dimension randomized Hadamard, to within about one point of downstream accuracy, and weighting after the rotation gains little. Around this search we build CurveTQ, a trellis codec with no rotation. The rotation also flattened the amplitude and the marginal shape of the weights, which we handle directly with a factored scale field and a closed-form quantile table. In either basis, error feedback passes each block's residual on to the next, so we store a start state per coding block that lets the trellis adapt to it. At two bits CurveTQ is 1-3 points higher in mean downstream accuracy than QTIP and Proteus on three 4-8B Instruct models, even after both are given our start state, which by itself lifts either baseline by 1-3 points. It also leads on a 35B mixture of experts, to our knowledge the first trellis-coded result on such a model. With no rotation to undo, our decoder is the fastest of the three at all tested batch sizes and bit widths.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "weight-only quantization", "trellis-coded quantization", "low-bit LLM", "Hessian-aware quantization", "model compression", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=5n4xeaG7bM", "pdf_url": "https://openreview.net/pdf/9c1c34a92cb7fcec350a20d6d36c1bf9dea42a4b.pdf", "updated_at": "2026-10-05T18:31:10.356Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ktEn3Rmz2G", "number": 11810, "title": "KVFetch: Temporal Prefetching for the Missing Half of KV Cache Compression", "abstract": "As context windows scale to tens or hundreds of thousands of tokens, KV cache compression has become essential for efficient LLM inference. Existing methods fall into three families: score-based eviction, summary compensation, and offload-and-recall. Yet all three decide what to keep or recall by content relevance to the current query. We show this shared design is structurally incomplete. A cache supports two access modes: associative lookup by content and sequential traversal by position; current compressors implement only the first. The gap matters in practice: retrieval-augmented generation, code completion, and structured-data extraction all require the model to reproduce identifiers, field values, or code tokens verbatim from the context. Under compression, content-based eviction retains the head of such a sequence but discards its continuation, causing verbatim copying to break irreversibly midway, a failure we call sequential forgetting. This failure resists better scoring, larger budgets, summary compensation, and dynamic re-scoring; it is the dominant source of remaining quality loss under compression. We propose KVFetch, a training-free, drop-in framework that opens a temporal recall channel for any score-based compressor. It demotes evicted candidates to a quantized cold tier, detects active copying through a monotone read pointer, and prefetches positional successors into fixed-size hot-tier slots without increasing attention cost. On RULER-16K under an iso-budget control, KVFetch recovers verbatim copying from 0.8 to 78.4 and raises the 13-task average by +8.4, with gains concentrating on tasks that require sequential access. On LongBench, where no task requires sequential access, the channel remains dormant and imposes no cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "long-context inference", "temporal addressing", "sequential forgetting", "verbatim copying", "training-free" ], "forum_url": "https://openreview.net/forum?id=ktEn3Rmz2G", "pdf_url": "https://openreview.net/pdf/53c5158af11ae1693dfc2e7d3444b63c53ecac70.pdf", "updated_at": "2026-10-05T18:31:17.103Z", "categories": [ "kv-cache" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8UJzSYuF9s", "number": 11912, "title": "Prox: Training-Free FFN Activation Sparsification in LLMs via Approximate Intermediate-State Salience", "abstract": "Feed-forward networks (FFNs) dominate memory traffic and computation in large language model (LLM) inference, making them a primary target for activation sparsification.\nHowever, existing training-free methods suffer substantial model-quality degradation at high FFN sparsity due to limitations in their channel-selection strategies.\nWe observe that the SwiGLU intermediate state provides a highly effective channel-selection signal, but obtaining it requires costly dense computation.\nTo address this, we present \\emph{Prox}, a training-free framework for sparse SwiGLU FFNs based on the insight that sparse execution needs only the channel mask induced by the intermediate state, which can be constructed from the magnitude ranking of its entries (i.e., channel salience) rather than their precise values.\nProx works in two stages: Stage1 approximates the ranking via input sparsity and quantized proxy weights to construct the channel mask; Stage2 computes the selected channels exactly using full-precision weights, enabling sparse execution of all three projections.\nExtensive evaluations across ten LLMs from six model families show that Prox consistently outperforms training-free baselines at various sparsity levels, achieving up to a $2.07\\times$ end-to-end decoding speedup at 70\\% FFN sparsity, and is compatible with quantization and sparse attention.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Efficient Inference", "Activation Sparsity", "Feed-Forward Networks", "Training-Free Sparsification" ], "forum_url": "https://openreview.net/forum?id=8UJzSYuF9s", "pdf_url": "https://openreview.net/pdf/6b1154c2b13a079aa64403a6a662c66632fee203.pdf", "updated_at": "2026-10-05T18:32:21.039Z", "categories": [], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bxV8FvKGca", "number": 11931, "title": "Scaling Influence Functions in LLMs through Eigenbasis-Corrected One-Bit Gradient Projection", "abstract": "Influence functions estimate how individual training examples affect the behavior of large language models (LLMs). Analyzing how training data influence different behaviors of an LLM involves repeated influence computation. Reusing stored training gradients reduces the computational cost, but storing full gradients is prohibitively expensive at LLM scale. We study how to compress these gradients while preserving influence estimates for future queries that are unknown at storage time. Through a worst-case analysis, we characterize the optimal fixed-dimensional linear representation and propose eigenbasis-corrected one-bit gradient projection (EOGP) to approximate it at scale. Specifically, EOGP uses EK-FAC to reduce gradient dimensionality, then applies PCA within the retained subspace to learn compression directions from the training gradients. We then apply one-bit quantization to the resulting coordinates, allowing more coordinates to be retained within a fixed storage budget. On GPT-2, EOGP predicts retraining outcomes more accurately than the evaluated compression baselines while using one-sixteenth of their per-example storage. On OLMo 2 SFT models from 1B to 32B parameters, EOGP remains competitive with the baselines allocated over 100 times as much storage per example. Code is available at https://anonymous.4open.science/r/EOGP_anonymous-70FA.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "influence functions", "training data attribution", "large language models" ], "forum_url": "https://openreview.net/forum?id=bxV8FvKGca", "pdf_url": "https://openreview.net/pdf/a5f81d88a98005192a56ed47f945dd69df42b5e9.pdf", "updated_at": "2026-10-05T18:32:31.167Z", "categories": [ "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IYnHo278PI", "number": 11977, "title": "Neural Dynamics as the Composition of Quantized Units", "abstract": "Deep learning is commonly interpreted at two levels: the macroscopic, through aggregate trends in loss summarized by scaling laws, and the microscopic, through neurons, features, and circuits. A central challenge is understanding how these levels connect, so that we can explain how elementary computations compose and collectively shape macroscopic behavior. To this end, we study an intermediate abstraction in which training is described as the ordered acquisition of \\emph{quanta}: reusable computations acquired suddenly and binary-activated across examples to reduce loss. By approximating population-gradient updates, we derive quanta's acquisition dynamics. This yields an acquisition priority governed by \\emph{demand}, how frequently a computation is required across examples, and \\emph{conditional complexity}, how difficult that computation is to acquire given those already available. In a Boolean compositional task, we derive predictions for acquisition order and show how staggered discrete acquisitions can produce smooth aggregate loss and, under certain geometries of quanta composition, give rise to scaling laws. We then train a Transformer to map numerals to English number names and recover candidate quanta from its checkpoint trajectory. From these units, we construct a model that preserves much of the Transformer's behavior while exposing interpretable latent computations and acquisition dynamics consistent with the theory. Separately, the quanta structure can serve as training targets to improve transformer generalization. Together, these results suggest the quanta abstraction can provide useful computational atoms for studying a variety of macroscopic phenomena.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "learning dynamics", "compositional learning", "neural scaling laws", "mechanistic interpretability", "science of deep learning" ], "forum_url": "https://openreview.net/forum?id=IYnHo278PI", "pdf_url": "https://openreview.net/pdf/cea7baa6e312bc66347fb8dba0837e591e2e222d.pdf", "updated_at": "2026-10-05T18:32:54.791Z", "categories": [ "training", "theory" ], "matched_terms": [ "binary", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "onh0OoRi1h", "number": 12021, "title": "TernRefine: Gradient-Guided Fixed-Capacity Refinement of Ternary LLMs", "abstract": "Post-training ternarization produces a compact discrete model, but the resulting\nternary assignment need not be optimal for the target task. We introduce TernRefine,\na refinement method that reallocates side-state occupancy after quantization while\npreserving the ternary codebook and the exact number of side states in every\nquantization group. TernRefine computes a single task gradient at the deployed\nquantized model, uses the first-order score −⟨G, ∆Q⟩ to rank legal relocations,\nand selects how many relocations to apply using a separate validation split. We\nfind that the point at which the gradient is evaluated matters substantially. In a\ncontrolled study over five fitting offsets, gradients computed at the quantized model\nand at the corresponding full-precision model produce patches with less than 2%\nJaccard overlap and lead to opposite average outcomes on held-out data. Ranking\nwith the quantized model gradient improves WikiText-2 in all five runs, whereas\nranking with the full-precision gradient degrades it on average. TernRefine also\nimproves WikiText-2 and C4 on OPT-350M, Llama-2-7B, and Qwen3-8B after\naffine ternarization. Starting from PT2 models, all three fitted patches improve both\ndatasets on Llama-2-7B and Qwen3-8B, and the selected models show positive\naverage transfer on standard zero-shot tasks. Overall, these results show that\nternary models can retain useful task structure after quantization and that gradients\nevaluated at the deployed discrete state provide an effective signal for refining that\nstructure.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "model quantization", "ternary quantization", "post-training quantization", "model compression", "discrete optimization", "task-aware refinement" ], "forum_url": "https://openreview.net/forum?id=onh0OoRi1h", "pdf_url": "https://openreview.net/pdf/572897874f2023b485874fcd13cf9d7091b152ab.pdf", "updated_at": "2026-10-05T18:33:14.264Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization", "quantized", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TA78JcVQac", "number": 12048, "title": "I-DiT: Integer-Only Quantization for Diffusion Transformers", "abstract": "Diffusion transformers achieve remarkable generative performance but remain challenging to deploy on resource-constrained hardware due to their high computational and memory costs.\nThese devices often rely on specialized AI accelerators whose efficiency comes primarily from low-precision integer-only datapaths. \nHowever, existing diffusion quantization methods still rely heavily on floating-point operations,\nand integer-only frameworks are not readily applicable to modern diffusion transformers due to fixed quantization scales, lack of tiled attention support, and sensitivity to activation outliers.\nFor this, we present I-DiT, the first framework for fully integer-only inference of diffusion transformers\nwithout costly fine-tuning.\nFirst, we propose a dynamic scale adjustment method for integer-only frameworks, that adapts to varying activation ranges during runtime.\nSecond, we introduce log2-based online softmax that enables numerically stable and memory-efficient tiled attention in integer.\nLastly, we reformulate dyadic requantization to support finer-grained activation quantization, localizing the impact of outliers while preserving integer-only execution. \nExperiments across PixArt, SD3, and FLUX demonstrate that I-DiT preserves generation quality while enabling efficient fully integer-only inference for large-scale diffusion models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Transformers", "Integer-Only Quantization", "Edge AI", "Post-Training Quantization", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=TA78JcVQac", "pdf_url": "https://openreview.net/pdf/6b09da5c4e9db7a3bf5d8074a0603969940a6eb0.pdf", "updated_at": "2026-10-05T18:33:53.551Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SB0L6kLa11", "number": 12066, "title": "Same Recipe, Different Weights: Quantization and Decision Stability Across Backends", "abstract": "Can one weight-only recipe be reapplied across tensor backends without changing its checkpoint, decisions, or cost? We evaluate symmetric INT8 and grouped INT4 on ten pretrained checkpoints spanning text, vision, audio, dual-encoder, and generative vision-language architectures in one multi-backend implementation on an A100. All 30 cross-backend FP32 state pairs match exactly; none of the 60 independently quantized pairs does. Exact PyTorch packed-state transplantation succeeds for eight ViT/CLIP target interventions and preserves all 800 held-out image decisions per intervention, yet a prespecified 99% calibration-agreement selector misses its holdout target on ViT and CLIP. Across three language-model families, all FP32 and INT8 16-token continuations agree across backends on eight fixed prompts, whereas minimum pairwise INT4 agreement is 5/8 on Qwen3 and Gemma3. Image-conditioned Qwen3-VL continuations agree across all backends for FP32 and INT8, but the worst INT4 pair agrees on only 15/20 images. Median stored-variable bytes fall to 0.254 and 0.184 of FP32 under INT8 and INT4, but portable direct-eager INT4 execution is slower across all three backends. These findings separate recipe reproducibility, exact state transfer, observed decision stability, and deployment cost without treating a changed tensor hash as a changed prediction.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "weight-only quantization", "cross-backend reproducibility", "numerical equivalence", "decision stability", "Keras 3", "PyTorch", "JAX", "TensorFlow", "inference benchmarking" ], "forum_url": "https://openreview.net/forum?id=SB0L6kLa11", "pdf_url": "https://openreview.net/pdf/15064696502925791baf75f49256a96d53746862.pdf", "updated_at": "2026-10-05T18:33:52.523Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "int4", "int8", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MLakehNzt0", "number": 12072, "title": "Role-Aware Quantization for Reliable Transformer Inference", "abstract": "Quantization is widely used to reduce Transformer memory usage and accelerate inference. Prior work has commonly used uniform activation quantization, assigning the same bit-width across activations. However, Transformer components perform distinct functions, including query--key routing, value transport, feed-forward computation, and normalization. We show that these functional roles differ in sensitivity to quantization error and therefore require different precision levels to preserve predictions. To allocate precision efficiently under a fixed budget, we introduce Role-Aware Reliability (RAR) allocation. RAR quantizes one layer--role activation site at a time and measures the resulting change in the next-token distribution using Kullback--Leibler (KL) divergence. The resulting functional risk atlas pinpoints where low precision most strongly distorts predictions. Building the functional risk atlas for a 7B checkpoint requires only a small sample of calibration text and approximately one to two minutes. Our allocation rule, RAR-log, uses the atlas to assign higher precision to high-risk sites and lower precision to low-risk sites under a fixed cost-weighted average bit-width budget. By accounting for how quantization errors in different roles affect predictions, RAR complements SmoothQuant-style activation smoothing. Our theoretical analysis explains how quantization errors from different sites combine and proves bounds on the resulting changes in model predictions. With the same target precision budget, the combined policy, Smooth-RAR-log, removes an average of 94.1\\% of the language-modeling degradation caused by uniform 6-bit activations, compared with 87.4\\% for smoothing alone. The combined policy improves on smoothing alone in all 24 paired model--corpus evaluations, covering three corpora and eight Transformer checkpoints ranging from 124M to 7B parameters. Even with packed activations and 8-bit quantized weights, Smooth-RAR-log lowers next-token prediction loss compared with smoothing alone across the evaluated models. Smooth-RAR-log also improves average accuracy on multiple-choice benchmarks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Model Compression", "Post-Training Quantization", "Activation Quantization", "Mixed-Precision Quantization", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=MLakehNzt0", "pdf_url": "https://openreview.net/pdf/bf2a7d4464cdc0cd8423fdb4180f1bcf0acfd2fa.pdf", "updated_at": "2026-10-05T18:33:56.736Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "6-bit", "8-bit", "bit-width", "low precision", "mixed-precision", "quantization", "quantized", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mQ9mqwGRJ7", "number": 12075, "title": "The Precision Cliff: An Information-Theoretic Lower Bound on Quantization for Iterative Compositional Retrieval", "abstract": "Low-bit quantization degrades multi-hop reasoning faster than single-hop recall, yet no theory has said why deeper reasoning needs more bits or how many more. We answer both with a single closed-form information-theoretic floor. For any post-training compression scheme whose error placement is independent of the stored content and whose distortion is deterministic—conditions met by GPTQ, AQLM, AWQ, and QuIP#—the minimum bits per entry for reliable $T$-hop retrieval is $b^*_{\\mathrm{any}}(T, N) = R(\\delta_T)$ with $\\delta_T = 1 - 2^{-1/T}$, and an explicit construction saturates this floor to within $0.002$ bits. A direct corollary is the $2\\times$ separation law: deep reasoning demands twice the bit-rate of single-hop recall at capacity saturation (predicted $2.40$ for $N{=}4096$; measured $2.397$). Capacity utilization $\\eta$ transfers this bits-per-entry floor to bits-per-parameter: a synthetic sweep traces a monotone depth penalty in $\\eta$ (Spearman $\\rho = -0.93$, $p<0.001$), and four LLM families (Llama-2-7B, Qwen3-8B, Phi-4-mini, InternLM3-8B) sit on the same $\\eta$-curve at production scale with wall premiums of $1.05$–$1.55\\times$—already forcing a $5$–$55\\%$ upward correction to any $T{=}5$ quantization floor calibrated at $T{=}1$. A chain-given control (accuracy $0.99$ read-aloud vs. $0.46$ sequential lookup at 3 bits) isolates iterative retrieval as the causal mechanism, ruling out attention-routing and generation-noise alternatives.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "multi-hop reasoning", "information theory", "rate-distortion", "precision cliff", "large language models", "compositional retrieval", "error amplification" ], "forum_url": "https://openreview.net/forum?id=mQ9mqwGRJ7", "pdf_url": "https://openreview.net/pdf/b88cadd0a573e52a29f93857d0bd4a9a54f9aab7.pdf", "updated_at": "2026-10-05T18:33:57.236Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "apiSDBM2nN", "number": 12240, "title": "C-PTQ: Unified Channel-wise Sensitivity and Scaling for Post-training Quantization of MLLMs", "abstract": "Multimodal large language models (MLLMs) require huge memory and computational costs, which limit their practical deployment. Post-training quantization (PTQ) techniques offer an efficient solution for model compression and inference acceleration. Yet, the quantized model faces performance degradation due to outlier channels, which are highly sensitive to quantization and substantially impair activation fidelity and task accuracy.\nTo protect these salient channels during quantization, existing PTQ methods leverage modality- or token-level metrics to guide channel-wise scaling (CWS) of LLM decoders. \nHowever, these orthogonal measurements fail to capture channel-wise impacts on task-specific loss, and the misalignment between importance and scaling factors ultimately leads to suboptimal performance.\nTo address this issue, we propose C-PTQ, a unified channel-wise PTQ method that harmonizes task-specific loss perturbation and quantization error. Motivated by second-order derivatives, we design a Fisher-weighted objective as a tractable Hessian approximation, seamlessly injecting task sensitivity into the scaling process. Notably, we achieve state-of-the-art performance without auxiliary modules like LoRA, thereby maintaining high efficiency.\nExperiments on Qwen2.5VL, InternVL2 and LLaVA-OV across 8 benchmarks demonstrate our effectiveness in both weight-only and weight-activation settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training Quantization", "Multimodal Large Language Models" ], "forum_url": "https://openreview.net/forum?id=apiSDBM2nN", "pdf_url": "https://openreview.net/pdf/47af72cd4f8037dc8fd32859f03c75b2b5b44ef2.pdf", "updated_at": "2026-10-05T18:35:49.846Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pyfEe77RFm", "number": 12327, "title": "Beyond Oracle Gains: Measuring the Learnability of Adaptive RAG Routing", "abstract": "Retrieval-augmented generation (RAG) exposes multiple inference-time compu-\ntational choices, including retrieval depth, generator size, and numerical preci-\nsion. Oracle analyses suggest that selecting these configurations separately for\neach query can yield substantially better quality–cost trade-offs than using a sin-\ngle fixed configuration. However, an oracle observes the realized outcome of every\ncandidate configuration and therefore does not establish whether the correspond-\ning routing decisions are predictable before execution.\nWe study the distinction between oracle opportunity and routing learnability in\nadaptive RAG. Our analysis considers retrieval depth, generator capacity, and in-\nference precision across four multi-hop question-answering datasets. Within a\ntargeted four-configuration space, a retrospective oracle achieves 0.423 F1 at an\naverage estimated cost of 36.57, while individual fixed configurations obtain only\n0.203–0.258 F1. A learned router provides useful low-cost operating points but\nreaches only 0.249 F1 even at an average cost of 57.59.\nTo determine whether this gap is primarily caused by insufficient supervision, we\nconstruct 2,000 additional train-only queries. Increasing routing supervision from\n1,421 to 3,421 queries improves retrieval-depth prediction by 0.072 ROC-AUC\nand model-size prediction by 0.042, but decreases precision-routing ROC-AUC\nby 0.033. These effects also vary substantially across datasets.\nOur results show that large oracle headroom does not necessarily translate into an\nequally learnable routing policy. We argue that adaptive RAG should therefore be\nevaluated in terms of both available oracle opportunity and the fraction of that op-\nportunity that can be predicted and recovered before paying the cost of candidate\nconfigurations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "retrieval-augmented generation", "adaptive RAG", "model routing", "adaptive computation", "inference efficiency", "retrieval depth", "quantization" ], "forum_url": "https://openreview.net/forum?id=pyfEe77RFm", "pdf_url": "https://openreview.net/pdf/212849dfb1a111d52ae0c21097ec1299d175ee32.pdf", "updated_at": "2026-10-05T18:36:37.797Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8ikHPPoJc0", "number": 12340, "title": "FIT to Forget: Robust Continual Unlearning for Large Language Models", "abstract": "While large language models (LLMs) exhibit remarkable capabilities, they increasingly face demands to remove the influence of privacy-sensitive, copyrighted, or harmful content.\nExisting unlearning methods primarily focus on \\emph{single-shot} scenarios, whereas real-world deletion requests arrive \\emph{continually}.\nNa\\\"ively applying these methods to sequential requests leads to severe utility degradation and catastrophic forgetting.\nTo address this, we propose \\fit, a robust framework for high-volume continual unlearning that resists both catastrophic forgetting and post-unlearning recovery.\n\\fit stabilizes sequential updates through three synergistic mechanisms: redundancy \\underline{F}iltering, \\underline{I}mportance-aware adaptive algorithm selection, and \\underline{T}argeted layer attribution.\nFurthermore, to facilitate rigorous evaluation, we introduce \\textbf{PCH}, a unified benchmark spanning \\textbf{P}ersonal information, \\textbf{C}opyright, and \\textbf{H}armful content, alongside two symmetric metrics, Forget Degree (F.D.) and Retain Utility (R.U.), that quantify forgetting-utility trade-offs.\nExperiments across five LLMs (up to 14B parameters) show that \\fit maintains a strong forgetting--utility balance throughout long request streams.\nEven after hundreds of randomly ordered requests, \\fit preserves competitive downstream (\\eg, GSM8K and MMLU) performance and resists relearning and quantization recovery attacks.\\footnote{Our code is available at \\url{https://anonymous.4open.science/r/FIT_Code-7304/}.}", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language model", "Continual unlearning", "Machine unlearning" ], "forum_url": "https://openreview.net/forum?id=8ikHPPoJc0", "pdf_url": "https://openreview.net/pdf/0a6edf1976d71283ea7fd42d3c2dfb928a350a39.pdf", "updated_at": "2026-10-05T18:36:43.680Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4Z2UegYSL2", "number": 12360, "title": "$R^2$-Omni: Reducing Structural and Precision Redundancy in Omni-LLMs for Long-Form Generation", "abstract": "Omni-modal large language models (Omni-LLMs) typically process large numbers of multimodal tokens from audio-visual inputs, incurring substantial computational costs and a large memory footprint for the key--value (KV) cache.\nAggressive token pruning can alleviate this problem while maintaining competitive performance on multiple-choice tasks, which require only a single option as output. However, we observe substantial quality degradation when generating long-form, open-ended responses.\nWe attribute this gap to query-guided selection, which can retain sufficient evidence for answer selection while discarding details needed for comprehensive generation. This motivates reducing overlapping information across tokens, termed structural redundancy, while preserving diverse content. However, the retained representations may still contain precision redundancy due to unnecessarily high numerical precision in their cached values.\nTo address these two levels of redundancy in token representations, we introduce a training-free framework named \\textbf{R$^2$-Omni}.\nFirst, Structure-Aware Token Consolidation (STC) consolidates redundant representations using spatiotemporal similarity in video and local continuity in audio. Then, Semantic-Sensitivity Guided Precision Allocation (SSPA) allocates value precision to the retained representations according to the estimated reduction in attention-output error under quantized keys, with query attention and audio-visual alignment providing a semantic prior. \nExtensive experiments on long-form generation benchmarks demonstrate an average \\textbf{5.0$\\times$} reduction in the memory footprint while preserving \\textbf{97.98\\%} of baseline generation quality and achieving a \\textbf{1.48$\\times$} end-to-end speedup, highlighting the benefits of jointly reducing structural and precision redundancy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Omni-modal Large Language Models", "Token pruning", "KV Cache Compression" ], "forum_url": "https://openreview.net/forum?id=4Z2UegYSL2", "pdf_url": "https://openreview.net/pdf/102914d01ba8a841e69faa3042727c4c3f1a0400.pdf", "updated_at": "2026-10-05T18:36:58.329Z", "categories": [ "kv-cache", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Kn2mPpj3ZO", "number": 12470, "title": "From Bits to Beliefs: Recovering Owner Identity for Black-Box Verification of Large Language Models", "abstract": "Open-weight large language models (LLMs) can be copied, modified, and redeployed behind black-box APIs, making post-release ownership verification difficult. Existing black-box fingerprinting methods largely treat verification as a detection problem, asking whether registered query-response behaviors remain observable in a suspect model. Such evidence can degrade under fine-tuning, pruning, quantization, model merging, and serving-time prompt changes. We propose SimPrint, a recoverable semantic fingerprinting framework that instead formulates ownership verification as identity recovery from noisy black-box behavior. SimPrint encodes a private owner identity into an error-correcting codeword and distributes its bits across natural binary question-answering probes. It selectively implants only probes that must deviate from the base model through a low-interference batch update, and later parses suspect-model responses into bits or erasures to recover the registered identity. Because verification requires only input-output queries, SimPrint remains applicable when model weights and activations are inaccessible. Experiments on three open-weight LLMs show reliable identity recovery in both clean and modified settings, robustness to fine-tuning, pruning, quantization, model merging, and serving-time perturbations, and comparable downstream utility.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Fingerprinting", "Ownership Verification", "Black-box Verification" ], "forum_url": "https://openreview.net/forum?id=Kn2mPpj3ZO", "pdf_url": "https://openreview.net/pdf/72b397865ce6ecff3d75e0557d6bcd915025ac12.pdf", "updated_at": "2026-10-05T18:37:54.850Z", "categories": [ "theory" ], "matched_terms": [ "binary", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "B3TYCMjqzj", "number": 12488, "title": "QuantWM: Temporally Consistent 2-Bit KV Cache Quantization for Video World Models", "abstract": "Video generation based world models achieve long-range temporal consistency by storing KV cache during generation, but the continuously growing cache makes KV cache memory become a major deployment bottleneck, which motivates low-bit quantization study for efficiency. Existing 2-bit KV cache quantization methods can achieve nearly quantitatively lossless performance on conventional video benchmarks such as VBench, however, when applied to video world models, we find that they still cause severe temporal flickering and visual degradation. Meanwhile, deeper investigates show that Key quantization produces smaller reconstruction errors than Value, but surprisingly leads to much larger output degradation. We trace this discrepancy to attention in video world models: Key perturbations can change the attention logits, \\textit{i.e.}, \\textit{QK$^\\top$}, and shift the temporal-spatial tokens selected by Queries. These observations motivate us to explicitly preserve attention logits and temporal-spatial token selection during KV cache quantization to alleviate the visual degradation problem. To address this issue, we present \\textbf{Quant\\textit{WM}}, a training-free 2-bit KV cache quantization framework for video world models. \\textbf{Quant\\textit{WM}} introduces two complementary techniques to mitigate the attention shifts. Firstly, \\textit{quantization-sensitivity-aware clustering (QSAC)} jointly considers historical Query sensitivity and residual ranges to select INT2-friendly Key centroids, which reduces quantization errors in channels that are more critical to attention. In addition, \\textit{principal-subspace attention compensation (PSAC)} restores the remaining Key errors along the dominant Query subspace using low-rank projections, which provides a direct and efficient correction to stabilize attention logits. Extensive experiments on LingBot-World-v2, HY-World 1.5, Matrix-Game-2, Longcat-Video and Causal-Forcing demonstrate that \\textbf{Quant\\textit{WM}} significantly improves visual quality and temporal consistency, while outperforming existing methods across image and video quality metrics with up to \\(6.20\\times\\) KV cache memory compression and limited additional overhead. The project page will be released soon.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model quantization", "KV cache", "world model", "video generation" ], "forum_url": "https://openreview.net/forum?id=B3TYCMjqzj", "pdf_url": "https://openreview.net/pdf/b915c9e3c26138a616d19df10dda5be3621ac13c.pdf", "updated_at": "2026-10-05T18:38:18.764Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "2-bit", "int2", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RQPH4EysHa", "number": 12495, "title": "RoQAT: A Robust Quantization-aware Training Method for LLMs Extremely Low-bit Downstream Reasoning", "abstract": "Quantization-aware training (QAT) can largely recover the capabilities of low-bit large language models (LLMs). However, we discover that although achieving nearly lossless performance on general benchmarks, there still exists a tremendous gap on downstream reasoning capabilities at extremely low-bit, especially at 2-bit. We argue that this degradation is closely related to two challenges in extremely low-bit QAT: 1) directly perform low-bit training will cause severe initial degradation. Although progressive quantization can partially stabilize training, existing methods typically ignore sensitivity differences during progression; 2) the weights may converge to the sharp regions of the loss landscape, which will limit the robustness and generalization of the quantized model. Motivated by these observations, we propose \\textbf{RoQAT}, a robust QAT framework for preserving downstream capabilities of extremely low-bit LLMs. RoQAT first introduces \\textit{sensitivity-adaptive progressive quantization} (SAPQ), which considers distinct quantization sensitivity across groups. Specifically, RoQAT dynamically assigns each group a different progression rate to the target bit-width according to the block quantization sensitivity along the actual progressive training direction, which allows more sensitive groups to approach the target bit-width more conservatively. Moreover, RoQAT incorporates \\textit{quantization-aware sharpness minimization} (QSAM) to establish flatter loss landscapes around low-precision weights, which further enhances downstream generalization. It performs sharpness-aware optimization in the weight space during block-wise training and the scale space during end-to-end refinement. Extensive experiments on Qwen3 instruction-tuned and Qwen3.5 hybrid-attention LLMs demonstrate that RoQAT consistently outperforms strong QAT baselines on downstream benchmarks while remaining stable general performance at extremely low-bit.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model compression", "Quantization-aware training", "Downstream reasoning" ], "forum_url": "https://openreview.net/forum?id=RQPH4EysHa", "pdf_url": "https://openreview.net/pdf/e5cd5b7a253660fa96d7dd58e17e578f1c942a47.pdf", "updated_at": "2026-10-05T18:38:20.494Z", "categories": [ "qat", "reasoning", "theory" ], "matched_terms": [ "2-bit", "bit-width", "low-bit", "low-precision", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eEicXkAWDk", "number": 12529, "title": "Where the FP4 Pretraining Gap Sits, and Why That Decides What Transfers Across Scale", "abstract": "Training a language model in 4-bit floating point (FP4) is faster than in BF16 but ends at a higher loss. Recipes narrow this gap by keeping some computations, or the end of training, in higher precision, and these choices are usually tuned on small models. We ask where the gap comes from in one FP4 recipe at three model sizes, and find that the answer changes with scale. In a small model nearly all of the gap comes from rounding in the forward pass; as the model grows this share falls, and rounding the gradients, alone and together with the forward pass, takes over. At every size a similar share of the gap is the cost of the rounding still in progress, which a switch to BF16 removes at once, while the rest stays in the weights. Since higher precision in one part of training removes at most the share of the gap that arises there, where the gap sits decides what transfers across scale: a higher-precision forward pass removes most of the gap in a small model but less than half in a large one, while a short switch to BF16 at the end removes the same share at every size. A few runs at the target scale measure these shares before any recipe is trained.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "FP4 training", "NVFP4", "low-precision training", "quantization", "mixed precision", "precision allocation", "LLM pretraining", "scaling" ], "forum_url": "https://openreview.net/forum?id=eEicXkAWDk", "pdf_url": "https://openreview.net/pdf/1699e69e7095aecfd1a685384c5b2c9347cab529.pdf", "updated_at": "2026-10-05T18:38:41.637Z", "categories": [ "ptq", "training" ], "matched_terms": [ "4-bit", "fp4", "low-precision", "mixed precision", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qFBUcM4Ujz", "number": 12599, "title": "RAFP: Identifying LLM Lineages via Rare-Region Fingerprints", "abstract": "Large language models (LLMs) are increasingly adapted and redistributed through finetuning, parameter-efficient tuning, and quantization, making it difficult to verify whether a suspect model is derived from a particular pretrained model.\nExisting fingerprinting approaches often rely on modifying the protected model or construct behavioral signatures without explaining why they should persist under downstream adaptation.\nWe introduce RAFP, a non-invasive framework for identifying LLM lineages through rare-region fingerprints. Our key insight is that downstream adaptation is concentrated on common high-density language behaviors, while low-probability prompt regions receive weak optimization signal and exhibit limited gradient alignment under finetuned distribution. We show theoretically that, under mild assumptions on downstream coverage and gradient alignment, the likelihood drift of behaviors in these rare regions remains bounded during finetuning.\nBuilding on this insight, RAFP searches for rare prompt--response pairs whose behaviors are shared across models in the same lineage but remain discriminative from unrelated models, without modifying model parameters. The resulting fingerprints support black-box ownership verification and can be generated post hoc for pre-trained models. Experiments across four LLM families and multiple downstream adaptations, including supervised finetuning, LoRA, quantization, prompt-template variation, and decoding changes, show that RAFP achieves strong fingerprint persistence and substantially outperforms prior fingerprinting baselines in black-box settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Fingerprinting; Model Lineage; Model Provenance; Black-box Verification; Robustness to Downstream Tasks" ], "forum_url": "https://openreview.net/forum?id=qFBUcM4Ujz", "pdf_url": "https://openreview.net/pdf/bcd68b812a348b548136a4ea87072c0f2ed73277.pdf", "updated_at": "2026-10-05T18:39:17.075Z", "categories": [ "training", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tEPKtV0ypd", "number": 12687, "title": "Test-Time Adaptation of Quantized ViTs via Single-Pass Quantizer-Aligned Recalibration", "abstract": "Post-training quantization is a standard route to fitting vision transformers (ViTs) into edge compute and memory budgets, yet quantized models become especially brittle under distribution shift. Test-time adaptation (TTA) addresses such shifts without labels, but most existing approaches are poorly aligned with the constraints of quantized inference. Prevailing TTA methods recover accuracy through backpropagation, while backprop-free methods often still incur overhead from extra forward passes or parameter updates, and lightweight feature- or logit-level methods recover only part of the loss. Across these approaches, a quantization-specific failure mode that amplifies the drop is not directly targeted: under shift, activations occupy frozen quantizers' calibrated ranges differently, distorting their *code distribution*. We propose *Quantizer-Aligned Recalibration* (QuAR), a *single-pass* TTA method tailored to quantized ViTs that neither backpropagates nor updates any model parameters. QuAR recalibrates activations at the input to a frozen quantizer, mapping the test stream's running per-channel statistics back toward the source calibration. On ImageNet-C with ViT-B, QuAR achieves the highest mean accuracy among state-of-the-art backprop-free TTA methods at 3-, 4-, 6- and 8-bit weight/activation precision, outperforming the strongest baseline by $2.28$ points at 8 bits and $4.00$ at 3 bits, with $46$% lower latency and a memory overhead of only $0.17$ MB ($0.01$% of peak inference memory). Analysis and diagnostics trace the gain to a reduced per-channel mismatch at these quantizers, which restores the code distribution the baselines leave unchanged or distort further. A single fixed configuration remains ahead across continual streams, non-i.i.d. label shift, seven out-of-distribution suites, and three other backbones.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "test-time adaptation", "distribution shift", "quantized neural networks", "vision transformers" ], "forum_url": "https://openreview.net/forum?id=tEPKtV0ypd", "pdf_url": "https://openreview.net/pdf/d49d8278024d4266a861d12ef9f812868a72032a.pdf", "updated_at": "2026-10-05T18:39:56.731Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "8-bit", "quantization", "quantized", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6Ec8ltjjAk", "number": 12732, "title": "Async-SNN: Efficient Spike-Driven Large Language Models with Asynchronous Computing Paradigm", "abstract": "Large language models (LLMs) are increasingly deployed in edge scenarios where power supply can be strict, intermittent, and time-varying. Spiking neural networks (SNNs), with their event-driven and asynchronous computation paradigm, offer a promising path toward power-adaptive LLM inference. However, existing spiking LLMs largely inherit synchronization-intensive operators from conventional Transformers, such as RMSNorm and Softmax Attention, whose global statistical dependencies fundamentally prevent true neuron-level asynchronous execution. In this work, we propose Async-SNN, a framework for constructing fully spiking LLMs with asynchronous computing capability at the billion-parameter scale. We first build a ready-to-convert Async-LLM by replacing RMSNorm with Hardtanh Dynamic Normalization (HDN) and Softmax Attention with Normalized Linear Attention (NLA), thereby eliminating major sources of synchronization. We then introduce a coarse-to-fine ANN-to-SNN conversion pipeline that combines calibration, supervised fine-tuning based quantization-aware training, and equivalent ST-BIF neuron conversion to preserve task performance. Finally, we propose greedy adaptive firing (GAF), which dynamically reallocates neuron firing budgets across layers and time-steps under instantaneous power constraints. Experiments on 0.6B and 1.7B models show that Async-SNN consistently outperforms existing spiking LLM baselines and maintains strong perplexity and zero-shot accuracy even when only 10% of neurons are allowed to be active per time-step.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Large Language Models", "Asynchronous Computing Paradigm" ], "forum_url": "https://openreview.net/forum?id=6Ec8ltjjAk", "pdf_url": "https://openreview.net/pdf/4b1e485d939133912e54adbe84cf4c1d13e34082.pdf", "updated_at": "2026-10-05T18:40:20.665Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "G1TNQTH7F1", "number": 12816, "title": "Bridging Shannon Surprisal and Latent Geometry: A Unified Framework for VLM Storage, Attention Routing, and Data Pruning", "abstract": "As Vision-Language Models (VLMs) expand into unconstrained, in-the-wild continuous spatiotemporal domains, uniform vector processing creates severe KV-Cache and computational attention bottlenecks. We propose a sparsity-driven methodology that bridges these geometric latent constraints with Shannon's Information Theory. By modeling dense background representations as a von Mises-Fisher (vMF) distribution, we derive that local surprisal exhibits an affine relationship with a vector's cosine similarity with the inverse global mean ($-\\bar{v}$). To overcome representation degeneration---where vectors are constrained within narrow anisotropic boundaries---we formulate an exponential weighting function equipped with a Dimensionality Compensator ($\\alpha$) and a Structural Domain Shifter ($\\beta$). This formulation drives a dual-phase optimization framework. First, in static environments, our Surprisal-Weighted Mixed-Precision (SWMP) strategy achieves a 57.50\\% memory reduction by quantizing low-surprisal representations. Second, for continuous video streams, we introduce Temporal SWMP, leveraging an Exponential Moving Average (EMA) to transition the framework into an Information-Theoretic Attention Router. We further observe strong metric coupling between adjacent-difference routing and a second-difference kinematic evaluation proxy. Under extreme Iso-KV Cache constraints ($\\sim$3\\% of frames), Temporal SWMP exceeds uniform sampling in recall while exhibiting selection behavior distinct from adjacent-difference routing. Together, these results support an information-theoretic criterion for attention routing in resource-constrained continuous VLM deployments across unpredictable real-world environments.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Models (VLMs)", "Representation Degeneration", "Information Theory", "Surprisal", "Attention Monopoly", "Surprisal-Weighted Mixed-Precision (SWMP)" ], "forum_url": "https://openreview.net/forum?id=G1TNQTH7F1", "pdf_url": "https://openreview.net/pdf/9fb3bff8132614b0d3ccc983da3ce8cd0f0b4213.pdf", "updated_at": "2026-10-05T18:41:00.411Z", "categories": [ "kv-cache", "hardware", "multimodal" ], "matched_terms": [ "mixed-precision", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QJTmwyJziM", "number": 13013, "title": "Prefix-Point: An Adaptive Non-Uniform Format for Diffusion Model Quantization", "abstract": "Post-training quantization of diffusion models is challenging because\nactivation distributions vary substantially across layers and denoising\nsteps. Prior work mitigates this with step-aware scaling for INT\nquantization, but scaling adjusts only the range of a distribution and not\nits shape, so it cannot follow activations that shift between heavy-tailed shapes, with\nmost values near zero, and light-tailed shapes spread evenly across the\nrange. Non-uniform\nformats represent such shapes better, yet existing ones fix a\nsingle codebook at design time from an assumed distribution and reuse it\nthroughout the network and therefore degrade\ngeneration quality sharply at low precision ($\\le 6$ bits). We propose \\underline{P}refi\\underline{x}-\\underline{P}oint (PXP), a format\nthat divides the value range into exponent levels and sets the resolution of\neach level independently, allocating bit-width where it reduces quantization\nerror most, and fitting this allocation separately to the distribution of\neach layer and denoising step. Since the resulting\ndesign space grows exponentially with bit-width, we cast level allocation as\na multiple-choice knapsack problem and solve it exactly by dynamic\nprogramming. Its floating-point-like structure keeps decoding and the\nmultiplier cheaper than in existing non-uniform formats. Across\nclass-conditional, text-to-image and video diffusion, PXP lowers FID by up to\n$4.5\\times$ over the strongest existing non-uniform format while being up to\n$6\\times$ faster.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Model", "DiTs", "Non-Uniform Formats" ], "forum_url": "https://openreview.net/forum?id=QJTmwyJziM", "pdf_url": "https://openreview.net/pdf/2469028023ae698dd31a2a9a1bfa610c64a49c5b.pdf", "updated_at": "2026-10-05T18:42:45.425Z", "categories": [ "ptq", "diffusion", "hardware" ], "matched_terms": [ "bit-width", "low precision", "quantization" ], "review_status": "abstract", "summary": "Prefix-Point 认为 step-aware scale 只能调整激活范围，不能跟随去噪过程中分布形状的改变。PXP 将数值范围拆成 exponent levels，逐层逐步骤配置各 level 的分辨率，把指数增长的分配空间写成 multiple-choice knapsack 并用动态规划精确求解，同时保留较简单的解码和乘法结构。", "contribution": "跨类别条件、文生图和视频扩散任务，摘要报告相对最强既有非均匀格式，FID 最高改善 4.5 倍，同时速度最高提升 6 倍。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。摘要中的 ≤6 bit 描述低精度问题背景，未逐项列出实验位宽；6 倍速度数字对应的执行环节与硬件也未注明。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "逐层、逐去噪步的扩散模型激活非均匀量化格式", "bit_width": "摘要未注明", "evidence": [ { "location": "官方摘要", "claim": "相对最强既有非均匀格式，FID 最高改善 4.5 倍。" }, { "location": "官方摘要", "claim": "摘要报告最高 6 倍加速，未注明具体执行环节。" } ] }, { "id": "N3WWNTcVLT", "number": 13042, "title": "Certified Nonlinear Computing for Scaled FP8 Datapaths", "abstract": "Large Language Models (LLMs) demand substantial computation for inference, motivating the adoption of efficient FP8 matrix engines. However, nonlinear operations such as SwiGLU, RMSNorm, and softmax-based attention still commonly rely on wider intermediate arithmetic before FP8 quantization. Premature rounding of these intermediates can alter the final output codes after downstream composition and shared-scale selection.\nIn this paper, we propose a sufficient-state framework for efficient nonlinear computation in scaled-FP8 pipelines. The framework constructs intermediate states that preserve the information needed to determine final element codes and shared scale bytes. We instantiate this principle through directed states for SwiGLU, exact boundary predicates over whole-row statistics for RMSNorm, and interval aggregates with adaptive refinement for softmax attention. These constructions resolve final quantization decisions through consumer composition and shared reductions. For SwiGLU, mathematical verification and exhaustive finite-state replay establish exact outputs under the declared numerical contract; for attention, interval bounds certify output cells when they are uniquely determined.\nBased on this framework, we design and implement complete hardware endpoints for SwiGLU, RMSNorm, and softmax attention. Hardware experiments in Nangate45 demonstrate that the SwiGLU semantic ROM reduces area by 54--61\\% and gate energy by 41--53\\% relative to polynomial controls. Under matched complete-row configurations, RMSNorm reduces area by 4.1\\% and gate energy by 23.0\\%, while attention achieves reductions of 58.8\\% and 79.6\\%, respectively, relative to their corresponding controls. Joint replacement experiments across three language models on GSM8K, HumanEval, and GPQA-Diamond demonstrate task performance comparable to native execution.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-precision inference", "FP8 quantization", "Nonlinear operators", "Rounding semantics", "Sufficient state representations", "Hardware–software co-design" ], "forum_url": "https://openreview.net/forum?id=N3WWNTcVLT", "pdf_url": "https://openreview.net/pdf/65a04405fccfb0ddcd75df483239168d012e1b33.pdf", "updated_at": "2026-10-05T18:42:59.151Z", "categories": [ "ptq", "training", "hardware", "reasoning" ], "matched_terms": [ "fp8", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tZvulD2yEV", "number": 13190, "title": "UDA-SVD: Upstream- and Downstream-Aware Low-Rank Approximation and Adaptive Rank Allocation for LLMCompression", "abstract": "The rapid growth of large language models (LLMs) has intensified the demand for efficient compression, and low-rank factorization provides a simple and broadly compatible solution. \nHowever, in sequential low-rank compression, the effectiveness of compressing each module is shaped by two factors: \\ding{182} upstream compression perturbs the activations received by the current module; and \\ding{183} approximation errors introduced at the current module propagate downstream and can have unequal effects on the final prediction. Existing methods typically model these effects in isolation. Conventional layer-wise approaches account for neither explicitly, whereas cumulative error-compensation methods primarily address upstream deviations and loss-aware methods focus on downstream sensitivity. Moreover, uniform rank allocation ignores differences in rank sensitivity across modules, leading to inefficient capacity allocation. \nWe propose UDA-SVD, an upstream- and downstream-aware framework for low-rank LLM compression. UDA-SVD captures both activation deviations accumulated through upstream compression and the sensitivity of the language-modeling objective to compression errors at the current module. By integrating these two signals into a unified objective, UDA-SVD provides a common optimization basis for module-wise low-rank approximation and adaptive rank allocation under a fixed parameter budget. \nAcross multiple compression ratios and model architectures, UDA-SVD consistently achieves the highest downstream accuracy among the compared SVD-based methods while substantially reducing language-modeling perplexity, including a nearly 40\\% reduction on WikiText-2. These results demonstrate the effectiveness of combining upstream drift, downstream sensitivity, and adaptive rank allocation, highlighting UDA-SVD as a promising framework for effective low-rank LLM compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization" ], "forum_url": "https://openreview.net/forum?id=tZvulD2yEV", "pdf_url": "https://openreview.net/pdf/a6f6914e79139b6db23fb01c92016dfeabd5d8c9.pdf", "updated_at": "2026-10-05T18:44:35.289Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jlv3Ua4jo5", "number": 13330, "title": "RosKV: Robust Spectral Key-Value Cache Compression for Time-Series Foundation Models", "abstract": "Autoregressive time-series foundation models (TSFMs) enable zero-shot long-horizon forecasting, but real-world deployment on resource-constrained hardware is bottlenecked by a growing key--value (KV) cache. While caching past states eliminates quadratic recomputation, persistent memory scales linearly with look-back length and horizon. Furthermore, TSFM prefill importance is \\emph{delocalized in time} yet \\emph{localized in frequency}: seasonal cycles span many look-back tokens, so token eviction and uniform time-domain quantization are ill-matched, whereas a small fraction of RFFT bins carries most K/V energy. We present \\textbf{\\speckv{}}, a memory-first, training-free spectral compressor for frozen TSFMs designed for memory-constrained deployments where persistent cache capacity is the primary barrier. RosKV (i) applies a fixed head-dimension rotation so NF4/NF2 block scales are not dominated by outliers on the few kept bins, (ii) transforms along the look-back with an RFFT, (iii) filters K and V separately by spectral energy, and (iv)~stores adaptive mixed-precision coefficients, reconstructing an approximate dense prefill at every decode step before attention. Across nine datasets, four horizons, and four backbones (TimeMoE, Chronos, Moirai, TimesFM), RosKV consistently outperforms NLP token eviction (SnapKV, PyramidKV), coordinate quantization (KIVI), and frequency baselines (FreqKV), staying within $+3.6\\%$ of Full-Cache aggregate MSE (with matching rankings confirmed on GIFT-Eval). On dense-token architectures where persistent KV storage forms a critical inference bottleneck (TimeMoE), RosKV slashes persistent cache memory by $55\\%$ on average ($27.5$\\,MB vs.\\ $60.5$\\,MB) and up to $78\\%$ at short horizons ($9.5$\\,MB vs.\\ $42.5$\\,MB), while maintaining competitive decoding latency ($383$\\,ms vs.\\ No-Cache $810$\\,ms).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV Cache", "TSFM" ], "forum_url": "https://openreview.net/forum?id=jlv3Ua4jo5", "pdf_url": "https://openreview.net/pdf/79c5edc4d3970fb788f282edab96489e226c7eeb.pdf", "updated_at": "2026-10-05T18:45:55.710Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "m9Kc3ca0p4", "number": 13335, "title": "SpecSpin: Spectral Truncation and Rotation for 4-bit Zero-Shot Time-Series Foundation Models", "abstract": "Time-series foundation models (TSFMs) forecast zero-shot across domains, but they are expensive to serve, and 4-bit weight quantization, the standard remedy borrowed from large language models, is not free for them: the autoregressive loop compounds rounding error over long horizons. We show that TSFM attention has spectral structure that a quantizer can exploit. Along the key/time axis, pre-softmax score rows are smooth and recency-biased, so most of their energy sits in the lowest DCT frequencies; along the input-feature axis, the q/k/v projections concentrate their energy in a small set of DCT directions. We introduce \\textbf{\\specspin{}} (Spectral Spin quantization), a training-free, calibration-free 4-bit recipe built on both observations. \\specspin{} truncates each q/k/v projection to its highest-energy DCT directions, applies a Haar rotation inside the kept subspace, quantizes all linear layers to 4-bit NormalFloat (NF4), and low-pass filters every attention score row before the softmax. On three architecturally distinct TSFMs (Chronos, TimeMoE, Moirai), six benchmarks and four horizons, \\specspin{} is the best 4-bit method on overall average for all three backbones and lowers average MSE below the BF16 model on each ($-0.9\\%$, $-5.5\\%$, $-5.2\\%$, versus $-0.1\\%$, $-0.6\\%$, $-3.5\\%$ for uniform 4-bit quantization). Its advantage grows with the horizon, reaching $-9.0\\%$ (TimeMoE) and $-7.7\\%$ (Moirai) at $H{=}720$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "TSFM" ], "forum_url": "https://openreview.net/forum?id=m9Kc3ca0p4", "pdf_url": "https://openreview.net/pdf/38f076ba4d63d25aaae3406a2d0abd0681875710.pdf", "updated_at": "2026-10-05T18:45:57.080Z", "categories": [ "ptq" ], "matched_terms": [ "4-bit", "quantization", "quantizer", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "57CoQFnJz1", "number": 13407, "title": "When the Second GPU Helps: Topology-Aware Mode Selection for Narrow-Interconnect LLM Serving", "abstract": "Modern LLM serving assumes NVLink-class interconnects, yet most multi-GPU deployments (consumer workstations, PCIe-only cloud VMs) link cards at a few GB/s. On such a narrow link, does a second GPU help, and in which mode? Replica places a full model copy per card, roughly doubling throughput when the context fits one card; staged split partitions the layers, extending reach but moving the per-token attention state (KV cache, growing with context) over the slow link every step. We show that this decision reduces to a calibrated discrete decision rule over two measurable coefficients—portable after per-engine recalibration: the per-token transport tax τ and the replica's fixed-overhead gap ΔA. Their ratio gives a crossover context ctx* = ΔA/τ; we are explicit about identifiability: the decision is robust to ±15% coefficient error (first flip ≈31%), while the point estimate of ctx* is the fit's least-identified parameter (LOCO 120–1,281). We measure both at four link settings across three hosts (6.65/13.3/17.6/25.4 GB/s), four engines, two models, up to k = 4 no-NVLink GPUs: the split tax is sub-linear (flat through k = 3), rejecting the shared-root (k−1)² prediction on distinct-bridge hosts, while replica scaling is near-linear (×3.89). Because τ is system-controllable, the coefficients also serve as an interface: a per-token cost reduction (an fp8-KV plugin, quality-failing in the tested configuration) measurably shifts ctx* rightward; we identify where the crossover breaks (fast links, ΔA < 0). The abstraction is embodied in an automatic decision layer that gates on memory, calibrates online, and drives llama.cpp within its 32,768-token admission window while reaching 65,536-token contexts on two 12 GB consumer GPUs through its own 4-bit staged path, with no manual quantization or RoPE tuning; the reach is not exclusive—manual vLLM and ExLlamaV3 configurations also admit 65K, and are faster there.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM serving; narrow interconnect; PCIe; consumer multi-GPU; model parallelism; KV cache placement; cost-model calibration; mode selection" ], "forum_url": "https://openreview.net/forum?id=57CoQFnJz1", "pdf_url": "https://openreview.net/pdf/dc20d50ce6b51c684529c1e0112abcbbfaf44b6a.pdf", "updated_at": "2026-10-05T18:46:57.752Z", "categories": [ "ptq", "training", "kv-cache", "hardware" ], "matched_terms": [ "4-bit", "fp8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "n4tcebRHmB", "number": 13421, "title": "RONA: ROTATION AND NOISE-AWARE ABSORPTION FOR LOW-BIT UNIVERSAL MULTIMODAL EMBEDDING", "abstract": "Universal multimodal embedding (UME) encodes text, images, video, and visually rich documents into a shared space for cross-modal retrieval. Although reasoning-enhanced methods improve embedding quality, their additional inference-time computation limits low-latency, high-throughput retrieval, leaving direct encoders as the more deployment-friendly choice. Their large backbones, however, still incur substantial serving cost. We study whether post-training quantization (PTQ) can compress UME without sacrificing retrieval accuracy. Across 78 MMEB-V2 tasks, strong PTQ methods remain near full precision at W4A16, W8A16, W4A8, and W8A8, yet collapse at W4A4, a failure not explained by layer-wise reconstruction error. We show that low-bit UME must preserve retrieval geometry and that directions safe under clean activations become harmful when coupled with activation quantization noise. We therefore propose RONA, which combines function-preserving rotations with a group-aligned, quantization-noise-aware safe subspace for weight reparameterization. A per-layer absorption coefficient, calibrated against grouped-GPTQ weight reconstructions and query--target embedding geometry, controls how strongly each layer uses these safe directions while the backbone remains frozen. RONA achieves 73.55 at W4A4, exceeding the strongest PTQ baseline by 3.65 points and retaining 94.4% of full-precision performance. Gains of 6.20 and 2.29 points on two 2B backbones demonstrate its effectiveness across model scales and architectures.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Universal Multimodal Embedding; Post-Training Quantization; Low-Bit Quantization; Embedding Geometry; Multimodal Retrieval" ], "forum_url": "https://openreview.net/forum?id=n4tcebRHmB", "pdf_url": "https://openreview.net/pdf/bb609953acdf08fafd0b77fee5e08992248a30d9.pdf", "updated_at": "2026-10-05T18:47:06.758Z", "categories": [ "ptq", "hardware", "reasoning", "multimodal" ], "matched_terms": [ "low-bit", "ptq", "quantization", "w4a16", "w4a4", "w4a8", "w8a16", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WS1ZR4uzit", "number": 13435, "title": "QDSB: Quantized Diffusion Schrödinger Bridges", "abstract": "Learning generative models in settings where the source and target distributions are only specified through unpaired samples is gaining in importance. Here, one frequently-used model are Schrödinger bridges (SB), which represent the most likely evolution between both endpoint distributions. To accelerate training, simulation-free SBs avoid the path simulation of the original SB models. However, learning simulation-free SBs requires paired data; a coupling of the source and target samples is obtained as the solution of the entropic optimal transport (OT) problem. As obtaining the optimal global coupling is infeasible in many practical cases, the entropic OT problem is iteratively solved on minibatches instead. Still, the repeated cost remains substantial and the locality can distort the global transport geometry. We propose quantized diffusion Schrödinger bridges (QDSB), which compute the endpoint coupling on anchor-quantized endpoint distributions and lift the resulting plan back to original data points through cell-wise sampling. We show that the regularized optimal coupling is stable w.r.t. anchor quantization, with an error controlled by the quality of the anchor approximation. In real-world experiments, QDSB matches the sample quality of existing baselines, requiring substantially less time. Code and data are available at github.com/anon1248/qdsb.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "generative modeling", "diffusion models", "Schrödinger bridges", "optimal transport", "quantization" ], "forum_url": "https://openreview.net/forum?id=WS1ZR4uzit", "pdf_url": "https://openreview.net/pdf/42e5ea715ed87f5bfca2c9be82cd356f100803ee.pdf", "updated_at": "2026-10-05T18:47:20.298Z", "categories": [ "diffusion" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jKVexf3iJ2", "number": 13471, "title": "Channel-group Hierarchical Allocation for Mixed Precision", "abstract": "Low-precision quantization can substantially reduce the memory and computation costs of Transformer inference, but aggressive 4-bit floating point (FP4) quantization often degrades model quality due to highly non-uniform quantization sensitivity across the model. Mixed-precision quantization addresses this challenge by assigning higher precision to sensitive regions while keeping less sensitive regions in lower precision. Existing fine-grained mixed-precision methods often select high-precision regions directly from local or global sensitivity scores, which can over-concentrate the high-precision budget in a subset of the model while leaving other sensitive regions insufficiently protected. We introduce CHAMP (Channel-group Hierarchical Allocation for Mixed Precision), a post-training framework combining FP4 and 8-bit floating point (FP8) that hierarchically distributes a global high-precision budget across model components and Transformer layers and constructs the resulting fine-grained channel-group precision mask using sensitivity-aware selection. CHAMP combines sensitivity- and size-aware module allocation, layer-specific adjustment, and local--global channel-group selection. Across language and video generation tasks, CHAMP provides the largest improvements when the FP8 budget is limited, outperforming prior fine-grained sensitivity-based mixed-precision allocation in this regime. CHAMP narrows the gap to full FP8 in our primary language-model settings, improves video frame-level fidelity, maintains comparable generation quality, and achieves higher theoretical throughput.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Mixed-Precision Quantization", "Low-Precision Inference", "Large Language Models", "Diffusion Transformers", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=jKVexf3iJ2", "pdf_url": "https://openreview.net/pdf/ccc2be31b8a08ec1b17bccf2c867d1d0b5a62306.pdf", "updated_at": "2026-10-05T18:47:34.523Z", "categories": [ "ptq", "training", "hardware", "diffusion", "theory" ], "matched_terms": [ "4-bit", "8-bit", "fp4", "fp8", "low-precision", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "72GrwpTMb5", "number": 13547, "title": "Highly Expressive Activation Functions Based on Learnable Soft Quantization", "abstract": "Activation functions (AFs) provide the crucial nonlinearity that underpins the expressiveness of deep neural networks (DNNs). However, widely used AFs such as ReLU and GELU are simple and non-adaptive, limiting the flexibility of DNNs to capture intricate structures in data. In this paper, we introduce Soft Quantization Activation Functions (SQUAFs), a family of trainable, highly expressive AFs constructed as differentiable relaxations of generalized scalar quantizers. We show theoretically that SQUAFs can approximate any continuous 1D function defined over a closed interval with arbitrary precision. Extensive evaluations on standard image classification benchmarks show that SQUAF yields consistent gains over strong fixed and trainable AFs. Moreover, the advantage of SQUAFs is especially pronounced when a task involves modeling complex signals. For partial differential equation (PDE) solving, Physics-Informed Neural Networks (PINNs) equipped with SQUAFs attain the lowest relative error across nine out of ten PDE problems compared to strong baseline AFs. For implicit neural representations (INRs), SQUAF also consistently improves upon existing INR methods across sixteen 2D image representation cases and four 3D shape representation cases, preserving fine-scale details in the fitted signals. Additionally, we show, through knowledge distillation and model depth reduction experiments, that SQUAFs enable smaller models to outperform larger ones trained with conventional AFs, suggesting that increased AF expressiveness can compensate for reduced network size, thus achieving model compression. These results establish SQUAF as a general-purpose, parameter-efficient AF with particular strength in modeling complex signals.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "activation function", "quantization", "nonlinearity" ], "forum_url": "https://openreview.net/forum?id=72GrwpTMb5", "pdf_url": "https://openreview.net/pdf/897c8fa8491b7fd20f1b4d33940c6ad88f34b195.pdf", "updated_at": "2026-10-05T18:48:25.867Z", "categories": [ "theory" ], "matched_terms": [ "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RAULpdQ0Km", "number": 13558, "title": "Constructive State Abstraction and Finite Identification in Autoregressive Transformers", "abstract": "Understanding which internal information is computationally necessary in a Transformer is a basic question in mechanistic interpretability and state abstraction. We study the state information needed to preserve a fixed-weight, pre-normalized autoregressive Transformer's computation under a declared class of internal interventions. From the weights, we construct explicit updates for residual coordinates in the joint read--write subspace, the squared norm of its omitted complement when normalization requires it, and cache entries grouped by key with their multiplicities and observable value sums. This state reproduces every finite generation distribution without reconstructing the original residual vector. If the centered terminal readout is nonzero, finitely many single-pass interventions identify the state from the probability ratio of one fixed token pair, even for a rank-one readout. Equality of the constructed states is therefore equivalent to equality of observations under all admissible controlled continuations. Controlled cache insertion separates the attention numerator and denominator. We show that residual resets are unnecessary and characterize the information retained when all executed queries lie in specified subspaces. These identification results yield minimum state dimensions for continuous and affine encoders on the stated local domains, and a sharp first-order approximation error under a specified observation metric. Checkpoint analyses examine ranks and sensitivity, while controlled FP64 experiments support fixed-token-pair reconstruction and its dependence on the available interventions, but reveal numerical instability in cache inversion. A separate quantized-checkpoint study finds equal candidate-scalar readings with unequal outputs and intervention effects, distinguishing causal influence from state sufficiency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "autoregressive Transformers; mechanistic interpretability; state abstraction; finite identification" ], "forum_url": "https://openreview.net/forum?id=RAULpdQ0Km", "pdf_url": "https://openreview.net/pdf/5112d102d921a201b79c16b3410dc42cb28a3e9c.pdf", "updated_at": "2026-10-05T18:48:19.872Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RouA5nA2Ny", "number": 13697, "title": "SIEVE: Estimate-Aware Clustering in the Query-Induced Space for Sparse Attention", "abstract": "Sparse attention uses only a small part of the KV cache at each decoding step, while cluster-based indices rank and estimate groups of keys through their centroids. Existing methods cluster keys by angular similarity, even though attention depends on the inner products between keys and decoding queries. Keys that are close in angle can therefore induce very different logits, causing important keys to be diluted in both retrieval and estimation. We introduce **SIEVE**, which clusters keys in the score space induced by the query distribution. Its objective minimizes residual logit error under a robust query-induced metric and places additional weight on keys that are likely to become important retrieval targets. We show that this objective controls errors in attention estimation. Across long-context and long-reasoning benchmarks, SIEVE outperforms training-free sparse-attention baselines at equal budget and remains close to full attention while using only a small fraction of the KV cache.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sparse attention; KV cache; long-context inference; vector quantization; maximum inner product search" ], "forum_url": "https://openreview.net/forum?id=RouA5nA2Ny", "pdf_url": "https://openreview.net/pdf/c1b99ca6fe149cc97b88480d3480fef7fc39dff7.pdf", "updated_at": "2026-10-05T18:49:29.663Z", "categories": [ "kv-cache", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cVoOhvqOiI", "number": 13733, "title": "Where Low Precision Bends Gradients", "abstract": "Low-precision training reduces training time and memory use for transformers. Realizing these gains requires deciding which tensors, operations, and matrix-multiplication passes retain higher precision. Existing precision-allocation methods have focused on local reconstruction error or forward-loss sensitivity. These criteria measure changes in tensor values or the forward loss, but do not directly measure the resulting change in the parameter gradient that drives optimization. We reveal this mismatch by measuring how quantization errors in forward and backward computation change the full model's parameter gradient: Similar local errors at different tensor locations produce gradient shifts whose magnitudes differ by more than tenfold. These shifts mainly rotate, or bend, the gradient. To preserve gradient direction during low-precision training, we introduce bending-guided precision allocation. The method allocates high precision to selected operation groups or individual matrix-multiplication passes where quantization rotates the gradient. Using only a small amount of calibration data, the proposed method determines how much high precision is needed and where to allocate high precision to bring the gradient directions of 4- and 8-bit training close to the bfloat16 (BF16) reference. In GPT-2 continued pretraining, our calibration probe ranks precision allocations before training, closely matching the allocations' final-loss ranking with a Spearman correlation of 0.95, compared with 0.35 for local reconstruction error. Under BF16 mixed-precision evaluation, a model trained in 8-bit floating point (FP8) with our allocation matches the loss of a BF16-trained model. These allocations use 11\\% to 24\\% less high-precision activation-cache memory than a fixed policy that stores the output activations of the token embedding, output head, and every normalization layer in high precision. In matrix-multiplication simulations of NVIDIA's NVFP4 format, our probe identifies forward passes as the main source of bending. Executing forward multiplications in BF16 uses high precision for just one third of the transformer-block matrix-multiplication work and closes 99\\% of the loss gap to BF16 training. At a matched high-precision compute budget, allocating BF16 to selected forward matrix multiplications closes 52\\% of the gap, compared with at most 20\\% for NVIDIA-style block protection. The method also improves low-precision training across four vision backbones.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-precision training", "Mixed-precision training", "Neural network quantization", "Gradient-based optimization", "Numerical stability", "Large language models" ], "forum_url": "https://openreview.net/forum?id=cVoOhvqOiI", "pdf_url": "https://openreview.net/pdf/249ea5af3486316d41a7d48bdc952da7ebcf5c56.pdf", "updated_at": "2026-10-05T18:49:46.122Z", "categories": [ "ptq", "training" ], "matched_terms": [ "8-bit", "fp8", "low precision", "low-precision", "mixed-precision", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wgpR2Y7DnS", "number": 13836, "title": "VeilMoE: Hierarchical Routing Envelopes for Privacy-Aware Mixture-of-Experts Inference", "abstract": "Sparse Mixture-of-Experts (MoE) models activate only a few experts per token, but input-dependent routing exposes expert activity that can reveal information about the input. We present \\method, which keeps the route used for computation inside a trusted controller while exposing a coarser execution structure to the accelerator. The controller maps the native Top-$k$ route to a hierarchical envelope; experts in each touched group receive the same quantized capacity and run through transformed SwiGLU kernels. Under the stated execution assumptions, we prove exact request-local semantics in real arithmetic, characterize the exposed routing structure, and bound slot expansion. On OLMoE-1B-7B, $a=16$ reduces routing-only Top-1 token reconstruction from 37.29\\% to 4.34\\% and Banking77 intent inference from 84.38\\% to 27.47\\%, with a utility change of $-0.22$ percentage points. A trainable attacker using raw rows and the envelope has low accuracy without target-epoch training examples, although it still uses 256 target-epoch requests for checkpoint selection; target-epoch training substantially increases accuracy. Full-decoder Qwen utility changes by $-0.19$ to $-0.03$ percentage points at the tested settings. With matched grouped-expert backends, OLMoE serving costs 7.10$\\times$ Plain latency in prefill and 1.83$\\times$ at 64-row cached decode. The privacy guarantees cover routing structure; transformed values and physical timing remain separate sources of leakage.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Privacy", "MOE" ], "forum_url": "https://openreview.net/forum?id=wgpR2Y7DnS", "pdf_url": "https://openreview.net/pdf/6309eb45008d767c64dbde77a9ae92cc0fa6ad74.pdf", "updated_at": "2026-10-05T18:50:36.155Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DC8YwjcpYE", "number": 13944, "title": "Not All Calibration Data Are Equal: Optimizing Calibration Data for Post-Training Quantization", "abstract": "An often overlooked source of performance variation in post-training quantization (PTQ) is the calibration data used to estimate quantization-related statistics. Even under the same model and quantization configuration, different calibration data can lead to substantially different performance. We propose CalibSelect, a calibration data assessment and selection framework based on model responses. CalibSelect introduces two metrics: the calibration activation range score (CARS), which measures the spread of activation responses while mitigating the influence of extreme samples, and the layer sensitivity discrimination score (LSDS), which measures how clearly and stably calibration data reveal differences in quantization sensitivity across layers. Based on these metrics, CalibSelect adopts a two-stage selection strategy that first identifies several promising calibration data sets from the candidate data pool and then refines them through indicator-guided local optimization. CalibSelect is compatible with different PTQ methods and does not modify their underlying algorithms. Experiments on Qwen2.5-7B and LLaMA3.1-8B using AWQ and AutoRound show that CalibSelect consistently outperforms random sampling and existing selection methods. The largest reduction in quantized model perplexity reaches 8.23\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Calibration Data", "Low-Bit Quantization", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=DC8YwjcpYE", "pdf_url": "https://openreview.net/pdf/fb6ef46c16b1a8b0eac4fd26ead4cb0771b485bc.pdf", "updated_at": "2026-10-05T18:52:06.571Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "o9QQ5aTWFw", "number": 13948, "title": "QSpar: Co-Designing Sparse Attention and Quantized Weights for Long-Context Self-Speculative Decoding", "abstract": "Reducing KV-cache traffic alone is insufficient to eliminate drafting overhead in long-context self-speculative decoding. Under a fixed GPU memory budget, longer contexts constrain the feasible batch size, limiting the amortization of weight access across requests. A sparse-attention-based drafter further reduces draft-side KV-cache traffic but leaves repeated access to the target-sized drafter weights unchanged, exposing a substantial weight-side bottleneck. Low-bit weight quantization reduces this cost, but naively combining it with sparse attention degrades acceptance length. We address this runtime–acceptance trade-off through two key observations: token flips concentrate at low-margin positions, and the BF16 target token usually remains among a few top-ranked draft candidates. Based on these observations, QSpar combines a weight-quantized sparse drafter with a candidate Reranker and a margin-aware Reducer. The Reranker recomputes candidate logits using BF16 LM-head weights, while the Reducer selectively retains alternatives under a bounded candidate budget. Final verification uses the original full-attention BF16 model to preserve the exact target-model generation. Under a fixed KV-cache budget, QSpar achieves $1.49\\times$ average and $1.53\\times$ maximum end-to-end speedups over Vegas, and up to $2.21\\times$ over vLLM.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Speculative Decoding", "Quantization", "Sparse Attention", "KV Cache", "Efficient LLM Inference" ], "forum_url": "https://openreview.net/forum?id=o9QQ5aTWFw", "pdf_url": "https://openreview.net/pdf/9dd1adeae353fefeb999a92ec35c05004f141f00.pdf", "updated_at": "2026-10-05T18:52:05.436Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hdBfu8lIC7", "number": 13985, "title": "WAMPTQ: Beyond Weight Quantization via Task-Utility-Guided Precision Allocation for Efficient World Model Deployment", "abstract": "World action models (WAMs) enable action-conditioned prediction and planning, but efficient deployment faces two challenges. Low-bit sensitivity varies across tasks, while reconstruction proxies do not directly measure native task utility. Prediction and context processing also create large transient tensors, so smaller packed artifacts need not reduce peak memory or latency. We introduce WAM post-training quantization (WAMPTQ), which couples task-utility-guided precision selection with liveness-aware execution. WAMPTQ constructs legal 4- and 8-bit weight maps under a matched packed-size budget and selects the final map using native task utility. An optional large language model (LLM) Agent narrows the legal proposal region before scoring. WAMPTQ then freezes the artifact and applies quantized epilogue fusion (QEF) and virtual-context consumer fusion (VCF) to remove avoidable rescaling and context-assembly intermediates. We separate within-pool selection error from candidate-coverage loss and derive an exact live-set criterion for when local savings reduce the global memory peak. Across 15 tasks from four WAM families, the largest observed quality drop relative to FP32 is 2.22 percentage points, with median end-to-end speedup and peak-memory reduction of $2.36\\times$ and $71.5\\%$, respectively. The code will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "World Action Models", "Post-Training Quantization" ], "forum_url": "https://openreview.net/forum?id=hdBfu8lIC7", "pdf_url": "https://openreview.net/pdf/3665259975bf083c03b729ab455c95f213d09076.pdf", "updated_at": "2026-10-05T18:52:32.368Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "8-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dthmf9Frb2", "number": 14085, "title": "Modulo Folding: Rethinking AI-Generated Image Detection Beyond Low-Bit Planes", "abstract": "Reliable detection of AI-generated images is increasingly important for media forensics and trustworthy visual content. \nRecent studies often use image-based reconstruction errors as an important feature for determining whether an image is AI-generated. However, these approaches typically incur high computational costs and also fail to capture intrinsic noisy features present in the raw images.\nThe recently proposed LOTA method assumes that low-bit planes provide the key discriminative cue and uses bit-plane guided noise generation for detection, which achieves good performance on AI-generated image detection.\nHowever, the underlying mechanism of LOTA remains unclear.\nIn this paper, we reveal that LOTA's bit-plane guided noise generation is mathematically equivalent to taking the image modulo eight. This equivalence shows that the essential ingredient is not the bit-plane representation itself but periodic modulo folding. Indeed, we observe that a broad range of modulo periods achieve comparable performance. We provide a theoretical explanation for this phenomenon: modulo folding compresses the large-amplitude semantic component into a small remainder while preserving small-amplitude artifacts, thereby substantially increasing the relative variance and information of artifacts. Based on this finding, we propose Multi-Period Modulo Folding Detector (MP-MFD), which generalizes the single-period modulo eight design of LOTA to multiple periods and injects the resulting folded images into different stages of a backbone network to capture complementary multi-level artifact cues. MP-MFD requires no reconstruction or iterative sampling and treats LOTA as a special case. Extensive experiments show that MP-MFD achieves SOTA results. These results validate the effectiveness of modulo folding.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "AI-generated image detection" ], "forum_url": "https://openreview.net/forum?id=dthmf9Frb2", "pdf_url": "https://openreview.net/pdf/d54ca70f07b649808b3fbb0d544d8490a66c6dcc.pdf", "updated_at": "2026-10-05T18:53:22.673Z", "categories": [ "ptq", "reasoning", "theory" ], "matched_terms": [ "low-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "n92Fgk1rH4", "number": 14119, "title": "IDScore: Task-Conditioned Likelihood Correction for Low-Bit QAT", "abstract": "Quantization-aware training (QAT) is a key approach for optimizing quantized neural networks. \nHowever, deterministic quantization introduces piecewise-constant mappings that make direct gradient-based optimization difficult, so QAT relies on surrogate gradients whose quality is critical to effective training.\nWe propose Identification-Score (IDScore), a correction that improves surrogate gradient estimation in QAT with little additional training cost. \nThis improvement comes from a new perspective: motivated by quantized system identification, IDScore exploits additional structure information from the quantizer itself, beyond the external signals used by existing methods.\nConcretely, IDScore constructs a local likelihood score from a task-selected neighboring target, the latent parameter, and quantization thresholds, and uses it to form a bounded residual correction.\nOur theory explains why IDScore provides a meaningful correction direction and how its correction strength should be allocated across parameters. Formally, we prove calibrated target-direction consistency and constrained-optimal likelihood allocation, and connect the resulting correction to threshold progress and a local sufficient condition for task-loss descent.\nExperiments on vision and language tasks show that IDScore consistently improves the tested LSQ and QuEST baselines at low additional cost. \nEvaluations span W1A1-W4A4 precision and models with up to 430M non-embedding parameters.\nIt also yields further gains when combined with FOGZO or CAGE.\nControlled ablations support the likelihood-based allocation mechanism characterized by our theory.\nTogether, these results provide both an efficient approach with practical potential for low-bit training and a broader perspective on how quantization structure can inform backward optimization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Aware Training", "Low-Bit Neural Networks", "Efficient Training", "Gradient Estimation", "Likelihood-Based Learning" ], "forum_url": "https://openreview.net/forum?id=n92Fgk1rH4", "pdf_url": "https://openreview.net/pdf/69e6cc1d7229bfd80763d0e7eae550769c413901.pdf", "updated_at": "2026-10-05T18:53:45.136Z", "categories": [ "ptq", "qat", "training" ], "matched_terms": [ "low-bit", "qat", "quantization", "quantized", "quantizer", "w1a1", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dAhF6NUdjU", "number": 14180, "title": "CARE-VLA: Capacity Allocation via Response Evaluation for 1-Bit VLA Quantization", "abstract": "Vision-language-action (VLA) models are rapidly advancing embodied intelligence, but their large parameter footprints remain a major obstacle to deployment on resource-constrained robotic platforms. One-bit post-training quantization (PTQ) offers an attractive route to extreme model compression, yet preserving model behavior at this precision remains challenging. To the best of our knowledge, vector quantization has not yet been explored for VLA models in the 1-bit regime, where deciding which concrete reconstruction deserves scarce additional capacity becomes critical. We introduce Capacity Allocation via Response Evaluation (CARE-VLA), a post-training vector-quantization framework that allocates capacity among concrete low-bit reconstruction candidates according to their response-level distortion. Rademacher-projected reverse-mode pullbacks efficiently estimate native-response sensitivity without materializing full Jacobians, allowing concrete reconstruction errors to be evaluated directly in response space. Complete-policy component interventions further calibrate heterogeneous native-response gains onto a common action-discrepancy scale, enabling reconstruction candidates across vision, language, action, and auxiliary modules to compete under a shared bit budget. At 1.09 bits per weight (BPW), CARE-VLA achieves 91.0\\% and 88.9\\% mean LIBERO success on $\\pi_{0.5}$ and OpenVLA-OFT, respectively, and 69.7\\% visual-matching and 53.2\\% variant-aggregation success on SIMPLER with CogACT. Executed $\\pi_{0.5}$ linear weights achieve $14.5\\times$ storage compression relative to 16-bit storage, including codec overhead. On the RTX 4090, the quantized $\\pi_{0.5}$ model achieves a measured $2.09\\times$ speedup. Our work advances extreme VLA compression toward practical deployment on resource-constrained robotic hardware.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision-language-action models", "post-training quantization", "vector quantization", "policy-response evaluation", "capacity allocation", "model compression" ], "forum_url": "https://openreview.net/forum?id=dAhF6NUdjU", "pdf_url": "https://openreview.net/pdf/9bb2b81dba93bf14589c9ab0cc544cb75c5eff6f.pdf", "updated_at": "2026-10-05T18:54:27.524Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "1-bit", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OC4ErIw9RL", "number": 14250, "title": "One Spectrum, Two Resources: Data-Memory Scaling in Autoregressive Prediction", "abstract": "How much learned memory is needed to benefit from more data? We show that\nthe two resources are governed by one predictive-energy spectrum in a\npositive-entropy autoregressive retrieval source. Each coordinate contributes\nits query probability times the squared radius of its unknown logit.\nWriting \\(\\mu\\) for the resulting energy spectrum, we prove the minimax law\n$\\mathfrak R^*_{\\rm value}(n,B)\\asymp_R\n\\Phi_\\mu(n^{-1})+\\Phi_\\mu(\\tau_B),\n\\Phi_\\mu(t)=\\int\\min\\{x,t\\}\\,\\mu(\\mathrm dx),$\nfor \\(n\\) prediction blocks and a learned state with at most \\(2^B\\) values.\nData set the resolution \\(1/n\\); memory sets the level \\(\\tau_B\\) reached\nby optimal bit allocation. The complete curve also recovers the positive\nspectrum. Energy-dimension pairing is essential: two causal sources with\nidentical block-energy and block-dimension marginals have different data and\nmemory exponents. A masked query-key attention head learns the route and\nvalues, realizing the law with explicit routing, format, and arithmetic errors. Further\nresults give exponent-adaptive allocation, finite-precision realization,\nand compute-precision laws under two-sided arithmetic assumptions.\nExperiments recover the data-memory collapse and coupling exponents,\nexplain the routing and allocation mechanisms, and examine weight-only\nquantization across six pretrained-model scales.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Scaling Laws; Memory-Constrained Learning; Autoregressive Prediction" ], "forum_url": "https://openreview.net/forum?id=OC4ErIw9RL", "pdf_url": "https://openreview.net/pdf/6db11465e9ab09c2359b5b64c6c4b73fd5b8418a.pdf", "updated_at": "2026-10-05T18:55:14.631Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BQw5Oc71cg", "number": 14307, "title": "WUSH-KV: KV Cache Quantization with Data-Adaptive Transforms", "abstract": "KV cache memory and bandwidth costs grow with context length and batch size, which limits efficient long-context inference. To address this bottleneck, we introduce WUSH-KV for low-bit KV-cache quantization. It adapts WUSH, which constructs a data-aware transform from second-order statistics of both factors in a matrix product to reduce quantization error. WUSH-KV uses calibration data to construct separate key and value transforms, with the value transform folded into the model weights and the key transform applied after RoPE. The transforms can be paired with clipped quantizers. For one such quantizer, QuEST INT, we show that, under mild assumptions, the WUSH transform is near-optimal. With this quantizer, WUSH-KV reduces layerwise reconstruction error and achieves the lowest end-to-end perplexity among other tested transforms. For end-to-end evaluation, we integrate WUSH-KV into SGLang using OSCAR-style percentile-clipped affine quantization. At 2-bit, it outperforms the OSCAR transform on Qwen3-8B over downstream benchmarks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV Cache", "Quantization", "LLM" ], "forum_url": "https://openreview.net/forum?id=BQw5Oc71cg", "pdf_url": "https://openreview.net/pdf/f61270e42d15b37b8f9e8edcc8ac6abb2c695023.pdf", "updated_at": "2026-10-05T18:55:53.560Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WVEhEexg1S", "number": 14401, "title": "Full-Stack NVFP4: Stable, Role-Aware 4-Bit LLM Pretraining Beyond Linear Projections", "abstract": "NVFP4 pretraining typically targets Transformer projections while leaving optimizer states, optimizer computation, and attention in higher precision. Extending NVFP4 beyond projections is not a uniform quantization problem: each module has a distinct numerical role and failure mode. We present $\\textbf{Full-Stack NVFP4}$, a role-aware framework instantiated by four independently composable recipes. $\\textbf{Subspace-Protected LoRA-SVD}$ retains a compact BF16 projection subspace within full-shape NVFP4 computation; gradient decoupling and periodic SVD realignment stabilize its low-rank factors, reducing the linear-only loss gap from $\\textbf{1.40\\%}$ to $\\textbf{0.61\\%}$. $\\textbf{NVFP4-State AdamW}$ transforms persistent momentum states before storage, $\\textbf{NVFP4 Root}$ stabilizes direct NVFP4 Newton--Schulz iterations, and $\\textbf{Selective FP4 Attention}$ protects softmax-sensitive products in BF16. On 3B pretraining with 64B tokens, BF16 and Full-Stack NVFP4 reach losses of $\\textbf{2.267}$ and $\\textbf{2.286}$, a $\\textbf{0.838\\%}$ gap, with similar aggregate zero-shot results. Native four-block measurements on one RTX 5090 show 2.50--2.83$\\times$ Root speedups over optimized BF16 and 37.9--42.5\\% lower AdamW peak memory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model pretraining", "low-precision training", "4-bit quantization", "NVFP4", "optimizer-state quantization" ], "forum_url": "https://openreview.net/forum?id=WVEhEexg1S", "pdf_url": "https://openreview.net/pdf/f5cdd72550a4e39d692448c092539cf22047c741.pdf", "updated_at": "2026-10-05T18:56:29.822Z", "categories": [ "training" ], "matched_terms": [ "4-bit", "fp4", "low-precision", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mte3AkyV9U", "number": 14402, "title": "ApexQuant: Data-Free Elastic Quantization by Residual Re-Isotropization", "abstract": "We introduce ApexQuant, a calibration-free quantization method that recursively re-quantizes\nthe residual error, serving as a refinement layer on top of existing quantizers. We establish\nthat a fresh random rotation returns each residual to the uniform distribution on the\nhypersphere, which characterizes the rate of progressive error decay across successive passes.\nThis result lets us determine, before any weight is read, how many passes a layer needs for a\ntarget weight-space error. Every prefix is itself a valid lower-rate model, so one artifact\nserves several precisions. We instantiate ApexQuant with three interchangeable stages, scalar,\n$E_8$ and trellis, and validate it on four open-weight LLMs and on Earth-observation and medical\n domains where in-distribution data is often\nunattainable as imagery arrives under restrictive licences or due to patient material under privacy\nconstraints. Progressive re-isotropization comes within a few percent of full precision at\nfour bits and gives the best two-bit arm we measure, in a completely data-free setting.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Data-free quantization", "LLMs" ], "forum_url": "https://openreview.net/forum?id=mte3AkyV9U", "pdf_url": "https://openreview.net/pdf/2f5d9bcd0bc3d7e7a0a4af0f3a11714c1d4b83cc.pdf", "updated_at": "2026-10-05T18:56:30.832Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantizers", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "o4vMcydmWh", "number": 14448, "title": "Which Computations Drive Visual-to-Text Transfer? A Causal Ledger for Vision-Language Models", "abstract": "Vision-language models can incorporate visual evidence into text states, but how much\neach decoder block contributes to the answer-relevant visual effect remains difficult\nto quantify. We introduce reciprocal nested interventions that decompose a common\nimage-conditioned answer effect into ordered block contributions. By replacing visual\nrows at successive decoder boundaries while retaining text states, we measure how much\nof the visual effect has entered text states by each depth. These contributions sum\nexactly to the total effect, yielding a layerwise ledger of cumulative transfer\nand residual visual dependence. On Qwen3.5-0.8B and Qwen3.5-4B, just two dominant writer blocks carry \n$83\\\\%$ and $71\\\\%$ of the total visual effect, respectively. This rapid accumulation establishes \nearly causal closures at $K12$ (0.8B) and $K16$ (4B). The ledger directly guides \ncompression: closure determines when visual tokens can be evicted, while writer share prioritizes which blocks' $K/V$ inputs to protect in BF16\nunder quantization. These predictions depend on task and readout: transfer schedules\nshift across tasks, and generation can still require late visual access beyond an\nimmediate-answer closure.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-language models", "Mechanistic interpretability", "Causal interventions", "Visual-to-text transfer" ], "forum_url": "https://openreview.net/forum?id=o4vMcydmWh", "pdf_url": "https://openreview.net/pdf/d58f8fa2601544978bf93b4255ac6dcce74ccdca.pdf", "updated_at": "2026-10-05T18:57:02.932Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "K0XSYEV1Dz", "number": 14536, "title": "Autonomous Model Quantization with Explorer--Reviewer Agents", "abstract": "Model quantization is critical for efficient deployment of large language\nmodels, but developing a strong quantization recipe requires decisions\nabout methods, data, and experimental evidence. We study frontier coding\nagents for this task.\n\nWhen given a quantization method and its target configuration,\na strong standalone agent can set up the environment, debug the\nimplementation, and reproduce the reported results. In this setting,\nhowever, the researcher still chooses the method. Autonomous quantization\noptimization instead requires the agent to select methods, design\nexperiments, and decide which directions to pursue. Although the agent\ncan develop competitive recipes, we observe four problems during\nautonomous exploration: overlooked prior work, inadequate experimental\ndesign, unsupported conclusions, and premature stopping.\nBased on these observations, we propose an Explorer--Reviewer system.\nThe Explorer develops and executes candidate improvements, while an\nindependent Reviewer evaluates research decisions at four\ncheckpoints: proposals, experimental plans, conclusions, and requests to\nstop. The agents communicate through a shared research log. Our study covers\n2-bit weight-only quantization of Llama-3-8B, Qwen3-8B, and Mistral-7B.\nOn Llama-3-8B, Qwen3-8B, and Mistral-7B, our system\nimproves average zero-shot accuracy across seven tasks by 0.88, 0.60,\nand 0.19 percentage points, respectively, over the standalone agent.\nIt also outperforms the evaluated published quantization baselines\non this metric for all three models.\nCase studies show how the Reviewer guides data selection, requests\nstronger experimental controls, and challenges unsupported conclusions.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model quantization", "LLM agents", "model compression", "multi-agent systems" ], "forum_url": "https://openreview.net/forum?id=K0XSYEV1Dz", "pdf_url": "https://openreview.net/pdf/5bbc8c584b5d8dfbe2f477fd47192b7e1d73ef9c.pdf", "updated_at": "2026-10-05T18:57:56.162Z", "categories": [ "hardware" ], "matched_terms": [ "2-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a5apIxiOUC", "number": 14691, "title": "C-TQIM: Cost-Aware Terminal Quantization for Inversion-Free Diffusion Watermarking", "abstract": "Image watermarks must survive image processing while preserving the intended output. We study this trade-off with frozen generator and variational autoencoder (VAE) weights, no additional neural watermark decoder, and detection through one VAE encoding without diffusion inversion. We propose C-TQIM, a cost-aware terminal quantization index modulation (QIM) method that embeds a fixed keyed pilot between sampling and VAE decoding. Offline calibration estimates the RGB distortion cost and bias-preserving mean squared error of each candidate discrete cosine transform (DCT) direction under the fixed Identity-calibrated reader used at deployment. A joint criterion selects 144 carriers for nearest-lattice embedding, and an analytic detector aggregates their affine-corrected, pilot-aligned scores. On Stable Diffusion 3.5 Medium, with 1,000 threshold-calibration hosts and 4,000 matched test hosts, C-TQIM achieves a macro AUC of 0.9946 and macro true-positive rate (TPR) of 95.40\\% across 13 conditions. The maximum observed false-positive rate (FPR) is 0.775\\%, with paired PSNR of 33.36\\,dB and LPIPS of 0.0212. In a separate 800-host comparison matched by development PSNR, joint selection improves macro TPR over cost-only selection by 4.45 percentage points (95\\% paired interval: 3.85--5.04). C-TQIM provides a directly readable fixed-key presence channel without changing the sampling process or training a watermark-specific neural decoder.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion models; Image watermarking; Quantization index modulation; Inversion-free detection; Cost-aware carrier selection." ], "forum_url": "https://openreview.net/forum?id=a5apIxiOUC", "pdf_url": "https://openreview.net/pdf/79dcf7b67870956f674eb014295fca7f10b495a8.pdf", "updated_at": "2026-10-05T18:59:12.132Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zgXd84JH28", "number": 14820, "title": "QuantForge: Discovering Residual Decompositions for MXFP4 Post-Training Quantization", "abstract": "Four-bit post-training quantization can reduce the memory demands of large language models, but preserving accuracy under strict MXFP4 W4A4 requires coordinating several design choices. Coordinate transforms change block-encoding errors, which in turn affect the residuals propagated through the network. The useful algorithmic decomposition is therefore not fully known before search. LLM-driven program evolution offers a way to explore these choices, but performance scores alone do not explain which design should change next. We introduce QuantForge, a PTQ discovery system that records competing explanations, selects controls that distinguish them, and checks that successor code implements the resulting conclusions. This residual compilation guides program revisions while retaining useful programs even when their original explanations are rejected. Remeasuring the revised program reveals the next error to address. This process discovers HiRes, a fixed MXFP4 quantizer that shapes coordinates, refines legal code assignments, and recovers errors along attention and MLP paths. Each stage acts on residuals measured after the preceding stage has executed. Across seven tasks, HiRes achieves the lowest seven-model Robust Fit (0.09300) and the lowest quantized Fit-7 at 32B. In matched-budget comparisons of LLM-driven program evolution, each with 240 evaluator calls, QuantForge reaches a held-out transfer target in six of eight runs, compared with three each for textual memory and reflection memory, and one for score-only evolution, despite evaluating fewer new programs. These results show that QuantForge improves the discovery of transferable PTQ algorithms by turning controlled evidence into subsequent program changes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Machine Learning", "Large Language Models", "Model Efficiency", "Automated Algorithm Design" ], "forum_url": "https://openreview.net/forum?id=zgXd84JH28", "pdf_url": "https://openreview.net/pdf/4d238270b6c64030eac250afaced6ee039dfb319.pdf", "updated_at": "2026-10-05T19:00:13.415Z", "categories": [ "ptq" ], "matched_terms": [ "mxfp4", "ptq", "quantization", "quantized", "quantizer", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a6UFdIpape", "number": 14821, "title": "SONU: Communication-Efficient Orthonormalized Updates via Sketching", "abstract": "As Large Language Model (LLM) training scales across distributed systems, gradient synchronization can incur substantial communication overhead, particularly in bandwidth-constrained settings. Gradient compression reduces this cost by communicating compact approximations of the gradient. Existing approaches, including sparsification and quantization, are typically designed to preserve the gradient under entrywise or Euclidean reconstruction criteria. These criteria, despite effective in coordinate-wise preconditioning optimizers, need not preserve the information used by matrix-orthonormalized optimizers such as Muon, whose update depends on the gradient’s spectral geometry rather than its individual entries. Such discrepancy makes preserving the geometry an ideal target of the gradient compressor. We propose Sketched Orthonormalized Updates (SONU), a communication-efficient optimizer built on two-sided gradient sketching that facilitates synchronizing gradients in a low-dimensional space. SONU forms a generalized Nyström surrogate of the gradient from the sketches, which coincides with the dense gradient up to the sketched subspaces, and extracts the paired singular vectors without dense orthonormalization. To prevent a fixed complement from being persistently omitted, SONU maintains an evolving sketch, in which the sketching subspace is updated with the orthonormalized directions and are gradually transported across iterations. Empirically, on chain-of-thought, agentic function-calling, and instruction-following tasks, SONU attains lower perplexity per communicated byte than the other communication-efficient orthonormalized update algorithms. Furthermore, SONU matches the perplexity trajectory of Dion, an efficient Muon variant which synchronizes the full gradients before applying orthonormalized updates.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "communication-efficient; orthonormalized updates" ], "forum_url": "https://openreview.net/forum?id=a6UFdIpape", "pdf_url": "https://openreview.net/pdf/b4fbb73866490b3bf30965b659da044b00fa24a6.pdf", "updated_at": "2026-10-05T19:00:13.480Z", "categories": [ "ptq", "training", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "kaA2itb1S8", "number": 14863, "title": "Rethinking Quantized LLM Fine-Tuning: An Adaptive-Grouping Alternating Zeroth-Order Algorithm in Continuous Mapping Space", "abstract": "Fine-tuning quantized Large Language Models (LLMs) is crucial for model adaptation on resource-constrained edge devices, where zeroth-order (ZO) optimization has emerged as a memory-friendly alternative to first-order methods. Existing ZO fine-tuning methods typically focus on either perturbing discrete quantized weights or coarse-grained (per-tensor/channel) continuous scales. In this paper, we rethink this fundamental choice and recognize that while fine-tuning should ideally target the dequantized weights, the discrete nature of weights restricts optimization to a finite search space, and tuning coarse-grained scales is limited to a constrained continuous subspace that lacks the expressive capacity of the full parameter space. To mitigate these issues, we propose to break down scales and zero-points to a fine-grained level and fine-tune them in a continuous mapping space. Motivated by this, we propose CoZO, an adaptive-grouping alternating zeroth-order algorithm that optimizes these fine-grained parameters while adaptively balancing search-space expressiveness and estimation stability. Crucially, we provide a theoretical convergence guarantee for CoZO. Extensive experiments demonstrate that CoZO achieves superior stability and effectiveness, consistently outperforming discrete-perturbation and scale-tuning baselines under stringent memory constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantized Large Language Models", "Zeroth-Order Optimization", "Memory-Efficient Fine-Tuning" ], "forum_url": "https://openreview.net/forum?id=kaA2itb1S8", "pdf_url": "https://openreview.net/pdf/46babaaa99d1681d2bbfdb4bf368a58c77c0cf80.pdf", "updated_at": "2026-10-05T19:00:26.408Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zgzxPkapxR", "number": 14871, "title": "Fine-Tuning Low-Bit Models with Gradients in Quantized Code Space", "abstract": "Fine-tuning low-bit models aims to adapt a quantized model while keeping the final deployed checkpoint in the same low-bit form. This setting is practically important as it reduces memory and inference cost for storage and deployment.\nUnder this constraint, adaptation becomes an optimization problem over quantization codes and scales. \nExisting\ncontinuous low-bit training is efficient, but it can be distorted by straight through estimation error or by post-quantization gap; discrete search is deployment-faithful, but can be computationally inefficient under a finite training budget.\nWe propose code surrogate gradient to find the steepest direction of a local continuous surrogate in code space to accelerate optimization, and perform guided search to preserve deployment faithfulness.\nExperiments across arithmetic reasoning, instruction following, and structured language understanding show that GradCodes consistently improves fine-tuning low-bit models across different quantization datatypes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantized LLM", "Optimization", "Search" ], "forum_url": "https://openreview.net/forum?id=zgzxPkapxR", "pdf_url": "https://openreview.net/pdf/901eeba5505a78f143361825b40c0cb277b68171.pdf", "updated_at": "2026-10-05T19:00:31.599Z", "categories": [ "training", "hardware", "reasoning" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6hhRn8lxmq", "number": 14933, "title": "From Relevance to Representatives: Equal-Rate Evidence Quantization for Long Video Understanding", "abstract": "Long video understanding often requires selecting useful visual observations under a limited frame budget. Uniform sampling distributes observations by elapsed time, whereas global relevance ranking can concentrate them in locally redundant high-scoring regions. In this paper, we propose Equal-rate Evidence Quantization (E²Q), a training-free frame selector for long video understanding that separates where to allocate the budget from which observation to retain. E²Q uses normalized frame–question relevance to allocate the frame budget, partitioning the ordered candidates into consecutive intervals with approximately equal relevance mass and retaining the highest-scoring frame from each interval. A budget-dependent projection handles overly concentrated candidate weights before partitioning, enabling relevance-adaptive intervals without an additional relevance–coverage coefficient. We characterize this construction and bound frame-level representation error in terms of within-interval variation and interval-mass imbalance. Across four long-video question answering benchmarks and three models, E²Q achieves the highest accuracy among training-free selectors in ten of the twelve model–benchmark settings. Controlled shared-score comparisons further show consistent gains when the relevance signal is fixed and isolate the respective contributions of relevance-mass allocation and within-interval representative selection to these gains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Long Video Understanding", "Keyframe Selection", "Multimodal Large Language Models" ], "forum_url": "https://openreview.net/forum?id=6hhRn8lxmq", "pdf_url": "https://openreview.net/pdf/8c2277ad075b2ed62453f2eb600697f879319e04.pdf", "updated_at": "2026-10-05T19:01:09.886Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "elcVIsGXqC", "number": 14958, "title": "FastSNN: Accurate Post-Training Compression for Accelerating Spiking Neural Networks", "abstract": "Spiking neural networks (SNNs) have shown promising performance on a wide range of perception tasks, making them attractive for edge deployment. However, these powerful SNNs require massive memory and computational resources, often exceeding the tight budgets of edge devices. Model compression is therefore essential, and post-training compression (PTC), which efficiently compresses a pretrained model without retraining, is particularly appealing. Existing PTC methods for SNNs minimize the local reconstruction error of each layer's signals, but this objective overlooks the discrete and temporal nature of spike transmission. In SNNs, a weight perturbation introduced by compression alters the transmitted signal only through spike flips, i.e., a spike is either inserted or deleted. We find that model performance after PTC depends less on how many flips perturbations cause than on whether flips are balanced. Building on this flip-balance perspective, we propose FastSNN, a PTC framework that jointly accounts for signal reconstruction and flip balance during SNN compression. By modeling the relation between weight perturbations and spike flips in a unified form, FastSNN applies to a broad range of compression types, including pruning and quantization. Across multiple datasets and SNN models, FastSNN outperforms the state-of-the-art method by 9.09\\% on average under 87.5\\% unstructured pruning and delivers up to a 4.2$\\times$ inference speedup on a prototype neuromorphic chip.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks; Post-Training Compression" ], "forum_url": "https://openreview.net/forum?id=elcVIsGXqC", "pdf_url": "https://openreview.net/pdf/a2c892e92a7512622af608f6cc2f9dff7c869291.pdf", "updated_at": "2026-10-05T19:01:20.601Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Pzd6dgwKKE", "number": 14980, "title": "RDKV: Rate-Distortion Bit Allocation for Joint Eviction and Quantization of the KV Cache", "abstract": "Large language models (LLMs) have shown strong performance across diverse tasks, but their inference with long input contexts is bottlenecked by memory size and bandwidth. The Key-Value (KV) cache size grows linearly with sequence length and needs to be re-read from off-chip high-bandwidth memory (HBM) to on-chip memory at every decoding step, resulting in memory-bound inference. Existing methods reduce the cache by either eviction or quantization, but typically treat the two in isolation. In this paper, we cast KV cache compression as a rate-distortion problem, under which eviction and quantization are two end-points of the same bit allocation scheme. This exposes the need to optimize them jointly, motivating our method, RDKV (Rate-Distortion KV cache compression). RDKV derives the weight of each token or channel from the distortion that compression induces on the attention computation. Based on these weights, it assigns each token or channel a bit-width ranging from full precision down to zero bits via Lagrangian relaxation. Experiments on LongBench, RULER, and InfiniteBench show that RDKV outperforms the best evaluated baseline by 9.1% on average. On LongBench it recovers 97.81% of full-cache accuracy with only 2.48% cache retention. Compared with full-cache FlashAttention-2 decoding, it achieves $4.5\\times$ decode speedup and $1.9\\times$ peak memory reduction with 128K context length, while maintaining comparable performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "KV cache compression", "efficient inference", "quantization", "long-context inference" ], "forum_url": "https://openreview.net/forum?id=Pzd6dgwKKE", "pdf_url": "https://openreview.net/pdf/4bfcb72f8ba0edf3a2de7bfe514204fdeedb380e.pdf", "updated_at": "2026-10-05T19:01:34.829Z", "categories": [ "kv-cache" ], "matched_terms": [ "bit-width", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qc0IiOjoyz", "number": 14996, "title": "ReAlignQ: Post-Compression Refinement for Joint Sparsification and Quantization of LLMs", "abstract": "Quantization and sparsification compress large language models by restricting\nnumerical precision and nonzero support, respectively. Support selected in\ncontinuous weight space may be poorly matched to the values that its retained\ncoordinates can realize on the quantization grid. Moreover, criteria applied\nto individual coordinates overlook interactions among simultaneous changes in\nsupport membership and quantized values. We therefore propose ReAlignQ, a\nrefinement framework that jointly selects support exchanges and legal discrete\nvalues for incoming weights under fixed quantization scales and a prescribed\nsparsity budget. For each linear weight matrix, ReAlignQ models interactions\namong screened candidates using an empirical gradient Gram surrogate built\nfrom gradients of individual calibration samples. It retains the resulting\ngradient signatures only for screened candidates, which reduces the storage\nneeded to model their interactions. We further implement sparse and quantized\noperators with exact bitmap masks. These operators execute refined nonuniform\nmasks without changing them. We also provide theoretical results that\ncharacterize support selection over legal codes, bound the empirical gradient\nGram interactions omitted by diagonal screening, and establish surrogate\ndescent while preserving sparsity for accepted exchanges. Across Llama and\nQwen models and multiple sparsity levels, ReAlignQ consistently improves\nlanguage-modeling quality\nwithout retraining, changing parent scales, or relaxing sparsity budgets. On\nLlama 3 8B with an EIWR W8A16 parent at 70\\% sparsity, it reduces WikiText-2\nperplexity from 35.48 to 25.57 and improves mean zero-shot accuracy by 2.6\npercentage points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Joint Sparsification and Quantization" ], "forum_url": "https://openreview.net/forum?id=qc0IiOjoyz", "pdf_url": "https://openreview.net/pdf/001d54430b9687a0bde676abe27832771d712937.pdf", "updated_at": "2026-10-05T19:01:50.435Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "quantization", "quantized", "w8a16" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YqUjb0f1UC", "number": 15134, "title": "Adaptive Inverted-Index Routing for Granular Mixtures-of-Experts", "abstract": "Sparse Mixture-of-experts (MoE) models enable scalable transformer architectures by activating only a subset of experts per token. Recent evidence suggests that performance improves with increasingly granular experts, i.e., many small experts instead of a few large ones. However, this regime substantially increases routing cost, which can dominate computation and memory. We introduce adaptive inverted-index routing for MoE (AIR-MoE), an inverted-index-inspired routing architecture based on vector quantization (VQ). In a first stage, AIR-MoE performs coarse shortlisting by assigning tokens to VQ codewords to construct a candidate set of experts. In a second stage, fine scoring computes exact routing scores restricted to this shortlist. This two-stage procedure approximates true top-K routing while avoiding full expert scoring and, in contrast to prior work, imposing no structural constraints on expert parameters. AIR-MoE serves as a drop-in replacement for standard routers and requires no modifications to the model architecture or loss function. We further provide a lower bound on the mass recall achieved by AIR-MoE that yields insights into the inner workings. Empirically, AIR-MoE improves the perplexity–FLOPs trade-off over existing MoE routers in regimes with up to more than a million tiny experts, and achieves consistent perplexity improvements over the best baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixture-of-experts", "transformer", "large language model", "inverted index" ], "forum_url": "https://openreview.net/forum?id=YqUjb0f1UC", "pdf_url": "https://openreview.net/pdf/11916f3894028cb34812b6a6b536c725241bd092.pdf", "updated_at": "2026-10-05T19:03:01.799Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NveQO5VAGV", "number": 15245, "title": "MxQBound: Efficient Mixed-Precision Quantization for MoE-LLMs Guided by Expert Importance and Error Bound Guarantees", "abstract": "While mixed-precision quantization reduces the memory overhead of Mixture-of-Experts (MoE)-Large Language Models (LLMs), existing methods often mis-estimate expert importance, causing suboptimal precision allocation and severe performance degradation under aggressive quantization. We propose MxQBound, an efficient and robust Mixed-precision Quantization framework for MoE-LLMs, guided by holistic assessment of expert importance and equipped with quantization error Bound guarantees. With only a single forward pass, MxQBound directly assesses each expert’s importance from its output norms, supported by a provable upper bound on quantization error. Guided by the importance assessment, MxQBound jointly optimizes the per-expert bit-widths of weights and activations to minimize quantization errors under an overall bit-allocation budget. It further adapts each expert’s activation bit-width according to its input strength during inference with negligible overhead. Extensive experiments on diverse MoE-LLMs and benchmarks show that, under comparable compression ratios, MxQBound outperforms state-of-the-art (SOTA) methods by 4.08-22.85% in model performance and 2.2-141× in computational efficiency. Under aggressive W4A4, MxQBound incurs an average accuracy drop of only 15.07\\% relative to the full-precision baseline, whereas SOTA methods collapse to near-zero accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "Large Language Model Quantization", "Post-Training Quantization", "Mixed-Precision Quantization", "Weight-and-Activation Quantization" ], "forum_url": "https://openreview.net/forum?id=NveQO5VAGV", "pdf_url": "https://openreview.net/pdf/e43d017230d90f84e45fbd815e19ab4f09b03309.pdf", "updated_at": "2026-10-05T19:04:01.219Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "bit-width", "mixed-precision", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Jx6ZFrztDl", "number": 15266, "title": "Weight-Adjusted Gradients Reveal Parameter Importance and Failure Modes in LLMs", "abstract": "Understanding which parameters are influential in Large Language Models (LLMs) is central to improving their efficiency, reliability, and interpretability. We introduce Weight-Adjusted Gradients (WAG), a simple yet effective approach for estimating parameter importance that explicitly captures the interaction between model weights and first-order gradient information and identifies parameters that disproportionately influence model behavior, such as those responsible for collapse phenomena in LLMs. Across a range of models and settings, we show that WAG surfaces a tiny but critical subset of parameters (≤ 0.5 parts per million or 0.00005% of model size) whose modification leads to dramatic degradation in performance, indicating a novel failure mode. These findings also reveal a previously underexplored interplay between weights and gradients, suggesting that parameter importance cannot be fully understood through either signal alone. We demonstrate the practical utility of WAG across several diverse applications, such as expert allocation in Mixture-of-Experts (MoE) architectures, targeted unlearning, mixed-precision quantization, and layer selection for knowledge editing. In sum, WAG can serve as a unified approach for analyzing, debugging, and controlling LLMs, and opens new directions for principled parameter-level interpretation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "parameter importance", "model collapse", "large language models" ], "forum_url": "https://openreview.net/forum?id=Jx6ZFrztDl", "pdf_url": "https://openreview.net/pdf/c50842cea8acb2d94e8e3637a3ecf75f39644705.pdf", "updated_at": "2026-10-05T19:04:18.452Z", "categories": [ "training" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eg16jDMikI", "number": 15271, "title": "Softmax Reparameterization For Output-Head Quantization", "abstract": "Large vocabularies make language-model output heads a substantial inference cost in small language models. We introduce softmax reparameterization, a post-training method that searches over functionally equivalent output heads before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL. For linear-\nsoftmax heads, these shifts preserve full-precision predictions exactly and require no decoder retraining; a rank-one correction extends the construction to nonlinear logit paths. Across seven output heads and three quantizers, W4 gains are largest where baseline quantization substantially distorts predictions: test KL falls by 93% on XGLM under RTN and by 73–77% on Phi, BLOOM, and BLOOMZ under activation-weighted MSE. Heads with low baseline error change little; at W2, used as a compression stress test, benefits extend more broadly. On Phi, the gains persist under stronger GPTQ calibration, and independent Phi and BLOOM holdouts reproduce the improvements. Frozen WikiText-selected coefficients also transfer to C4 and OpenWebMath. Residual analysis on Phi shows how fidelity\ncan improve despite greater total logit error: the selected representative reduces error on likely outputs and lowers its Fisher-weighted cost. For shift-compatible heads, the shift adds no inference operation. With the decoder held in BF16, a packed W4 Phi output head reduces batch-one generation latency by 10.8%, and reparameterization preserves this speedup.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Inference", "Quantization", "PTQ" ], "forum_url": "https://openreview.net/forum?id=eg16jDMikI", "pdf_url": "https://openreview.net/pdf/edfb16959e35b959974ba36f61c701323ed6fcef.pdf", "updated_at": "2026-10-05T19:04:20.426Z", "categories": [ "ptq", "theory", "hardware" ], "matched_terms": [ "ptq", "quantization", "quantizers" ], "review_status": "reviewed", "summary": "针对大词表语言模型的输出头权重量化，利用 softmax 对公共 logit 偏移的不变性：从每个词表行减去 t 倍行均值，量化后用验证集相对 BF16 源模型的 KL 选择 t。主方法对每个量化器分别搜索含原始头与均值中心化的 14 个候选；线性 softmax 在量化前保持分布完全等价，decoder 与输入 embedding 均保持 BF16，无需重训。", "contribution": "把输出头的加性等价类变成量化前可搜索的表示，而非固定做均值中心化或只优化权重重建误差；用概率加权的相对 logit 误差解释 KL 改善。非线性 soft-cap 路径需要秩一补偿。附录另探索按量化组选择系数，并以单模型、相同候选数的实验检验扩展收益。", "limitations": "实验主要隔离输出头，不能推广为整个 decoder 的量化效果。部署仅测 Phi 的部分量化器；Gemma 4 秩一补偿未测 packed 开销，GPTQ 未测时延。保留 BF16 输入 embedding 并新增独立量化头，使 tied Phi 常驻权重从 7.17 增至 7.47 GiB，并非整模内存下降。英文选出的系数在 BLOOM 的 FLORES 阿拉伯语/RTN 与印地语/RTN、AW-MSE 上反而增加 KL；W2 仍可能明显偏离源分布。主测试集参与过方法开发，另有独立 holdout；分组九模型实验仍是探索且搜索预算更多，KL 降低也不保证 PPL 降低。有限验证搜索仅保证不劣于同验证集候选，不能保证跨域泛化，本文读稿未独立复现实验。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "32c8b9f1cf548d18c099074ad0f041b7a98c3b20a99965bed76c2a845a9e892a", "reviewed_pdf_url": "https://openreview.net/pdf/edfb16959e35b959974ba36f61c701323ed6fcef.pdf", "reviewed_source_updated_at": "2026-10-05T19:04:20.426Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "语言模型输出投影头的权重；BF16 decoder 与输入 embedding 固定。", "bit_width": "主实验 W4、group size 128；W2 为压力测试，附录补充 W3。实际 Phi 部署为 packed INT4/Marlin。", "evidence": [ { "location": "PDF p.3–4 / Section 2.1–2.2 / Equations 1–3", "claim": "公共行向量偏移保持线性 softmax 的全精度分布；每个量化器以真实验证 KL 从含 t=0、t=1 的 14 点标量网格选表示。" }, { "location": "PDF p.5 / Table 1 / Section 3.1", "claim": "七头 W4 主实验：XGLM RTN KL 2.13→0.143，Phi GPTQ 0.158→0.059；误差本来低的头改善较小，W2 并非恢复源级保真度的保证。" }, { "location": "PDF p.6–7 / Section 3.3 / Table 3", "claim": "匹配 16,352-token Phi packed 质量评估：AW-MSE PPL 31.65→15.40、KL 0.97→0.28；GPTQ PPL 13.67→12.23、KL 0.15→0.06；BF16 PPL 11.65。" }, { "location": "PDF p.7 / Table 3; p.24–25 / Appendix G.1", "claim": "A10G、vLLM/Marlin、64-token 贪心生成：shifted AW-MSE B1 1013.4 ms，BF16 1136.6 ms；GPTQ 未测延迟。包含 prefill、关闭 prefix cache，三次 warmup、七次重复、两轮独立重载并反转运行顺序。" }, { "location": "PDF p.7–8 / Section 4 / Tables 4–5", "claim": "Fisher 二次型刻画源概率加权的相对 logit 误差；Phi 可以在总 logit 重建误差增加时降低 KL。该二次近似用于诊断，选系数仍用真实 validation KL。" }, { "location": "PDF p.14–15 / Appendix B", "claim": "对 Gemma 4 soft-cap，需在非线性变换之前恢复公共偏移的秩一项；Phi 线性头可将偏移折入权重而无需新推理操作。" }, { "location": "PDF p.15–16 / Appendix C.1 / Table 7", "claim": "独立未用于开发的 holdout 上，Phi AW-MSE KL 0.943→0.249、PPL 23.41→11.93；该评估路径的 BF16 PPL 是 9.33，与部署表的 11.65 分开解释。" }, { "location": "PDF p.16 / Appendix C.2 / Table 8", "claim": "冻结英文系数的 FLORES 短上下文评估：BLOOM 阿拉伯语 RTN KL 增加 67.3%，印地语 RTN 增加 13.4%、AW-MSE 增加 82.7%；这些结果不是翻译任务评测。" }, { "location": "PDF p.6 / Section 3.3; p.25 / Appendix G.2", "claim": "tied Phi 保留 BF16 输入 embedding，新增独立 packed 输出头，常驻模型权重约 7.17→7.47 GiB；输出头带宽下降可与常驻权重增加同时发生。" }, { "location": "PDF p.30–31 / Appendix J.1–J.2 / Tables 30–31", "claim": "分组扩展保持线性 softmax 等价性；九模型 scalar/grouped 候选数为 14/78，GPTQ 有五个改善和四个略退化，因此不能把额外收益全归因于参数空间。" }, { "location": "PDF p.32 / Appendix J.3", "claim": "独立 SmolLM3 匹配预算实验双方均 91 候选，16 篇此前未用文章；分组相对精细标量 KL RTN 0.8923→0.8008、AW-MSE 0.3404→0.2848；仅单模型探索，未测 packed 时延。" }, { "location": "PDF p.32 / Table 32", "claim": "九模型分组实验中 KL 改善不总伴随 PPL 改善，例如 Gemma 3 AW-MSE 的 PPL 38.086→38.176；不可混用其 BF16 参照与部署评估。" } ], "results": [ "主实验七个输出头、RTN/AW-MSE/GPTQ、W4 G128，decoder 固定 BF16。XGLM RTN 的 test KL 由 2.13 降至 0.143；Phi GPTQ 由 0.158 降至 0.059。Gemma 3/4 与 Qwen3.5 原本 W4 误差小，改善也小。另一个未用于开发的 holdout 上，Phi AW-MSE KL 0.943→0.249、PPL 23.41→11.93，其同路径 BF16 PPL 为 9.33。", "Phi-4-mini 的 vLLM/Marlin INT4 实际部署在相同 16,352 个 WikiText token 上：AW-MSE 加偏移的 KL 0.97→0.28、PPL 31.65→15.40；GPTQ 加偏移的 KL 0.15→0.06、PPL 13.67→12.23，同部署 BF16 PPL 11.65。", "A10G、64-token 贪心生成、含 prefill、关闭 prefix cache：AW-MSE 加偏移的 B1 延迟 1013.4 ms，对比 BF16 1136.6 ms 降低 10.8%；B16 为 1210.3 对比 1335.8 ms，降低 9.4%。加偏移前的 W4 AW-MSE 为 1014.8/1210.2 ms，表明偏移保留输出头压缩带来的提速。GPTQ 未测延迟。", "扩展实验按 G128 分组搜索：九模型 RTN/AW-MSE 的 KL 均低于标量搜索，但候选预算为 78 对 14，GPTQ 五模型改善、四模型略退化。另在 SmolLM3、双方各 91 个候选与 16 篇此前未使用的 WikiText 文章上，分组搜索相对精细标量搜索的 KL 为 RTN 0.8923→0.8008、AW-MSE 0.3404→0.2848；未测分组方法的 packed 延迟。" ], "reading_basis": "已阅读全文 32 页官方投稿 PDF，逐项核对正文 Sections 2–4、Tables 1–5，以及 Appendices A–J 的量化约定、非线性修正、独立 holdout、跨域/多语言转移、强 GPTQ 控制、部署时延与 tied-memory 账目、分组搜索与匹配预算实验。数值取自当前投稿原稿，不混用预印本或不同评估路径的 BF16 分母。" }, { "id": "lgAbT33ZED", "number": 15322, "title": "Controlling Symmetry Drift in Forward-Only Continual Learning", "abstract": "Edge devices need to adapt in the field despite limited access to backpropagation, making node perturbation an appealing alternative based on local activity, neuron-level perturbations, and scalar loss feedback. Yet persistent perturbation noise can drive parameter growth over long data streams and undermine a learner's ability to adapt to later tasks. We show that, to leading order in a frozen-state analysis, non-gradient perturbation noise produces a mean-zero diffusion along loss-invariant rescaling symmetries, whereas gradient updates have no first-order component along these directions.\nWe introduce gauge-damped node perturbation (\\ours{}), which periodically corrects symmetry-charge displacement through local reciprocal rescaling of each hidden neuron's incoming and outgoing weights. For positively homogeneous activations, each positive rescaling preserves the current function exactly, making the correction prediction-unbiased; the unclipped update contracts the targeted charge for an isolated layer pair, while the guarantee depends on the architecture. Across continual-learning streams, \\ours{} substantially controls charge displacement on long ReLU runs and can improve late-task accuracy; tanh experiments and simulated quantized writes further show that the benefit depends on the activation and write protocol.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Lifelong Learning", "Forward propagation", "Node perturbation" ], "forum_url": "https://openreview.net/forum?id=lgAbT33ZED", "pdf_url": "https://openreview.net/pdf/a12cdd9b2ac48c8d0a05463ba55e04b2b4b3c2cb.pdf", "updated_at": "2026-10-05T19:04:48.570Z", "categories": [ "training", "hardware", "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Mcwviwkd5X", "number": 15428, "title": "Reference Accuracy and Predictive Value in Low-Precision Training", "abstract": "Direct low-precision write-back can erase nonzero optimizer proposals. A high-precision reference can evaluate that event on a projected state, but aggregate agreement need not imply improvement over simpler controls. In a controlled grid, 55/72 cells have measured and predicted post-initialization crossings spanning $384\\times$; 52/55 lie within 15%, while four cells disagree on crossing category. A reused-source NeoX audit passes temporal skill against a constant, yet an outcome-blind secondary comparison favors a historical template (RMSE 0.00360) over the corrected source (0.00438). A prospective Pythia-410M checkpoint-restart stage evaluates 15 holdout targets across three ages. Its arithmetic-margin selector meets study-specific absolute error/coverage requirements on every target but has higher error than always predicting unchanged on the same accepted events. Every unit-gain source correction also loses to its template. Weak residual alignment limits even a separate hindsight-optimal scalar gain to less than 3.4% RMSE improvement per observed comparison, below the fixed 10% requirement; this is a post-outcome geometric result. The complete original protocol already rejects the forecast, and finalization includes a disclosed analysis-binding repair. Matched rounding interventions separately show large fixed-run policy contrasts. These empirical results characterize this reference construction and its controls, rather than a general failure of reference forecasting.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-precision training", "numerical precision", "reference-based prediction", "selective classification", "empirical evaluation" ], "forum_url": "https://openreview.net/forum?id=Mcwviwkd5X", "pdf_url": "https://openreview.net/pdf/e4b77512dd89084963d1531773117d582bb72a3b.pdf", "updated_at": "2026-10-05T19:05:44.381Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SfzHeH3XnN", "number": 15433, "title": "SEAOTTER: Sensor Embedded Autoencoding with One-Time Transcode for Efficient Reconstruction", "abstract": "In robotics, wearable, and remote-sensing systems, vast amounts of visual data are easily captured at high resolution using low-cost, low-power hardware. Yet, limited bandwidth and on-device compute resources prevent full utilization when transmitted via conventional codecs like JPEG/MPEG. Newer codecs, like AV1/AVIF, improve the rate-distortion trade-off, but demand far more resources for encoding, impractical without custom ASICs. Recent asymmetric autoencoders deliver high quality under extreme power and bandwidth constraints, but add prohibitive decoding cost and use bespoke formats that ignore decades of infrastructure built around standards like JPEG. To address these limitations, we introduce a split-inference compression framework based on a Sensor Embedded Autoencoder paired with a One-Time Transcode for Efficient Reconstruction (SEAOTTER). Because the sensor, cloud, and consumer stages face very different power and bandwidth budgets, SEAOTTER combines the compactness of a learned latent with the broad usability of a standard JPEG file. Since naive transcoding degrades performance, we propose a learnable JPEG color and quantization transform that enables increased accuracy for global, dense, and vision-language-based perception. Using SEAOTTER, we train both general-purpose and task-aware transcoding pipelines for a pre-trained, frozen encoder. At a compression ratio of 200:1 and compared to AVIF, we observe 7x faster encoding, 3.5x faster decoding, and +8% ImageNet top-1 accuracy, while retaining compatibility with JPEG infrastructure. Our code is available at URL REDACTED FOR ANONYMITY.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "image compression", "JPEG proxy", "representation learning", "autoencoding", "transcoding", "split computing", "cloud robotics" ], "forum_url": "https://openreview.net/forum?id=SfzHeH3XnN", "pdf_url": "https://openreview.net/pdf/4b70e09260edfee99d8c00e9e7d2c69239999e6f.pdf", "updated_at": "2026-10-05T19:05:45.842Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9LOd0o5L5x", "number": 15444, "title": "Compress Now, Contaminate Later: One Quantizer for Every Huber Level", "abstract": "In binary hypothesis testing with compressed observations, the encoder\nmay have to be fixed before anyone knows a bound on the contamination.\nWe study this setting where the decoder learns a common Huber contamination\nbound only after encoding, and can then choose its test but not undo the\nencoder's merges. For known finite laws, $O(1+\\log(1/J))$ output symbols\nsuffice to keep a constant fraction of robust Hellinger information at\nevery feasible level, where $J$ is comparable to clean squared Hellinger\ndistance. We also characterize the cost of using fewer symbols, which\ncan be steep. On a geometric family\nwith $L$ evidence scales, the best uniform sample overhead\n$\\mathcal R_K$ satisfies $\\log(1+\\mathcal R_K)\\asymp1+L/K$, even though\na bit selected separately for each level has constant overhead. An\nadjustable-growth version of our construction attains this order. The\nexponential penalty applies to one identical memoryless channel where a known\nschedule of different one-bit encoders has $O(L)$ overhead on the same\nfamily. Learning the encoder from clean observations is harder still. No\nfinite training budget depending only on clean difficulty preserves every\nraw-feasible level with a fixed output budget. A fitted test does succeed\nabove an estimation-error threshold, and independent calibration\ncertifies tests through simultaneous moment bounds. Experiments measure\nthe alphabet--information trade-off, compare designs that target clean\ninformation, and quantify what certification costs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "information-preserving quantization", "robust hypothesis testing", "Huber contamination", "communication-constrained inference" ], "forum_url": "https://openreview.net/forum?id=9LOd0o5L5x", "pdf_url": "https://openreview.net/pdf/c65536de9c81ca7b9c5eba6d26a7c4e36902f426.pdf", "updated_at": "2026-10-05T19:05:57.547Z", "categories": [ "ptq" ], "matched_terms": [ "binary", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ozl2QtZgxL", "number": 15459, "title": "CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches", "abstract": "Autoregressive video diffusion models maintain a growing attention key--value\n(KV) cache, making memory a key bottleneck for long video generation.\nRecently, clustering-based quantization has emerged as an effective method to\ncompress video KV caches by grouping similar key and value vectors and\nquantizing their residuals. However, when\nlater frames attend to the compressed history, small reconstruction errors can\ncompound as the video grows. Beyond the within-layer similarity exploited by\nclustering, we find strong dependencies across cache layers in video models.\nThis motivates Coral-KV: Calibrated\nOffline, Residual-coded, and Adapted\nLocally---a KV-cache compression framework that combines reusable\ncross-layer predictors with low-bit residual quantization.\nWe use reconstructed cache states to predict keys and values in subsequent\nlayers, then quantize the resulting prediction residuals.\nWe construct compact block-diagonal-plus-low-rank (block-DPLR) predictors:\nblock-diagonal components model channel relationships within each\nattention head, while a global low-rank term models dependencies between heads.\nCoral-KV exposes a flexible fidelity--memory trade-off: it can reuse its\npredictors across videos or refit lightweight block-diagonal predictors to the\ncurrent video for higher fidelity. \nAcross extensive evaluations on both\nHY-WorldPlay and MAGI-1, our reusable configuration\nimproves PSNR over QVG-4 by 0.32dB and 4.95dB,\nrespectively, while using 48\\% and 37\\% less cache memory.\nAgainst QVG-Pro, our video-adapted configuration\nimproves MAGI-1 PSNR by 3.77~dB while using 33.7\\% less cache memory.\nCoral-KV also maintains higher reconstruction fidelity over extended\ngeneration and reduces temporal discontinuities while retaining comparable\ngeneration latency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "video diffusion", "kv cache compression", "quantization", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=Ozl2QtZgxL", "pdf_url": "https://openreview.net/pdf/2543eb27a0ac95d90fb2513744abc6ab7c1b1dec.pdf", "updated_at": "2026-10-05T19:06:07.793Z", "categories": [ "ptq", "kv-cache", "hardware", "diffusion" ], "matched_terms": [ "low-bit", "quantization", "quantize", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W9gciOf15Z", "number": 15481, "title": "HyperHadamard: Layout-Aware Hadamard Transforms on Tensor Cores", "abstract": "Low-bit quantization is an appealing way to cut the memory and bandwidth cost of large language models. But activation outliers make few-bit quantization unstable, and rotation-based methods address them by mixing channels before quantization. The Hadamard transform is the rotation of choice because it admits a fast algorithm. Tensor-core FWHT kernels made these rotations cheap enough to deploy, but at large rotation dimensions existing kernels sustain only 42--77\\% of achievable memory bandwidth, even though the transform is memory-bound. We identify intermediate-layout materialization as a source of excess instructions and dependency stalls. We propose HyperHadamard, which treats butterfly stages as rotations along the binary axes of a hypercube and assigns those axes to tensor-core register layouts at compile time, so intermediate transposes become metadata rather than data movement. At rotation dimensions large enough for these overheads to dominate, HyperHadamard is up to 1.59$\\times$ faster than HadaCore, the tensor-core FWHT used in vLLM, and stays within 1.7\\% of achievable memory bandwidth. Together with a gather remap and a fused sign load, a true-FP32 instantiation of the same schedule speeds up a billion-dimensional randomized gradient projection used for training-data selection by 1.75$\\times$ while remaining bit-exact. The same design supports non-power-of-two transforms built from available catalog Hadamard matrices, whereas prior fast kernels expose only power-of-two and selected fixed-factor sizes. Integrated into vLLM, a Llama-3.1-8B QuIP W4A16 model produced by llm-compressor runs 14.7$\\times$ faster end to end than vLLM's dense GEMM path and 1.9$\\times$ faster than the same factorization assembled from Dao-AILab kernels and cuBLAS, with no measurable change in perplexity or zero-shot accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Hadamard Transform", "GPU kernel optimization", "tensor cores", "quantized LLM inference" ], "forum_url": "https://openreview.net/forum?id=W9gciOf15Z", "pdf_url": "https://openreview.net/pdf/c1a355df0c8d2f91434e0815620c009222f47cdc.pdf", "updated_at": "2026-10-05T19:06:15.392Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "binary", "low-bit", "quantization", "quantized", "w4a16" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2hbehF9EBG", "number": 15574, "title": "CAJQUANT: LARGE LANGUAGE MODEL QUANTIZATION VIA COMPENSATION-AWARE JOINT GRID SELECTION", "abstract": "Large language models (LLMs) impose substantial memory and inference costs, motivating weight-only post-training quantization (PTQ) for efficient deployment. However, preserving low-bit accuracy while keeping quantization overhead low remains challenging. In column-wise quantization with error compensation, the candidate grid determines the quantization errors in the current column, which drive compensation updates to weights in later columns. Static group-wise grid evaluation overlooks these candidate-dependent updates by scoring grids on the weights before intra-group compensation. Moreover, fixing the zero point during clipping search can overlook grids with smaller quantization errors for weights outside the clipping interval. To address these issues, we propose Compensation-Aware\nJoint Grid Selection for LLM Quantization (CAJQuant), which jointly selects the clipping ratio and zero point using compensation-aware grid evaluation. This evaluation measures reconstruction loss under optimal compensation while accounting for the weight updates induced by each candidate grid. Experiments on Llama models from 7B to 70B parameters demonstrate strong performance at 2, 3, and 4 bits, with substantial improvements in model quality at 2 bits. In this setting, CAJQuant outperforms all evaluated baselines in WikiText-2 perplexity and average downstream accuracy across all four models. On a single NVIDIA A100-80GB GPU, CAJQuant completes\n2-bit quantization of Llama-2-70B with a group size of 128 in approximately 5 hours.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "quantization" ], "forum_url": "https://openreview.net/forum?id=2hbehF9EBG", "pdf_url": "https://openreview.net/pdf/9409deb2753238b072fd5f9ef98372d2b8a1fdfb.pdf", "updated_at": "2026-10-05T19:07:20.287Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TH1X1ziwZc", "number": 15621, "title": "Density-Matched Power-of-Two Quantization: Allocating Logarithmic Levels for Multiplier-Free LLM Inference", "abstract": "Large language model (LLM) inference spends most of its compute in the multiply-accumulate (MAC) operations of linear projections.\nPost-training quantization (PTQ) reduces this cost with low-bit integer weights and activations, leaving a multiplier in every MAC.\nPower-of-Two (PoT) weight quantization replaces that multiplier with a shifter but spaces its levels a factor of two apart, degrading accuracy at low bit-widths.\nFiner fixed-ratio grids narrow the gap but remain suboptimal, since they spend the same number of levels per doubling of magnitude, wherever the weights lie.\nWe introduce **Density-Matched Power-of-Two (DMPoT)** quantization, which formulates logarithmic level allocation as a continuous point-density optimization problem.\nWe derive the density that minimizes quantization mean squared error (MSE) and use it to guide the construction of a multi-piece family of fractional PoT grids whose deployable members support folded shift-add hardware.\nThree results follow.\n(i) The derived density concentrates quantization resolution on the upper flank of the weight distribution while leaving the bottom of the range sparse.\n(ii) The optimal grids at successive bit-widths nest as sets of levels into one chain, and two bit-width laws fix the operating regime at four to five bits: a one-adder-per-weight bound and a width-minimality result at INT-matched MSE.\n(iii) Every deployed 4-bit level is an integer shift of $1$ or of $3$, so a folded datapath computes the single addition once per activation, shared by all output channels.\nAcross seven models and two configurations, DMPoT matches the uniform integer baseline in all fourteen cells, never more than $0.09$ perplexity above it, with mean gaps of $-0.008$ and $+0.038$.\nThe synthesized MAC array with 4-bit weights and activations occupies $0.702\\times$ the area of its integer counterpart.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models (LLMs)", "Post-Training Quantization (PTQ)", "Power-of-Two (PoT) Quantization", "Hardware-Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=TH1X1ziwZc", "pdf_url": "https://openreview.net/pdf/f3a1eb63431bcd1ec8229e8d8a335fc4c750c212.pdf", "updated_at": "2026-10-05T19:07:51.807Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "bit-width", "low bit", "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TsV05bN2qh", "number": 15726, "title": "Elastic-QVG: Efficient Long Video Generation with Elastic KV-Cache Quantization", "abstract": "Auto-regressive video generation uses a growing KV-cache to retain long-range context. Existing KV-cache quantization reduces the memory cost per token but still allows total memory usage to grow with context length, whereas eviction bounds memory by discarding cached context entirely. To bridge this gap, we present **Elastic-QVG, a training-free framework that progressively lowers cache precision without re-quantizing quantized caches.** A bit-level sliding window extends conventional token-level windowing to individual bit-planes, keeping recent chunks at higher precision while truncating older chunks to fit a fixed memory budget. Each chunk is quantized once, and its precision can subsequently decrease as new context arrives or the available memory budget shrinks. A truncation-exact quantizer enables hardware-friendly truncation directly on the quantized KV-cache, with no additional error relative to direct lower-bit quantization. Its nested quantization grids make dropping bit-planes exact, while separately allocated bit-plane storage and fused dequantization turn this property into memory savings without repacking the cache. Across LongCat-Video, HY-WorldPlay, and LingBot-World, Elastic-QVG improves fidelity over static QVG and retains VBench scores close to BF16 at 13–15× compression. It reduces KV memory by up to 12.7× with less than 5% latency overhead on H100, and avoids KV offloading on an RTX 5090 to achieve a 4.3× speedup over offloaded BF16-cache inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Elastic Quantization", "Video Generation", "KV Cache" ], "forum_url": "https://openreview.net/forum?id=TsV05bN2qh", "pdf_url": "https://openreview.net/pdf/1fd761c88762a2990bfcaf7c29cae6f180db407d.pdf", "updated_at": "2026-10-05T19:08:23.154Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization", "quantized", "quantizer", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FMWoUsitKJ", "number": 15787, "title": "HorizonKV: Three-State KV Cache Compression under a Shrinking Generation Horizon for Visual Autoregressive Models", "abstract": "Visual autoregressive (VAR) KV caches consume massive GPU memory—over 80 GB for Infinity-8B at 1024×1024 resolution. Existing compression relies on binary pruning: each KV state is either kept at full precision or discarded entirely. Our analysis shows that this forced choice is unnecessary: preserving all tokens at lower precision often proves more effective than discarding a fraction of them, demonstrating that fidelity reduction is not information deletion. This insight points to a different path: rather than deciding only whether to keep a state, we should also decide how faithfully to keep it as its future role evolves. To this end, we propose HorizonKV, which replaces binary decisions with a three-state lifecycle {Native (full precision), K8V8 (8-bit quantized), and Drop (released)} governed by the monotonically shrinking remaining generation horizon. Specifically, a small offline calibration first decouples future demand from compression sensitivity for each state. Afterwards, we allocate a global physical-byte budget via Lagrangian relaxation, enforcing irreversible $\\texttt{Native}\\rightarrow\\texttt{K8V8}\\rightarrow\\texttt{Drop}$ transitions as future responsibility decreases. Extensive experiments across Infinity-2B and Infinity-8B show that HorizonKV consistently improves the memory–quality frontier: at only 1% KV storage, it approaches full-cache visual quality and achieves roughly 4× additional compression over the state-of-the-art pruning baseline HeatKV at comparable fidelity, while preserving semantic generation quality on GenEval and DPG-Bench.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Visual Autoregressive Models", "KV Cache Compression", "KV Cache Quantization", "Multi-State Memory Management", "Horizon-Aware Scheduling", "Memory-Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=FMWoUsitKJ", "pdf_url": "https://openreview.net/pdf/f6a236bc95fbb9c3f915dc8c86377931add24b90.pdf", "updated_at": "2026-10-05T19:09:04.076Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "8-bit", "binary", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PdHbXU64LT", "number": 15793, "title": "ReQuant: Fixed-Grid Discrete Refinement for Post-Training Quantization", "abstract": "Post-training quantization (PTQ) is widely used to reduce the memory and computational cost of large language models. Existing PTQ methods typically initialize quantized models with heuristic rules or greedy optimization, and usually treat the resulting integer assignments as final. We introduce ReQuant, a backpropagation-free fixed-grid refinement stage *within* PTQ that optimizes an executable quantized model while preserving its format. Agnostic to the PTQ initializer, ReQuant starts from an existing feasible quantized model and iteratively revisits its discrete weight assignments on the fixed quantization grid. Accepted updates strictly reduce the mean squared reconstruction error and remain on the original grid. ReQuant thus integrates with existing PTQ pipelines as a plug-and-play post-processing stage. Experiments across diverse model families, bit-widths, and downstream tasks show that ReQuant consistently improves quantized models from heterogeneous PTQ initializers, with especially large gains on simple initializers and lower bit-widths. Notably, ReQuant can refine a simple round-to-nearest initialization across multiple sweeps until it approaches or surpasses GPTAQ under the same quantization format.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Large language models", "Discrete optimization" ], "forum_url": "https://openreview.net/forum?id=PdHbXU64LT", "pdf_url": "https://openreview.net/pdf/ea04e185d8922a30ed57ae853a71db3b17d899c7.pdf", "updated_at": "2026-10-05T19:09:03.356Z", "categories": [ "ptq" ], "matched_terms": [ "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ReOTai8HD3", "number": 15829, "title": "DynamicKV: Predictive and Reversible KV-Cache Precision Allocation for Agentic LLM Inference", "abstract": "Agent workflows and long-chain reasoning continually retain and access growing histories, making the KV cache a major GPU memory cost during inference. Existing KV quantization methods typically determine precision when historical KV states enter the cache or reduce bit widths unidirectionally during generation. Analysis of real generation traces shows that earlier KV states can be reused after a period of limited attention; moreover, the effect of demoting the same KV region on subsequent inference can increase or decrease across generation stages. To anticipate changes in future demand and adapt quantization under a limited budget, we introduce DynamicKV, a predictive and reversible KV precision allocation method. DynamicKV partitions historical KV states into regions aligned with underlying storage blocks and periodically estimates two signals: $\\pi$, a score for significant next-interval attention, and $d$, a score for significant damage from demotion to low-bit precision. Their product, $\\pi d$, serves as a gated ranking score for dynamic precision allocation under a fixed proportional GPU KV-byte budget. Each region retains a low-bit base on the GPU and a separable residual in host memory. When the region becomes important again, activating the residual restores an approximately high-bit representation, enabling repeated demotion and restoration. On agent workflow tasks ($\\tau^2$-bench and BFCL V4), DynamicKV improves success rates by 5.0–10.5 percentage points over the strongest compression baseline for each model. On AIME 2024/2025 and MATH-500, it improves accuracy by up to 3.3 percentage points over the strongest compression baselines. On a single A100-80GB GPU, DynamicKV supports $2.5\\times$ the concurrency of BF16 at 32K and improves aggregate output throughput by 4.7–44.3% at sequence lengths of 16K–32K.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization; mixed precision; long-chain reasoning; LLM agents; predictive control" ], "forum_url": "https://openreview.net/forum?id=ReOTai8HD3", "pdf_url": "https://openreview.net/pdf/2eac300027846a5740c3b23bc5f97db59914f4e4.pdf", "updated_at": "2026-10-05T19:09:27.132Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "low-bit", "mixed precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tyflex1iqI", "number": 15832, "title": "Discovering Parameterized Motor Skills for Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models have shown strong potential for general-purpose robotic manipulation, yet their learned behaviors often remain tied to specific task contexts.\nA key observation is that semantically different manipulation tasks can share recurring short-horizon motor behaviors.\nInspired by unsupervised skill discovery in reinforcement learning (RL), we formulate reusable behavior in VLA as discovering, parameterizing, and planning reusable motor skills.\nWe introduce \\method{}, a framework that discovers recurring local motion patterns from multi-task demonstrations and converts them into detachable parameterized skill modules around a frozen task-adapted VLA.\nSpecifically, a factorized vector-quantized variational autoencoder identifies recurring motion patterns, and each discovered skill is encoded as a low-rank weight update around the frozen supervised fine-tuning (SFT) VLA policy. \nThis parameterization provides compact and detachable skill modules that can be selectively activated while preserving the frozen base policy and its native action interface.\nDuring execution, a vision-language router performs skill planning\nfrom the current observation and language instruction by choosing\neither a parameterized skill or the base policy.\nExperiments on LIBERO and RoboTwin 2.0 show consistent improvements over SFT references, while cross-task skill reuse and real-world evaluation demonstrate that learned skills can transfer across task contexts.\nOur results point toward a VLA paradigm in which increasingly capable high-level reasoning can leverage reusable motor skills, rather than relearning low-level behaviors for each new task. Code is available at https://anonymous.4open.science/r/0D2F.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision–Language–Action Models; Skill Discovery; Low-Rank Adaptation" ], "forum_url": "https://openreview.net/forum?id=tyflex1iqI", "pdf_url": "https://openreview.net/pdf/0e8e2595e44cf2027510f54e0da1a86222c76790.pdf", "updated_at": "2026-10-05T19:09:34.386Z", "categories": [ "reasoning", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Dr5z2dsaJq", "number": 15846, "title": "HHR: Hierarchical Hash Retrieval for Efficient LLM Generation", "abstract": "Efficient long-context inference is essential for large language models (LLMs), yet it poses a severe computational bottleneck. Hash-based retrieval offers an efficient alternative by encoding queries and keys into binary codes and using Hamming distance for key selection. However, this leads to a critical mismatch between Hamming distance and attention relevance. Query-Key logits depend jointly on\ndirectional similarity and feature magnitudes, whereas hash binarization discards magnitude information, causing both false-positive retrieval of low-logit keys and false-negative omission of high-logit keys. To address these failures, we propose Hierarchical Hash Retrieval (HHR), a coarse-to-fine framework that progressively improves retrieval accuracy through Geometry-Aware Key Routing (GKR) and Learned Hash Projection (LHP). GKR learns a head-wise orthogonal transformation to redistribute feature magnitudes and derive more discriminative page-level logit bounds, enabling effective pruning of low-logit keys while preserving important candidates. LHP then learns a head-wise projection space that aligns Hamming distance with the true Query-Key relevance ranking for fine-grained retrieval. By combining GKR and LHP, HHR suppresses false positives and recovers false negatives, substantially improving the fidelity of hash-based\nsparse attention. Extensive experiments across diverse LLMs and benchmarks demonstrate that HHR achieves superior performance over existing methods. For example, on LongBench, HHR improves the average score by 1.10 points and, at a context length of 128K, achieves up to a 3.30× decoding speedup and a 2.83× end-to-end speedup for Llama-3.1-8B-Instruct.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Hash", "Sparse Attention" ], "forum_url": "https://openreview.net/forum?id=Dr5z2dsaJq", "pdf_url": "https://openreview.net/pdf/d8f5c428da525423120d557ad8445b6353ec45ca.pdf", "updated_at": "2026-10-05T19:09:42.045Z", "categories": [], "matched_terms": [ "binarization", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ttnaG7TM9A", "number": 15861, "title": "MiST: Memory-inspired Token-Efficient Spiking Transformer", "abstract": "Transformer-based Spiking Neural Networks (SNNs) combine the high performance of Transformers with the energy efficiency of SNNs, offering a promising paradigm for intelligent edge computing. However, they remain bottlenecked by the token-intensive processing inherited from conventional vision transformers, which hinders their deployment on resource-constrained edge devices. Inspired by the key-value memory mechanism of the human brain, we propose MiST, a \\underline{m}emory-\\underline{i}nspired token-efficient \\underline{s}piking \\underline{t}ransformer that exploits the inherent efficiency of transformer-based SNNs while maintaining competitive performance. Specifically, MiST decouples the full feature map into a memory repository and a small set of propagated query tokens, enabling forward inference to operate on only a fraction of the original tokens. The core component in MiST is the bidirectional spike-driven cross-attention (Bi-SDCA) module, which comprises two pathways: a forward pathway that retrieves semantic content from the repository into query tokens, and an inverse pathway that propagates enriched token representations back to update the repository. Furthermore, we observe that the memory repository in MiST exhibits high semantic disorder during early training, which slows convergence. To address this, we propose a semantic-guided self-distillation (SGSD) strategy that drives the repository to form compact semantic anchors. Extensive experiments on image classification and semantic segmentation show that MiST achieves accuracy comparable to token-intensive baselines using only 20$\\sim$40\\% of tokens, establishing a new efficiency frontier for spiking transformer architectures.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "spiking quantization; spiking pruning; spiking token pruning" ], "forum_url": "https://openreview.net/forum?id=ttnaG7TM9A", "pdf_url": "https://openreview.net/pdf/f092b503448322308b6ed0150a4f8224c7b8c232.pdf", "updated_at": "2026-10-05T19:09:54.706Z", "categories": [ "kv-cache", "hardware", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HaN5beKJR3", "number": 15864, "title": "GLF-Q: Global-Local Feature-based Quantization for Vision Transformers", "abstract": "Post-training quantization (PTQ) efficiently compresses Vision Transformers (ViTs) without retraining, yet suffers severe accuracy degradation at low bit-widths. Existing optimization-based PTQ methods guide block reconstruction via either soft logits or second-order Hessian proxies. Logit supervision is prone to overfitting on limited calibration data, while Hessian approximations incur structural truncation errors. To address these limitations, we propose \\textbf{GLF-Q}, a novel PTQ framework guided by Global-Local Feature alignment. GLF-Q propagates quantized block outputs through downstream full-precision layers to align penultimate-layer representations under local output regularization, providing downstream feature supervision without explicitly approximating the Hessian or using a Taylor expansion. Furthermore, offline Hadamard transformations are introduced with zero runtime overhead to disperse activation outliers across channels, effectively contracting dynamic ranges and reducing quantization errors. Meanwhile, optimizing this loss via a Straight-Through Estimator (STE) achieves rapid convergence, bypassing continuous relaxation rounding formulations such as AdaRound. Extensive experiments across representative ViT architectures demonstrate that GLF-Q with standard uniform quantizers substantially outperforms state-of-the-art methods under 3-bit quantization on image classification. In addition, GLF-Q exhibits strong out-of-domain calibration robustness and achieves speedups under 8-bit GPU deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Compression", "Post-training quantization", "Vision Transformer" ], "forum_url": "https://openreview.net/forum?id=HaN5beKJR3", "pdf_url": "https://openreview.net/pdf/653f4ad50d9a088ad449ddb3f730d2930190af39.pdf", "updated_at": "2026-10-05T19:09:52.364Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "3-bit", "8-bit", "low bit", "ptq", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QwpFlHjX2g", "number": 15947, "title": "PACE: Precision Allocation with Cost Estimation for Mixed-Precision Quantization", "abstract": "Mixed-precision quantization reduces model storage by\nassigning different bitwidths under a fixed budget. Its effectiveness depends\non accurately estimating which precision decisions best preserve model\nquality. Existing methods often rely on indirect sensitivity measures that do\nnot reflect the actual output damage caused by different quantization\ncandidates. We introduce PACE (Precision Allocation with Cost\nEstimation), a second-order\nmixed-precision quantization method grounded in a stationary output-fidelity\nobjective. PACE efficiently predicts the output damage of concrete precision\ncandidates and globally allocates the available bit budget according to their\nestimated benefits. Experiments validate the consistency between predicted\nand measured quantization damage and compare PACE across three 7--8B model\nfamilies, two perplexity benchmarks, six zero-shot tasks, and average 6-bit\nand 3-bit settings. Under the average 6-bit setting, PACE ranks first or second\nacross all nine evaluations, with WikiText2 perplexity increases of at most\n0.095, while remaining robust under the more challenging average\n3-bit setting. On Qwen3-8B, PACE is approximately $490\\times$ faster than\nSFMP, the strongest comparable-quality baseline. Its packed representations\nprovide approximately $2.6\\times$ compression under W6 and more than\n$4.6\\times$ compression under W3.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "mixed precision", "post-training quantization", "large language models", "memory-efficient ML" ], "forum_url": "https://openreview.net/forum?id=QwpFlHjX2g", "pdf_url": "https://openreview.net/pdf/f1d07a0260c674279e317719ad8fab41d79bedc6.pdf", "updated_at": "2026-10-05T19:10:34.873Z", "categories": [ "ptq" ], "matched_terms": [ "3-bit", "6-bit", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aDeYmEcvv2", "number": 15982, "title": "E-PMQ: Expert-Guided Post-Merge Quantization with Merged-Weight Anchoring", "abstract": "Low-resource deployment constraints have made model quantization essential for\ndeploying neural networks while preserving performance. Meanwhile, model\nmerging has become an increasingly practical low-resource strategy for\nintegrating multiple task- or domain-specialized experts into a single\nmodel without joint training or multi-model serving. Together, quantization and model merging enable an efficient low-resource\ndeployment pipeline by integrating multiple experts into one low-bit model. We formulate this setting as Post-Merge Quantization (PMQ).\nWe show that directly applying post-training quantization (PTQ) to a merged\nmodel is unreliable because two distinct deviations are coupled: the\nquantization deviation introduced by low-bit reconstruction and the\nexpert-relative merging deviation inherited from model merging. To\nmitigate these deviations, we propose E-PMQ, an expert-guided PMQ\nframework that uses source expert weights to provide expert-guided output\ntargets during layer-wise calibration, together with merged-weight anchoring to\nstabilize the calibration and preserve the integrated behavior of the merged\nmodel. Experiments on CLIP, FLAN-T5, and Llama show consistent gains\nover naive PMQ across vision and language, from 8-task CLIP-ViT-B/32 to\n20-task CLIP-ViT-L/14 and up to 8B-scale LLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "model merge", "CLIP", "LLM" ], "forum_url": "https://openreview.net/forum?id=aDeYmEcvv2", "pdf_url": "https://openreview.net/pdf/5cf9e4aa4783cdeeab0c17209f554bbfa857c8c4.pdf", "updated_at": "2026-10-05T19:10:52.301Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "d8YA40OtHW", "number": 16102, "title": "Tokens, Numbers, or Words? Benchmarking How Large Language Models Express Uncertainty", "abstract": "As large language models (LLMs) are increasingly deployed in high-stakes applications, robust uncertainty estimation is essential for ensuring the safe and trustworthy deployment of LLMs. We present a large-scale study of uncertainty estimation in LLMs, evaluating 80 models spanning open- and closed-source families, dense and Mixture-of-Experts (MoE) architectures, reasoning and non-reasoning modes, quantization variants and parameter scales from 0.6B to 671B. Focusing on three representative single-pass methods that require no access to hidden states, including token probability-based uncertainty (TPU), numerical verbal uncertainty (NVU), and linguistic verbal uncertainty (LVU), we systematically evaluate uncertainty calibration and selective classification using the challenging MMLU-Pro benchmark, which covers both reasoning-intensive and knowledge-based tasks. Our results show that, on MMLU-Pro, LVU outperforms TPU and NVU on average and for most models, offering stronger calibration and discrimination. We also find that high accuracy does not imply reliable uncertainty, and that model scale, post-training and reasoning ability influence estimation performance, whereas the effects of quantization are small and method-dependent. Notably, LLMs exhibit better uncertainty estimates on reasoning tasks than on knowledge-heavy ones, and calibration and error ranking are only partially aligned. These findings highlight the need for multi-perspective evaluation and position LVU as a promising, though not universally superior, tool for improving the reliability of LLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "uncertainty estimation", "calibration", "LLM" ], "forum_url": "https://openreview.net/forum?id=d8YA40OtHW", "pdf_url": "https://openreview.net/pdf/b26d5be3d7ea7541758ff07f5ad12098ed19f68d.pdf", "updated_at": "2026-10-05T19:11:53.065Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gcrBHtTrPN", "number": 16127, "title": "Exact, Reversible, and Verifiable Lifecycle Control for Routed Parametric Memory", "abstract": "Deleting, correcting, or reverting what a deployed language model knows is today an optimisation problem on entangled weights. Unlearning objectives and model editors move parameters until a behavioural probe is satisfied. The result is approximate, depends on the order of operations, and can be verified by no one but the operator. The problem disappears when knowledge is placed rather than optimised. Block-Rollback stores injected knowledge in the content-routed memory blocks of a hierarchical-memory language model whose backbone, the anchor, stays frozen. Every block is a deterministic function of the documents it holds and a public seed. Insert, edit, delete, and revert therefore reduce to rebuilding one block from an append-only log, and each operation lands element-wise on the counterfactual model: the one that never saw the deleted fact, only ever saw the corrected fact, or stopped at the reverted step. We test this on five public anchors from GPT-2-XL to Llama-2-7B and on two anchors pretrained from scratch with the memory pathway, across the KnowEdit, TOFU, and MUSE benchmarks. More than 17000 locality checks find no violation, and every rebuilt block matches its counterfactual exactly. On the same base, gradient editors halve their locality within ten edits and cannot revert, and optimisation-based unlearners remain separable from the counterfactual by membership, extraction, and jailbreak probes. Block-Rollback matches it on every probe, including after int8 and NF4 quantisation. A third-party certificate rejects 10000 forged deletions at the cost of rebuilding one block. Per-operation cost tracks the touched block rather than the corpus or the record's age. A read-only routing gate keeps general capability at the anchor's values through 300 edits.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "parametric memory", "knowledge editing" ], "forum_url": "https://openreview.net/forum?id=gcrBHtTrPN", "pdf_url": "https://openreview.net/pdf/eb14e07c7124907cb1067de7abdc2ba97152baba.pdf", "updated_at": "2026-10-05T19:11:55.690Z", "categories": [ "training" ], "matched_terms": [ "int8", "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yY5TDfUS8V", "number": 16178, "title": "ViGuardQ: Quantized Video Deepfake Detection for Edge Devices", "abstract": "The growing use of edge devices, e.g. smartphones, for video conferencing and facial biometric authentication creates an increasing demand for efficient video deepfake detection. However, the substantial computational and memory requirements of existing detectors hinder their deployment on resource-constrained devices. Although quantization offers an effective means of model compression, conventional methods do not explicitly account for the subtle spatial artifacts and temporal inconsistencies essential to video deepfake detection, which can be distorted by low-precision representations. To address this challenge, we propose ViGuardQ, a post-training quantization framework tailored to video deepfake detection on edge devices. First, Dual-Evidence-Constrained (DEC) Rotation mitigates forgery-evidence relevant quantization distortion by identifying a sensitive channel subspace using within-frame and cross-frame differential decision gradients, and optimizing orthogonal rotations to preserve more forgery-evidence sensitive information. Then Evidence-guided Consistent Response (ECR) Calibration refines the quantizer's consistency in response to evidence under aligned and opposite perturbations. They together improve the preservation of and responses to subtle forgery evidence and greatly improve quantization accuracy. Extensive experiments on six benchmark datasets and eleven deepfake detectors demonstrate that ViGuardQ outperforms state-of-the-art quantization baselines. Specifically, we achieve 83.59% AUC scores at a lower 13.4% compression rate, compared with 68.07\\% achieved by the strongest baseline, PTQ4VM, at a 14.95% compression rate on the DFD dataset. Real-world deployment on edge devices further demonstrates our practical effectiveness. Our code will be released upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Deepfake Detection; Quantization; Edge Deployment" ], "forum_url": "https://openreview.net/forum?id=yY5TDfUS8V", "pdf_url": "https://openreview.net/pdf/bddd32100bc9650b07b243f02856526d3035c368.pdf", "updated_at": "2026-10-05T19:12:21.335Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "low-precision", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ptNmbmCdj3", "number": 16181, "title": "Spiking, Fast and Slow: The Anatomy of Loss Spikes in Language Model Pretraining", "abstract": "Loss spikes in language model pretraining can waste substantial compute and derail otherwise healthy runs. A primary driver of loss spikes is the growth and divergence of output logits in the final language modeling (LM) head. In this work, we divide these training instabilities into two types: fast spikes that abruptly increase loss before recovering, and slow terminal divergences that emerge only after prolonged training, which we call time-bomb spikes. We build on observations that these instabilities stem from fundamental properties of the softmax cross-entropy loss and its interaction with Adam’s preconditioner. We observe that this interaction leads to LM head weight growth that may be harmless under exact arithmetic but can be destabilizing under finite-precision training and propose a simple optimization adjustment that ameliorates the issue. Furthermore, we show that various mitigation techniques from the literature (including our own) largely act on the shared underlying problem of logit growth by reducing quantization error in the LM head. Our work connects fast and time-bomb spikes and an array of prior methodological proposals through a common lens of analysis: controlling last-layer logit growth reduces finite-precision rounding error which is essential for stabilizing long pretraining runs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "loss spikes", "LLM pretraining", "training instabilities", "long horizon training" ], "forum_url": "https://openreview.net/forum?id=ptNmbmCdj3", "pdf_url": "https://openreview.net/pdf/f67e4dca06fc75a3ecbd8a4f1763826a63056b7f.pdf", "updated_at": "2026-10-05T19:12:23.488Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "js21L1V6ap", "number": 16357, "title": "Value Channels Gaussianize Rounding Noise: Tensor Geometry of KV-Cache Quantization", "abstract": "A quantized key-value cache is written once and read by many future, possibly adaptive, queries. We show that, within a factor of two, the uniform read error is governed by one invariant: the expected injective norm of a random sign tensor that pairs the attention family with the Euclidean value space. Three reductions isolate it. Convex order removes the encoder: among conditionally unbiased encoders on fixed grids, independent stochastic rounding is optimal for every convex loss. Duality replaces the query family by its convex hull, and symmetry replaces the bit budget by its orbits. A structure theorem then reduces this invariant, for every attention family and up to absolute constants, to two scalar quantities. It is a Bernoulli counterpart of Chevet's inequality for a Euclidean factor: the error is the worst fixed-query fluctuation, amplified by $\\sqrt{d}$ in value dimension $d$, plus a flat-direction width in which the value channels Gaussianize the bounded rounding noise. When each token's values share one grid, Bernoulli and Gaussian noise models therefore differ by a factor of order at most $\\max\\\\lbrace 1,(\\log n/d)^{1/2}\\\\rbrace$ on $n$ tokens, and softmax attention with logarithmic-dimensional keys attains this factor in every dimension. Across independent token groups, fluctuations add in $\\ell_2$ and selections in $\\ell_1$. Optimal precision is a convex program, rounded within a factor of eight, whose optimum equalizes the contributions to the worst-case error of all coordinates stored above the minimum precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "stochastic rounding", "uniform error", "Bernoulli processes", "injective tensor norm", "Chevet inequality", "precision allocation" ], "forum_url": "https://openreview.net/forum?id=js21L1V6ap", "pdf_url": "https://openreview.net/pdf/0795691a71e25bebdf0d497e845b89c58f309c74.pdf", "updated_at": "2026-10-05T19:13:41.692Z", "categories": [ "ptq", "kv-cache", "theory" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YZqOG55KLK", "number": 16525, "title": "PepSAGE: Hierarchical All-Atom Peptide Generation with Continuous Local-Geometry States", "abstract": "Peptides can engage protein surfaces that are difficult to target with small molecules, making pocket-conditioned peptide design a valuable yet challenging therapeutic goal. Generating such peptides at all-atom resolution requires the joint modeling of global backbone pose, amino-acid identity, and residue-internal atomic conformation. Existing representations face a persistent tension between expressive local detail and geometric validity: Cartesian coordinates expose rigid-body redundancy, torsion variables are chemically grounded but less convenient as a uniform learned interface across residue types, and discrete structural tokens may discard fine conformational variation. We introduce a geometry-state generation formulation for pocket-conditioned all-atom peptide design, instantiated as PepSAGE (Peptide Structure-Aware Generation). Rather than directly generating unconstrained atom coordinates, relying on residue-specific torsions, or discretizing local structure into tokens, PepSAGE represents residue-internal conformation with a shared six-dimensional\ncontinuous state ($z \\in \\mathbb{R}^6$), while modeling residue pose and amino-acid identity through separate channels. This state is learned through a reconstruction interface that is frozen during generation, allowing PepSAGE to generate continuous pre-quantization local geometry and recover all-atom structures through a geometry-constrained reconstruction pathway. Under a unified training budget and leakage-controlled evaluation, PepSAGE achieves the best backbone, all-atom, and side-chain RMSD, the lowest clash, and the highest rotamer recovery among competitive baselines, with additional analyses showing a tunable quality--diversity trade-off. Controlled ablations show that the continuous local-geometry channel drives side-chain recovery, while geometry-constrained reconstruction improves geometric validity, supporting continuous local geometry as an effective internal state for structured all-atom peptide generation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Deep Generative Models", "Biomolecular Design", "Representation Learning" ], "forum_url": "https://openreview.net/forum?id=YZqOG55KLK", "pdf_url": "https://openreview.net/pdf/1bee481f5e92135d271a14d509586596637bf0ea.pdf", "updated_at": "2026-10-05T19:14:57.767Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UuAAvOn17o", "number": 16568, "title": "Attend to Your Own Thoughts: Calibration Matters for 1.58-Bit Reasoning LLMs", "abstract": "In this paper, we present ScaleQ-1.58, a ternary post-training quantization (PTQ) framework for reasoning LLMs. Its core insight stems from an empirical finding: although modern LLMs are typically trained to have chain-of-thought reasoning capabilities, in the PTQ regime, even recently proposed differentiable ternarization still leads to performance collapse on challenging mathematics and coding tasks. We reveal that this is caused by conventional calibration schemes that ignore the model's reasoning process. Motivated by this, we present Attend to Your Own Thoughts (AYOT), a new and simple calibration method, which constitutes the core contribution of this paper. Given a pre-trained target LLM, AYOT uses its self-generated reasoning traces and final answers on a proper set of calibration samples, together with the corresponding questions, as context inputs during ternarization. ScaleQ-1.58 is built by combining AYOT with differentiable ternarization, and demonstrates several favorable scaling properties. With only 4M calibration tokens: (1) Qwen3-1.7B ternarized by ScaleQ-1.58 reaches over 90.52% of the performance of the prior best BitNet b1.58 2B4T averaged over 4 mathematics and coding tasks, and our ternary Qwen3-4B shows an absolute gain of 8.97%, while requiring 1,000,000$\\times$ fewer calibration tokens for ternarization; (2) ScaleQ-1.58 generalizes well to 8 dense and MoE architectures, with performance significantly improving as model scale increases from 1.7B to 235B parameters; (3) ScaleQ-1.58 exhibits strong generalization across 13 tasks of varying difficulty levels, including challenging mathematics, coding and scientific logic reasoning, as well as basic commonsense reasoning and language generation. Intriguingly, ScaleQ-1.58 shows continuously improving performance as the number of calibration tokens increases. Besides, AYOT generalizes well beyond 1.58-bit quantization to other bit-width settings. With the default 4M calibration tokens, ScaleQ-1.58 requires only 4 to 240 hours on 8 A100-80GB GPUs to ternarize 1.7B to 235B reasoning LLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "post-training quantization", "quantization-aware training", "ternary quantization", "reasoning tasks" ], "forum_url": "https://openreview.net/forum?id=UuAAvOn17o", "pdf_url": "https://openreview.net/pdf/72c255bce33869ef285cdd417161a247e5f65236.pdf", "updated_at": "2026-10-05T19:15:15.767Z", "categories": [ "ptq", "reasoning", "training", "hardware" ], "matched_terms": [ "1.58-bit", "b1.58", "bit-width", "ptq", "quantization", "ternary" ], "review_status": "reviewed", "summary": "普通文本校准容易丢失三值模型的数学和代码推理能力。AYOT 从领域问题取样，让待量化的高精度目标模型生成自己的思维链与答案，把完整上下文交给 CAT-Q 的可微三值重构和跨层滑窗优化。默认数学/代码各 1024 条、每条 2048 tokens，约 4M 校准集 tokens，重复优化 60 epochs；推理权重只保留三值与组尺度。", "contribution": "把领域覆盖和目标模型自身轨迹作为校准的两个关键变量，直接比较无思维链、强教师思维链与自身思维链。官方原稿还在 AWQ、GPTQ、SliderQuant 的 2/3/4-bit 设置测试迁移，并分析校准预算、上下文长度和领域混合的取舍。", "limitations": "自身轨迹校准仍与 FP16 存在较大差距，数学/代码提升可能伴随语言建模或其他领域退化；极低位的部分 AWQ/GPTQ 结果仍崩溃。作者明确校准量及耗时高于常规 GPTQ/AWQ。4M 校准与 BitNet 4T 从头训练的百万倍 token 比较不代表总训练成本等价节省，预训练及轨迹生成也需成本。真实吞吐表缺硬件型号、输入长度与重复误差，不能把 8×A100 校准服务器当作部署测试硬件；代码/模型声明将在发表后发布。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "feb665bfde42e98d05f430551f96d693285c7fbcd65a51b0785bae36179dfe84", "reviewed_pdf_url": "https://openreview.net/pdf/72c255bce33869ef285cdd417161a247e5f65236.pdf", "reviewed_source_updated_at": "2026-10-05T19:15:15.767Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "预训练推理 LLM 的权重三值 PTQ；激活为 16 bit，MoE router 保留全精度。AYOT 是校准策略，ScaleQ-1.58 是 AYOT 与 CAT-Q 的组合。", "bit_width": "主设置 W1.58A16，三值 {-1,0,+1} 加组尺度，group 128；另把 AYOT 用于 AWQ/GPTQ/SliderQuant 的 W2A16、W3A16、W4A16。部署表的 CPU/GPU 格式分别为 TQ1_0/TQ2_0。", "evidence": [ { "location": "PDF p.1，Title；p.4–5，§2.1–2.2", "claim": "原投稿副标题是 Calibration Matters for 1.58-Bit Reasoning LLMs；AYOT 使用目标模型自己生成的思维链与答案，并组合 CAT-Q。" }, { "location": "PDF p.5，§2.2；p.7，Table 4", "claim": "同一 Qwen3-4B 校准比较的五任务均值：领域问答 17.43%、强教师轨迹 20.06%、自身轨迹 45.60%。" }, { "location": "PDF p.6，Table 1；p.7，Table 2", "claim": "Qwen3-4B 的 Math-500 FP16 为 96.80%，W1.58A16、4M 校准为 58.40%，仍有明显精度差距。" }, { "location": "PDF p.8，Table 8", "claim": "GPTQ W4A16 Math-500 91.80→94.80%，C4 PPL 21.95→22.62；W2 AWQ/GPTQ 的 Math-500 都为 0，迁移收益不是无条件成立。" }, { "location": "PDF p.16–17，Appendix C.1–C.2", "claim": "默认 2048 样本×2048 tokens，数学/代码各半，60 epochs、batch 3；group 128，MoE router 保持全精度。" }, { "location": "PDF p.15，Appendix A；p.18，Appendix E / Table E", "claim": "作者承认比常规 GPTQ/AWQ 的校准量更大且更慢；8×A100-80GB 上 Qwen3-4B 三值化为 8 小时。" }, { "location": "PDF p.18，Appendix F / Table F", "claim": "真实 llama.cpp 部署的 CPU TQ1_0 为 37.35、GPU TQ2_0 为 320.34 tokens/s；batch 1、生成 512 tokens；未给出部署硬件型号。" }, { "location": "PDF p.3，Figure 1；p.10，Reproducibility Statement", "claim": "4M tokens 用于已预训练模型的三值化，而 BitNet 对照为 4T 从头训练；原稿声明代码和模型将在发表后发布。" } ], "results": [ "§2.2 / Table 4：Qwen3-4B、W1.58A16、约 4M 校准 tokens，Math-500/GSM8K/Omni-MATH/HumanEval+/MBPP+ 五项均值，领域 QA 17.43%，DeepSeek-R1-671B 轨迹 20.06%，自身轨迹 45.60%；Math-500 的三值结果 58.40%，FP16 96.80%。", "Table 8：迁移并非所有指标都改善。例如 GPTQ W4A16 的 Math-500 91.80%→94.80%，C4 PPL 却由 21.95→22.62（越低越好）；W2 AWQ/GPTQ 的 Math-500 在两种校准下都为 0。", "Appendix E / Table E：8×A100-80GB、约 4M 校准集 tokens、60 epochs，三值化 Qwen3-4B 需 8 小时；1.7B–235B 模型需 4–240 小时。4M 是校准集大小，不能等同只处理一次 4M tokens，也不等同从零预训练成本。", "Appendix F / Table F：真实 llama.cpp 的 Qwen3-4B、batch 1、生成 512 tokens，CPU TQ1_0 为 1.01 GB / 37.35 tokens/s，GPU TQ2_0 为 1.17 GB / 320.34 tokens/s；Q4_K_M 对照分别为 2.32 GB / 19.31 和 2.32 GB / 225.91。论文未给出该部署表的 CPU/GPU 型号。" ], "reading_basis": "实际阅读本次取得的 27 页官方投稿 PDF，首页确认短副标题：正文 §2–3、Tables 1–8、Appendices A/C/D 的限制与实现、E/F 的成本和部署、G/H 的长度统计与八个定性例子。直接查看 PDF p.7/8/18 图像核对校准消融、跨位宽与吞吐表；本笔记仅依据官方原稿，数字为作者报告，未独立复现。" }, { "id": "7Y0QIJSohA", "number": 16589, "title": "Relation Matters: Learning Local Relations for Binary Infrared-Visible Image Fusion", "abstract": "Infrared--Visible Image Fusion (IVIF) integrates thermal targets with spatial details, but full-precision models require substantial computation and storage. Binarization reduces these costs, yet binary IVIF faces three connected challenges: binary representation, where binarization can obscure local structural differences; feature propagation, where binary processing can degrade the magnitude information needed for reconstruction; and capacity-compatible training, where the binary network must learn complementary fusion patterns within its limited representational capacity. To address these challenges, we propose SDRTNet, a binary IVIF framework that jointly addresses binary representation, feature propagation, and capacity-compatible training. Sliding Directional Relation Transform (SDRT) makes local directional differences explicit by decomposing overlapping $2\\times2$ patches into a base component and four directional relation components for separate binary processing. Local Statistical Alignment (LSA) then uses the local means and variances of raw SDRT features to calibrate binary convolution outputs before spatial reconstruction. During training, Binary Vector-Quantized Knowledge Distillation (BVQKD) organizes a full-precision teacher's fusion knowledge into a shared codebook and guides the binary student through codeword-assignment distributions, without adding inference cost. Experiments on four fusion benchmarks and two downstream tasks show competitive fusion performance with reduced computational and storage costs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Image fusion", "Infrared image", "Visible image", "Binary Neural Network" ], "forum_url": "https://openreview.net/forum?id=7Y0QIJSohA", "pdf_url": "https://openreview.net/pdf/feb8abd3a66cd402b228eca8141faca08bece248.pdf", "updated_at": "2026-10-05T19:15:33.005Z", "categories": [ "ptq" ], "matched_terms": [ "binarization", "binary", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qEXbAU67Kh", "number": 16660, "title": "SHield: Watermarking 4D Gaussian Splatting via Spherical Harmonic Coefficient Perturbation", "abstract": "4D Gaussian Splatting (4DGS) models are computationally expensive to train but distributed as easily copied parameter files. \nExisting watermarks protect these assets by fine-tuning the scene against a neural decoder to extract bit strings from rendered video, a heavy process vulnerable to the temporal dynamics of dynamic scenes. \nHowever, if the threat model targets the redistributed file itself, verification can be vastly simplified. \nBecause the canonical Spherical-Harmonic (SH) coefficients of 4DGS are time- and view-invariant, parameter-space perturbations can be recovered via deterministic arithmetic. \nWe propose **SHield**, the first non-blind, decoder-free watermarking approach that natively embeds full-resolution RGB images into 4DGS files. \nSHield tiles a zero-mean payload across the SH coefficients of a key-selected Gaussian subset, applying a brief anchored optimization to preserve visual fidelity. \nTo verify ownership, it geometrically re-associates the key using canonical Gaussian centres to survive file tampering and recovers the payload in closed form. \nAcross 30 scenes (D-NeRF, HyperNeRF, DyNeRF), SHield guarantees near-perfect payload recovery (Pearson $r \\geq 0.991$) on all 150 tested models with a mean rendering penalty of $\\leq 0.13$ dB. \nHeld-out evaluation yields perfect detection on 400 true trials and zero false positives across $1{,}600$ wrong-subset null trials. \nWe also strictly bound the method's robustness: SHield survives INT8 quantization and $30\\%$ opacity pruning for the image payload, and on aggressive SH-degree-1 truncation bit payloads are still decodable.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "4D Gaussian Splatting", "Gaussian Splatting", "Digital Watermarking", "Model Watermarking", "Spherical Harmonics", "Ownership Verification", "Dynamic Scene Representation", "Neural Rendering" ], "forum_url": "https://openreview.net/forum?id=qEXbAU67Kh", "pdf_url": "https://openreview.net/pdf/7bd43005fa440c9150a2dacca0b095a0653b6fe6.pdf", "updated_at": "2026-10-05T19:16:11.239Z", "categories": [ "theory" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6DDK7uyqqZ", "number": 16670, "title": "SteerQuant: Steering Quantization Error with Action-Guided Scaling in World–Action Models", "abstract": "World--action models (WAMs) jointly generate future world states and actions through iterative denoising, using shared weights\nto process heterogeneous semantic streams of video, proprioceptive, and action tokens. Quantization reduces inference cost, but comparable numerical errors in different streams can have markedly different effects on final actions, making numerical accuracy alone insufficient for reliable control. We introduce **SteerQuant**, a 4-bit quantization framework for WAMs that steers errors toward computations with less influence on final actions. It maps how each stream's quantization errors affect final actions and uses this map to guide shared channel scaling. Activation scaling is further calibrated for each stream and denoising step to accommodate changes in activation ranges and action impact. This adapts quantization to different stream requirements without duplicating weights or increasing bit-widths for selected streams. To reduce the extra kernel launches and memory traffic introduced by scaling, we develop **Rudder**, a 4-bit inference engine for WAMs that fuses scaling and output compensation into low-bit kernels. Under W4A8 and W4A4, SteerQuant maintains mean LIBERO success within 0.8 percentage points of full precision, while delivering up to $2.23\\times$ denoising speedup over BF16 across three WAMs with reduced peak GPU memory usage. On a real dual-arm robot, W4A8 deployment achieves a $1.35\\times$ end-to-end inference speedup while maintaining average task success relative to BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "world-action models", "post-training quantization", "low-bit inference", "action impact" ], "forum_url": "https://openreview.net/forum?id=6DDK7uyqqZ", "pdf_url": "https://openreview.net/pdf/66e3e7d9d371b40c3ff04244bc1515d04ad65eea.pdf", "updated_at": "2026-10-05T19:16:17.178Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "low-bit", "quantization", "w4a4", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lTJxs8Lkz1", "number": 16717, "title": "OUTPUT FORMAT AND INDEX LOOKUP DEGRADE INCONTEXT TOOL SELECTION AT LARGE CATALOG SIZES", "abstract": "When an LLM selects a tool from a large in-context catalog, how much of its\ndegradation with catalog size is due to output format and index lookup, as opposed\nto selection itself? Across Qwen2.5 (0.5B–14B, 4-bit) on the Berkeley FunctionCalling Leaderboard (BFCL), a zero-learning lexical-overlap heuristic overtakes\nin-context LLM selection at a finite, measurable catalog size |V |\n∗\n. This crossover\ngrows with scale (8.7 at 1.5B to 500.0 at 14B; α = 1.784), then more slowly\nbeyond 14B (365.3 at 32B, upper bound resolved rather than right-censored; 552.6\nat 72B, 95% CI [346.6, >650], upper bound right-censored — both precisionmatched, App. B.24). Two frontier closed models (GPT-4o, Claude Sonnet 5,\nsmall-n) show no crossover at all through |V |=650 (App. B.26, B.27), limiting\npractical generality.\nForcing name-based rather than index-based output recovers a substantial share of\nthe lost performance at high |V | (3B: +37.5pp, 7B: +41.5pp, both at |V |=234;\n3B peaks at +50.5pp at |V |=64), but performance still declines with |V | under\nname emission, so selection difficulty and output-format cost are not cleanly separated — a direct log-likelihood control that removes output format entirely does\nnot reveal an easier underlying selection ability, if anything the reverse at lowto-moderate |V | (App. B.21). Extending the grid to |V |=650 resolves 3B’s own\nname-emission crossover (|V |\n∗=350, 12.2× index-emission’s point estimate, CI\n[9.4×, 15.1×]); 7B stays right-censored, so its ratio is an upper bound only, since\na hashed-name control could not cleanly rule out lexical-shortcut copying as the\nsource of the gain. A further test removes the task entirely, asking only whether\nthe model can locate a named, given answer’s index: accuracy still declines significantly with |V | at both scales (3B: −70.0pp; 7B: −41.5pp), though at 3B the\ntask is counterintuitively harder than ordinary selection at |V |≤16 — near-miss\ncounting fits well (74.2%, 2.6× chance, at 8; 41.0%, 3.1× chance, at 16), not\na bug. Two further controls narrow the picture: |V |, not raw prompt length, is\nthe primary driver (partially so at 3B), and the near-miss pattern is not specific\nto one binding direction; we do not claim a fully clean causal isolation of this\nmechanism.\nA second mitigation, filtering the catalog before selection, closes most of the gap,\nand is effective across a range of k rather than only at a model’s own crossover\n(App. B.9). Most of this gain is achievable through candidate ranking alone, without truncation — a cheaper, more directly deployable finding, interacting with\nthe output-format effect (§4.5). Every |V |\n∗\nreported here is specific to lexicallytransparent catalogs: a direct opacity check collapses the crossover roughly 3×\n(App. B.7), and the overall finding replicates directionally on an independent\nbenchmark and model family. We report every number here, including the ones\nthat complicate the story, as measured.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "tool selection", "function calling", "output format sensitivity", "zero-learning baseline", "catalog size", "model scale", "agentic tool use" ], "forum_url": "https://openreview.net/forum?id=lTJxs8Lkz1", "pdf_url": "https://openreview.net/pdf/2287958baf0b0bacaeaa60a03c30168d2fdeabf2.pdf", "updated_at": "2026-10-05T19:16:38.442Z", "categories": [], "matched_terms": [ "4-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "utXuOW0dBw", "number": 16751, "title": "LoPRo: Enhancing Low-Rank Quantization via Permuted Block-Wise Rotation", "abstract": "Post-training quantization (PTQ) enables effective model compression with minimal accuracy loss. While most weight-only PTQ methods target the challenging sub-3-bit regime and often require fine-tuning, we propose LoPRo—a fine-tuning-free algorithm that improves residual matrix quantization via block-wise permutation and Walsh-Hadamard transforms to align columns of similar importance. We also introduce a mixed-precision, fast low-rank decomposition based on rank-1 sketching (R1SVD) to reduce quantization costs. Experiments show LoPRo outperforms existing fine-tuning-free PTQ methods at 2- and 3-bit (with its vector variant LoPRo$_v$ delivering the best 2-bit results), matching fine-tuned accuracy while achieving up to 5$\\times$ quantization speedup over GPTVQ. On Mixtral-8x7B, it finishes quantization in under 2.5 hours, reducing perplexity by 0.75 over MoEQuant++ (3.17 vs.\\ 3.00 avg.\\ bits) and boosting BoolQ accuracy by 5.4 points. LoPRo also achieves higher accuracy than other low-rank methods at much lower ranks, with low latency overhead ($\\le$10.3\\% at batch 1, lower for larger batches).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "low-rank", "rotation" ], "forum_url": "https://openreview.net/forum?id=utXuOW0dBw", "pdf_url": "https://openreview.net/pdf/01b060179c70eea3d90be7f5b87e8f54fb3fa05c.pdf", "updated_at": "2026-10-05T19:16:57.862Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "3-bit", "mixed-precision", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "z8Y4Oua8RH", "number": 16809, "title": "BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration", "abstract": "Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose \\textbf{BitNest}, a bit-nested speculative decoding framework that embeds a low-precision draft directly into the higher-precision target representation. Instead of deriving a draft from a predefined target, BitNest first constructs a strong low-precision base and then recovers the higher-precision target through residual refinement, enabling both models to share a single physical weight representation. BitNest further extends this progressive-precision design to the KV cache for long-context inference. Across multiple 7B--8B edge friendly LLMs and diverse workloads, BitNest achieves an average speculative acceptance rate of \\textbf{95.2\\%} while closely preserving higher-precision model quality, and delivers \\textbf{1.48--1.61$\\times$} end-to-end speedup over FP16 autoregressive decoding. On the LLaMA models supported by all representative self-speculative baselines, BitNest also achieves consistently competitive or higher decoding speedup.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Speculative Decoding", "LLM Quantization", "KV Cache Quantization", "Large Language Model Inference" ], "forum_url": "https://openreview.net/forum?id=z8Y4Oua8RH", "pdf_url": "https://openreview.net/pdf/994c1444076b9e59049ec2c4d7dfb07cf24510b6.pdf", "updated_at": "2026-10-05T19:17:22.757Z", "categories": [ "kv-cache" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AWj9yXEzcr", "number": 16818, "title": "Rethinking the Quantization Error Lifecycle in LLMs", "abstract": "Post-training quantization (PTQ) is essential for efficient deployment of large language models (LLMs), yet aggressive low-bit quantization can severely degrade model accuracy. Existing GPTQ-style compensation-based methods sequentially quantize weight columns, typically determining the quantization decision of the current column using round-to-nearest (RTN) and then updating the remaining weights to compensate the induced error. However, the RTN-based quantization decision is made independently of subsequent compensation, thereby treating error formation and error compensation as separate stages. In this work, we uncover the overlooked coupling between error formation and error compensation and formalize this coupling through the quantization error lifecycle, a systematic perspective on compensation-based LLM quantization. Building on this perspective, we derive a shared compensation representation that reveals complementary information for error formation and error compensation. From this shared representation, we develop CARE-Q, a unified compensation-based quantization framework comprising DiCR for compensation-aware rounding and ODPP for enhanced error compensation. Extensive experiments across diverse LLMs and quantization settings show that CARE-Q consistently outperforms leading compensation-based quantization methods and complements representative transformation-based approaches, including QuaRot and FlatQuant, while introducing no additional inference-time overhead. Our code is provided in the supplementary material and at https://anonymous.4open.science/r/CARE-Q/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "LLM Quantization", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=AWj9yXEzcr", "pdf_url": "https://openreview.net/pdf/5492b3f93a1734e2c7cdcfc3674ee1a8f365d4fc.pdf", "updated_at": "2026-10-05T19:17:26.262Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FZvhlChEep", "number": 16888, "title": "Beyond Reconstruction: Fisher-Regularized Post-Training Quantization for LLMs", "abstract": "Large language models must be compressed to be deployed under realistic memory and latency budgets, and post-training quantization (PTQ) is dominant because it needs no retraining. Existing PTQ methods minimize a per-layer reconstruction error that approximates perplexity, but this proxy is not aligned with downstream task loss. Since a deployed model is typically dedicated to one task, the most direct way to preserve its task accuracy is to draw the calibration data from that task, yet this does not close the gap. We argue that the mismatch lies in the objective rather than in the data, and we introduce a simple yet effective modification that augments the PTQ objective with a Fisher-information regularizer preserving the weight directions to which the loss is most sensitive. We instantiate the regularizer in two forms, a \\emph{Task Fisher} computed from a labeled task training split and a \\emph{Generic Fisher} computed from the generic calibration data already used by PTQ, the latter removing any dependence on task labels and yielding a single quantized model that serves all tasks rather than a separate quantization per task. The regularizer requires no architectural changes and integrates into any per-layer PTQ solver as a diagonal loading of the per-layer curvature. Across five LLMs from 1B to 70B parameters and 2--4 bit widths, Task Fisher and Generic Fisher improve zero-shot accuracy and perplexity to a comparable degree, showing that neither task labels nor per-task quantization is necessary. Adding the label-free Generic Fisher to standard PTQ solvers raises multiple-choice accuracy by up to 6.3 points and generative-task accuracy by up to 9.4 points over the unregularized baselines, and brings the same improvement to a state-of-the-art solver.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "large language models", "Fisher information", "model compression" ], "forum_url": "https://openreview.net/forum?id=FZvhlChEep", "pdf_url": "https://openreview.net/pdf/b32cfe7892b5b6d5f909f7ec6c18d6521a3bd3b7.pdf", "updated_at": "2026-10-05T19:18:01.220Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4 bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UAUwDIekdq", "number": 16975, "title": "Spiking DGCNN: Max-First Graph Spiking Networks for Efficient Point Cloud Learning", "abstract": "Spiking neural networks (SNNs) offer sparse, event-driven computation for energy-efficient learning, but the naive spiking conversion of graph-based point cloud networks remains poorly understood. EdgeConv relies on neighbourhood-wise max aggregation to select discriminative local responses. Simply replacing LeakyReLU with spiking neurons in EdgeConv quantizes features before aggregation, weakening credit assignment and causing severe performance degradation. We therefore propose Spiking DGCNN, built on a Max-First Spiking Rule that performs neighbourhood selection on continuous-valued features before spike generation. We further introduce an Adaptive Pseudo-Temporal Expansion-Compression (APTEC) module that supports lightweight pseudo-temporal evidence accumulation through leaky spiking integration, adaptive thresholding, and binary spike compression. Ablation experiments show that Max-First Spiking Rule improves Spike-before-Max baseline from 79.09% to 91.77% at T=1 and from 84.11% to 92.18% at T=4 on ModelNet40. The complete model with APTEC achieves 92.38% accuracy, closely matching the 92.9% of DGCNN, while reducing the operation-level energy estimate from 11.23 mJ to 0.22 mJ.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Point Cloud Processing", "Efficient Computing", "Brain-inspired Computing" ], "forum_url": "https://openreview.net/forum?id=UAUwDIekdq", "pdf_url": "https://openreview.net/pdf/e3f20b13683f9e1fb5e4008f6411b4a3e1473616.pdf", "updated_at": "2026-10-05T19:19:10.346Z", "categories": [], "matched_terms": [ "binary", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZOZCV5K6hs", "number": 16996, "title": "SPECULATIVE CACHE LINE PUNCTURING: GRANULAR MEMORY REDUCTION FOR LLM INFERENCE", "abstract": "Long-context speculative inference frequently suffers from severe memory bottlenecks caused by rapidly expanding Key-Value (KV) cache footprints. Existing eviction and quantization strategies often introduce intra-cache fragmentation or disrupt token-index alignment between draft and target models. In this paper, we introduce Speculative Cache Line Puncturing (SCLP), a fine-grained runtime memory-management primitive tailored for speculative execution. Rather than evicting coarse blocks or reallocating memory, SCLP applies an in-place sub-block validity mask to deactivate redundant attention states while strictly preserving underlying tensor layouts and positional coordinate systems. We evaluate SCLP on the Qwen-2.5 architecture (7B target with 1.5B draft) across three distinct hardware tiers (NVIDIA A40, A100 SXM 80GB, and H100 SXM 80GB) using 500-sample evaluations across 5 independent random seeds (2,500 cumulative evaluation prompts per benchmark). Our findings reveal a clear operational dichotomy: on short-context conversational flows (ShareGPT), SCLP maintains low kernel overhead and yields an acceptance rate improvement of 53.2% vs. 42.9% (+10.3%, $p < 0.001$) while leaving short-context memory footprint unperturbed; conversely, on long-context retrieval workloads (RULER, 4k–16k tokens), SCLP achieves a sustained 48.2% reduction in active KV cache memory (0.33 GB vs. 0.65 GB per sequence at 4k context) and delivers a +12.7% boost in speculative acceptance rate (61.5% vs. 48.8%, $p < 0.001$), nearly doubling concurrent serving batch capacity under fixed VRAM constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "speculative decoding", "KV cache", "memory optimization", "large language models", "systems" ], "forum_url": "https://openreview.net/forum?id=ZOZCV5K6hs", "pdf_url": "https://openreview.net/pdf/b7570de7cf93d692122144850a947de814b33a43.pdf", "updated_at": "2026-10-05T19:19:21.844Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dcENiU588x", "number": 17227, "title": "LZIP-P2P: LOSSLESS COMPRESSION SUPERCHARGED GPU COMMUNICATION", "abstract": "The rapid growth of large language models (LLMs) has made GPU communication a critical bottleneck. While prior work reduces communication volume via quantization or lossy compression, these approaches introduce numerical errors that can degrade convergence, accuracy, and stability. We present LZIP-P2P, a design that integrates lossless compression directly into GPU point-to-point (P2P) communication primitives without compromising numerical correctness. LZIP-P2P employs a split-send pipeline that exposes transmissible data early and overlaps compression with communication, while preserving high GPU efficiency by operating on large data blocks. In real workloads, LZIP-P2P accelerates RL weight synchronization by up to 47.5% without application changes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "machine learning systems", "GPU communication", "lossless compression", "distributed LLM training" ], "forum_url": "https://openreview.net/forum?id=dcENiU588x", "pdf_url": "https://openreview.net/pdf/78570d9d58cb7b2856b0df8c08aa67af654029ea.pdf", "updated_at": "2026-10-05T19:21:46.880Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qOstx3qCfu", "number": 17315, "title": "SinkMark: Robust Model Fingerprinting via Attention-Sink-Induced Hallucinations", "abstract": "Large Vision-Language Models (LVLMs) are increasingly released and adapted for downstream applications, making it challenging to verify whether a suspicious black-box model is derived from a copyrighted checkpoint. In this work, we propose SinkMark, a robust model fingerprinting framework that turns model hallucinations from an undesirable failure mode into a behavioral ownership signal. Our key idea is to construct private image-text probes that induce a predefined target hallucination in the source model and its derivatives, while remaining inactive on unrelated models. To achieve this, SinkMark dynamically identifies a shared attention sink across multiple decoder layers, promotes cross-layer agreement and attention concentration at the sink, and binds its hidden representation to the target hallucination semantics. We jointly optimize a bounded image perturbation and a natural-language suffix using PGD and GCG, respectively. To improve persistence under downstream model modifications, we further introduce robust min–max optimization with temporary adversarial parameter updates and predecessor-guided target induction. Importantly, SinkMark does not permanently modify the released model and requires only black-box access during ownership verification. Experiments demonstrate that SinkMark provides reliable and model-specific target hallucinations, while remaining robust to fine-tuning, pruning, and quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Vision-Language Models; Model Fingerprinting; Black-box Ownership Verification; Attention Sink; Robust Hallucination" ], "forum_url": "https://openreview.net/forum?id=qOstx3qCfu", "pdf_url": "https://openreview.net/pdf/44a79a2f93ac382b92089a40b4ca00b55bcdf052.pdf", "updated_at": "2026-10-05T19:22:35.697Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "osd72MlzL5", "number": 17376, "title": "Scalar Smoothness Is One Point on a Besov Spectrum", "abstract": "Studies of neural network generalization and robustness often summarize smoothness with a single number, such as a Lipschitz constant or a Hessian curvature statistic. However, Lipschitz constants summarize global worst-case sensitivity and Hessian statistics summarize local curvature; neither captures how model responses vary with the size, direction, and location of perturbations encountered in real-world deployment. To measure smoothness under these perturbations, we introduce a finite-difference roughness spectrum that profiles model responses across these conditions. Our roughness spectrum generalizes scalar smoothness measures, bringing Lipschitz sensitivity and directional Hessian curvature into a common framework across perturbation scales. We connect this spectrum to classical Besov theory, which describes how a function's variation changes with scale. Our theoretical analysis also proves that a larger output shift can cause less task damage than a smaller one, and that local curvature alone can miss the blocks most vulnerable to finite perturbations. In image classification, Roughness Source Control (RSC) suppresses margin roughness along noise directions and improves noise robustness in a comparison of 80 checkpoints across four model--dataset combinations. In post-training quantization, we identify blocks that need higher precision by measuring task-loss increases from perturbations matched to the size and direction of each block's quantization error. Our mixed-precision allocation achieves lower perplexity than the Hessian-based HAWQ-V2 baseline under the same GPTQ weights and bit budgets in all 14 settings across seven held-out large language models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "neural network smoothness", "model robustness", "Besov spaces", "finite differences", "perturbation sensitivity", "post-training quantization" ], "forum_url": "https://openreview.net/forum?id=osd72MlzL5", "pdf_url": "https://openreview.net/pdf/41bdba0ac427071d79d4bedc7867b72be3d4a7b7.pdf", "updated_at": "2026-10-05T19:23:05.579Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yqb7fK5WNq", "number": 17425, "title": "MDMixQ: Modality-Decoupled Mixed-Precision Quantization for Multimodal MoE Compression", "abstract": "Post-training quantization is widely adopted for reducing the memory footprint of large-scale Mixture-of-Experts (MoE) models, and expert-level mixed precision further improves the accuracy-compression trade-off by allocating heterogeneous bit-widths across experts. In multimodal MoE settings, however, we find that existing mixed-precision pipelines overlook inter- and intra-modality routing non-uniformities that skew expert bit allocation and degrade accuracy under compression. To address these failure modes, we propose MDMixQ, a mixed-precision quantization framework comprising three targeted components: Modality-Decoupled Calibration (MDC), which isolates a clean text-routing signal from multimodal inputs; Salience-Guided Filtering and Rectification, which retains only confident text-side routing events and rectifies sign-inconsistent logits; and Dynamic Mixed-Precision Allocation, which ranks experts by the refined salience scores under a global bit budget and applies modality-aware activation quantization during MoE inference. At a 2.52-bit average expert budget, the best MDMixQ configurations retain 96.29%, 96.05%, and 95.33% of FP16 performance on Qwen3-VL-30B-A3B, Kimi-VL-16B-A3B, and Qwen3-Omni-30B-A3B, respectively, outperforming prior expert-priority methods across two model families. The framework also transfers to the MXFP format with 99.72% performance retained on Qwen3-VL-30B-A3B, confirming compatibility with emerging hardware-oriented quantization standards.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "Post-Training Quantization", "Multimodal Large Language Models", "Mixed-Precision Allocation" ], "forum_url": "https://openreview.net/forum?id=yqb7fK5WNq", "pdf_url": "https://openreview.net/pdf/188eab3dbefcf37b8ba4227d94e529ecb6f7fc21.pdf", "updated_at": "2026-10-05T19:23:39.093Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cc1sddTsQo", "number": 17442, "title": "CoCaQ: Coupling of Compression and Quantization for Flexible LLM Deployment", "abstract": "The practical deployment of Large Language Models (LLMs), especially on edge devices, is heavily constrained by their massive model sizes and prohibitive computation costs, which are generally addressed by either model compression or quantization. In practice, these two kinds of techniques can be applied concurrently for flexible LLM deployment, especially under extremely low memory budgets, which, however, are typically optimized separately and so lead to suboptimal performance. Particularly, we observe that the joint of model compression and quantization is challenged by the effect of compressed structures on quantization and the ability of quantization to account for distribution shifts induced by compression. To address the above issues, we propose a new Coupling of Compression and Quantization ($\\textbf{CoCaQ}$) method to perform model compression and quantization in a coupling manner, aiming to achieve effective and flexible LLM deployment by unifying the two techniques. To this end, our CoCaQ is developed with two key components, including a quantization-aware model compression (QaMC) and an activation stabilization for quantization (ASQ). First, our QaMC constructs quantization-friendly low-rank representations based on activation statistics and optimizes the layer-wise compressed structures with a quantization-aware loss, aiming to identify the robust substructures of LLM for the subsequent quantization. Furthermore, our ASQ introduces an efficient orthogonal rotation regularization for low-rank weight matrices, enabling low-bit quantization to better tolerate activation amplification induced by the preceding model compression. Based on the proposed QaMC and ASQ, our CoCaQ couples model compression and quantization by fully considering the mutual influence between the two, resulting in an effective yet flexible solution for LLM deployment. Experiments are conducted across LLaMA and Qwen models, and the results show that our CoCaQ achieves better memory-accuracy trade-offs than state-of-the-art compression-only and quantization-only methods, while clearly outperforming the counterparts with joint of compression and quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "model quantization", "model compression" ], "forum_url": "https://openreview.net/forum?id=cc1sddTsQo", "pdf_url": "https://openreview.net/pdf/2131091f1e441400a0293486a27d97c61c08cb4e.pdf", "updated_at": "2026-10-05T19:23:42.574Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "uo5OpAXPn0", "number": 17458, "title": "DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization", "abstract": "Complex reasoning and agentic applications increasingly rely on long-context inference, where growing KV caches increase both memory usage and decoding overhead. Hybrid models reduce these costs by combining Softmax Attention with Gated DeltaNet (GDN) or Kimi Delta Attention (KDA), which maintain fixed-size recurrent states. These states are commonly stored in FP32 and consume substantial GPU memory, while their updates are limited by memory bandwidth. Quantization can reduce both storage footprint and memory traffic, but we find that uniform INT8 and FP8 degrade complex reasoning accuracy, while INT4 and NVFP4 collapse it to near zero. To our knowledge, this is the first study of post-training recurrent-state quantization for GDN and KDA. Our analysis reveals that outliers in GDN and KDA states are concentrated in particular key channels and value dimensions. Learned decay influences how much quantization error is retained. We find that largely the same GDN heads and KDA key channels exhibit slow decay across tasks. Based on these insights, we propose DAMP, which jointly considers quantization error and decay-based error retention to select high-risk key channels offline. Under a fixed storage budget, it retains these channels in FP16 and stores the remainder in INT8. We evaluate DAMP on Qwen3.6-35B, Kimi-Linear-48B and Kimi-K3 across six reasoning and code generation benchmarks. At 9.9 bits per state value, DAMP maintains average accuracy close to FP32. In SGLang, DAMP reduces recurrent-state storage by 69.1\\%, accelerates the recurrent-state update kernel by up to $2.59\\times$, and lowers full-model time per output token by up to 19.0\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Efficient LLM Inference; Hybrid Attention; Recurrent-state Mixed-precision Quantization" ], "forum_url": "https://openreview.net/forum?id=uo5OpAXPn0", "pdf_url": "https://openreview.net/pdf/0367e12f2922b6c746845f60b8c044e7a66091b5.pdf", "updated_at": "2026-10-05T19:23:47.707Z", "categories": [ "ptq", "training", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "fp8", "int4", "int8", "mixed-precision", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "11I4efr8nC", "number": 17461, "title": "Elastic Prediction from the First Time-Step: Spiking Neural Networks with Temporal Membrane Biasing", "abstract": "Spiking neural networks (SNNs) naturally support elastic inference, where salient inputs trigger early predictions, and additional time-steps progressively refine the output, enabling fast responses and improved accuracy without repeated inference or model retraining.\nHowever, existing SNN methods are not well-suited for elastic inference: ANN-to-SNN conversion methods suffer from poor early-time-step accuracy, whereas directly trained SNNs perform well at specific time-steps but often degrade when inference continues, limiting the accuracy refinement.\n\nIn this work, we identify the root cause of the early-stage performance degradation as a *membrane-threshold mismatch* between membrane potentials and firing thresholds. \nTo mitigate this mismatch, we propose a temporal membrane biasing (TMB) method that applies sustained calibration across time-steps while preserving Quantized Artificial Neural Network (QANN) equivalence. \nIn addition, we develop a bias-only fine-tuning strategy that enables efficient optimization with minimal training cost. \nAcross multiple tasks, our work markedly improves low time-step accuracy (*e.g.,* 70.15\\% top-1 accuracy improvement at the first time-step on ImageNet). Combining elastic inference and adaptive methods, TMB achieves 24.3\\% operation and 24.8\\% time-step reduction with $<0.1$\\% accuracy degradation compared to the same SNN without TMB.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Network", "Elastic Inference", "Accuracy-Operation Trade-off" ], "forum_url": "https://openreview.net/forum?id=11I4efr8nC", "pdf_url": "https://openreview.net/pdf/e429fcefda178f8e5da7e0ccffb9f620baa51e38.pdf", "updated_at": "2026-10-05T19:24:01.091Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LhoYVUdwzO", "number": 17471, "title": "Mixed precision inference for Diffusion language models", "abstract": "Diffusion Language Models (DLMs) have drawn increasing attention as an alternative to classical Large Language Models architecture due to their potential to significantly increase inference speed. However, they lack in practice some key optimizations. One such feature is mixed-precision inference, i.e., using a quantized or a full precision model at different steps of the generation process. This hybrid design enables one to benefit from the accelerated inference of quantized models while maintaining an output quality similar to that of original floating-point precision. As existing approaches are not applicable to this setup, we propose here the first mixed precision framework specifically targeted for DLMs. By modeling the quantization error, we are able to design a training-free, adaptive mechanism to successfully exploit the speed of a quantized model and the precision of a regular one. Our results show we can achieve a 66\\% speed increase without significant degradation of the model's performance. We also demonstrate this can be combined with other acceleration systems to further speed up the sampling process.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion large language models", "quantization" ], "forum_url": "https://openreview.net/forum?id=LhoYVUdwzO", "pdf_url": "https://openreview.net/pdf/93a215f2042d27ffaa0a5885606d35abcfd54393.pdf", "updated_at": "2026-10-05T19:24:00.702Z", "categories": [ "diffusion" ], "matched_terms": [ "mixed precision", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mg9hCSolvk", "number": 17472, "title": "D-Quant: Driftable Entropy Coding for KV Cache Quantization", "abstract": "The KV cache has become a major bottleneck in deploying LLMs, as its memory footprint grows linearly with sequence length and batch size, imposing substantial pressure on both memory capacity and bandwidth.\nAmong various KV cache compression techniques, quantization is particularly attractive due to its effectiveness and ease of deployment.\nHowever, most existing methods rely on fixed-width quantization, where a $b$ bit representation is inherently limited to $2^b$ quantization levels.\nAs the bit width decreases, the number of available levels shrinks exponentially, leading to severe information loss and rapid performance degradation.\nWe further observe that fixed-width quantization fails to exploit the highly non-uniform distribution of KV cache.\nAfter rotation and normalization, KV values approximately follow a normal distribution, with most values concentrated near the center and only a small fraction appearing in the tails.\nNevertheless, fixed-width coding allocates the same number of bits to frequent and rare symbols.\nEntropy coding naturally exploits such non-uniformity by assigning shorter codewords to frequent symbols and longer ones to rare symbols, substantially reducing the average number of bits required for representation.\nHowever, its variable-length output is not suited to highly parallel attention kernels, where efficient dequantization and computation rely on regular memory layouts and fixed-stride accesses.\nTo bridge this gap, we propose D-Quant, a flexible KV cache quantization framework that introduces a drift mechanism to convert entropy-coded representations of each token into fixed-size bitstreams, enabling regular memory access and parallel dequantization within attention kernels.\nD-Quant also provides greater flexibility than conventional quantization methods, allowing the bit budget and number of quantization levels to be configured more freely rather than being constrained by fixed bit widths and group sizes.\nFurthermore, since the rotated and normalized KV values closely follow a normal distribution, the probability model for entropy coding can be directly derived from the standard normal distribution, making D-Quant entirely calibration-free.\nExtensive experiments show that D-Quant preserves near BF16 performance at only $2.26$ bits per element while substantially outperforming existing baselines.\nD-Quant further achieves up to a $7\\times$ reduction in KV cache memory footprint and a $3.5\\times$ improvement in decoding throughput.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "entropy coding" ], "forum_url": "https://openreview.net/forum?id=mg9hCSolvk", "pdf_url": "https://openreview.net/pdf/c11f2b4a4e96bb95a27031a7c1b15419d46a4b64.pdf", "updated_at": "2026-10-05T19:24:01.121Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "bit width", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IsoNXwco3c", "number": 17475, "title": "UNISPIKE: Boosting the Performance of Spiking Neural Networks with Hybrid Training", "abstract": "Spiking neural networks (SNNs) offer an efficient inference paradigm with two desirable properties: addition-only computation and step-by-step inference. Addition-only computation removes expensive multipliers and reduces arithmetic energy, while step-by-step inference enables temporally progressive outputs for faster response and early exiting. However, achieving high accuracy while preserving these properties remains challenging. Several accuracy-oriented SNN operators, such as SEW-addition, temporal attention, and temporal addition, improve representation learning by introducing multi-valued activations or temporal synchronization, thereby violating addition-only computation or step-by-step inference.\n\nTo improve accuracy under these property-preserving constraints, we propose UniSpike, a hybrid-training framework that initializes an SNN from a quantized ANN and further fine-tunes it with back-propagation-through-time (BPTT). The key design is Unified Clip, a unified ANN-side activation that replaces operators incompatible with spike-based inference, including softmax, layer normalization, and GeLU. Its quantized form is equivalent to the ST-BIF spiking neuron, providing effective ANN-to-SNN initialization for SNN fine-tuning. Built upon Unified Clip, we design UniFormer, a spike-compatible vision transformer tailored for accurate, addition-only, and step-by-step SNN inference. On ImageNet-1K, UniSpike achieves 80.8% accuracy at 4 time-steps, outperforming the SOTA addition-only and step-by-step direct-training method Spike-Driven Transformer V2.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Nerual Network", "Hybrid Training", "Energy Efficiency" ], "forum_url": "https://openreview.net/forum?id=IsoNXwco3c", "pdf_url": "https://openreview.net/pdf/44b43780584e4c757afbeff4ddf4e8879a411f62.pdf", "updated_at": "2026-10-05T19:24:01.470Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7AWQLrUQ0R", "number": 17636, "title": "Searching the Quantized Weight Space for Adaptation: Task-Aware Joint Quantization and Low-Rank Initialization", "abstract": "Quantized low-rank fine-tuning enables memory-efficient adaptation of large language models by training low-rank adapters on a frozen low-bit backbone. Since the backbone remains fixed during fine-tuning, the quantized weights selected at initialization define the starting point from which the adapter must learn. Existing methods mainly select quantized weights to reconstruct the pretrained model’s weights or activations. However, reconstruction quality before fine-tuning does not necessarily identify the best starting point for adaptation. We therefore view initialization as a search for quantized weights that perform well after adaptation. However, this poses a major challenge: the adapted model is not yet available during the calculation of the quantized initialization. To address this challenge, we propose \\emph{Task-Aware Quantized Initialization} (TAQI), which uses target-task output gradients measured before fine-tuning to guide this\nsearch. TAQI weights layer-wise reconstruction errors by gradient energy and\nmixes the resulting task-sensitive curvature with general calibration\ncurvature. We show that the task-weighted error bounds the energy of local first-order task-loss perturbations. Using IBM Granite-4.0-H-350M at 4 bits, TAQI improves the primary metric by 0.45--1.33 points over GPTQ-intrinsic across medical, financial, and scientific question-answering benchmarks. These results highlight the promise of adaptation-oriented quantized initialization for quantized low-rank fine-tuning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization;Quantized Low-Rank Adaptation;Parameter Efficient Fine-tuning" ], "forum_url": "https://openreview.net/forum?id=7AWQLrUQ0R", "pdf_url": "https://openreview.net/pdf/5f98420d9a47f1b7cda44daf72f3ed0e9c3a064d.pdf", "updated_at": "2026-10-05T19:25:26.565Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ETAQbepFQT", "number": 17656, "title": "Cache the Encoder Within: Compact, Reusable Memory across LLM Queries", "abstract": "Repeated queries over shared documents incur redundant encoding,\nwhile caching model states introduces persistent storage costs.\nBuilding on CoMem's intermediate-state interface, EncBank treats\na pretrained LLM's lower layers as a reusable document encoder\nand compactly stores their outputs for an adapted upper-layer reader.\nA self-distilled suffix adapter is shared across storage precisions\nwithin each backbone, without quantization-specific retraining.\nAcross five benchmark suites on three Qwen backbones spanning\ndifferent sizes and full-attention and hybrid architectures,\n4-bit storage keeps each reported benchmark aggregate within\none score point of native-precision EncBank.\nIn a fixed Qwen3-8B workload, it retains 28.1\\% of the\nnative-precision persistent GPU store.\nSeparate native-precision controls yield a $1.40\\times$\nselected-pack prefill speedup over same-evidence, same-adapter\ntext replay, at a 3.12-point RULER accuracy cost.\nA native-precision Qwen3.8-27B configuration also passes\n70 of 89 Terminal-Bench 2.1 tasks.\nEncBank thus combines reusable computation with compact memory,\nwhile task fidelity and end-to-end benefits remain dependent\non the workload, preparation costs, and reuse frequency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "persistent memory", "intermediate representations", "representation reuse", "memory quantization", "efficient inference", "retrieval-augmented generation" ], "forum_url": "https://openreview.net/forum?id=ETAQbepFQT", "pdf_url": "https://openreview.net/pdf/f3c27d95d9f75c55096812b5a4d57e8cc6514457.pdf", "updated_at": "2026-10-05T19:25:34.240Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OMsudQ7BPk", "number": 17747, "title": "RetinaPack: Few-Shot Continual Retinal Pre-Screening with Offline Neuromorphic Inference", "abstract": "Retinal pre-screening in rural and resource-constrained settings requires portable, offline and privacy-preserving inference that can adapt to local acquisition shifts without large-scale retraining. A key difficulty is that current-stage labels can be obtained immediately, whereas prognostic outcomes mature months later, so local adaptation must improve staging without destabilizing longitudinal risk estimates. We introduce RetinaPack, a participant-disjoint longitudinal benchmark of 13,040 participants, and RetinaLPA-2S, a few-shot continual adaptation method for four-stage age-related macular degeneration (AMD) pre-screening. RetinaPack combines adjudicated change and stable pairs with crossed camera/operator repeats to separate disease-related variation from acquisition nuisance and define a pathology-referenced protected subspace. RetinaLPA-2S updates only its orthogonal complement through a 4,096-parameter two-sided low-rank adapter, while a frozen competing-risk branch preserves fixed-input prognosis during stage-only adaptation. A quantized encoder runs on the Akida AKD1000 and signed adaptation remains on the local host, enabling network-independent inference, local updating and feature replay without retinal-image storage. With 10 labels per stage, RetinaLPA-2S reaches 82.20 domain-balanced candidate F1 with 1.07-point forgetting, 1.68 points above BiLoRA-style FA. On 920 staging queries, it achieves 83.78 macro-F1 and 95.33% referral sensitivity. The complete image-to-report path requires 2.150 J, 650.4 ms P95 latency and 224 MiB peak memory, reducing processing energy by 34.35% relative to CPU INT8.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "retinal pre-screening", "age-related macular degeneration", "continual learning", "few-shot adaptation", "parameter-efficient adaptation", "neuromorphic inference", "offline inference", "longitudinal risk prediction" ], "forum_url": "https://openreview.net/forum?id=OMsudQ7BPk", "pdf_url": "https://openreview.net/pdf/b9404e918ee327263b4a729ced36108f208e3065.pdf", "updated_at": "2026-10-05T19:26:19.560Z", "categories": [ "hardware" ], "matched_terms": [ "int8", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VseZJ6MGKd", "number": 17796, "title": "Quantized One-Step Generative Modeling via Cell-Margin Quantization-Aware Training", "abstract": "Low-bit quantization and one-step sampling can substantially reduce the cost of flow-matching models, yet existing quantization-aware training methods typically minimize velocity-field reconstruction error, although generation quality is ultimately determined by the terminal distribution after ODE integration. We introduce CM-QAT, which trains a quantized student for one-step sampling with one function evaluation (1-NFE) to preserve the teacher's terminal geometry in a frozen feature space. Specifically, CM-QAT constructs a public nested partition and uses a differentiable prototype-margin objective to encourage each source sample to reach the correct terminal cell. Our counterexamples show that small velocity-$L_2$ error is neither necessary for terminal cell correctness nor sufficient for a distribution-free cell-risk guarantee without boundary-margin control. We derive conditional feature-space $W_1$ bounds from recursive cell-histogram discrepancies and relate cell-crossing probability to endpoint displacement and boundary margins. Experiments on Fashion-MNIST, CIFAR-10, and class-conditioned ImageNet-256 show that CM-QAT outperforms recent low-bit flow-generation methods and enables generation with models quantized to as low as 2-bit \\& 3-bit weights and sampled in a single step (1-NFE).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Flow Matching", "Quantization-Aware Training", "Generative Modelling", "Optimal Transport" ], "forum_url": "https://openreview.net/forum?id=VseZJ6MGKd", "pdf_url": "https://openreview.net/pdf/bb7202f79e95b28592740440f1533dd08dbb834d.pdf", "updated_at": "2026-10-05T19:26:42.564Z", "categories": [ "ptq", "qat", "diffusion" ], "matched_terms": [ "2-bit", "3-bit", "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5R8LqE51Hr", "number": 17827, "title": "Better Fixed-State Fidelity Need Not Improve Closed-Loop Performance in Quantized Neural Policies", "abstract": "Does closer action agreement with a full-precision policy imply better closed-loop\nperformance after quantization? We hold trained source policies fixed, vary calibration\ndata or compiler settings, and execute compiled policies on Edge TPU\nand Hailo accelerators in simulated control tasks with paired initial conditions.\nIn a complete, prespecified eight-rule Hailo FetchReach family, one artifact has\na 16% smaller fixed-state mean action-error norm than another yet incurs a 34-\nfold larger estimated mean return loss relative to FP32. An early–middle contrast\nselected from that family and locked before replication favors early-state calibration\nin seven of eight independently trained FetchReach policies, and outcomeinformed\naudits retain large calibration effects under alternative compiler settings.\nWithin the original family, a full-budget, AdaRound-enabled compiler intervention\nreduces action root-mean-square error (RMSE) on calibration-disjoint states\nfor all eight rules, but mean return decreases for five; retrospective family-adjusted\nintervals place all five changes below zero but within the prespecified 0.5-return\npractical band. In a separate prospective benchmark of eight fresh FetchReach\npolicies and 128 Hailo artifacts, a secondary analysis finds lower RMSE on a\ncalibration-disjoint panel in 53 of 64 policy–rule comparisons, 22 of them with\nlower observed mean return; these are descriptive contrasts clustered within policies.\nPost-outcome episode resampling retains 16 of the 53 RMSE improvements\nwith adjusted RMSE-change intervals below zero; the count varies across bootstrap\nseeds. Changing action coordinates or evaluation states changes which\ncomparisons reverse, and reversals also occur when both executables are scored\non the same visited states. We provide an execution-bound qualification protocol\nand provenance-recording harness that compare paired task outcomes against\ntask-specific tolerances. The results distinguish fidelity-based candidate selection\nfrom task-tolerance approval and support evaluating the measured executable on\nits intended execution path rather than fixed-state fidelity alone.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "reinforcement learning", "closed-loop evaluation", "neural policy deployment", "execution-bound qualification", "reproducibility" ], "forum_url": "https://openreview.net/forum?id=5R8LqE51Hr", "pdf_url": "https://openreview.net/pdf/e0c8d5f43ba5b20802fcd7b7cb22465ce3521498.pdf", "updated_at": "2026-10-05T19:26:57.031Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Bp5sMtU8BH", "number": 17844, "title": "Repurposing Influence Diagnostics for Low-Bit 3D Object Detection Models", "abstract": "3D object detection is essential for instance-level scene understanding, yet deploying computationally intensive detectors on resource-constrained devices often requires low-bit quantization. However, reducing bit width remains challenging because 3D detectors exhibit activation distributions in which high-magnitude outliers are not uniformly informative. While some outliers carry task-relevant information, others arise from incidental or nuisance patterns. Such uninformative outliers can unnecessarily broaden the activation range, leading to coarse discretization and feature distortion under low-bit quantization. To address this issue, we repurpose influence diagnostics for quantization, where influence reflects how strongly an individual activation affects task-relevance. Our approach builds on the statistical perspective that informative outliers are crucial for preserving the properties of the data distribution. This perspective aligns with the classical notion of influential observations, motivating us to identify informative outliers by assessing their impact on task-relevant information. To make this influence task-aware, we use activation gradients to capture the relevance of individual activations to the task objective. Because these gradients contain both dominant task-relevant patterns and incidental variations, we model their shared structure with a low-rank representation that retains the principal patterns to be preserved. We then assess activation influence by measuring how strongly each activation affects the resulting gradient structure. Experiments on the nuScenes benchmark demonstrate improvements across W8A8, W4A16, W4A8, and W4A4 settings, with larger gains at lower bit widths. These results suggest that activation influence provides an effective criterion for identifying informative outliers in low-bit quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "3D Object Detection", "Efficient AI", "Statistical Signal Processing", "Regression Diagnostics" ], "forum_url": "https://openreview.net/forum?id=Bp5sMtU8BH", "pdf_url": "https://openreview.net/pdf/be16930aed4aafb5de57f2f4e7a3ceb5aad66610.pdf", "updated_at": "2026-10-05T19:27:07.347Z", "categories": [ "training" ], "matched_terms": [ "bit width", "low-bit", "quantization", "w4a16", "w4a4", "w4a8", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lOAjZwcSJk", "number": 17875, "title": "KVarN: Variance-Normalized KV-Cache Quantization for Improved Reasoning in LLMs", "abstract": "Test-time scaling improves language-model reasoning but increases KV-cache memory during long-horizon generation. Existing KV-cache quantization methods are commonly evaluated in static prefill settings, which do not capture feedback from repeatedly attending to a quantized cache. We show that this feedback increases attention-output error over time and identify token-scale distortion as a major contributor to the largest key-quantization errors. We introduce KVarN, a calibration-free, linear KV-Cache quantization method to effectively suppress token scale errors and error accumulation. At an effective storage cost of 2.25 bits per element, KVarN achieves the best overall quality against strong quantization baselines on AIME 2024, MATH-500, HumanEval, and IFEval. It also reduces attention-output error under our pseudo-decode diagnostic. In end-to-end vLLM measurements across three models, KVarN improves decode throughput by 10–28\\% over the 16-bit KV-cache baseline; at a nominal 3-bit precision, it achieves 1.44–2.00× the throughput of the vLLM TurboQuant implementation. Code for the KVarN method is given in the supplementary.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "kv-cache", "inference", "LLM", "reasoning", "compression" ], "forum_url": "https://openreview.net/forum?id=lOAjZwcSJk", "pdf_url": "https://openreview.net/pdf/9fb1b367036f80c317311875964a0bd35fa584a4.pdf", "updated_at": "2026-10-05T19:27:25.963Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "3-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZoVv8NH8sp", "number": 17886, "title": "DraftAttention2: Fast Video Diffusion with Low-Resolution-Guided Mixed-Precision Attention", "abstract": "Video generation has broad applications in content creation and entertainment. Diffusion transformers have advanced the quality of generated videos, but attention over spatiotemporal tokens becomes increasingly expensive as video resolution and duration increase. We present DraftAttention2, a training-free framework that uses the low-resolution draft attention map to jointly select attention blocks and assign their numerical precision. Specifically, spatial 2D average- and max-pooled queries and keys capture complementary regional statistics to estimate block importance, and a shared ranking assigns higher precision to important blocks, lower precision to less important retained blocks, and skips the rest under configurable budgets. Our analysis separates sparsification error from attention-weighted quantization error, establishing when recovering skipped interactions with low-bit computation tightens the output-error bound. This analysis motivates retaining more interactions at low precision while reserving higher precision for blocks with larger attention mass. To translate these fine-grained assignments into practical speedups, we further develop fused operand preparation and a single attention kernel with precision-specific phases, sharing data movement, softmax statistics, and output accumulation across precisions. Experiments demonstrate that our method achieves a superior quality–efficiency trade-off over existing efficient video generation methods. Notably, its advantage is particularly pronounced for few-step video diffusion, where jointly combining sparsity with 4- and 8-bit mixed-precision computation substantially improves generation quality while retaining significant acceleration.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "video diffusion", "diffusion transformers", "sparse attention", "mixed-precision attention", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=ZoVv8NH8sp", "pdf_url": "https://openreview.net/pdf/f1a336c375054c17ea5f8b9aa051bf4e1f3aac96.pdf", "updated_at": "2026-10-05T19:27:28.980Z", "categories": [ "hardware", "diffusion", "theory" ], "matched_terms": [ "8-bit", "low precision", "low-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AlEeRqfTB4", "number": 17923, "title": "Predicting Compression Gains Before Encoding: Attribute Models for 2D Gaussian Images", "abstract": "Gaussian image representations, which encode an image as a set of 2D Gaussians, have become a practical representation for image compression. Compressing this set involves two interlinked design choices: the spatial organization of the Gaussians and the probability model for their attributes. Such choices are typically made empirically, by building candidate coders and comparing total file size. A controlled study of lossless coding of quantized Gaussian sets shows that this practice can mislead: among the structures tested, a quadtree hierarchy codes positions most compactly, yet loses overall because of the attribute model paired with it. We therefore propose a predict-before-encoding protocol: it measures, on a selection split, whether the information a quadtree cell carries about each attribute outweighs the cost of learning a separate model per cell, and fixes the answer before any evaluation image is encoded. In a prospective study on GaussianImage representations, the predicted gain over conditioning every attribute, 1.414 bits per Gaussian, closely matches the realized gain of 1.361, and the protocol also correctly forecast the one comparison the chosen coder loses.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Gaussian Splatting", "Image Compression", "Entropy Coding", "Information Theory", "Minimum Description Length", "Adaptive Coding" ], "forum_url": "https://openreview.net/forum?id=AlEeRqfTB4", "pdf_url": "https://openreview.net/pdf/aeb8c1dfdf107e6c5601406e0af09f395ebf7fc6.pdf", "updated_at": "2026-10-05T19:27:46.300Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DMru42Hepp", "number": 17947, "title": "Unlocking the Critic: Reward-Free Policy Optimization for LLM Post-Training", "abstract": "Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions therefore provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while substantially cutting compute and memory overhead. This makes RFPO particularly well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "reinforcement learning for LLMs", "LLM post-training", "policy optimization", "PPO", "value function", "critic", "reward-free reinforcement learning", "long chain-of-thought reasoning" ], "forum_url": "https://openreview.net/forum?id=DMru42Hepp", "pdf_url": "https://openreview.net/pdf/5348fba5845c110ede91acab387cf51451891eda.pdf", "updated_at": "2026-10-05T19:28:01.112Z", "categories": [ "ptq", "reasoning", "theory" ], "matched_terms": [ "binarized", "binarizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "918uH63DhR", "number": 18003, "title": "Spectral-TCAV for Pairwise Concept-Based Interpretability in Raman Classifiers", "abstract": "Testing with Concept Activation Vectors (TCAV) interprets neural network predictions in terms of human-defined concepts, but has not been applied to Raman spectroscopy, where interpretability is usually limited to individual spectral bands or perturbation responses. We introduce Spectral-TCAV, adapting TCAV to Raman spectra for bacterial species classification, and show that it recovers spectrally meaningful concepts underlying model predictions. We further extend the framework from single spectral regions to pairs of regions, decomposing their interaction into three contrasts: Sum (joint magnitude), Diff (relative dominance), and Tradeoff (antagonistic behavior, isolated by masking for opposite-sign changes). Each contrast is binarized and evaluated with independent linear CAVs and TCAV directional-derivative scores tested against null distributions from randomly sampled regions. Across classes, the same region pair can diverge in significance and direction across contrasts, showing that Sum, Diff, and Tradeoff capture complementary information beyond a single concordant-versus-discordant label.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "interpretation", "explainable AI", "concepts", "chemistry", "interpretability", "TCAV", "learned representations" ], "forum_url": "https://openreview.net/forum?id=918uH63DhR", "pdf_url": "https://openreview.net/pdf/d8ed1c7ed039f8b35fc408a760f285e3681fe7fa.pdf", "updated_at": "2026-10-05T19:28:33.205Z", "categories": [], "matched_terms": [ "binarized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cXZSKKEsHb", "number": 18055, "title": "Rambling on Repeat? Exact First-Exit Laws and Minimal Intervention for Verbatim LLM Loops", "abstract": "Modern LLM generations occasionally enter verbatim repetition and run to their output caps, wasting user token budgets, increasing request latency, exhausting computing power, and exposing a system vulnerability. Existing repetition penalties and $n$-gram bans are global heuristics that distort healthy text but neither estimate a detected loop's persistence nor certify exit. We combine serving-stack signals and rigorous first-exit theory to quantify and control this behavior with guarantees on minimum intervention cost. Temperature-$T$ decoding samples the Boltzmann distribution over token energies, making the runner-up gap $\\Delta$ an activation energy. Hazard products along one greedy trace give the exact finite-horizon distribution of exit time and phase at any temperature. If a measured lap repeats indefinitely, escape obeys an Arrhenius law governed by the cycle's minimum gap, with exits concentrating at that weakest phase. The parameter-free law is validated across 37 orbits from six families up to a 284B MoE model including its 3-bit quantization. Inverting it yields two training-free, event-triggered controllers for a user-specified risk tolerance $\\delta$: a provably minimal-temperature pulse that exits within a chosen horizon, and a stepwise minimum-KL guard that contains every detected trajectory within $L$ tokens at minimum compatible distortion, each with successful-exit probability at least $1-\\delta$. Deployed as a vLLM logits processor, the untuned guard raises Qwen3-4B sustained escape from 0.71 to 0.91 and reduces cap exceedance on 710 matched natural-workload prefixes from 30.7% to 5.1%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "LLM degeneration", "decoding-time control", "Arrhenius law", "minimum-KL intervention", "vLLM" ], "forum_url": "https://openreview.net/forum?id=cXZSKKEsHb", "pdf_url": "https://openreview.net/pdf/fe20c358a38901d614336e65b374ec6cee91bdc2.pdf", "updated_at": "2026-10-05T19:29:05.362Z", "categories": [ "hardware" ], "matched_terms": [ "3-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "b6NjjIgfsT", "number": 18075, "title": "How Extractable Are Edge LLMs?", "abstract": "Edge large language models (Edge LLMs) are commonly quantized for deployment on devices with limited memory and compute. At the same time, their text interfaces allow an attacker to collect responses and train a student model, creating a risk of black-box behavioral extraction. For such attacks, collecting more target responses provides additional supervision; however, slow on-device inference makes large-scale data collection costly. We therefore investigate whether an attacker can extract instruction-following behavior more effectively under a limited target-query budget. To assess this risk, we introduce Coordinated Linguistic Instruction Querying (\\textbf{CLIQ}), a structured probing framework designed to reduce repeated probing of similar behaviors by coordinating queries across semantic regions. Specifically, CLIQ partitions a public instruction pool and generates region-conditioned queries before observing any target response. We compare it with Original Queries (\\textbf{OQ}), which are randomly sampled from the same pool and used without modification. Under a nominal 500-query budget, CLIQ improves a 4B student's BERT-F1 from \\textbf{77.92} with OQ to \\textbf{82.44} on Alpaca evaluation references. Furthermore, experiments with two phone-deployed Gemma teachers show improved student performance over OQ while target-side generation speeds remain comparable. Taken together, these results demonstrate that functional instruction-following behavior remains transferable from quantized edge models under the evaluated query budgets: the cost of collecting target responses \\textbf{does not, by itself, establish resistance to behavioral extraction}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "edge LLMs", "quantization", "model extraction", "knowledge extraction", "query-efficient distillation", "structured querying", "security" ], "forum_url": "https://openreview.net/forum?id=b6NjjIgfsT", "pdf_url": "https://openreview.net/pdf/c32521520f360a325568ef1fa0dca3c8f01d654d.pdf", "updated_at": "2026-10-05T19:29:12.482Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FMo2cupFU2", "number": 18101, "title": "Freeze or Dither: Rounding-Aware Precision Floors for Low-Precision Training", "abstract": "Convergence theory for low-precision training has produced only upper\nbounds, in a worst-case relative-error model blind to the rounding rule. We\ngive the first lower bounds for training with floating-point weight\nstorage, organized by one principle: a coordinate survives\nround-to-nearest (RN) when its update magnitude either exceeds the local grid\nspacing $u$ or carries noise at that scale, and freezes or floors when it does\nneither --- freeze or dither.\nSign-type updates can never dither. Below an explicit threshold they freeze\nbitwise, weight growth stops at a hard ``binade wall,'' and their\nmovement-weighted noisy floor is exactly $\\sqrt{\\pi/2}/2\\cdot u\\sigma/L$\nunder both RN and stochastic rounding (SR), for arbitrary predictable\nstep-size schedules, so at the stationary floor SR provably cannot help sign\ndescent. SGD self-dithers: RN's rounding bias collapses as\n$e^{-2\\pi^2(\\eta\\sigma/u)^2}$, which puts homogeneous-noise SGD--RN on the\nuniversal scale. But un-ditherable low-noise coordinates escape the rescue,\nand under heterogeneous noise with a shared constant step SGD--RN floors at\n$\\Theta\\big((u\\sigma^2/L^2)^{2/3}\\big)$, strictly worse than SGD--SR, so no\nrounding-blind analysis can be tight for both rules. A representation floor\nseparately forces $M=\\Omega(\\log T)$ mantissa bits; with the $O(\\log T)$\nsufficiency of tang2026convergence the requirement is $\\Theta(\\log T)$.\nThe floors are measured to their constants, and the wall is visible in real\ntraining: in a $124$M-parameter GPT-2 storing weights in bf16 with no master\ncopy, RN pins every LayerNorm gain at exactly its initialization of $1.0$ (a binade boundary) in $8$ of $9$ (optimizer, seed) cells, SR pins\nnone; at the tuned budget the held-out RN/SR ratios order, in every seed, as\nditherability predicts, and SR recovers the fp32 master's loss at a third of\nthe weight memory. All code, logs and figures behind every number reported here\nare in the supplementary material and will be released publicly upon\nacceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-precision training", "quantization", "stochastic rounding", "round-to-nearest", "lower bounds", "convergence analysis", "mantissa bits" ], "forum_url": "https://openreview.net/forum?id=FMo2cupFU2", "pdf_url": "https://openreview.net/pdf/3b9c562f5a13eb902f7906df933e19c0c40634b6.pdf", "updated_at": "2026-10-05T19:29:24.242Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wrjRMndnIC", "number": 18179, "title": "ShadowMoE: Graph-Conditioned Shadow Expert Representations for Mixed-Precision MoE Quantization", "abstract": "Sparse mixture-of-experts (MoE) models expose a structured trace of computation through token-level routing: experts are not used independently, but repeatedly cooperate across layers. We ask whether this routing topology contains information about expert vulnerability to model compression. We introduce ShadowMoE, a topology-aware representation framework that converts token-level\ntop-k routing traces into weighted cross-layer expert graphs and learns relational representations of individual experts. A graph-restricted encoder propagates information only along observed routing paths, while leakage-controlled structural objectives test whether the learned representation captures non-local routing information. We further probe whether these representations predict expert-level quantization residuals and use them to construct budget-constrained mixed-precision policies. On Qwen3-30B-A3B and Mixtual V0.1, ShadowMoE achieves a strict link-prediction AUC of 0.9432, compared with 0.8378 for a graph-free MLP, and its expert representations explain R2 = 0.5405 and 0.5953 of FP8 and INT4 local quantization residuals, respectively. At an average 5.0-bit budget, the resulting INT4/FP8 allocation retains 99.43% and 98.33% of FP8+activation quantized accuracy on the two models, respectively. Importantly, independently\nestimated expert risks do not reliably compose into scheme-level degradation: the correlation between predicted local risk and complete-scheme loss is only.\n0.0580. This negative result suggests that compression sensitivity in MoEs is\nrelational rather than purely expert-wise. ShadowMoE therefore provides both\na practical compression mechanism and a framework for studying how routing\ntopology encodes functional dependencies among experts. Anonymous artifacts:\nhttps://anonymous.4open.science/r/topology_0723-70B3.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixed-precision quantization", "Graph Transformer", "Expert-Aware Model Compression" ], "forum_url": "https://openreview.net/forum?id=wrjRMndnIC", "pdf_url": "https://openreview.net/pdf/7accd99c59abad67de70d2ad160953f9241ff0ef.pdf", "updated_at": "2026-10-05T19:30:08.343Z", "categories": [ "training" ], "matched_terms": [ "fp8", "int4", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DQ8EGp92sS", "number": 18241, "title": "Preserving Regularity in Finite-Precision Transformer Approximation", "abstract": "Transformers are the dominant architecture for large language models and are widely used in multimodal foundation models. Their substantial storage requirements have made low-bit quantization an important tool for practical deployment. Yet quantitative Transformer approximation theory predominantly treats weights as unconstrained real numbers, leaving a gap between theoretical guarantees and finite-precision representations. This question goes beyond approximation accuracy: quantization can increase a model's sensitivity to small input perturbations even when the approximation error remains small. It is therefore important to understand whether finite-precision Transformers can simultaneously achieve accurate approximation, controlled storage complexity, and preservation of a prescribed regularity bound.\nIn this paper, we construct a finite-precision Transformer that approximates matrix-valued $\\lambda$-Hölder functions on $[0,1]^{d\\times n}$ while preserving a prescribed componentwise $\\ell^1$-Hölder bound $L$. Its hidden feature weights are discrete at $\\{0,\\pm1,\\pm\\tfrac12\\}$ and target-independent, while target dependence is confined to the output coefficients. This separation allows us to directly control the effects of quantization on both approximation accuracy and regularity.\nFor $D=dn$, the resulting Transformer achieves uniform error $\\varepsilon$ with $\\mathcal O(\\varepsilon^{-D/\\lambda})$ nonzero parameters and a common $\\mathcal O(\\log(1/\\varepsilon))$-bit fixed-point format for all parameters. \nThese guarantees concern parameter storage under exact forward evaluation. \nAs an application, empirical risk minimization over a structured finite-precision Transformer class achieves expected squared prediction risk $\\mathcal O((\\log(em)/m)^{2\\lambda/(2\\lambda+D)})$ under sub-Gaussian noise, matching the classical Hölder minimax exponent up to a logarithmic factor.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Approximation theory; Transformer networks; Finite-Precision Networks; Nonparametric Regression; Expressive Power" ], "forum_url": "https://openreview.net/forum?id=DQ8EGp92sS", "pdf_url": "https://openreview.net/pdf/086e34e316c75adce9fc5bb0e1ef57f7e866df63.pdf", "updated_at": "2026-10-05T19:30:41.526Z", "categories": [ "hardware", "multimodal", "theory" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mxWajphwkT", "number": 18374, "title": "SparseSliderQuant: Sparse-Quantized LLM with Cross-Block Error Coordination", "abstract": "Sparse quantization is an attractive compression target for Transformer decoders because it simultaneously reduces numerical precision and arithmetic density. However, aggressive configurations such as W4A4 quantization with 2:4 structured sparsity expose a fundamental gap between local compression and cross-block consistency. To bridge this gap, we propose SparseSliderQuant, a unified rewrite framework that coordinates sparsification and quantization errors across Transformer blocks. SparseSliderQuant first employs sliding group alternation to iteratively optimize sparse masks and quantization scales within overlapping block windows, enabling the two compression processes to adapt to each other rather than being optimized independently. It then performs dual-source mismatch calibration by explicitly decomposing pruning and quantization residuals, accumulating their mismatch in a residual buffer, and redistributing it to the remaining unpruned weights. Finally, cross-block gradient anchoring uses next-block output drift to jointly guide mask and scale updates, aligning local reconstruction with downstream behavior throughout the decoder. SparseSliderQuant introduces no auxiliary modules or additional inference-time overhead. Experiments on LLaMA and Qwen models show that it consistently preserves language-model performance across model scales and a broad range of language understanding and reasoning tasks under W4A4 quantization with 2:4 structured sparsity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Structured Sparsity", "Model Quantization", "Cross-Block Optimization" ], "forum_url": "https://openreview.net/forum?id=mxWajphwkT", "pdf_url": "https://openreview.net/pdf/d54d098801322b083c6b32f814cfb1794475d137.pdf", "updated_at": "2026-10-05T19:32:01.501Z", "categories": [ "ptq", "training", "reasoning" ], "matched_terms": [ "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fiXL6fkcoV", "number": 18386, "title": "RACC: Coupling Disagreement Routing with Receiver-Conditioned Messages under Bit Caps", "abstract": "A scarce multi-agent channel must choose both which receiver warrants a packet and what receiver-specific information that packet should contain. Receiver-Aware Coupled Communication (RACC) makes these choices from one sender-side receiver prediction: a thresholded forward KL between sender and predicted-receiver policies activates a directed edge, while the predicted receiver state and identity condition its binary $B$-bit message. On three hard SMAC maps and three packet widths, RACC exceeds BVME in all nine cells by 1.4--3.6 win-rate points under common widths; after independently matching mean charged payload bits, the gains remain 1.0--3.0 points and every paired interval excludes zero. Traffic-matched interventions expose the contribution of each connection: replacing discrepancy routing with a random gate loses $3.8 [2.4,5.1]$ pooled tactical points, sharing one sender payload across receivers loses $4.8 [3.2,6.3]$, and shuffling receiver identities loses $5.5 [3.9,7.1]$. At $B{=}32$, RACC reaches 75.4%, 65.1%, and 90.8% SMAC win rate, 3.2--4.8 points above an encoder-, quantizer-, and width-matched always-send control, and leads every evaluated communication baseline in two- through five-player Hanabi. On held-out SMACv2 scenarios spanning 5, 8, and 11 agents, two graph densities, and three packet widths, it averages $3.7 [2.5,4.8]$ points above BVME. Receiver prediction can therefore serve as a shared computational interface between packet allocation and personalized content when asymmetric information makes directed bits consequential.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "world models and rl", "RACC" ], "forum_url": "https://openreview.net/forum?id=fiXL6fkcoV", "pdf_url": "https://openreview.net/pdf/ac62f456e2de2bd5db4dc7ead9ecc1477725ef6c.pdf", "updated_at": "2026-10-05T19:32:14.175Z", "categories": [], "matched_terms": [ "binary", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a7wnbAxueO", "number": 18387, "title": "PairCast: Direct Counterfactual Packet Value for Multirate Swarm Communication", "abstract": "Freeze a swarm one control cycle before an occluded neighbor becomes collision-relevant: decoding a short update can redirect several receivers, while an equally novel packet can leave every action unchanged. Scheduling with an integrated controller and packet format therefore requires the control consequence of each available rate, not packet salience alone. Counterfactual packet value (CPV) is the delivery-adjusted, sender-local return gain per attempted bit between successful decoding of a candidate prefix and continued silence. PairCast estimates CPV for nested 1--8-bit prefixes by branching a permutation-equivariant neighborhood world model into paired multi-step futures, then admits the highest-density prefix through a shared reservation price and a one-step-lagged occupancy signal. On 20-agent formation under occlusion, 30-agent dynamic coverage, and 10-versus-3 pursuit-evasion, it reaches 95.9--97.4% normalized return with roughly 48% of Dense Broadcast traffic; at the matched 48.2% operating point, it leads Adaptive self-triggering by 1.9--3.4 taskwise points and a SchedNet-style multirate scheduler by 2.1 pooled points [1.0, 3.1]. Consequence estimation is predictive on 45000 held-out interventions (0.66 Spearman correlation, 78.4% sign accuracy, and 0.18 normalized decision regret), while replacing it with uncertainty scoring costs 4.8 return points and reducing its paired horizon from five steps to one costs 2.1. Shared-prefix execution retains a 1.6-point advantage within an 8.9 ms A100 p99 envelope, and matched-load gains remain 1.8--2.2 points throughout the communication-scarce 40--60% regime. Directly valuing how a decode changes the team's future provides a practical scheduling principle for multirate coordination under shared-medium contention.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "world models and rl", "PairCast" ], "forum_url": "https://openreview.net/forum?id=a7wnbAxueO", "pdf_url": "https://openreview.net/pdf/8f599059792bc429766d594fb707211d02020a1a.pdf", "updated_at": "2026-10-05T19:32:14.285Z", "categories": [], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aTtR1G9Xnv", "number": 18434, "title": "CoSR: Memory-Efficient FP8 Training for Multitask Reinforcement Learning", "abstract": "FP8 reduces compute and memory costs in large-scale deep learning, but bootstrapping, nonstationary targets, and closed-loop feedback between policies and data distributions make deep reinforcement learning sensitive to low-precision perturbations.\nHigh-capacity multitask critics further increase training-state memory costs.\nWe find that deterministic writeback in FP8-resident online actor--critic training can suppress updates smaller than the quantization step, degrading long-term performance.\nIndependent stochastic rounding (SR) preserves these updates in expectation but ignores error interactions in forward and backward matrix operations and Adam state updates, allowing locally unbiased noise to accumulate into substantial optimization perturbations.\nWe propose CoSR, an operator-aware joint SR framework.\nCoSR uses forward/backward channel correlations and the coupled effects of Adam's first and second moments on update directions to coordinate correlations among rounding errors while preserving elementwise rounding probabilities and unbiasedness.\nIt requires no per-parameter compensation buffers, reducing persistent-state memory overhead; major weight matrices and Adam states reside in FP8.\nAcross diverse reinforcement-learning benchmarks, including DeepMind Control Suite and Meta-World, CoSR achieves performance comparable to the high-precision baseline while reducing training-state memory and peak GPU memory by 74\\% and 63\\%, respectively.\nThese savings enable wider critics, yielding substantial multitask performance gains under the same hardware budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Precision Training", "Deep Reinforcement Learning", "FP8", "Memory-Efficient Training", "Multitask Reinforcement Learning" ], "forum_url": "https://openreview.net/forum?id=aTtR1G9Xnv", "pdf_url": "https://openreview.net/pdf/20162873f2355ebb263ec88c18631dfb46d6b9ca.pdf", "updated_at": "2026-10-05T19:32:34.055Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp8", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "m4y8Kcquau", "number": 18491, "title": "RExMap: Counterfactual Recovery for Model-Edit Rollback after Composition and Serving Drift", "abstract": "A revoked edit can disappear from the live weights while its influence survives inside every later update optimized in its presence. RExMap recovers the corresponding counterfactual history: its ledger unwinds the retained suffix, removes the revoked update, re-estimates each retained request on the corrected state, and evaluates the recovered model on the edited fact, 500 clean prompts, 200 attacked probes, and an off-target entity neighborhood. In 24,000 Llama-3.1-8B transactions, exact all-conditions restoration falls from 97.4% at depth 1 to 82.8% at depth 20 in FP32, 68.7% under INT8, and 60.9% with dual overlap and non-LIFO revocation. Re-estimating retained descendants recovers 12.0 points over direct subtraction at depth 20, identifying state-dependent descendants as a correctable source of rollback error. A transaction-state score predicts behavioral validity with 0.914 AUROC and a 2.7% false-valid rate; routing uncertain cases through checkpoint replay raises restoration by 22.0 points over uniform subtraction to 97.8%, with a 60.1% direct fast path and 0.31 s median rollback. The depth and serving-representation ordering recurs across a reversible ROME-style writer, Mistral-7B-v0.3, and MQuAKE. RExMap therefore turns reversibility from a write-time algebraic property into a behaviorally testable recovery decision for evolving edited models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "memory editing and lifelong learning", "RExMap" ], "forum_url": "https://openreview.net/forum?id=m4y8Kcquau", "pdf_url": "https://openreview.net/pdf/cd1e32524c0a9e3a7b8cf873e291a554c0996853.pdf", "updated_at": "2026-10-05T19:32:47.241Z", "categories": [], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Lb0T1d1L63", "number": 18546, "title": "Action-Contract Preservation: Detecting and Repairing Hidden Decision Drift in Quantized LLMs", "abstract": "A quantized checkpoint enters an action-bearing system through a controller whose score boundaries dispatch tools, invoke retrieval, abstain, or stop generation. Standard benchmark recovery does not determine whether those boundaries still implement the FP16 action policy. We formulate this deployment question as Action-Contract Preservation (ACP): a paired, family-wise test that freezes action semantics, arbitration, operating-point selection, report examples, and repair permissions, and separately checks controller preservation and externally labeled action error. A boundary--perturbation decomposition then isolates two causes of drift---reference requests near a frontier and quantization error larger than their available margin---and motivates BoundaryLock, which combines sensitivity-guided 3/4/8-bit allocation with signed-margin training. Applied to four Llama and Qwen checkpoints, ACP yields explicit qualification decisions. At 90% BFCL recall, it rejects GPTQ with threshold search on Llama-3.1-70B and Qwen-2.5-72B: false tool triggers rise from 8.2% to 11.9$\\pm$0.5% and from 7.6% to 10.9$\\pm$0.4%, respectively, although MMLU stays within 0.8 points of FP16. The paired audit also detects drift in coupled routing and labeled retrieval and termination endpoints. Against fully crossed QAT with threshold search---matched on data, 500 updates, parameter-weighted bit budget, and threshold rule---BoundaryLock lowers frontier crossing by 1.8--2.3 points and false tool triggers by 0.8--1.7 points across all checkpoints; it retains reported benchmark quality and 1.73--1.84$\\times$ FP16 throughput. ACP therefore turns a benchmark-close replacement into an explicit action-contract decision, and its geometry supplies an effective target for low-bit repair.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "foundational causal formal and efficient methods", "Action-Contract Preservation" ], "forum_url": "https://openreview.net/forum?id=Lb0T1d1L63", "pdf_url": "https://openreview.net/pdf/b26ce6f26bcc686049b0340673a840e40ed48c47.pdf", "updated_at": "2026-10-05T19:33:00.262Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "8-bit", "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lKwZOLWjjE", "number": 18576, "title": "TurboT2VA: Fast Large-Scale Text-to-Video-Audio Generation via Score-Regularized Consistency Distillation", "abstract": "Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Compared with discrete-time consistency, continuous-time consistency avoids finite-step discretization errors in the training target and provides more faithful supervision along the teacher trajectory, making it particularly attractive for few-step distillation. However, scaling continuous-time consistency to a 19B joint T2VA model is challenged by modality-imbalanced optimization, numerically fragile trajectory-tangent estimation, and the quality--diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive dCM$\\rightarrow$sCM$\\rightarrow$sCM+DMD curriculum, where dCM establishes a stable initialization, sCM more faithfully learns the continuous teacher trajectory while preserving diversity, and DMD improves perceptual quality. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\\times$768, achieving a 20.1$\\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths without retraining. Under the high-resolution deployment setting at 1024$\\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\\times$ generator-only speedup over the original teacher.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Text-to-video-audio generation", "consistency distillation", "distribution matching", "multimodal generation", "inference acceleration", "quantization." ], "forum_url": "https://openreview.net/forum?id=lKwZOLWjjE", "pdf_url": "https://openreview.net/pdf/3fad52e68daeb662075e9aa7501cba7f0b053779.pdf", "updated_at": "2026-10-05T19:33:08.125Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "quantization", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6Wg0BsKrDq", "number": 18603, "title": "E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models", "abstract": "Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored. We propose **Enhanced Mixture-of-Experts (E-MoE)**, which builds the reverse process as a mixture of factorized distributions over a *discrete shared latent* given by the expert-routing decisions of a mixture-of-experts backbone, without increasing the number of active parameters relative to the factorized baseline. On synthetic multi-modal benchmarks, binarized MNIST and LM1B, E-MoE substantially improves few-step generation over factorized MDM baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Discrete diffusion language models", "Masked diffusion models", "Mixture-of-experts", "Non-factorized generation", "Few-step generation", "Latent variable models" ], "forum_url": "https://openreview.net/forum?id=6Wg0BsKrDq", "pdf_url": "https://openreview.net/pdf/f89ca3c4e50200af78dc3706f34988f9a59a19d8.pdf", "updated_at": "2026-10-05T19:33:21.098Z", "categories": [ "diffusion", "multimodal" ], "matched_terms": [ "binarized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xWE4ZqljZe", "number": 18623, "title": "MUQuant: Mask State Adaptive Dual-Center Quantization for Diffusion Language Models", "abstract": "Block-diffusion language models repeatedly process partially masked blocks during generation. Within a decoding step, masked and unmasked tokens can have distinct activation centers, making a shared quantization origin poorly suited to either population. Calibration on fully unmasked text also misses the activation mixtures encountered during denoising. **MUQuant** addresses these two sources of W4A4 error with a state-conditioned representation. **MuZero** estimates separate centers for masked and unmasked tokens, and **MuCal** calibrates weights on the reconstructed activations collected along denoising rollouts. **MuEngine** then computes the center and residual contributions in one packed four-bit multiplication, avoiding separate weight passes. Across three diffusion language-model families, this approach improves quantized accuracy while preserving the decoding procedure. On Nemotron-Labs-Diffusion-8B, the evaluated NVFP4 implementation achieves up to **3.24× speedups over BF16**.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Large Language Models" ], "forum_url": "https://openreview.net/forum?id=xWE4ZqljZe", "pdf_url": "https://openreview.net/pdf/5cc615e155b9b91f7d5feb72a6b89c0b8a0ea656.pdf", "updated_at": "2026-10-05T19:33:36.160Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "nvfp4", "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mdXHKyhjvR", "number": 18641, "title": "Functional Subspace Watermarking for Large Language Models", "abstract": "Model-side watermarking protects the ownership of large language models (LLMs) by leveraging model weights or internal representations. However, these internal signals often undergo complex distortions during practical model modifications, such as fine-tuning, quantization, pruning, and knowledge distillation, making reliable watermark extraction extremely challenging. Although model-side watermarking has been widely studied, existing methods still lack sufficient robustness against parameter-level and representation-space perturbations. To address this issue, we propose Functional Subspace Watermarking (\\texttt{\\textbf{FSW}}), which embeds multi-bit watermark information into a low-dimensional functional backbone of LLM internal representations. Specifically, \\texttt{\\textbf{FSW}} solves a generalized eigenvalue problem to jointly model functional sensitivity and compression invariance, thereby extracting watermark directions that are both task-critical and stable under common model modifications. We further introduce an adaptive spectral truncation strategy to select an appropriate spectral band that balances robustness and model utility, while incorporating a vector consistency constraint to mitigate representation drift during watermark injection. Experiments show that \\texttt{\\textbf{FSW}} achieves 100\\% bit recovery accuracy across diverse LLM backbones and, while maintaining competitive utility on most evaluated backbones, preserves watermark detectability under LoRA fine-tuning, noise perturbation, pruning, quantization, and backbone-preserving distillation. In addition, we further evaluate FSW under untargeted representation perturbation and homogeneous output-only distillation. The results show that, under the tested budgets, FSW retains statistically detectable watermark signals across multiple challenging configurations, demonstrating its robustness to common model modifications while characterizing its practical robustness boundary.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model watermarking", "model ownership verification", "functional subspace", "robust watermarking" ], "forum_url": "https://openreview.net/forum?id=mdXHKyhjvR", "pdf_url": "https://openreview.net/pdf/d43b74c413abed0daa4297d5ddf49f71761e8ce5.pdf", "updated_at": "2026-10-05T19:33:38.718Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8iTqrW1sxY", "number": 18853, "title": "Stable Directions, Different Effects: Intervention Fidelity Across Model Precision", "abstract": "Activation interventions can be characterized on one numerical realization of a language model and later applied to a\nquantized version. Preserving standard model quality or direction-level representation similarity does not guarantee that the\nintervention effect is preserved. We formalize this as cross-precision intervention fidelity and define a signed interventionfidelity gap between native and transferred steering effects. Across several language models, precision changes, model\nsizes, steering estimators, and prospectively held-out concepts/tasks, we compare a prespecified directionalignment/simple-geometry baseline with local functional sensitivity. Intervention effects can change even when directions\nremain strongly aligned, whereas a standard first-order local approximation tracks the signed gap closely in the strongest\nheld-out evaluations. The novelty is not the Taylor expansion itself, but the intervention-fidelity problem and the empirical\naccuracy of this local account across the tested replication axes. A correction frozen before fresh evaluation improves\nreconstruction of native intervention effects and a prespecified sign-based decision. Finally, a prospective parsimony test\nfavors the compact first-order approximation over the richer predictor. Within the tested normalized-transfer regime,\nintervention reliability should therefore be evaluated functionally rather than inferred from ordinary quantization quality or\ndirection similarity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "activation steering", "model quantization", "intervention fidelity", "mechanistic interpretability", "representation interventions", "large language models", "cross-precision transfer" ], "forum_url": "https://openreview.net/forum?id=8iTqrW1sxY", "pdf_url": "https://openreview.net/pdf/6743ca5541343dcd06c2f8ec4134f5892f9baf8e.pdf", "updated_at": "2026-10-05T19:35:20.269Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nEuM4oO9hI", "number": 18873, "title": "CoBIT: Unified Optimization of Codebooks, Assignments, and Activation Composition for Extreme Low-Bit LLM Quantization", "abstract": "Post-training quantization is essential for efficient large model inference, yet preserving accuracy at ultra-low budget below 4 bits remains challenging. Clustering-based quantization provides adaptive non-uniform representations, but existing methods use limited information to construct codebooks and assign weights, while activation quantization is typically handled separately. We present \\textit{CoBIT}, a post-training quantization framework that optimizes weight clustering and activation quantization under a common layer-output reconstruction objective. For weights, CoBIT alternates Hessian-aware codebook construction with error-compensated assignment. For activations, CoBIT selects a per-token exponent--mantissa composition based on layer-output error. We introduce \\textit{PhaseLUT} to replace runtime composition search with a lookup, and a hardware-aware kernel that integrates composition selection, activation quantization, and low-precision computation. Across all four general-purpose models, CoBIT improves average accuracy over the strongest baseline in each setting by 4.8\\% and 35.9\\% points with 4-bit activations and 3- and 2-bit weight clustering, respectively. With 4-bit weight indices, it also achieves up to $1.53\\times$ end-to-end decoding speedup over dense FP16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Quantization", "Efficiency" ], "forum_url": "https://openreview.net/forum?id=nEuM4oO9hI", "pdf_url": "https://openreview.net/pdf/77702c2a91ab9bfeded5ab585e6d29bb165fe2a4.pdf", "updated_at": "2026-10-05T19:35:42.846Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "4-bit", "low-bit", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bITW1JwxpG", "number": 19041, "title": "XOR-Trellis: Ultra-Low-Complexity Dequantization and Curvature-Aware Hadamard-Free LLM Quantization", "abstract": "Trellis-coded quantization enables high-dimensional compression of large\nlanguage model (LLM) weights at ultra-low bit widths without the exponentially\nlarge codebooks required by conventional vector quantization. Practical\ndeployment, however, presents two challenges: reconstructing compressed weights\nat sufficient parallel throughput to avoid making dequantization an inference\nbottleneck, and maintaining quantization accuracy without costly incoherence\ntransformations. We address these challenges with two complementary techniques.\nFirst, we introduce an ultra-low-complexity trellis dequantizer that uses a\nstructured, hardware-efficient state-to-value mapping while preserving diverse\nreconstruction choices for trellis search. Second, we reformulate discrete\ntrellis path optimization with a curvature-aware objective that reflects model\nsensitivity directly in the original coordinate space. Together, these\ntechniques enable high-quality ultra-low-bit trellis quantization with\ninexpensive, highly parallel runtime reconstruction and without relying on\nHadamard-based incoherence processing.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Ultra-Low-Bit Quantization", "Vector Quantization", "Trellis-Coded Quantization", "Hardware-Efficient Inference", "Curvature-Aware Quantization" ], "forum_url": "https://openreview.net/forum?id=bITW1JwxpG", "pdf_url": "https://openreview.net/pdf/769d75acd1621e47877bf47d849c9cb29b7bed31.pdf", "updated_at": "2026-10-05T19:37:06.975Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low bit", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "v5Z8GhEgnv", "number": 19158, "title": "Text Scores Do Not Establish Performance on Lexically Non-Diagnostic Speech Tasks: A Qwen2-Audio Quantization Case Study", "abstract": "Text-output scores alone do not show whether quantization preserves performance when the transcript does not reveal the target label. We evaluate fixed 6- and 7-bit Qwen2-Audio-7B-Instruct allocations on 508 English-to-German FLEURS utterances and on 512 RAVDESS emotion clips from 16 speakers. The provenance-bound FLEURS rerun has selected-minus-FP16 BLEU and chrF intervals that all include zero. On RAVDESS, where the same two sentences occur equally often with every emotion label, canonical-label accuracy is -3.71% for 6 bit and -1.17% for 7 bit relative to FP16. The 6-bit speaker interval excludes zero and an exact two-sided sign-flip test gives p=0.0148; the 7-bit interval includes zero. Same-budget controls do not identify either selected allocation as best. This case study shows that a translation score does not establish performance on a speech task whose target cannot be recovered from the transcript.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "speech-language models", "quantization", "speech translation", "emotion recognition", "audio evaluation", "cross-task evaluation" ], "forum_url": "https://openreview.net/forum?id=v5Z8GhEgnv", "pdf_url": "https://openreview.net/pdf/b8fa12a37d933ebd04c54240c76c031c4b355b3b.pdf", "updated_at": "2026-10-05T19:38:03.131Z", "categories": [], "matched_terms": [ "6 bit", "6-bit", "7 bit", "7-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BBFrNtanEm", "number": 19225, "title": "The Geometry of Robust Symbolic Computation in Neural Networks", "abstract": "When can a neural network reliably carry out multi-step symbolic computation? Correct predictions on canonical symbolic inputs are not sufficient, because intermediate neural representations may fall outside regions in which subsequent modules preserve the intended symbolic semantics. We study this problem for fixed encoders and deterministic decoders, characterizing correct execution by containment of reachable neural states in the intended decoder cells. We establish three main results. A multi-step inner-reachability certificate treats the endpoint as a function of the disturbance sequence and can detect full-dimensional reachable sets even when every individual disturbance is rank deficient. For bounded encodings with the $\\ell_\\infty$ norm, we obtain an exact closed-ball capacity bound and a matching finite-ReLU construction for arbitrary finite deterministic transition families, without a network-size constraint. Finally, polyhedral module certificates reduce compositional verification to linear optimization over activation regions. An explicit separation shows that identical canonical tables and identical global Lipschitz constants can coexist with different robustness properties. Outer-radius bounds and quantization-aware recurrences provide complementary sufficient conditions. Experiments with learned arithmetic modules distinguish canonical accuracy from representation robustness and terminal accuracy from pathwise correctness. Additional tests find correct sampled executions despite inconclusive regional inclusion checks, and full numerical rank in multistep endpoint Jacobians under low-rank disturbances, but no positive inner-radius estimate centered at the target encoding.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Neural-symbolic computation; robustness; reachability analysis; compositional verification" ], "forum_url": "https://openreview.net/forum?id=BBFrNtanEm", "pdf_url": "https://openreview.net/pdf/002cd4c67b69128de70e0e5809bd4d0d21945243.pdf", "updated_at": "2026-10-05T19:38:47.447Z", "categories": [ "qat", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SHyLNNoz46", "number": 19290, "title": "G$^2$PTQ: Improving LLM Post-Training Quantization with Generalized Gradient Compensation", "abstract": "Post-training quantization (PTQ) is a practical approach to reducing the memory and computational footprint of large language models (LLMs) without retraining. GPTQ-based methods have become the de facto standard, yet they suffer from two complementary limitations. Methods with local, layer-wise objectives lack global supervision; while methods with global objectives fix their Hessian estimates at the start and ignore first-order gradients, so their guidance grows stale as quantization proceeds. This paper presents G$^2$PTQ, a unified PTQ framework with Generalized Gradient Compensation that integrates both first- and second-order information under a globally supervised, block-wise optimization objective. By refreshing gradient and Hessian estimates before quantizing each Transformer block, G$^2$PTQ avoids the staleness of prior global methods. Furthermore, to stabilize the exact first-order compensation, we introduce a trust-region scaling mechanism that dynamically bounds the gradient step to prevent exploding weight updates. Finally, we derive efficient implementations for block-wise Hessian approximation and exact gradient compensation. Experimental results on various model families and bit-widths demonstrate that G$^2$PTQ enables better alignment with the full-precision model, outperforming state-of-the-art baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Quantization", "Model Compression", "Large Language Model" ], "forum_url": "https://openreview.net/forum?id=SHyLNNoz46", "pdf_url": "https://openreview.net/pdf/3f2d3dbe44af6331ead32d689f93d97a020412a6.pdf", "updated_at": "2026-10-05T19:39:23.648Z", "categories": [ "ptq", "training" ], "matched_terms": [ "ptq", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SUjdBtZZ27", "number": 19325, "title": "Instance-Customized LoRA Generation via Discrete Priors for Efficient LLM Adaptation", "abstract": "As a mainstream Parameter-Efficient Fine-Tuning (PEFT) method, Low-Rank Adaptation (LoRA) significantly reduces the optimization and deployment costs of large language models by introducing trainable low-rank matrices. Despite the progress, existing LoRA and its variants typically learn a universal adapter at the entire task or dataset level, failing to achieve fine-grained personalized adaptation for individual instances, and thus limiting the performance in complex scenarios. In this paper, we propose InLoRA, a novel dynamic, instance-customized low-rank adapter considering individual idiosyncrasies. In particular, our InLoRA first maintains a discrete codebook with vector quantization to store the prior patterns of adapters. More importantly, we retrieve from the codebook for individual instances and then utilize lightweight projections to generate instance-customized yet well-structured adapters, which would be combined with task-specific adapters for efficient fine-tuning. Extensive experiments demonstrate that our InLoRA not only outperforms state-of-the-art PEFT methods, but also matches or even surpasses the performance of full fine-tuning on multiple benchmarks. The code and pre-trained models will be made publicly available.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Parameter-Efficient Fine-Tuning", "Low-Rank Adaptation", "instance-customized adaptation" ], "forum_url": "https://openreview.net/forum?id=SUjdBtZZ27", "pdf_url": "https://openreview.net/pdf/89a6da78e2297dcb451d0855fb286a6c87ebd897.pdf", "updated_at": "2026-10-05T19:39:48.075Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ot6rWMzuu6", "number": 19471, "title": "Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling", "abstract": "Post-training quantization (PTQ) is a practical path to deploy large diffusion models, but quantization noise can accumulate over the denoising trajectory and degrade generation quality. We propose Q-Drift, a principled sampler-side correction that treats quantization error as an implicit stochastic perturbation on each denoising step and applies the paired drift adjustment prescribed by generalized marginal-preserving SDEs. Unlike prior sampler-side corrections, which aim to cancel the quantization error of each individual sample, Q-Drift aims to preserve the distribution. Q-Drift estimates a timestep-wise variance statistic from calibration, in practice requiring as few as 10 paired full-precision/quantized calibration runs. The resulting sampler correction is plug-and-play with common samplers, diffusion models, and PTQ methods, while incurring negligible overhead at inference. Across six diverse text-to-image models (spanning DiT and U-Net), three samplers (Euler, flow-matching, DPM-Solver++), and two PTQ methods (SVDQuant, MixDQ), Q-Drift improves FID over the corresponding quantized baseline in most settings, with up to 4.79 FID reduction on PixArt-$\\Sigma$ (SVDQuant W3A4), while preserving CLIP scores. Code is available at https://anonymous.4open.science/r/Q-Drift.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion model", "quantization" ], "forum_url": "https://openreview.net/forum?id=Ot6rWMzuu6", "pdf_url": "https://openreview.net/pdf/43d7bc5bd94482e8055d5e51c4cabdda2b9e4514.pdf", "updated_at": "2026-10-05T19:40:56.215Z", "categories": [ "ptq", "qat", "diffusion" ], "matched_terms": [ "ptq", "quantization", "quantized", "w3a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "szkL8iUTnD", "number": 19595, "title": "EdgeEdit: An Edge-Oriented Multimodal Knowledge Editing Framework for On-Device MLLMs", "abstract": "On-device multimodal large language models (MLLMs) are increasingly enabling multimodal applications, yet updating their static encoded knowledge efficiently remains a critical challenge. In-context editing (ICE) provides an efficient and lightweight approach for fine-grained knowledge updates. However, on-device deployment typically demands joint model and input compression, introducing severe obstacles for ICE: model quantization impairs the model’s ability to utilize contextualized edited knowledge, while edit-agnostic token reduction inadvertently discards edit-critical visual evidence. To address these challenges, we propose EdgeEdit, an edge-oriented multimodal knowledge editing framework that coordinates offline model deployment with online inference. Specifically, EdgeEdit introduces Edit-Aware Mixed-Precision Quantization for offline model deployment, which conducts zeroth-order sensitivity profiling and budget-constrained dynamic programming to preserve contextualized edited knowledge utilization. For online inference, EdgeEdit employs EvidenceSaliency Guided Token Reduction, combining cross-modal saliency attribution with semantic relevance to progressively remove redundant visual tokens while retaining edit-critical visual evidence. Extensive experiments demonstrate that EdgeEdit maintains competitive editing performance while substantially reducing system costs. Compared with the strongest competing baseline for each efficiency metric, EdgeEdit achieves average reduction factors of 2.9×/1.5×/2.5× in memory usage, editing latency, and energy consumption on the server, and 2.0×/2.2×/2.0× on mobile devices, respectively.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "On-device Multimodal Large Language Models", "Multimodal Knowledge Editing", "In-Context Editing", "Mixed-Precision Quantization", "Visual Token Reduction" ], "forum_url": "https://openreview.net/forum?id=szkL8iUTnD", "pdf_url": "https://openreview.net/pdf/e9d6b0071c365f95ccda0b1317adb417065ba94a.pdf", "updated_at": "2026-10-05T19:42:00.314Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ILugpHmVWj", "number": 19699, "title": "BARQ: Balanced Codebook Refinement for Low-Bit LLM Quantization", "abstract": "As large language models (LLMs) grow in parameter count, model storage and parameter memory traffic have become major bottlenecks to efficient deployment.\nCodebook-based weight quantization reduces these costs, but imbalanced nearest-codeword assignments during fitting can leave some codewords insufficiently updated, limiting effective codebook utilization.\nTo address this limitation, we propose Balanced Assignment Refinement for Quantization (BARQ), which improves quantization quality through balanced fitting of existing codebooks.\nSpecifically, we first compute joint soft assignments between weight blocks and codewords through entropically regularized optimal transport with uniform marginals and curvature-weighted reconstruction costs, ensuring equal positive fitting mass for every codeword in the exact solution.\nWe then refine the codewords through an assignment-weighted barycentric update, which we prove minimizes the fitting objective for fixed assignments.\nFor finite Sinkhorn iterations, the implemented update retains this optimality provided all codeword masses exceed the denominator floor.\nFinally, we discard the soft assignments and use the refined codebook for standard hard nearest-codeword encoding, with our analysis establishing sufficient conditions for reducing hard-quantization distortion and evaluation loss.\nAcross multiple LLMs, BARQ achieves lower perplexity and higher mean zero-shot accuracy than the evaluated baselines at comparable bit budgets.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=ILugpHmVWj", "pdf_url": "https://openreview.net/pdf/2c4d1d8820b31fac0001aacc4e7337d639c548f7.pdf", "updated_at": "2026-10-05T19:42:47.245Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OoetS6pSkM", "number": 19730, "title": "STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization", "abstract": "Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors \ndepends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5× recurrent-state compression and reduces total serving memory by up to 68.7%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Gated Linear Attention", "Model Quantization", "Efficient Inference", "Recurrent State Compression" ], "forum_url": "https://openreview.net/forum?id=OoetS6pSkM", "pdf_url": "https://openreview.net/pdf/1d35e1d78802d58b903654fae9da0ce62d89d7d8.pdf", "updated_at": "2026-10-05T19:42:58.309Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "4-bit", "6-bit", "int8", "low precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mcts2oO1Og", "number": 19740, "title": "Pinpoint Quantization: A Mechanistic Account of Tail-Capability Failures in Quantized LLMs", "abstract": "Weight quantization is essential for deploying large language models (LLMs), yet quantization below $4$ bits can sharply degrade accuracy. Prior work has primarily characterized this degradation through model behavior or weight statistics, offering limited insight into its internal causes or how to enable targeted mitigation. Instead, we provide a *mechanistic account* of this failure by tracing capability-specific degradation to internal model components. We find that the damage is strikingly selective: it disproportionately impairs ''tail'' capabilities (e.g., niche languages, long-context processing, complex factual queries, and multi-hop reasoning) while largely preserving ''head'' capabilities. To explain this selectivity, we analyze module function and discover two predominant perturbation mechanisms: attention heads undergo *semantic direction shifts*, whereas MLP neurons undergo *activation-magnitude suppression*. Crucially, the mechanistic fragility is concentrated in a sparse set of key modules: restoring their weight precision recovers most of the accuracy loss on the degraded tasks, suggesting that precise bit allocation, rather than bit width alone, is critical to sub-$4$-bit quantization. As a practical application of these findings, we introduce Pinpoint Quantization, which uses causal-effect scores to guide group-knapsack bit allocation. Across $18$ benchmarks and dense and MoE models ranging from $4$B to $235$B parameters, it achieves near-lossless accuracy at an average weight precision below $4$ bits.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Mechanistic Interpretability", "Post-Training Quantization" ], "forum_url": "https://openreview.net/forum?id=mcts2oO1Og", "pdf_url": "https://openreview.net/pdf/0633623b0e63fa19a1fd0a5b8e08df9e492d8320.pdf", "updated_at": "2026-10-05T19:43:03.970Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "bit width", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QsgtyR4pxA", "number": 19794, "title": "Refine or Preserve? Acceptance-Gated Correction for Autoregressive Image Generation", "abstract": "Post-hoc refinement can repair degraded autoregressive image samples, but the same intervention can damage outputs that are already well tuned. We therefore ask a prior question: should a generated sample be refined at all? We introduce Acceptance-Gated Refinement (AGR), a two-level correction policy for frozen autoregressive image generators. An image-level router chooses between exact preservation and refinement, while a token-level gate proposes masked-token replacements and commits only edits that pass a refiner-score test. The router is derived from the gate's own initial refiner pass and calibrated per domain using tuned and severely degraded validation samples; no degradation label is required at test time. Across FFHQ, CelebA-HQ, and ImageNet, AGR bypasses at least 98.5% of tuned samples, with observed FID changes of only +0.06, +0.08, and −0.01, while reducing severe-input FID from 102.71 to 47.36, 81.26 to 30.17, and 96.25 to 58.09. Mild inputs excluded from router fitting also improve. A router fitted only on temperature shifts improves unseen 4-bit quantized samples while bypassing top-$k$ and top-$p$ truncated samples for which forced refinement is harmful. At matched edit rates, random editing degrades mild inputs whereas score-gated editing improves them. Across thirteen settings in three domains, refinement gain increases with the input's score deficit under the refiner. These results show that a refiner can provide not only candidate corrections, but also a practical signal for deciding when to refine and when to preserve.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "autoregressive image generation", "test-time refinement", "selective correction", "token-level gating", "pseudo-likelihood", "masked generative models", "abstention" ], "forum_url": "https://openreview.net/forum?id=QsgtyR4pxA", "pdf_url": "https://openreview.net/pdf/3b0ce2b791c170d504437014b70f0265e3ada6c4.pdf", "updated_at": "2026-10-05T19:43:42.526Z", "categories": [ "ptq" ], "matched_terms": [ "4-bit", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IQIVM8DEHx", "number": 19862, "title": "RampQ: Quantizing What Recovery Cannot Fix for Mixed-Precision LLM Quantization", "abstract": "Low-bit quantization reduces the memory footprint of large language models but often degrades model quality at aggressive precision. Mixed-precision quantization alleviates this problem by assigning different bit-widths based on module sensitivity or quantization error, while low-rank recovery adds a low-rank branch to recover the quantization error. Pre-recovery quantization error, however, does not necessarily reflect the error that remains after recovery. We show that quantization-error recoverability varies substantially across modules and bit-widths, such that recovery can reorder where higher precision is most beneficial. This motivates allocating precision according to the error that survives recovery. We propose RampQ, a Recovery-Aware Mixed-Precision Quantization framework that performs allocation in the recovered representation space. RampQ constructs bit-specific $W_q+AB$ representations, evaluates each precision choice using module sensitivity and post-recovery residual error, and selects representations under their joint memory cost. A model-level refinement further searches complete-model assignments using selective end-to-end perplexity feedback. Our evaluation results show that RampQ improves the quality memory trade off over mixed precision and low rank recovery baselines. Compared with the best baseline under comparable memory settings, RampQ reduces perplexity on average by 8.1\\% on WikiText-2 and 5.7\\% on C4, and improves average zero-shot accuracy by 1.67 percentage points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model Compression", "Recovery-Aware Quantization", "Mixed-Precision Bit Allocation", "Low-Rank Correction", "Post-Training Quantization" ], "forum_url": "https://openreview.net/forum?id=IQIVM8DEHx", "pdf_url": "https://openreview.net/pdf/d1d0122f5d07b4d81f43d762e71f548ad0be4430.pdf", "updated_at": "2026-10-05T19:44:11.520Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "mixed precision", "mixed-precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Tq9jZdZ0rt", "number": 19864, "title": "One Adapter, Every Resolution: Gated Low-Rank Adaptation for Remote Sensing VLMs", "abstract": "Remote sensing imagery spans ground sampling distances (GSDs) from centimeters to tens of meters, so both the visual evidence for a geographic concept and the questions it can support change with physical scale. Existing remote sensing vision-language models (RS-VLMs) either ignore GSD or encode it as a text token, applying one set of scale-agnostic parameters across this spectrum. We show that GSD is decodable from frozen visual features, but exploiting it requires scale-dependent adaptation. We therefore introduce ScaleEarth, which conditions low-rank adaptation on the continuous scalar $s=\\log_{10}(\\mathrm{GSD})$. Its core module, CS-HLoRA, gates the rank dimensions of a single shared LoRA with sigmoid functions of $s$ whose thresholds are initialized at object, structure, and semantic scales; the gates route gradients by scale during training and select the active adaptation subspace at inference, with the backbone frozen. When metadata is unavailable, a heteroscedastic head, SSE-U, estimates $s$ with calibrated uncertainty and falls back to a default scale when uncertain. To align supervision with this mechanism, we build GeoScale-VQA, 1.5M QA pairs in which every sample carries a resolved GSD and newly generated questions are conditioned on the same $s$. With an 8B backbone, ScaleEarth achieves an average score of 59.4 on XLRS-Bench ($+5.2$ over the strongest RS specialist). On OmniEarth-Bench, it achieves an average score of 40.71 ($+7.45$ over the strongest open-source baseline). The largest gains occur on scale-sensitive sub-tasks. The 4-bit variant can be deployed on a single 40 GB GPU and retains an average score of 57.4 on XLRS-Bench, with substantially lower adaptation and deployment costs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Remote sensing; VLM; Efficient AI" ], "forum_url": "https://openreview.net/forum?id=Tq9jZdZ0rt", "pdf_url": "https://openreview.net/pdf/9250bd715a492e4e799e9f5e77535932c8f26cae.pdf", "updated_at": "2026-10-05T19:44:27.838Z", "categories": [ "ptq", "training", "hardware", "multimodal" ], "matched_terms": [ "4-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DzzRKx58pI", "number": 19875, "title": "HSA: Hierarchical Sensitivity-Aware Quantization for Resource-Constrained Embodied Models", "abstract": "Embodied foundation models combine visual perception, language-conditioned reasoning, and action generation, but their memory footprint constrains deployment on resource-limited robotic platforms. Post-training weight quantization reduces this footprint, yet its perturbations can affect continuous action predictions and closed-loop execution. We propose Hierarchical Sensitivity-Aware Quantization (HSA), which augments a 4-bit AWQ backbone with BF16 residual blocks. An action-guided score combines a reconstruction-error proxy with action-output Jacobians, and a layer--channel--block hierarchy determines where the residuals are retained. Across LIBERO evaluations with VLA-Adapter, OpenVLA, and FastWAM, HSA improves average success over other methods by 2.55\\%--2.90\\% and remains within 0.30\\% of the original models. Meanwhile, it reduces GPU memory by 53.2\\%--67.2\\%. The quantized policy also matches the original policy in the real-robot setting under various conditions.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Vision-Language-Action Models", "Embodied AI" ], "forum_url": "https://openreview.net/forum?id=DzzRKx58pI", "pdf_url": "https://openreview.net/pdf/9a4b0a120a1b3f97187f36cb460a4092526e95bc.pdf", "updated_at": "2026-10-05T19:44:29.388Z", "categories": [ "ptq", "hardware", "reasoning", "multimodal" ], "matched_terms": [ "4-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6WATgwTWK8", "number": 20157, "title": "Inverse Affine Rectification for Low-Bit Post-Training Quantized Vision Transformers", "abstract": "Low-bit post-training quantization (PTQ) enables efficient deployment of vision transformers but can leave systematic distortions in the outputs of the resulting quantized model.\nWe introduce Inverse Affine Rectification (InvAR), a simple post-hoc output rectification method for fixed quantized models that requires no rerunning of quantization optimization.\nInvAR uses paired logits from the full-precision (FP) and quantized (Q) models on a small set of unlabeled calibration images. The key idea is to model quantization-induced output distortion as a forward mapping from FP logits to Q logits.\nRather than directly predicting FP logits from Q logits, InvAR estimates this distortion with a classwise affine model and analytically inverts it to rectify the quantized outputs.\nThis distinction is important because, in the presence of residual variation, the least-squares Q-to-FP predictor generally differs from the inverse of the fitted FP-to-Q affine map.\nUsing identical checkpoints and calibration images, vanilla InvAR and direct regression yield average Top-1 accuracy changes of +1.33 and $-1.07$ percentage points, respectively, across PTQ methods, ViT architectures, and bit widths.\nDirect inversion, however, can amplify estimation errors when calibration data are limited, especially for affine slopes near zero.\nInvAR therefore uses uncertainty-guided shrinkage to reduce classwise deviations from a shared affine reference before inversion.\nOnce paired logits are collected, fitting is non-iterative and requires $\\mathcal{O}(NC)$ operations for $N$ calibration images and $C$ classes.\nAt inference, InvAR adds only one multiplication and one addition per prediction output.\nExperiments on ImageNet-1K and COCO demonstrate overall improvements across multiple PTQ methods, ViT and CNN architectures, bit widths, calibration budgets, and vision tasks. Our code will be publicly released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training Quantization", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=6WATgwTWK8", "pdf_url": "https://openreview.net/pdf/265d736668f917fe50fca417c32c975917e3683d.pdf", "updated_at": "2026-10-05T19:46:22.718Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wHrRhsSB8Q", "number": 20250, "title": "SignRoundV2: Toward Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs", "abstract": "Extremely low-bit quantization is critical for efficiently deploying Large Language Models (LLMs), yet it often leads to severe performance degradation at 2 bits and even at 4 bits (e.g., MXFP4). We present SignRoundV2, a post-training quantization framework designed to maintain high performance even under aggressive compression. SignRoundV2 introduces (1) a simple yet efficient adaptive mixed-precision strategy that leverages gradient information and quantization-induced reconstruction errors to guide layer-wise bit allocation, and (2) a set of lightweight techniques, including loss filtering and a pre-tuning scale search, to improve tuning effectiveness in extremely low-bit regimes. Our approach takes a significant step toward closing the performance gap between quantized and full-precision models. Experimental results across diverse LLMs demonstrate that SignRoundV2 achieves near-lossless performance in mixed MXFP settings, narrowing the gap to $\\sim$1\\% at an average of 4.5 bits, while substantially improving accuracy in challenging 2-bit weight-only quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "Extremely low-bit quantization", "adaptive mixed-precision", "LLM" ], "forum_url": "https://openreview.net/forum?id=wHrRhsSB8Q", "pdf_url": "https://openreview.net/pdf/cbd07917533012cfa7ac369cdfab7cd6c15a1021.pdf", "updated_at": "2026-10-05T19:46:42.743Z", "categories": [ "ptq", "training" ], "matched_terms": [ "2-bit", "low-bit", "mixed-precision", "mxfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Avflih35YC", "number": 20368, "title": "Puro-2B: Poor Lab’s Qwen2-1.5B Trained on RTX 5090 within $5090", "abstract": "Language model pretraining has become almost synonymous with prohibitive cost, placing it out of reach for much of the academic and open-source communities. Although strong open-source efforts already exist, including open-weight models and open-source training recipes, a cost-efficient, hardware-accessible, and open-source pretraining recipe has long been missing. Even at a small scale, training Llama-3.2-3B costs over \\\\$1.5M, and reproducing SmolLM3-3B needs over \\\\$700K. In this report, we present an open pretraining recipe designed to lower this barrier.\nUsing this recipe, we train a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs. The models differ in token budgets and selected recipe variants. Our best model is trained for less than \\\\$6.9K in compute and approaches Qwen2.5-1.5B performance under our evaluation protocol. This cost efficiency comes from combining hardware selection, low-precision training, hyperball optimization, curriculum model averaging, and the data recipe. Beyond the recipe, we provide two additional results. First, across the Puro-2B collection, we derive a Puro Cost Scaling Law that relates training cost to average model performance. The fitted law shows that about \\\\$4.4K, less than \\\\$5,090, is sufficient to reach the performance of Qwen2-1.5B. Second, our end-to-end case study shows that the curriculum-based pretraining advantage persists under matched supervised fine-tuning, with higher performance on mathematical and general-purpose tasks. Access to the full pretraining pipeline rather than model weights alone enables such controlled studies. We release the full Puro-2B training recipe, including data, code, and model weights.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Pretraining", "Fully Open-Source", "Cost-Efficient" ], "forum_url": "https://openreview.net/forum?id=Avflih35YC", "pdf_url": "https://openreview.net/pdf/f452e501d630f2f093fdada437d990b66a026714.pdf", "updated_at": "2026-10-05T19:47:03.354Z", "categories": [ "training", "hardware", "reasoning" ], "matched_terms": [ "fp8", "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zrc7Pg12zB", "number": 20372, "title": "Memory-Efficient Autoregressive Video Generation via Hierarchical Mixed-Precision KV Cache Quantization", "abstract": "Autoregressive video generation models face a critical KV cache memory bottleneck as the demand for long video synthesis grows. Existing LLM quantization methods perform poorly when applied to video generation models due to the high dynamic range of video KV cache activations and their heightened sensitivity to quantization errors. We propose HiMix, a hierarchical mixed-precision KV cache quantization framework for memory-efficient autoregressive video generation. Specifically, HiMix first determines layer-wise bit budgets offline based on per-layer sparsity analysis, then performs runtime channel-aware quantization guided by Q-K saliency: important channels receive higher bit-widths (4-bit) while less important ones are quantized to 2-bit or 1-bit, with spatiotemporal Hadamard smoothing applied prior to quantization to reduce error. Experiments on the Self-Forcing and Causal Forcing models demonstrate that HiMix achieves nearly 7.1x compression ratio for KV cache while preserving video generation quality and long-range consistency comparable to the BF16 baseline, substantially reducing memory consumption.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "ARVD", "KV Cache" ], "forum_url": "https://openreview.net/forum?id=zrc7Pg12zB", "pdf_url": "https://openreview.net/pdf/372abd2ffcb7edb7ea58411aac991e279c66930a.pdf", "updated_at": "2026-10-05T19:47:04.693Z", "categories": [ "kv-cache" ], "matched_terms": [ "1-bit", "2-bit", "4-bit", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Y0Ww5DnHS4", "number": 20488, "title": "Fewer Bits, More Tokens: Rethinking KV Cache Quantization in Agentic Coding", "abstract": "Agentic coding has emerged as a prominent application of LLMs, but its multi-turn interaction trajectories continually expand the KV cache, imposing substantial memory demands during inference. KV cache quantization reduces per-token memory usage, but do these savings translate into more efficient agent trajectories? We find that low-bit quantization disproportionately amplifies three failures: repetitive action loops, improper tool use, and insufficient debugging, while increasing the actions and tokens required per task. To identify which attention heads are most sensitive to quantization around these failures, we measure changes in attention distributions and head outputs under controlled KV cache perturbations. Highly sensitive heads comprise only about 5% of all heads and cluster in a few layers. Guided by this localization, we introduce a simple and effective mixed-precision strategy that restores higher precision in selected layers while retaining low precision elsewhere. On SWE-bench Verified with Qwen3-Coder-30B-A3B-Instruct, increasing average precision from $3.0$ to $3.35$ bits substantially reduces the three behavioral failures and brings task resolution close to BF16. Despite higher per-token precision, observed KV cache usage per case decreases by 12% versus uniform 3-bit quantization. These results shift the focus of agent quantization from storing each token more cheaply to completing tasks more reliably with less memory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV Cache Quantization", "Agentic Coding", "Behavioral Degradation Analysis" ], "forum_url": "https://openreview.net/forum?id=Y0Ww5DnHS4", "pdf_url": "https://openreview.net/pdf/808e6c6028838eeda4029830a489b3025774b488.pdf", "updated_at": "2026-10-05T19:47:23.113Z", "categories": [ "kv-cache" ], "matched_terms": [ "3-bit", "low precision", "low-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hkk3JBeda2", "number": 20521, "title": "The Devil Is in the Reconstruction Loss Scale: Rethinking Optimization in LLM Quantization", "abstract": "Post-training quantization (PTQ) methods typically quantize a pre-trained large language model (LLM) sequentially using a small set of calibration samples, achieving high data and computational efficiency. Specifically, the model is partitioned into a series of units, such as transformer blocks, with one unit quantized at each stage. State-of-the-art PTQ methods are predominantly learning-based, optimizing auxiliary quantization parameters (e.g., scaling factors, rotation matrices, clipping thresholds, and adapters) via gradient descent to minimize a reconstruction loss, thereby mitigating the impact of outliers and reducing quantization errors. In learning-based PTQ, a common practice is to use mean squared error (MSE) as the reconstruction loss function, yet its induced optimization behavior remains largely unexplored. In this work, we take a holistic view of sequential quantization and systematically investigate how optimization evolves from the first quantization stage to the last, aiming for a deep understanding of optimization in learning-based PTQ schemes. Through extensive empirical studies spanning representative learning-based PTQ methods, model families, model scales, architectures, and quantization settings, we consistently uncover *Optimization Imbalance*: reconstruction loss magnitudes vary dramatically across stages, accompanied by highly uneven gradient magnitudes and parameter updates under MSE. We term the cross-stage range of loss magnitudes the *reconstruction loss scale*. Our analysis reveals that MSE translates the unexpectedly large reconstruction loss scale into highly uneven gradient magnitudes, which in turn lead to uneven optimization strength across quantization stages. This finding suggests a general principle for improving learning-based PTQ: optimization strength across stages should be decoupled from the reconstruction loss scale. Theoretically, we show that root mean squared error (RMSE) naturally realizes this principle through implicit gradient normalization, thereby enabling balanced optimization across quantization stages. Moreover, RMSE can be naturally defined over different feature groupings, yielding four variants at the sample, channel, token, and element levels. As a drop-in replacement for MSE, RMSE consistently improves quantization performance across language modeling, commonsense reasoning, mathematical reasoning, and coding tasks, with all four variants outperforming MSE and the gains becoming increasingly pronounced under more aggressive quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "post-training quantization", "learning-based quantization", "sequential quantization", "optimization imbalance" ], "forum_url": "https://openreview.net/forum?id=hkk3JBeda2", "pdf_url": "https://openreview.net/pdf/d348f91254f4176858f33472ee2a6369aaed06c4.pdf", "updated_at": "2026-10-05T19:47:27.732Z", "categories": [ "ptq", "training", "reasoning", "theory" ], "matched_terms": [ "ptq", "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SoTY9ukdnI", "number": 20534, "title": "ARSO: Reconstruction-Aware Rounding with Joint Scale Optimization for Weight-Only LLM Quantization", "abstract": "Post-training quantization (PTQ) compresses large language models (LLMs) into low-precision representations to reduce memory and accelerate inference without retraining. In weight-only PTQ, existing methods either rely on round-to-nearest (RTN), whose local decisions ignore reconstruction error, or employ costly iterative rounding optimization. Moreover, quantization scales are often reused for dequantization, coupling integer-grid construction with weight reconstruction. We identify reconstruction freedom as an important optimization resource. Scale Optimization (SO) exploits scale freedom by deriving reconstruction-optimal dequantization scales for fixed integer codes; eliminating this scale freedom further reveals a directional criterion over integer weights, enabling Reconstruction-Aware Rounding (AR). In sequential quantization, AR exploits future compensation freedom through a conditional reconstruction metric, while J-SO exploits prefix scale freedom by jointly updating processed-group scales. This stage-wise exploitation naturally yields ARSO, a PTQ method with single-pass reconstruction-aware rounding that achieves the lowest perplexity in 32 of 36 settings and the highest average zero-shot accuracy across all six models at 2 bits among the evaluated methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Weight-Only Quantization" ], "forum_url": "https://openreview.net/forum?id=SoTY9ukdnI", "pdf_url": "https://openreview.net/pdf/9e5c021679044a855e333f2054722079ddda6387.pdf", "updated_at": "2026-10-05T19:47:29.982Z", "categories": [ "ptq" ], "matched_terms": [ "low-precision", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "V75Yn3Yon2", "number": 20698, "title": "Colinearity Decay: Training Quantization-Friendly Transformers via Decoupled Weight-Space Regularization", "abstract": "Low-bit quantization enables efficient Transformer deployment, yet activation outliers complicate fully quantized inference. Quantization methods act after or during training: post-training methods adapt pretrained models, while many training-time approaches address outliers through architectural changes, or by regularizing activations. Yet minimizing activation magnitude is incomplete: excessive suppression can degrade full-precision accuracy without consistently improving quantized models. We argue that quantization-friendly training should target harmful structural artifacts that cause extreme activations. To this end, we introduce Colinearity Decay (CD), a weight-space regularizer for matrix pairs within Transformer blocks. CD penalizes detrimental cross-matrix alignment and mitigates extreme activations. The decoupled update leaves architecture and task loss unchanged, adding little training overhead and no inference-time cost. Across ImageNet-1K pre-training, object detection, downstream fine-tuning, and small-scale LLM pre-training, CD improves quantized task performance across algorithms and bit widths while preserving unquantized performance. These results support decoupled weight-space regularization for training quantization-friendly Transformers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Friendly Training; Model Compression; Transformers" ], "forum_url": "https://openreview.net/forum?id=V75Yn3Yon2", "pdf_url": "https://openreview.net/pdf/245c601efb1ccd10399c41829b17efba49ef888e.pdf", "updated_at": "2026-10-05T19:47:55.612Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GYm4mg9yGQ", "number": 20711, "title": "When Quantization Helps: Corrective Computation in Low-Bit Language Models", "abstract": "Low-bit quantization reduces deployment costs by approximating high-precision weights, introducing changes that can improve some predictions. We investigate which activation changes support lower target-task loss than the high-precision reference and whether their useful structure can be shared across inputs. Experiments on Llama (GPTQ, AWQ, NF4) and Qwen (GPTQ) reveal systematic input-level improvements despite lower average performance over the full evaluation set. Improved inputs exhibit stronger loss-reducing contributions per unit activation displacement. Stronger first-order loss reduction can outweigh comparable or larger loss increases beyond the linear approximation. Targeted interventions link these changes to predictions: replacing selected quantized activations with their high-precision values can reverse naturally corrected answers; strengthening favorable changes or suppressing harmful ones repairs errors. We construct useful local adjustments from quantization-induced changes and find contrasting shared structure: a rank-four subspace retains 95.3\\% of their cross-entropy improvement on HANS, whereas MMLU retains 47.8\\% even at rank 128. Mechanism-guided activation adaptation raises accuracy above the matched high-precision model by 7.41 percentage points on HANS test data and 2.42 on MMLU development data. Its benefit over ordinary supervised adaptation is larger on HANS, consistent with its stronger shared structure. Together, these results connect the loss-reducing contributions of quantization-induced changes to their effects on predictions and relate their shared structure to the benefits of mechanism-guided adaptation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-bit Quantization; Large Language Models; Mechanistic Interpretability; Causal Intervention; Parameter-Efficient Adaptation" ], "forum_url": "https://openreview.net/forum?id=GYm4mg9yGQ", "pdf_url": "https://openreview.net/pdf/8f8e95445f7cdb733bf966dedf83aaf13ec657e8.pdf", "updated_at": "2026-10-05T19:47:58.274Z", "categories": [ "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4dijQYTHI3", "number": 20782, "title": "GeoCoCo: Representation-Decoupled Co-occurrence Compression for MLLM Geospatial Tokens", "abstract": "Long visual token sequences (e.g., $10^4$-$10^5$) from ultra-high-resolution (UHR) remote sensing imagery increase the inference cost of multimodal large language models (MLLMs). Existing token compression methods typically depend on high-dimensional MLLM features, adding planning overhead and requiring feature computation before planning. Our observation study shows that low-dimensional image descriptors provide compression guidance with comparable accuracy and lower prefill time, while spatially regional grouping further improves accuracy. We therefore propose GeoCoCo, a feature-decoupled, plug-and-play framework for low-overhead, region-level visual token merging. Offline, GeoCoCo quantizes image descriptors into shared basic visual symbols and hierarchically merges them using spatial adjacency co-occurrence statistics aggregated across images to learn a shared compression prior. Online, this prior guides connected-region planning from input images and aggregation of the original model's visual token features without retraining the target MLLM. Planning is independent of MLLM features, enabling parallel CPU planning and GPU visual encoding. Experiments across XLRS-Bench-lite, LRS-VQA, and MME-RealWorld-RS and six MLLMs show that GeoCoCo reduces prefill time by up to $73.1\\%$ relative to uncompressed models, achieving the best performance--efficiency trade-off among the compared compression methods. We will release the relevant code for further research.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Ultra-high-resolution Remote Sensing", "Visual Token Merging", "Visual Token Pruning", "Multimodal Large Language Models" ], "forum_url": "https://openreview.net/forum?id=4dijQYTHI3", "pdf_url": "https://openreview.net/pdf/56befce5c3071b3173489d1e375ba31b4b8c83ae.pdf", "updated_at": "2026-10-05T19:48:08.618Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rlj9KWzi9l", "number": 20809, "title": "AdasignSGD: Turning 1-Bit Worker Votes into Adaptive Updates", "abstract": "Momentum-sign methods reduce synchronization traffic in distributed AI training by transmitting each worker’s local momentum with one bit per coordinate. However, these methods discard local momentum magnitudes and often fail to match the optimization quality of full-precision methods. We introduce AdasignSGD, which uses the majority sign to determine the update direction and the vote margin to set the update magnitude, taking larger steps on coordinates with stronger worker agreement. In C4 language-model pretraining, AdasignSGD outperforms all evaluated 1-bit worker-sign baselines in validation perplexity across 13 model and training-budget settings from 60M to 1B parameters. Its gap to dense AdamW narrows with model scale. At 1B, AdasignSGD closely matches AdamW in validation perplexity while delivering a $1.46\\times$ complete-step speedup over fused AdamW on 32 H200 GPUs. AdasignSGD uses one model-sized momentum state while incurring only 14.1% of FP32 dense-DDP model-sized traffic at 128 workers. On 128 P100 GPUs, AdasignSGD delivers a $4.39\\times$ complete-step speedup over AdamW for the 350M model. Together, these results show that one-bit worker votes can be turned into adaptive updates that combine strong optimization quality with a smaller memory footprint and faster distributed training steps at scale.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Distributed optimization", "1-bit optimization", "LLM pretraining", "Communication-efficient training" ], "forum_url": "https://openreview.net/forum?id=rlj9KWzi9l", "pdf_url": "https://openreview.net/pdf/388f85afddd45e2829e9e1e48c4cd63de29c286d.pdf", "updated_at": "2026-10-05T19:48:14.315Z", "categories": [ "hardware" ], "matched_terms": [ "1-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7ngGo6C7sB", "number": 20938, "title": "Safe4D: Causal Video Protection against Unauthorized 4D Reconstruction", "abstract": "Recent feed-forward geometry models recover cameras, depth and dynamic point clouds from ordinary video, either offline from a complete clip or online as frames arrive, turning a video into a renderable 4D asset within minutes. However, this capability raises serious concerns about privacy and intellectual property, since anyone who records a published video can recover the shape and motion of the people, products and places it shows. Existing protection methods perturb images or complete videos before release, but live broadcasts and video calls are released causally: each frame is published as soon as it is captured and cannot be revised, while viewers can record the stream and reconstruct it. Protecting video against 4D reconstruction under such causal release remains unexplored and raises three challenges: (i) temporal detachment: a perturbation fixed in image coordinates drifts off the moving content it protects and is averaged out by motion-aligned purification; (ii) purification uncertainty: the attacker may reconstruct from the raw or the purified video, and a perturbation tuned to one input leaves the other unprotected; and (iii) irrevocable release: each frame is transmitted as an 8-bit image as soon as it is protected, so an update weakened by rounding cannot be corrected later. To address these challenges, we propose Safe4D, the first video protection method against 4D reconstruction under causal release. Safe4D transports the perturbation committed at the previous frame along the estimated scene motion, so that the protection stays attached to the scene points it perturbs. Moreover, we introduce a purification-aware objective on a frozen streaming surrogate, which maximizes the weaker of the depth discrepancies on the direct and the temporally averaged stream, and a verified commit that accepts a sparse 8-bit update only if the quantized frame to be released improves this objective. Empirically, quantitative and qualitative results on four dynamic benchmarks show that Safe4D effectively disrupts the geometry recovered by streaming 4D reconstruction and remains robust under both causal and non-causal temporal purification.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "4D Reconstruction", "Video Protection" ], "forum_url": "https://openreview.net/forum?id=7ngGo6C7sB", "pdf_url": "https://openreview.net/pdf/88afdcd0ee26286ac4d2bf81e57aee13bd82d9f4.pdf", "updated_at": "2026-10-05T19:48:35.588Z", "categories": [ "ptq" ], "matched_terms": [ "8-bit", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mRLcwDPSTy", "number": 20964, "title": "Chameleon: Dynamic Format Adapter for Efficient Diffusion", "abstract": "A very popular way to enable image generation on memory-constrained accelerators using modern spatiotemporal generative models is post-training quantization (`PTQ`). However, every existing diffusion `PTQ` scheme shares an unstated commitment: the *number format* is fixed in advance, and only the scale, zero point, or the per-layer bit-width is allowed to move. This is unfortunate, because at a fixed bit-width the best format depends on the distribution being encoded, and that distribution differs from tensor to tensor: across weight channels, across layers, and — in a denoising model — along the *diffusion timestep*, where activations slide from heavy-tailed and noise-dominated at $t\\rightarrow T$ to tightly clustered and structured at $t\\rightarrow 0$. One format for the whole network is therefore mispriced for most of its tensors, and re-tuning a scale cannot repair a format mismatch. In this work, we propose a `PTQ` framework, which we call **Chameleon**, that holds the bit-width fixed and treats the number format itself as a discrete variable, chosen per weight channel and per *(layer, timestep bucket)* activation tensor. Chameleon draws activation formats from a palette of \\{`INT8`, `FP8 E4M3`, `FP8 E5M2`, `MXFP8`, `MXINT8`\\} and weight formats from \\{`INT8`, `MXINT8`\\} at $8$ bits or \\{`INT4`, `NF4`, `FP4 E2M1`, `MXINT4`, `MXFP4`\\} at $4$ bits, choosing activation formats ahead of time from two cheap statistics — empirical activation kurtosis and closed-form diffusion SNR $\\bar\\alpha_t/(1-\\bar\\alpha_t)$ — stored in a per-(layer, timestep bucket) lookup table, and weight formats offline by reconstruction error. An architectural fork adapts the same selection layer to each model family, so each model exercises the palette as far as its own statistics warrant: weights and per-(layer, bucket) activations on multi-step `SDXL`, weights alone where a single sampler step leaves no timestep axis. We evaluate Chameleon across three diffusion regimes — multi-step `SDXL`, single-step `SDXL-Turbo`, and the `PixArt`-$\\mathtt{\\alpha}$ Diffusion Transformer — on COCO-2014. Chameleon achieves the best `FID` for all six backbone $\\times$ bit-width settings, while preserving semantic alignment with the text prompts (`CLIP` within $0.24$ of the `FP16` reference in every setting, and the best of all quantized methods at $W_{4}A_{8}$). Adapting the format, rather than only its scale, recovers distributional fidelity at 4-bit weights while holding prompt adherence at the `FP16` level.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Diffusion models", "adaptive number formats", "Timestep aware activation quantization", "Efficient Edge Inference" ], "forum_url": "https://openreview.net/forum?id=mRLcwDPSTy", "pdf_url": "https://openreview.net/pdf/fbae86d38377a6439870982a884faa695c7f6cb5.pdf", "updated_at": "2026-10-05T19:48:40.084Z", "categories": [ "ptq", "diffusion", "hardware" ], "matched_terms": [ "4-bit", "bit-width", "fp4", "fp8", "int4", "int8", "mxfp4", "mxfp8", "ptq", "quantization", "quantized" ], "review_status": "reviewed", "summary": "Chameleon 在位宽固定时选择数值格式，以处理扩散模型权重通道、层与去噪时刻的分布差异。权重离线按 SQNR 选格式，激活用峰度和扩散 SNR 建立每层/时间桶 LUT，并为 UNet shortcut 固定 MXFP8；PixArt 额外联合搜索 group size 与格式。SDXL-Turbo 只有一步，继承 MixDQ 激活方案而只改权重格式，不能把三条架构路径都描述为相同的动态激活路由。", "contribution": "官方正文 Table 1 的 W4A8：SDXL FID 为 14.43（Q-Diffusion 21.41），SDXL-Turbo 为 21.29（MixDQ 24.50），PixArt-α 为 22.24（Q-DiT 23.53）。这是共享 COCO-2014 评测 harness 下的假量化质量结果，正文没有测量或主张实际推理加速。", "limitations": "所有主结果均为单 A100 上的 fake quantization，未测原生低精度内核的延迟/显存。正文 §4.7 明确 SDXL 权重侧 MX 使用不受约束的 per-block scale，比标准共享指数格式更有表达力，不能把该质量结果直接当作 E8M0 合规硬件结果；Turbo MX 权重则使用二次幂尺度，DiT 不选 MX 权重。实验仅覆盖图像生成，且不同 backbone 的 FID 不可横比；作者也没有验证低位 FID 优于 FP16 的统一解释。此次 13 页官方 PDF 未附 A.1–A.13，相关附录消融、格式合规代价及泛化细节无法从此文件核验。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "6379e22c50ee1975723247aadfba0a1973badc16e8e27c91da1c715a30d93ba8", "reviewed_pdf_url": "https://openreview.net/pdf/fbae86d38377a6439870982a884faa695c7f6cb5.pdf", "reviewed_source_updated_at": "2026-10-05T19:48:40.084Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "图像扩散模型权重及激活；权重通道/group 格式选择、激活层/时间桶格式选择", "bit_width": "名义 W8A8 / W4A8；SDXL-Turbo 九个 conv_shortcut 层保留 FP16，激活方案继承 MixDQ", "evidence": [ { "location": "PDF p.5–7，Sections 4.2–4.6", "claim": "权重 palette 随 8/4 bit 改变，激活由峰度/SNR 路由；Turbo 继承 MixDQ 激活，仅调整权重选择。" }, { "location": "PDF p.7，Section 4.6", "claim": "PixArt-α 的 FP8 E5M2 在每个时间桶占 97–100% 量化线性层，未重现 UNet 的三阶段路由。" }, { "location": "PDF p.8，Table 1", "claim": "W4A8 的 SDXL/Turbo/PixArt FID 为 14.43/21.29/22.24，对应先前基线为 21.41/24.50/23.53。" }, { "location": "PDF p.8，Section 5.1 / Table 1", "claim": "24,576 个 COCO-2014 captions、同一 harness；三架构的分辨率/采样器不同，FID 只可在相同 backbone 内比较。" }, { "location": "PDF p.7，Section 4.7", "claim": "SDXL weight-side MX 的 scale 不受二次幂/E8M0 约束；全文报告 fake-quant 数字。" }, { "location": "PDF p.9，Sections 5.3–6", "claim": "只报告校准成本和质量，明确没有效率声明；原生内核延迟/显存、视频及更低位范围仍待研究。" } ], "results": [ "PDF p.8，Table 1：SDXL W4A8 FID/CLIP=14.43/26.64；FP16=16.16/26.88；Q-Diffusion W4A8=21.41/26.27。", "PDF p.8，Table 1：SDXL-Turbo W4A8 FID/CLIP=21.29/26.85；FP16=21.63/26.63；MixDQ W4A8=24.50/25.88，九个 conv_shortcut 层仍为 FP16。", "PDF p.8，Table 1：PixArt-α W4A8 FID/CLIP=22.24/26.09；FP16=27.63/25.99；Q-DiT W4A8=23.53/25.73。", "PDF p.8，Table 1：三模型 W8A8 的 FID 分别为 14.23、20.93、24.33；采样步数与分辨率因架构而异，不能跨模型直接比较 FID。", "PDF p.9，Section 5.3：单 A100 校准耗时为 SDXL 45 分钟、PixArt-α 7 分钟、SDXL-Turbo 1 分钟；这是校准成本而非推理延迟。" ], "reading_basis": "实际逐页阅读本次已核验的 13 页官方原稿：正文 p.1–9、方法分支、Table 1 与局限；p.10–13 为参考文献。此文件不含正文所引用的附录 A，未借用其他版本补齐；未运行作者代码或复现实验。" }, { "id": "ZHTVUlqYOA", "number": 20999, "title": "Coordinating Bits and Priors: Compact and Faithful Decoding for Generative Image Compression", "abstract": "Generative image compression methods have recently shown strong perceptual quality at ultra-low bitrates. Pretrained generative models provide strong natural-image priors for ultra-low-rate image compression, but they are designed for image synthesis and generation rather than reconstructing a particular source from a decoded bitstream. This motivates two considerations: different stages of the generative prior contribute unequally to reconstruction, and the decoded evidence needs to be preserved as synthesis progresses toward pixels. We propose CFIC, a compact and faithful generative image compression framework built around these observations. A training-free analysis reveals that reconstruction importance varies substantially across stages, motivating asymmetric compaction of the one-step diffusion prior rather than uniform shrinking. Multi-scale evidence recovered from the same quantized representation is then coordinated with corresponding synthesis stages, allowing source-specific information to complement the generative prior without additional side information. Finally, we alleviate the decoding bottleneck inherited from the pretrained model by preserving this evidence hierarchically along the path to pixels. Experimental results demonstrate that our proposed method achieves state-of-the-art perceptual quality with a compact model size and lower computational cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Learned Image Compression; Generative Image Compression; Diffusion Model; Computer Vision" ], "forum_url": "https://openreview.net/forum?id=ZHTVUlqYOA", "pdf_url": "https://openreview.net/pdf/36e8637dfb371023e927c92f42928f4717b43c21.pdf", "updated_at": "2026-10-05T19:48:47.111Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Th5k7KD4ZB", "number": 21030, "title": "Fragility-Aware Adaptive Routing for Quantized Mixture-of-Experts Serving", "abstract": "Quantizing the experts of a Mixture-of-Experts (MoE) model is standard practice\nfor relieving memory pressure during serving, producing a pool of pre-quantized\nmodel copies: instances, whose quality cost varies sharply across requests. Existing\nsystems have two limitations: statically committing to a single instance for the\nworkload and relying on ad hoc heuristic request score predictors. Closing this\ngap, we propose an adaptive routing framework with theoretical guarantees that\nassigns each incoming request to an instance so as to maximize throughput within\na quality-degradation budget. Underlying the framework is a structural result: a\nsmall-noise expansion factors quality loss over experts into request-independent\nexpert fragility and request-dependent expert affinity. This factorization yields\nFWP (Fragility-Weighted Perplexity), a lightweight quality estimator read from a\nsingle prefill pass on the cheapest instance and transferred to every other instance\nby affine calibration. Building on FWP, we cast routing as a window-level linear\nprogram and prove that the induced per-request greedy policy is KKT-consistent\nwith the LP optimum. Empirically, on complete Qwen3-30B-A3B instances, FWP-\nguided allocation at a fixed quality budget comes within $1.5$% of an allocation\nthat knows each request’s true quality loss, well ahead of activation-frequency and\nprompt-length signals. Turning this signal into live serving, the LP-derived router\nover 4-bit, 4/8-bit mixed, and 8-bit instances meets every per-class quality budget at up\nto $1.20\\times$ the measured throughput of the only static deployment that satisfies all\nclasses.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixture-of-experts", "post-training quantization", "large language model serving", "adaptive routing", "constrained optimization" ], "forum_url": "https://openreview.net/forum?id=Th5k7KD4ZB", "pdf_url": "https://openreview.net/pdf/4de7f3aa1afe2c93a1b17eb2d8c0e37a1fcde80f.pdf", "updated_at": "2026-10-05T19:48:51.852Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "4-bit", "8-bit", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "T1fUcoIg6z", "number": 21064, "title": "GhostQuant: The Ghost that Appears Only After Quantization", "abstract": "Model quantization is widely used to deploy deep learning models on resource-constrained devices. However, it may introduce previously overlooked security risks. Attackers can implant backdoors that are inactive in full-precision model but get activated after quantization. \nIn Large Vision-Language Models (LVLMs), building such attacks is challenging. Visual backdoors are easily erased by low-bit quantization. Moreover, most existing methods support only static quantizers, limiting their real-world threat. They also noticeably degrade the model's full-precision utility, making the attack easier to detect.\nTo solve these, we propose GhostQuant, a stealthy quantization-aware backdoor for LVLMs. GhostQuant optimizes boundary-sensitive parameters under quantized constraints. It then applies minimal perturbations to activate the backdoor after quantization while preserving full-precision behavior.\n Experiments show that GhostQuant outperforms existing methods across diverse LVLMs. For example, on Qwen3VL-2B with Caption Hijacking, GhostQuant achieves an average ASR of 92.5\\% across 12 quantization methods while limiting the full-precision utility deviation to only 1.55\\%. In contrast, existing baselines support only three static quantizers and cause utility deviations of 12.31\\%--14.01\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Vision-Language Model; Backdoor; Quantization" ], "forum_url": "https://openreview.net/forum?id=T1fUcoIg6z", "pdf_url": "https://openreview.net/pdf/2c417895aa8ec40f31cff80d53402317660a0885.pdf", "updated_at": "2026-10-05T19:48:59.511Z", "categories": [ "qat", "multimodal" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ypARYhkNKF", "number": 21072, "title": "H-Scale: Hessian-Guided Scale Refinement for NVFP4 Sub-Byte LLM Inference", "abstract": "The NVIDIA Blackwell architecture, with native support for the ultra-fine-grained NVFP4 format, opens new opportunities for accelerating large language model (LLM) inference. NVFP4's micro-block design, such as a group size of 16, offers strong representational flexibility for capturing local weight distributions and isolating outliers, but it also introduces a large and highly sensitive space of per-group scaling factors. Existing post-training quantization (PTQ) methods primarily focus on refining quantized weight values, leaving this scale-selection step underexplored. To address this gap, we propose \\textbf{H-Scale}, a lightweight post-processing method for NVFP4 per-group scale refinement. Instead of minimizing plain weight reconstruction error, H-Scale selects hardware-valid group scales using a diagonal second-order proxy derived from calibration activations, thereby targeting layer output perturbation more directly. It is designed as a drop-in replacement for RTN-style scale selection in diverse NVFP4 pipelines, requires only modest offline calibration, and introduces strictly zero overhead at inference time. Under a fixed evaluation protocol, experiments on mainstream LLMs show that H-Scale generally improves a broad range of NVFP4 baselines and brings several variants closer to the BF16 reference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "AI Infrastructure", "LLM Quantization" ], "forum_url": "https://openreview.net/forum?id=ypARYhkNKF", "pdf_url": "https://openreview.net/pdf/abdf7b2df2f71dadc731e91475118bd253394803.pdf", "updated_at": "2026-10-05T19:49:01.914Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "nvfp4", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Loi7z1kAY2", "number": 21137, "title": "When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization", "abstract": "Weight-only post-training quantization (PTQ) relies heavily on reconstruction loss minimization to preserve model quality at low precision. We show that the weights favored by minimizing this loss need not yield better model performance on new tasks. In fact, we find that lower reconstruction loss can even degrade model performance on the same calibration data. Our analysis further shows that weights with lower reconstruction loss on calibration data can have higher loss than other weights when the distribution of input activations changes. Motivated by these observations and our analysis, we propose Distributionally Robust Quantization (DRQ), a post-hoc refinement process that minimizes worst-case reconstruction loss over a constrained set of input activation distributions. DRQ refines the integer codes representing quantized weights within the existing quantization grid, keeping quantization parameters and inference operators unchanged. Extensive experiments show that DRQ improves models quantized by six representative PTQ methods, including AWQ, GPTQ, and ParoQuant, and delivers gains across both dense and mixture-of-experts large language models. These results establish DRQ as a general post-hoc refinement framework for weight-only PTQ, achieving better downstream performance without adding inference overhead.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "model compression", "distributionally robust optimization", "discrete optimization" ], "forum_url": "https://openreview.net/forum?id=Loi7z1kAY2", "pdf_url": "https://openreview.net/pdf/5376853b56f5e28cb42f3b26e516d8d017587a5d.pdf", "updated_at": "2026-10-05T19:49:13.447Z", "categories": [ "ptq" ], "matched_terms": [ "low precision", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9xxrtedMRD", "number": 21155, "title": "PolyLoRA: Discrete-Continuous Transversal Updates for Quantized Diffusion Adaptation", "abstract": "Large-scale diffusion models are often customized via low-rank adaptation and then quantized for deployment to reduce memory costs. However, a fundamental mismatch remains. The adapter is optimized in continuous space but deployed after discrete integer merge. This leads to two failures. First is semantic facet sticking, where updates that reduce the training loss may not cross a quantization boundary and leave the deployed model unchanged. Second is asymmetric merging distortion, where even when a boundary is crossed, the base integer anchor non-uniformly suppresses or amplifies the update.\nWe propose PolyLoRA to reshape the adaptation trajectory along transversality in polyhedral space for deployable semantic changes. It introduces three lightweight mechanisms. Forward-Aligned Crossing promotes transverse boundary crossing in the actual pre-merge adapter, directly countering facet sticking. Merge-Anisotropy Calibration weights these crossings according to the integer-merge geometry, steering updates toward coordinates where post-merge leverage is strong to mitigate distortion. Semantic Transversality Shaping further uses an activation-derived Jacobian-Gram proxy to align merge-retained displacements with denoising-relevant response patterns, ensuring that survived updates actually benefit generation. These mechanisms are coupled with the denoising loss via a dual-barrier strategy at no inference overhead.\nBy enforcing transversality, PolyLoRA ensures that optimizer updates cross quantization boundaries and survive integer merge in semantically relevant directions. Under W4A8 DreamBooth, PolyLoRA further improves DINO and CLIP-I over IntLoRA by 0.0162 and 0.0083, respectively, while preserving the direct integer-merge inference path.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "Model Quantization", "Low-Rank Adaptation" ], "forum_url": "https://openreview.net/forum?id=9xxrtedMRD", "pdf_url": "https://openreview.net/pdf/682057de9ea340fa0ae8428a37337d7ba6563657.pdf", "updated_at": "2026-10-05T19:49:17.291Z", "categories": [ "ptq", "training", "hardware", "diffusion" ], "matched_terms": [ "quantization", "quantized", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OWsSUwyLyQ", "number": 21351, "title": "RAQ-MLA: Role-Aware Quantization for Multi-Head Latent Attention Caches", "abstract": "Multi-Head Latent Attention (MLA) has been adopted by recent frontier large\nlanguage models. Instead of per-head keys and values, MLA caches a compressed\nKV latent that is up-projected into keys and values at decode time, together\nwith a narrow positional key that carries position information into the\nattention score. Existing KV-cache quantization methods are designed around the\nkeys and values of standard attention and do not account for what MLA actually\ncaches, which costs accuracy at low bit widths. We analyze the MLA cache both\nin activation distribution and in how quantization error propagates to the\nattention output, and show that the two cached tensors play different roles:\nthe latent mainly supplies the values, whereas the positional key dominates the\nattention score. We therefore propose \\textbf{RAQ-MLA} (Role-Aware Quantization\nfor Multi-Head Latent Attention Caches), which fits an invertible transform to\neach cached tensor against the error that tensor delivers to the attention\noutput: attention-weighted value error for the latent and score perturbation for\nthe positional key. Without training, RAQ-MLA quantizes the full MLA cache to\nabout $2.5$ bits per element, a $6.5\\times$ increase in theoretical KV-cache\ncapacity, while remaining nearly lossless on reasoning and coding tasks across\nfour MLA models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "MLA", "KV cache", "Quanitzation" ], "forum_url": "https://openreview.net/forum?id=OWsSUwyLyQ", "pdf_url": "https://openreview.net/pdf/0ddb82d10b02bb40906cd54e38e083143a4675f4.pdf", "updated_at": "2026-10-05T19:49:48.708Z", "categories": [ "kv-cache", "reasoning", "theory" ], "matched_terms": [ "low bit", "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XnF8RSUPxE", "number": 21437, "title": "BRACE: Error-Guided Mixed-Precision LLM Quantization Without Backpropagation", "abstract": "Post-training quantization (PTQ) reduces the deployment costs of large language models (LLMs). Uniform-precision quantization remains a widely used approach, but it often incurs substantial quality degradation at ultra-low bit-widths. In contrast, mixed precision quantization offers a better accuracy–budget trade-off by selectively assigning precision. However, at ultra-low precision, a quantization error cannot be judged by its local magnitude alone. Its significance also depends on the stage-specific response it should induce. This observation motivates a common design principle for bit allocation, error compensation, and quantizer calibration: quantization error should not merely be minimized locally, but should guide the response taken at each stage. Our approach builds on this observation and introduces three key components: (i) Curvature-Conditioned Residual Transport (CCRT) allocates bits using a joint curvature-conditioned cost that combines the error committed at a group with the one-step correction burden transported to its immediate successor; (ii) Block-Anchored Ridge Compensation (BARC) pairs reference and quantized paths within each block to derive a ridge-stabilized correction for propagated error before applying GPTQ to the corrected center under a metric with a ridge term; and (iii) Residual-Feedback Grid Refinement (RFGR) converts the elementwise errors of curvature-weighted first-pass grids into weights for a second scale and zero-point fit at the assigned bit-width, with adjacent-bit error reductions providing the feedback signal. Built on a GPTQ backbone, BRACE uses only a small calibration set and local layer-wise statistics, without backpropagation or iterative retraining. Extensive experiments on language modeling and zero-shot commonsense reasoning across the Llama2, Llama3, Qwen3, and OLMo3 model families show that BRACE achieves state-of-the-art performance among comparable backpropagation-free PTQ methods under 2- and 3-bit weight-only quantization settings. Code will be released upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "mixed-precision quantization", "low-bit quantization" ], "forum_url": "https://openreview.net/forum?id=XnF8RSUPxE", "pdf_url": "https://openreview.net/pdf/8aa964f4d515469812609588f6433f60047de862.pdf", "updated_at": "2026-10-05T19:50:02.388Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "3-bit", "bit-width", "low bit", "low precision", "low-bit", "mixed precision", "mixed-precision", "ptq", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ndq5JE4YI8", "number": 21465, "title": "LOPA: Learning Local Subspaces for Privacy Protection in Split Federated Learning", "abstract": "Split Federated Learning (SFL) reduces the computational burden on clients through collaborative training between clients and a server. However, an honest-but-curious server can still reconstruct private images from uploaded intermediate representations. Existing defenses limit information exposure through activation dropping, quantization, or perturbation, but these operations do not necessarily suppress reconstruction-relevant information while preserving information needed for classification, and stronger restrictions may degrade task performance. We observe that low-rank channel projection combined with normalization and spatial pooling can reduce reconstruction quality, but random selection of subspace directions does not consistently preserve classification accuracy. To address this issue, we propose LOPA (Local Privacy-Aware Subspace Adaptation), which uses only local client data to initialize a task-relevant subspace based on between-class and within-class scatter statistics. It then calibrates the subspace using classification loss and distance correlation to preserve task-discriminative information while reducing the statistical dependence of representations on raw inputs. The clients' orthogonal projection matrices are aggregated into a shared subspace before training, which remains fixed during online training to filter uploaded representations without further subspace optimization. Experiments on CelebA, CIFAR-10, and FMNIST show that LOPA substantially degrades reconstruction quality under both the FORA and Inverse-Network attacks. Unlike the compared defenses, which trade accuracy for protection, LOPA matches undefended SFL on all three datasets. Furthermore, subspace initialization takes only seconds and requires less memory than client-side training itself.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Split federated learning", "privacy preservation", "differential privacy" ], "forum_url": "https://openreview.net/forum?id=Ndq5JE4YI8", "pdf_url": "https://openreview.net/pdf/2177cd6f2b2c73e64136c10fdb72fdba325ceb3b.pdf", "updated_at": "2026-10-05T19:50:09.201Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IzDTEuEZo8", "number": 21549, "title": "E8-QAT: Efficient 2-Bit Quantization-Aware Training with Implicit Lattice Codebooks", "abstract": "Low-bit quantization reduces the memory footprint of large language models, but preserving accuracy with efficient training remains challenging.\nScalar quantization encodes weights independently, while explicit-codebook vector quantization introduces costly search during training. We propose E8-QAT, a\n2-bit quantization-aware training framework that uses the $E_8$ lattice as an implicit\neight-dimensional codebook. It jointly quantizes weight groups through arithmetic\nprojection without storing or searching an explicit codebook during the training forward\npass. A 2-bit core encoding and sparse extension preserve the selected lattice\npoints exactly, adding less than 0.1\\% to core index storage. \nOn Qwen3.6-35B-A3B, E8-QAT matches the performance of the full-precision model in terms of the average score across the four tasks.\nE8-QAT improves over the state-of-the-art baseline by 2.03\\% on IFEval and by 1.51\\% on LiveCodeBench, under the same data, schedule, and bit width.\nThe 2-bit representation reduces peak memory by 80\\% and supports a $4\\times$ larger batch.\nAt each representation's largest feasible batch, E8-QAT reaches $1.97\\times$ the decoding throughput of BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Vector Quantization", "Quantization-Aware Training" ], "forum_url": "https://openreview.net/forum?id=IzDTEuEZo8", "pdf_url": "https://openreview.net/pdf/8550a57e1d3fa35b2e9a75f3e0e62da3095a758c.pdf", "updated_at": "2026-10-05T19:50:18.787Z", "categories": [ "qat", "training", "hardware" ], "matched_terms": [ "2-bit", "bit width", "low-bit", "qat", "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LjdcKlasnX", "number": 21590, "title": "SpheRoPE: Zero-Shot Optimization-Free 360 Panorama Generation with Spherical RoPE", "abstract": "We present a zero-shot, training-free and optimization-free framework for generating 360 panoramic images and videos by directly injecting spherical priors into pre-trained diffusion transformers. Existing methods either rely on costly fine-tuning on scarce panoramic data that limits generalization, or leverage multi-step optimization that incurs prohibitive inference latency. We observe that contemporary generative models natively exhibit some panoramic priors from large-scale training. However, these emergent capabilities are insufficient, as the models fundamentally fail to satisfy the rigorous topological constraints imposed by equirectangular projection (ERP). We introduce a zero-shot and optimization-free approach that resolves these constraints at inference time. Spherical RoPE replaces standard rotary position embeddings: low-frequency channels are re-parameterized as 3D Cartesian coordinates to natively encode the spherical manifold, while high-frequency channels are harmonically quantized to enforce exact periodicity. Coupled with complementary Semantic Distortion classifier-free guidance (CFG) that explicitly steers geometry, we avoid retraining and inherit the full creative breadth of state-of-the-art models. Our approach generalizes across diverse backbones and 360 generation modalities. We demonstrate this across text-to-panorama using Flux.1, Flux.2, and LTX-Video backbones, achieving competitive performance against baselines, all while remaining training-free. Our code will be publicly available.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "zero shot", "306 generation", "diffusion models" ], "forum_url": "https://openreview.net/forum?id=LjdcKlasnX", "pdf_url": "https://openreview.net/pdf/83080643c133c6bb49e79a19ae927e0d4a634375.pdf", "updated_at": "2026-10-05T19:50:27.883Z", "categories": [ "hardware", "diffusion", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nzAHo1Cp7l", "number": 21623, "title": "LayerShield-QAT: Accuracy-Preserving Integer Quantization for Physics-Grounded Edge Regression", "abstract": "Integer accelerators can make scientific inversion practical at the edge, yet regression models are unusually vulnerable to output discretization: errors that leave a classifier's decision unchanged can invalidate a continuous physical estimate. We present \\method, a physics-grounded multi-task regression and quantization workflow for gamma-ray well-logging inversion on 8-bit neural accelerators. The model maps 11 detector-window measurements to five coupled formation and mud properties, using shared representations and task-weighted objectives to prioritize formation density and photoelectric factor. Layer-wise diagnostics identify the input and output transformations as the dominant sources of quantization error; \\method \\ therefore quantizes the hidden representation to INT8 while retaining higher precision at the boundary layers. Training additionally injects systematic, random, and count-statistical perturbations to reduce the synthetic-to-field gap. On the reported synthetic benchmark, multi-task learning increases the fraction of predictions within application tolerances from 23.61\\% to 89.53\\% for formation density and from 19.82\\% to 88.15\\% for photoelectric factor, while reaching 98.18--99.79\\% on three auxiliary mud targets. Accuracy-preserved quantization and deployment on NPU SoC systems reports 3.5$\\times$ CPU and 23$\\times$ NPU speedups over a legacy C implementation, with stable field data trends. The results support a broader principle: for high-accuracy edge regression, precision should be allocated according to physical sensitivity and layer-wise error propagation rather than uniformly.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Regression Integer Quantization", "Accuracy-preserved regression QAT", "NPU", "8-bit quantization", "Physics", "Nuclear geophysics", "subsurface exploration." ], "forum_url": "https://openreview.net/forum?id=nzAHo1Cp7l", "pdf_url": "https://openreview.net/pdf/e07534003fd40ef3817636f90737070bd69392f6.pdf", "updated_at": "2026-10-05T19:50:32.218Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "8-bit", "int8", "qat", "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ccVdQhHOc0", "number": 21659, "title": "Maximum Activations in Open Large Language Models", "abstract": "The dynamic range of activations is a first-order constraint for low-bit quantization, activation scaling, and stable LLM inference. Prior work characterized outlier features and massive activations on pre-2024 LLaMA-style models, and the downstream activation-quantization stack inherits that picture without revisiting it for the post-LLaMA open-model boom. We ask the deployment-oriented question: how large can activations get in modern open LLMs, and how does this magnitude vary across families, generations, and training stages? Under a unified pipeline (5000-sample multi-domain corpus, family-specific tokenization, identical hooks across embeddings, hidden states, attention, MLP/MoE, SwiGLU gates, and final norm), we measure global and layerwise maxima on 27 checkpoints from 8 open families spanning dense, MoE, vision-language, intermediate-training, and instruction-tuned variants. We find that (i) global maxima span over nearly four orders of magnitude at comparable parameter counts, with Qwen3.5 and MoE checkpoints in the $10^2$--$10^3$ range and Gemma3-27B-it reaching $\\sim7\\times 10^5$; (ii) cross-family and cross-generation comparisons break simple monotonic scaling; and (iii) in two matched MoE--dense pairs, MoE checkpoints exhibit $14.0$--$23.4\\times$ lower peaks, while the residual stream carries the global maximum in 22/24 main-analysis checkpoints. A lightweight INT-8 sanity check shows that measured maxima co-vary with low-bit reconstruction error via activation-scale selection. We conclude that maximum activation magnitude is a protocol-conditioned model statistic associated with family, architecture, and training stage---not a simple byproduct of size---and should be measured and reported alongside any open-weight release before low-bit deployment. An anonymous code repository is linked at \\url{https://anonymous.4open.science/r/Max_act_llm-CB63/}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM" ], "forum_url": "https://openreview.net/forum?id=ccVdQhHOc0", "pdf_url": "https://openreview.net/pdf/3d99e6e36770c746161f5edc4f20c1093d5a82f0.pdf", "updated_at": "2026-10-05T19:50:39.240Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3NMGV5gAEg", "number": 21671, "title": "Faithful Rules Are Not Enough: Boolean Surrogates for Deep RL and an Audit of Their Concepts", "abstract": "Rule surrogates are among the most readable explanations of deep reinforcement learning (DRL) policies, and their quality is almost always judged by a single number: how often the rules reproduce the policy's actions. We ask whether that number says anything about the concepts the rules are written in. We first build a strong surrogate for vision-based policies. LUCID jointly trains a sparse autoencoder on the frozen policy's features, a learnable binarization bottleneck and a product t-norm logic layer, and extracts Disjunctive Normal Form (DNF) rules; once the selectors polarize, the extracted rules are invariant over a wide range of extraction thresholds, and a state-level certificate bounds the disagreement between the rules and the continuous surrogate. Across five vision-based tasks, from MiniGrid to Pong and Boxing, the rules agree with the PPO teacher on 96.8–100% of states and need about 2× fewer feature tests than a decision tree of matched fidelity on gridworlds and 28–46× fewer on Atari; controls show that joint training is what keeps them faithful after extraction. We then audit the concepts along two axes: alignment with simulator ground-truth factors, and causal effect under intervention on the frozen teacher. Under our criteria the two never hold together in any of the five environments, and on Dynamic-Obstacles rules with near-perfect fidelity rest on concepts whose manipulation moves the teacher away from the predicted action. VLM-generated concept labels, tested blind, predict their concepts above chance but well below certainty. Behavioural fidelity, factor alignment and causal efficacy are distinct properties, and rule-extraction work should report all three.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "explainable reinforcement learning", "post-hoc surrogate", "rule extraction", "disjunctive normal form", "sparse autoencoder", "concept discovery", "neuro-symbolic", "faithfulness evaluation", "causal intervention", "vision-language models" ], "forum_url": "https://openreview.net/forum?id=3NMGV5gAEg", "pdf_url": "https://openreview.net/pdf/d80cff431750d403d8514b9c0dc08c0358fbe6d9.pdf", "updated_at": "2026-10-05T19:50:40.932Z", "categories": [ "multimodal" ], "matched_terms": [ "binarization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GpBZtJO3oD", "number": 21685, "title": "NanoKV: High-Ratio KV Cache Compression for Autoregressive Video Generation", "abstract": "Autoregressive video models rely on historical key–value (KV) caches to preserve scene consistency over extended rollouts, making KV storage and repeated attention over history a growing memory and computational bottleneck. Aggressive compression can degrade generation quality, while reconstructing full-dimensional KV leaves attention computation largely unchanged. We introduce NanoKV, a frozen-backbone framework for generation-aware high-ratio KV compression. NanoKV exploits redundancy across tokens and attention heads by storing shared content in cluster centers and encoding joint head–channel residuals into low-bit coefficients. Crucially, NanoKV uses different fitting objectives for Keys and Values according to their distinct roles in attention: Key decoders are optimized to preserve downstream block outputs, whereas Value encoders and decoders are fitted for faithful reconstruction through the deployed quantization path. Building on the same compressed representation, the optional NanoKV-Fast variant reuses the stored coefficients for mixed-dimensional attention, reducing computation without modifying the video backbone. Across three autoregressive video backbones, NanoKV consistently improves the quality–storage trade-off. On HY-WorldPlay, NanoKV achieves 26.09× effective KV compression—nearly twice that of KVTC—while improving PSNR by 1.0 dB. NanoKV-Fast retains 20.53× effective compression and achieves an approximately 1.15× end-to-end speedup over KVTC. Code will be released at https://anonymous.4open.science/r/NanoKV-ICLR2027-187D.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "autoregressive video generation", "efficient inference", "low-dimensional attention", "quantization" ], "forum_url": "https://openreview.net/forum?id=GpBZtJO3oD", "pdf_url": "https://openreview.net/pdf/65571c845c16b2d2bd5241ff1571396942a2b17f.pdf", "updated_at": "2026-10-05T19:50:44.940Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "oGWaGHIPWm", "number": 21740, "title": "ThunderWorld: Efficient World Model Generation with Selective FP4 Refinement via Native GEMMs", "abstract": "We study whether world models can recover high-quality generation through dense four-bit floating-point (FP4) matrix computation without sparse matrix instructions or token/channel reordering. ThunderWorld is a training-free framework that corrects four-bit weight-and-activation (W4A4) error using additional dense FP4 products formed from quantization residuals. Block-aligned supports jointly govern residual storage, online preparation, and dense execution under a compute budget, preserving every base attention block. Controlled attention diagnostics support structure-informed selection over random and window policies at matched repair computation. On Cosmos3 Edge and Nano across RBench and PAI-Bench-G, a configuration capped at twice base W4A4 matrix work keeps mean task-score gaps within 0.08 RBench points and 0.37 PAI-Bench-G percentage points of BF16. It reduces LPIPS, a perceptual distance to same-input, same-seed BF16 generations, by 3.3–7.8% versus ARCQuant (2.0), with 1.17–1.88× its throughput on RTX PRO 6000 and Jetson AGX Thor. These results provide a numerical and systems basis for world-model inference on devices with native FP4 Tensor Cores.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "World Models", "FP4 Quantization", "Low-Precision Inference", "Video Generation", "Mixed-Precision Refinement", "Tensor Cores", "Efficient Inference", "GPU Systems" ], "forum_url": "https://openreview.net/forum?id=oGWaGHIPWm", "pdf_url": "https://openreview.net/pdf/e193746e2bc2bf55de4f96a07bf9c7edd06cf577.pdf", "updated_at": "2026-10-05T19:50:52.881Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "low-precision", "mixed-precision", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "78hwS03RCf", "number": 21741, "title": "Compression-Induced Representational Dissonance: Anticipating and Suppressing Prediction Flips", "abstract": "A compressed model can preserve benchmark accuracy while altering individual predictions, leading to potentially severe reliability failures. Across accuracy- preserving quantization, pruning, and low-rank compression of 58 vision models on 9 datasets, we find that models lose only 0.42 percentage points of accuracy on average while changing 9.4% of their predictions. We trace these changes to the interaction between compression error and the constraints it imposes, and to how they affect features of different energies. Compression disproportionately distorts low-variance feature directions, drowning them in noise. Whether these changes alter a prediction depends on the base model’s decision margin, linking differential representational change to decision instability. This relationship generalizes across compression families and model classes, allowing prediction flips to be anticipated before evaluating the compressed model on the target inputs. Across 453 configurations of vision and language models spanning quantization, sparsification, and low-rank, we predict the flip rates with a median relative error of 5.7%. And as we focus more on preserving representation during compression, we substantially restore decision fidelity and reduce flips at matched accuracy by 50 to 93%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model compression; efficient deep learning; representation learning; model reliability; quantization; knowledge distillation." ], "forum_url": "https://openreview.net/forum?id=78hwS03RCf", "pdf_url": "https://openreview.net/pdf/95937f1fe26e26f766771d349f68e955a961e8c3.pdf", "updated_at": "2026-10-05T19:50:52.084Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AsvvRkh77C", "number": 21821, "title": "Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding", "abstract": "Location encoders learn compact representations of latitude--longitude coordinates for downstream geospatial prediction tasks. Existing approaches are pretrained using continuous or naturally occurring supervision, such as imagery, climate variables, or species distributions, and therefore do not explicitly capture administrative boundaries or place names, despite their strong influence on many real-world geospatial phenomena. We introduce, to our knowledge, the first method for learning administrative boundary-aware location encoders, using reverse-geocoded text as a self-supervised training signal. Our approach can be applied to existing location encoder architectures and consistently improves their downstream performance. We further introduce RevGeoRQ, a custom architecture tailored to reverse-geocoding supervision, which yields additional gains and matches or exceeds contemporary location encoders on several downstream tasks. Finally, the use of human-interpretable reverse-geocoded strings and residual-quantized semantic IDs provides a more interpretable representation of geographic location than existing approaches.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "geospatial machine learning", "location encoders", "self-supervised learning", "reverse geocoding", "geospatial representation learning", "location embeddings", "geographic representation learning", "multimodal learning", "spatial machine learning", "geographic priors" ], "forum_url": "https://openreview.net/forum?id=AsvvRkh77C", "pdf_url": "https://openreview.net/pdf/ed0faef90d0c9f6db5e7c496455bc79bf7d3a14f.pdf", "updated_at": "2026-10-05T19:51:05.157Z", "categories": [ "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zbC2iH2uy9", "number": 21837, "title": "LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization", "abstract": "Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, joint QAT nearly matches FP32 VBench-7 (0.7403 versus 0.7409), yet LIBERO success collapses from 95.5% to 10.5%. Controlled decoder-only experiments show that activation quantize–dequantize operations alter the reconstruction signal and decoder Jacobian, redirecting the gradient returned to the encoder and inducing persistent latent drift. Based on this mechanism, we introduce LatentQuant, a two-stage NVFP4 QAT framework that first aligns the quantized encoder with its high-precision counterpart, then freezes it while adapting the decoder. Across Wan2.1 and Wan2.2, LatentQuant preserves near-baseline control and high reconstruction quality, achieving 95.75% success on LIBERO and 68.8% on RoboTwin. On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17×–1.26× end-to-end VAE speedups over BF16 cuDNN.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-aware training", "NVFP4", "video VAE", "world action models", "latent representation alignment" ], "forum_url": "https://openreview.net/forum?id=zbC2iH2uy9", "pdf_url": "https://openreview.net/pdf/9bb3880ebe46180599c05eeab0b2897a7c2a8659.pdf", "updated_at": "2026-10-05T19:51:07.261Z", "categories": [ "ptq", "qat", "training", "hardware" ], "matched_terms": [ "nvfp4", "qat", "quantization", "quantize", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "29c4ZlBv3g", "number": 21995, "title": "Making Analog Training Scale: Co-Designing Mapping, Optimizer, and Converters", "abstract": "Analog in-memory computing (AIMC) offers a transformative alternative for model training by executing matrix operations directly where weights are stored. However, scaling AIMC to train modern deep models remains an open challenge due to severe hardware non-idealities, including physical weights with finite dynamic range and write granularity, analog-digital converters with finite resolution, and noisy and asymmetric updates. Guided by the insight that gradient accumulation is sensitive to precision and rounding errors, we adopt a mixed-precision training paradigm: executing forward and backward matrix multiplications in the analog domain while computing weight gradients in the digital domain. To enable scalable training, we present a holistic system-algorithm co-design that co-optimizes weight mapping to ensure well-conditioned physical and logical weight profiles, couples a preconditioned optimizer with threshold-triggered open-loop pulsing to stabilize training trajectories, and aligns converter dynamic ranges to suppress quantization errors.\nEvaluated via hardware-calibrated architectural simulations calibrated with electrochemical RAM measurements, our framework scales Transformer training up to $123\\text{M}$ parameters; writing parameter count as $N$, its validation loss scales as $L\\propto N^{-0.231}$, comparable to $L\\propto N^{-0.238}$ for digital training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Analog AI; in-memory computing; mixed-precision training; Scaling law" ], "forum_url": "https://openreview.net/forum?id=29c4ZlBv3g", "pdf_url": "https://openreview.net/pdf/14c3d1af7e77ea891ee3022fabdce5571eb1944c.pdf", "updated_at": "2026-10-05T19:51:35.229Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NZAbSVNeqK", "number": 22016, "title": "Predictive Memory Complexity: Persistent-State Requirements for Continual Prediction", "abstract": "Continual world models accumulate an ever-growing history of experience, while the memory that can be persistently retained remains fundamentally bounded. This mismatch raises a fundamental question: how much information from the past must remain encoded in persistent state to preserve future predictive accuracy?\n\nWe introduce predictive memory complexity (PMC), the minimum persistent-state capacity required to attain a prescribed prediction risk. (i) Under stationary fully observed Gaussian dynamics, Theorem 1 gives the exact finite-memory law $R_B^\\star = q + a^2 D_{2^B}(v)$, showing that arbitrary recursive memory reduces to quantization of the current predictive state and yielding $B^\\star(d) = \\frac{1}{2}\\log_2\\frac{a^2v}{d-q} + O(1)$. (ii) Removing stationarity, Theorem 2 shows that memory tracks the scale of the predictive state: $B_t^\\star(d) = \\frac{1}{2}\\log_2 v_t + O(1)$, producing bounded, logarithmic, and linear growth for stable, critical, and unstable dynamics, respectively. (iii) Under partial observability, Theorem 3 bounds the excess prediction risk between quantization of the Bayesian predictive state and the better of two finite-memory mechanisms—history retention and recursive belief representation—thereby separating inferential from representational memory. (iv) Beyond Gaussianity, Theorem 4 preserves the exact quantization reduction while introducing the entropy-power factor $\\beta = e^{-2D_{\\mathrm{KL}}(P_X\\Vert\\mathcal{N}(\\mu,v))}$, showing that equal variance need not imply equal finite-memory difficulty. (v) Theorem 5 specializes the theory to an isotropic KV-cache model: with $R = B/(dL)$ bits per scalar, $D_L^\\star \\geq \\sigma_\\eta^2/(1-\\rho^2 2^{-2R})$, and half of the asymptotically available predictive gain occurs at $R_{\\mathrm{onset}} = \\frac{1}{2}\\log_2(1/(1-\\rho^2))$.\n\nMeasurements from Transformer key representations provide an empirical scale check for the stylized cache regime. The results show that persistent memory is a property of the predictive process rather than merely of the predictor architecture: its required capacity is jointly governed by predictive-state geometry, dynamical expansion, and inferential uncertainty. This provides a principled basis for understanding memory in continual world models, including what predictive information must persist, how accurately it must be represented, and when fixed-capacity memory is fundamentally insufficient.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "predictive memory complexity", "bounded-memory prediction", "persistent state", "predictive states", "rate-distortion theory", "sequential prediction", "partial observability", "KV-cache compression" ], "forum_url": "https://openreview.net/forum?id=NZAbSVNeqK", "pdf_url": "https://openreview.net/pdf/a8f926c1630a5d8cefa08cd226b7b89699474d1d.pdf", "updated_at": "2026-10-05T19:51:37.840Z", "categories": [ "kv-cache", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W7ENwItdpD", "number": 22059, "title": "Contextual Utility of Quantization Moves in Extreme Low-Bit LLMs", "abstract": "Accurate measurements of isolated weight replacements need not predict their joint effect on a quantized model. We study this gap for legal replacements at a fixed bitrate, separating isolated pricing error, composition error and changes in evaluation data. On Llama-3.2-1B and activation-ordered 3B GPTQ bases, option-KL composition residuals are 26× and 42× the sums of absolute midpoint-pricing errors; large residuals remain with exact-singleton prices. Exhaustive Qwen3-0.6B lattices reveal context-dependent effects: six of eight moves in the stratified lattice have statistically resolved ARC-KL sign reversals, excluding an exact monotone additive representation. Pair measurements improve prediction of unqueried combinations on both Qwen and Llama. Their value for selection depends on the criterion: the stratified Qwen front retains a quality-coverage advantage after refitting, whereas the 3B additive and quadratic fronts are nearly equal in quality at a one-standard-error tolerance. Low-order endpoint structure leaves continuous gradient responses undetermined, and shared-midpoint prices can differ from those of the executed subset. Support-matched controls show that restricting the number of replacements curbs the failures of large combinations, while additional endpoint search gives bank-dependent NLL/ARC trade-offs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization", "extreme low-bit quantization", "post-training quantization", "contextual utility", "quantization interactions", "GPTQ", "model compression" ], "forum_url": "https://openreview.net/forum?id=W7ENwItdpD", "pdf_url": "https://openreview.net/pdf/b73f5fc92f97c08e82880858ec1db8ff1f11cb7d.pdf", "updated_at": "2026-10-05T19:51:44.950Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bW0e9QqvKm", "number": 22061, "title": "Overcoming the Communication-Performance Tradeoff in LLM Pretraining", "abstract": "Communication-efficient distributed training algorithms (e.g., DiLoCo) have received considerable interest due to their benefits for training large language models (LLMs) in bandwidth-constrained settings, such as across datacenters and over the internet. While these local-update methods achieve communication reduction through reduced synchronization, they still require communicating dense model-sized pseudo-gradients, resulting in a communication bottleneck even for cross-datacenter links. While quantization is often applied to reduce the pseudo-gradient's size, in the context of LLM pre-training, existing approaches have not been able to leverage sparsification without incurring significant performance degradation. In this work, we introduce SparseLoCo, a communication-efficient training algorithm for LLMs that can effectively leverage Top-k sparsification and 2-bit quantization to reach extreme sparsity in the communicated pseudo-gradient, as high as 97-99%, while achieving lower final loss than dense DiLoCo. In our empirical study of language model pre-training, we demonstrate that SparseLoCo's improvements over DiLoCo in performance and pseudo-gradient compression are maintained across dense model scales (178M-2B), an MoE transformer (645M-A273M), increasing the number of workers, and increasing communication intervals.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "LLM pre-training", "distributed training", "communication efficiency", "gradient compression", "Top-k sparsification", "error feedback", "LocalSGD", "DiLoCo" ], "forum_url": "https://openreview.net/forum?id=bW0e9QqvKm", "pdf_url": "https://openreview.net/pdf/25e0b8f9b324d7d2ae6be689e2e656651262ec8e.pdf", "updated_at": "2026-10-05T19:51:45.492Z", "categories": [ "training" ], "matched_terms": [ "2-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CvRKdl4k2w", "number": 22122, "title": "OffQ: Taming Structured Outliers in LLM Quantization by Offsetting", "abstract": "Low-bit quantization is now a standard approach for accelerating large language model (LLM) inference, as it substantially reduces memory usage and compute cost.\nHowever, structural activation outliers remain a major obstacle to effective quantization.\nIn this paper, we introduce OffQ, a post-training quantization method that mitigates this issue through a novel offsetting mechanism.\nSpecifically, OffQ first identifies a low-dimensional outlier subspace in the activations using a tailored top-1 PCA, and then concentrates high-magnitude activations into a single channel by rotation.\nIt then absorbs this concentrated outlier energy into a shared offset, reducing the variance of the activations and making them more amenable to uniform-grid, uniform-precision quantization.\nThis strategy enables accurate W4A4KV4 quantization of LLMs while preserving efficient low-bit execution.\nExperiments across diverse LLM architectures and benchmarks show that OffQ consistently outperforms state-of-the-art baselines, improving accuracy while maintaining low-bit efficiency.\nCode will be released upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization", "singular defect", "W4A4KV4" ], "forum_url": "https://openreview.net/forum?id=CvRKdl4k2w", "pdf_url": "https://openreview.net/pdf/4986832aa660a0e57eb82d2a926cd376ddf4da04.pdf", "updated_at": "2026-10-05T19:51:52.748Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "quantization", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1e9jAnBoYb", "number": 22135, "title": "Quantization Enables Private Dense Retrieval Against Malicious Service Providers", "abstract": "Dense retrieval, the key component of Retrieval Augmented Generation (RAG), retrieves the most relevant documents by comparing dense vector representations of queries and passages from a large corpus. In privacy-sensitive applications, the server observes the query and controls which evidence is returned, creating both confidentiality and integrity risks. We formulate private dense retrieval as providing query privacy and retrieval integrity against a malicious server, and develop a two-round cryptographic protocol that provides both guarantees. Our protocol reduces private and verifiable retrieval to multiplication of a committed matrix by an encrypted vector and uses low-bit quantization to make this computation practical. We evaluate the resulting trade-off between cryptographic cost, retrieval quality, and downstream RAG accuracy across six embedding models, four language models, and corpora of up to 2.68 million passages. Our results show that, with a clipped quantizer, three-bit quantization largely preserves retrieval quality and downstream accuracy, while a private query over a corpus the size of a clinical reference requires one to three minutes of server time. These results suggest that private dense retrieval is already practical for moderately sized, privacy-sensitive corpora when minute-scale latency is acceptable.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Privacy-preserving protocols", "dense retrieval", "RAG", "quantization" ], "forum_url": "https://openreview.net/forum?id=1e9jAnBoYb", "pdf_url": "https://openreview.net/pdf/4561c49258d24da66ef0387e5ecd15543a28ae0f.pdf", "updated_at": "2026-10-05T19:51:54.548Z", "categories": [ "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yQ5OaQwfcn", "number": 22185, "title": "TACT: Tiered Activation Sparsity", "abstract": "Deploying LLMs on consumer-grade GPUs with insufficient memory to hold their weights can result in prohibitively slow inference, because decoding repeatedly transfers offloaded weights from system RAM or flash storage into GPU at much lower bandwidth than local GPU-memory access. Activation sparsity reduces these transfers by skipping weights associated with zero or near-zero activations. However, as more activation contributions are omitted, model quality eventually degrades rapidly, indicating that weights associated with small-magnitude activations collectively influence model quality sharply. In this work, we improve the trade-off between model quality and decoding performance when exploiting activation sparsity. Our key idea is to replace the binary choice of whether or not to read a weight with three options: fully retain it, approximate it using a compressed weight representation, or omit it entirely. TACT skips weights associated with activations closest to zero, reads approximate weights for smaller-magnitude activations, and reads original weights for the largest-magnitude activations. Smaller-magnitude activations attenuate the errors introduced by approximate weights, while compressed weight representations require fewer bytes to be transferred. With weights offloaded to CPU memory, \\method speeds up decoding by $3.86\\times$ on average (up to $5.57\\times$) with 16-bit weights and $2.06\\times$ (up to $2.74\\times$) with 4-bit weights, at a WikiText-2 perplexity increase of at most 10%. With weights offloaded to flash storage, the speedups are $3.31\\times$ on average (up to $4.81\\times$) and $1.54\\times$ (up to $2.03\\times$).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Activation Sparsity", "Deep Learning Performance", "Model Compression", "LLMs", "On-device Inference" ], "forum_url": "https://openreview.net/forum?id=yQ5OaQwfcn", "pdf_url": "https://openreview.net/pdf/3b35d5aaa1da881a2836b67f42322a0d2985869b.pdf", "updated_at": "2026-10-05T19:52:05.629Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vwsfTsPMgw", "number": 22216, "title": "LMTC: Quantizing KV Caches for How Attention Reads Them in Omni-Modal LLMs", "abstract": "Omni-modal Large Language Models (OmniLLMs) unify the understanding of\nvideo, audio, and text, yet their continuous streaming inputs generate an\noverwhelming volume of tokens, making the Key-Value (KV) cache a major\nmemory bottleneck. Existing KV cache quantization methods are suboptimal\nfor OmniLLMs because they either minimize simple reconstruction error or\nrely on query activations that are unavailable when streaming tokens are\ncached, ignoring both the attention reading mechanism and cross-modal\nheterogeneity. To address these challenges, we propose Logit-Metric\nTransform Coding (LMTC), a training-free and query-free KV cache\nquantization framework that minimizes the expected attention error using\nonly the query projection weights.\nSpecifically, LMTC quantizes Keys using a Modality-Conditioned Transform\nCoder guided by a Query-Free Logit Metric derived directly from query\nprojection weights, thereby encoding each modality in its own basis\nand allocating precision via reverse water-filling to the\ndirections attention reads. Simultaneously, LMTC handles Values through\nPhase-Offset Rounding, a metadata-free mechanism that offsets the\nrounding grids of spatio-temporally adjacent visual tokens so that\nrounding errors cancel out during attention aggregation rather than\naccumulate. Extensive evaluations across three OmniLLMs on eight\nbenchmarks show that LMTC is virtually lossless at 3 effective bits per\nscalar, retaining over\n99\\% of BF16 average accuracy. Compared to\nBF16 FlashDecoding-v2, LMTC achieves up to a $1.94\\times$ decoding\nspeedup, a $5.28\\times$ KV cache memory reduction, and a $2.08\\times$\nthroughput boost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "omni-modal large language models" ], "forum_url": "https://openreview.net/forum?id=vwsfTsPMgw", "pdf_url": "https://openreview.net/pdf/e49a067fb1b5079cec3c1195aedac41840580a8b.pdf", "updated_at": "2026-10-05T19:52:10.393Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantization", "quantizes", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "kJh6QRjbQW", "number": 22258, "title": "MaLaCo: Masked Latent Completion for Efficient Generative Recommendation", "abstract": "Semantic-ID (SID) recommenders retrieve items by generating short discrete identifiers, but masked-diffusion decoding repeatedly processes every beam hypothesis through the full history backbone. We introduce \\emph{masked latent completion} (MaLaCo), which factorizes retrieval into a single bidirectional history-completion pass followed by trie-constrained SID generation with a small autoregressive head. Beam-dependent computation is therefore confined to the head. On one A100, at beam width 50, MaLaCo reduces amortized generation cost from $13.097$ to $0.526$ ms/query at batch $20$ ($24.9 \\times $) and single-request latency from $24.453$ to $9.984$ ms ($2.4 \\times $), despite a higher total parameter count relative to LLaDA-Rec, a leading generative retrieval baseline. Across four datasets, MaLaCo improves Recall@$k$ over LLaDA-Rec at every reported cutoff \\(k\\geq10\\), with larger gains at deeper cutoffs. It trades Recall@1 for stronger deep recall on Industrial \\& Scientific, Musical Instruments, and Video Games, while also improving Recall@1 on Arts, Crafts \\& Sewing. A four-seed \\(2\\times2\\) architecture--encoder study confirms that this depth-dependent advantage persists under either item encoder, while the stronger encoder consistently benefits MaLaCo. We further introduce a collision-aware SID construction pipeline combining balanced residual quantization, last-level deduplication, learned candidate refinement, and catalog-level assignment. Across all four catalogs, it reduces the excess-collision rate from \\(18.12\\%\\)--\\(31.04\\%\\) to at most \\(0.0122\\%\\). On an item-arrival partition of Arts, the method inserts the newest quarter of the catalog without changing existing identifiers or introducing new collisions. After recommender retraining, the extended vocabulary obtains \\(0.04526\\) Recall@10 versus \\(0.04598\\) after complete vocabulary refitting, a $1.6\\%$ relative shortfall. Together, these results position MaLaCo as an efficient, high-recall candidate generator with collision-aware and stably extensible semantic identifiers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "generative recommendation", "semantic IDs", "masked diffusion", "efficient inference", "catalog expansion" ], "forum_url": "https://openreview.net/forum?id=kJh6QRjbQW", "pdf_url": "https://openreview.net/pdf/85001bed1ef2782368b057a7200d69ffabb5906c.pdf", "updated_at": "2026-10-05T19:52:14.662Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QAltVBXLtx", "number": 22337, "title": "Score Centering Stabilizes Off-policy Reinforcement Learning", "abstract": "Reinforcement learning (RL) of large language models is notoriously sensitive to small differences between training and inference engines, often referred to as the training-inference mismatch (TIM). However, completely eliminating TIM is impractical, as it would come at a major cost to rollout efficiency. In this paper, we show that the instability of RL under TIM is primarily caused by drift: a persistent bias between training and inference engines that accumulates with every training step. We derive an additive \"score centering\" correction term that stabilizes RL under TIM by canceling drift. When training models from 0.6B to 30B parameters, score centering alone matches or outperforms methods based on importance sampling under quantization, with the gap growing as the mismatch becomes more severe. Because the correction is additive, score centering also composes with importance sampling -- their composition outperforms pure importance-sampling baselines in our staleness experiments.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "reinforcement learning", "training-inference mismatch" ], "forum_url": "https://openreview.net/forum?id=QAltVBXLtx", "pdf_url": "https://openreview.net/pdf/b756701ab1cb922090b5af42772e00cf5414ea40.pdf", "updated_at": "2026-10-05T19:52:25.527Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "P5WG3ugXiS", "number": 22390, "title": "Delta-Matching: Closing the Final Gap of Native 8-bit Training for LLMs", "abstract": "Reliable FP8 attention remains a barrier to fully native 8-bit large language model training.\nWe derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics.\nOur stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss and downstream degradation at 1.67B and 5.29B.\nQK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it.\nThis pattern suggests accumulated optimization error that smaller models and short runs can conceal.\nWe propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions.\nIt enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs.\nAcross tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance.\nWe will release our implementation, trained models, and data recipes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Precision Training", "Quantization", "Attention", "Large Language Model", "Efficient AI", "Numerical Stability" ], "forum_url": "https://openreview.net/forum?id=P5WG3ugXiS", "pdf_url": "https://openreview.net/pdf/5ed98c872f949151f7187cae1d2bf4cc50704ca8.pdf", "updated_at": "2026-10-05T19:52:35.826Z", "categories": [ "training" ], "matched_terms": [ "8-bit", "fp8", "low-precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "S2KSOSEFaH", "number": 22473, "title": "TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning", "abstract": "Full-parameter fine-tuning of large language models (LLMs) incurs substantial optimizer state memory overhead, limiting the model sizes that fit on modern GPUs. Existing approaches either compress optimizer state, abandon first-order gradients, or change the update geometry while retaining dense state. The recently introduced Muon optimizer reduces optimizer memory through matrix-valued updates. Still, its geometry differs from AdamW and can lead to performance degradation when fine-tuning AdamW-pretrained models. To reduce optimizer memory without sacrificing accuracy or computational efficiency in LLM fine-tuning, we propose **T**ernary **A**bsolute-max **C**olumn-wise **O**ne-sparse optimizer, or **TACO**, which follows Muon's operator-norm steepest-descent view but takes the geometric route further. TACO computes the exact steepest-descent direction under a dimension-normalized $1\\to 1$ operator norm by selecting the sign of the largest magnitude entry in each column of two-dimensional weight matrices. This retains first-order gradients while making optimizer state memory nearly negligible. Our practical TACO optimizer maintains only a small set of low precision gradient components per column, reducing persistent optimizer state by $174\\times$ relative to AdamW8bit (from $27.7$ GB to $0.16$ GB) and peak training memory by $2.9\\times$ (from $80.6$ GB to $27.5$ GB) on OPT-13B, while achieving comparable accuracy and runtime. TACO further enables full-parameter fine-tuning of 30-32B-parameter models on a single 80 GB H100 GPU across multiple model families and tasks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLMs Fine-Tuning" ], "forum_url": "https://openreview.net/forum?id=S2KSOSEFaH", "pdf_url": "https://openreview.net/pdf/d6a8f0f5f6bb9b9da2b777b9ed0e532b546b28e1.pdf", "updated_at": "2026-10-05T19:52:50.473Z", "categories": [ "training", "hardware" ], "matched_terms": [ "low precision", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1W5aP7VZOd", "number": 22552, "title": "Computed Permutation Trellis Coding: Near-Optimal Quantization Designed for Native Hardware Acceleration", "abstract": "We propose Computed Permutation Trellis Coding\n(CPTC), a low-bit weight format whose decoder multiplies the\ntrellis state by two $16$-bit coefficients and uses the top bits\nof the products to index a $64$-byte table. Formats that round\neach value independently, such as INT4, FP4, and NF4, waste\n$0.6$ or more of their $4$ bits against the information-theoretic\nideal; trellis codes recover this loss by encoding values\njointly, but prior decoders either store a large table (Random\nPermutation Trellis Coding, RPTC) or replace it with a hash (QTIP) that fixes the distribution of the\ndecoded values. CPTC computes the table index instead: the index\ndistribution the two multiplications induce is a near-optimal\nrule for placing codepoints in a small table. One decoder with\n$16$-bit multipliers, a $64$-byte table and a coefficient pair loaded per\nconfiguration serves every bit width $K \\in [1, 6]$, reproduces\nthe standard integer and floating-point formats exactly, and loses at most\n$0.07$ bits against the information-theoretic ideal for all\n$K \\le 4$, matching RPTC run two state bits deeper from an\n$8{,}192\\times$ smaller table and exceeding QTIP's best\nvariant ($3.93$ vs.\\ $3.86$ effective bits at $K{=}4$). In LLM\nweight quantization without\nfine-tuning, CPTC has the lowest KL divergence to the unquantized\nmodel at every tested bit width on LLaMA 3.1 8B and 70B Instruct against\nQTIP's HYB (by up to $27\\%$) and 1MAD (by up to $5\\%$) codebooks,\nand at $3$--$4$ bits surpasses even YAQA's fine-tuned $70$B\nreleases. A fused persistent GPU Viterbi kernel encodes an\n8B-parameter model in about two GPU-minutes, and the decoder has\nbeen built into the weight-load path of a commercial\nsystolic-array accelerator, where, post-layout, it runs at native FP4/FP8\nthroughput for about $3\\%$ of the systolic array's area.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "trellis coding", "quantization", "LLM", "inference", "machine learning systems" ], "forum_url": "https://openreview.net/forum?id=1W5aP7VZOd", "pdf_url": "https://openreview.net/pdf/51edd405a38a7f365b14f6666b074b0b69b129d7.pdf", "updated_at": "2026-10-05T19:53:00.504Z", "categories": [ "training", "hardware" ], "matched_terms": [ "bit width", "fp4", "fp8", "int4", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nFLnHIUMlx", "number": 22607, "title": "AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation", "abstract": "Hadamard transforms have become a standard tool for stabilizing MXFP4 LLM training by spreading outliers within its 32-element scaling blocks. We show that their effectiveness depends on a simple geometric factor: how outliers are oriented relative to the scaling axis. Outliers perpendicular to the axis create within-block concentration that Hadamard mixing can effectively reduce, whereas outliers aligned with the axis enlarge the shared block scales, leaving a scale-driven component that rotation does not directly address. We call this the scaling-axis blind spot. Measurements of LLM training tensors reveal stable Row-wise, Column-wise, and None outlier patterns whose combinations vary across GEMMs; notably, 97.8--100\\% of weight-gradient GEMMs contain a blind-spot operand. We introduce \\textbf{AdaHOP}, an outlier-pattern-aware MXFP4 training framework that identifies these patterns in a short calibration phase, uses Inner Hadamard Transform (IHT) as the low-precision backbone, and selectively extracts only the blind-spot outlier rows or columns into a small BF16 path. With fused Triton kernels, AdaHOP-Lv1 executes less than 1\\% of GEMM FLOPs in BF16 while trailing BF16 by only 0.23 points in average zero-shot accuracy across models from 1B to 8B. On Llama3.1-8B, it reduces memory by $3.6\\times$ and accelerates end-to-end training by $ 1.45\\times $ over BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Precision Training", "Quantization for training" ], "forum_url": "https://openreview.net/forum?id=nFLnHIUMlx", "pdf_url": "https://openreview.net/pdf/2a1065f288e992959451269ac139392a8d1cec74.pdf", "updated_at": "2026-10-05T19:53:14.986Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "low-precision", "mxfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zNMUu0NUQx", "number": 22648, "title": "PulseQuant: Propagation-Guided Subspace Correction for 4-Bit Video Diffusion Transformers", "abstract": "Quantization errors in video diffusion transformers can be amplified or attenuated by subsequent denoising updates, making local reconstruction error an incomplete predictor of final impact. \nWe introduce PulseQuant, a 4-bit post-training quantization method that combines trajectory sensitivity with activation geometry to guide offline calibration. \nIsolated block--step interventions estimate propagation risk, which prioritizes sensitive trajectory states during row-radius selection. \nWith these radii fixed, response-subspace correction uses neighboring-code edits to reduce residual components along dominant activation directions. \nBoth stages preserve the original 4-bit weight representation. \nControlled interventions show that short-horizon propagated error predicts final latent error more reliably than immediate block-output error, supporting calibration beyond local reconstruction objectives. \nEvaluations on Wan models, Self Forcing, and MiniMax-H3 demonstrate improvements in key consistency and dense-reference metrics while remaining competitive on other attributes across model scales and generation paradigms.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Low-Bit Quantization", "Efficient Video Generation" ], "forum_url": "https://openreview.net/forum?id=zNMUu0NUQx", "pdf_url": "https://openreview.net/pdf/b4cdaae36c00668e937dee9c43bd6b0f7b088492.pdf", "updated_at": "2026-10-05T19:53:21.425Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "4-bit", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hT5MpfcmG8", "number": 22716, "title": "Efficient Residual-Centric Inference for Spacetime Autoregressive Video Generation", "abstract": "Spacetime autoregressive video models combine multiscale spatial generation\nwithin each clip with temporal autoregression across clips, avoiding repeated\nfull-resolution denoising while enabling previously generated visual states to\nbe reused as context. This design, however, introduces two complementary inference\nbottlenecks: attention computation becomes concentrated at late,\nhigh-resolution scales, while reusable history creates a persistent\nKV cache.\nA natural solution is to sparsify attention and compress historical KV, but\nthe two are typically optimized with different proxy objectives. We observe\nthat once omitted attention is replaced by a cheap pooled correction, the\nrelevant quantity is no longer raw attention importance or standalone tensor\nreconstruction error, but the output residual left unexplained by the\napproximate attention operator. Based on this observation, we introduce residual-centric inference,\nwhich provides a shared post-correction residual view for compute allocation\nand cache representation. We instantiate this principle with\nOutput-Sensitive Routing (OSR), which allocates token-level attention using\na proxy for post-correction residual risk, and Residual-Interaction\nQuantization (RIQ), which preserves pooled statistics and fits low-bit\nhistory representations to residual K--V interactions. On InfinityStar, OSR achieves up to\n1.86× end-to-end acceleration at native 720p/5s, while an orthogonal\nFFN FP8 optimization increases the complete-system speedup to 2.01×.\nAt 480p/10s, OSR+RIQ achieves up to 1.50× speedup and\n5.74× retained-history KV compression. Across both settings and tasks, these configurations largely preserve generation quality, with at most 0.31% relative degradation in VBench and VBench-I2V scores.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "efficient inference", "video generation", "visual autoregressive models", "sparse attention", "KV cache quantization", "low-bit quantization" ], "forum_url": "https://openreview.net/forum?id=hT5MpfcmG8", "pdf_url": "https://openreview.net/pdf/2ab8e0f1e90743a0928ea8538843c218ec91a7f3.pdf", "updated_at": "2026-10-05T19:53:30.211Z", "categories": [ "ptq", "training", "kv-cache", "diffusion" ], "matched_terms": [ "fp8", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nYra38pdbe", "number": 22744, "title": "Off-Diagonal Hessian Sensitivity for Post-Training Quantization of Large Language Models", "abstract": "Mixed-precision post-training quantization (MP-PTQ) is a promising approach for improving resource efficiency under constrained computational and memory budgets through effective bit allocation. \nMP-PTQ methods based on independently optimized layer-wise or module-wise surrogate objectives are computationally efficient, but primarily rely on local quantization errors and therefore do not explicitly account for cross-layer interactions under the task loss.\nIn contrast, search-based methods treat mixed-precision allocation as a black-box optimization problem, leading to costly exploration of a large search space.\nTo address these limitations, we propose Off-Diagonal Energy Sensitivity (ODES), which captures interactions within and across projections through the off-diagonal structure of the task-loss Hessian.\nStarting from a second-order expansion of the task loss, we separate its off-diagonal interaction term. Then, we derive a sufficient condition under which ODES-guided precision reallocation reduces the second moment of the off-diagonal task-loss contribution. To estimate ODES at LLM scale without materializing the Hessian, we develop a compressed probe-centered (CPC) estimator. CPC reduced estimator state memory by approximately \\(126\\times\\) while producing ODES estimates that closely matched those obtained with the direct Welford estimator.\nUsing the estimated ODES ranking, we reduce mixed-precision bit allocation to a one-dimensional search over the number of reallocated projection pairs and evaluate these candidates using HQQ as a lightweight quantization proxy. This reduced the total allocation time by 36--50\\% relative to search-based methods while maintaining comparable performance in downstream accuracy and perplexity. ODES-guided allocations also improved over uniform allocation with RTN, HQQ, and GPTQ, indicating that the ODES ranking transfers across different quantizers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "large language models", "mixed-precision quantization", "off-diagonal Hessian", "Hessian sensitivity profiling", "matrix-free Hessian estimation" ], "forum_url": "https://openreview.net/forum?id=nYra38pdbe", "pdf_url": "https://openreview.net/pdf/0ccf095e2c7e837222fcd17f125af9dfbd4de01e.pdf", "updated_at": "2026-10-05T19:53:33.202Z", "categories": [ "ptq" ], "matched_terms": [ "mixed-precision", "ptq", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ADnB3N5A2D", "number": 22768, "title": "Temporal Quantization for Adaptive Candidate Frame Construction in Video QA", "abstract": "Video language models (VLMs) operate under limited visual-token and computational budgets, requiring videos to be represented by a compact set of frames. A common pipeline first constructs a temporally subsampled candidate frame set and then selects the final input frames from it. While conventional pipelines rely on query-agnostic temporal sampling, recent work has increasingly explored adaptive frame acquisition guided by the query and partial observations. Yet, how a fixed observation budget should represent the video timeline before downstream selection remains less explicitly characterized. To address this, we introduce QUEST, a training-free method that formulates candidate construction as relevance-weighted temporal quantization, explicitly measuring how well the acquired observations resolve the timeline. By jointly considering estimated query relevance and temporal gaps between observations, QUEST allocates finer temporal resolution to relevant yet under-resolved regions while maintaining coverage elsewhere. The resulting candidate set can be used with existing frame selectors without modifying their selection rules. Extensive experiments across multiple VLMs, frame selectors, and video understanding benchmarks demonstrate improved downstream accuracy under matched observation and selection budgets.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Large Language Model", "Video understanding", "Retrieval" ], "forum_url": "https://openreview.net/forum?id=ADnB3N5A2D", "pdf_url": "https://openreview.net/pdf/a67c2c3d4b2e26d02af2153ad0b85cb0488e9d68.pdf", "updated_at": "2026-10-05T19:53:37.754Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HXmiKD1kS0", "number": 22773, "title": "RC-VQ: Recovering What Calibration Discards with Vector-Quantized Residual Correction", "abstract": "Calibration data decides what training-free LLM compression keeps. In activationaware\nfactorization and post-training quantization alike, a small calibration set\ndefines an importance measure over the weights, and whatever it does not certify\nis discarded as a residual. We argue that failing the calibration proxy does not\nmake weights unimportant, and that the discarded residual should be recovered.\nThe standard recovery tool, a low-rank adapter, is structurally mismatched to this\ntask. On LLaMA-2 7B the residuals are dense and spectrally flat. Quantization\nresiduals are statistically indistinguishable from a Marchenko–Pastur noise bulk,\nand truncation residuals form a gapless spectral continuum without isolated spikes,\nso a rank-r corrector recovers a vanishing fraction of either. Across four model\nfamilies, the energy a low-rank corrector recovers grows only linearly in its rank.\nWe propose RC-VQ, a full-rank vector-quantized corrector fitted to the residual\nwithout gradients or recovery data, which composes with SVD factors, INT4\nweights, and their stacked combination. On structural compression of LLaMA-2\n7B, LLaMA-3 8B, and Qwen3 8B, RC-VQ reduces perplexity degradation by\n3–13× against the strongest reported structured baseline at matched checkpoint\nsize. On the INT4-AWQ quantization gap of Qwen3.5 4B, it recovers 96% of the\ndistance to FP16 at three bits per parameter, a budget at which an identically sized\nlow-rank corrector recovers under 40% and a scalar residual branch needs 37%\nmore bits to match. Sweeping codebook objectives across structural, quantization,\nand stacked residuals, we show that each pure objective has a failure regime, while\nthe hierarchical direction-then-refinement objective, PolarVQ, never collapses,\nmaking it the robust default when residual geometry is unknown.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "vector quantization", "large language models", "low-rank factorization", "residual correction", "on-device AI" ], "forum_url": "https://openreview.net/forum?id=HXmiKD1kS0", "pdf_url": "https://openreview.net/pdf/481b57bafa6bf18e0bfe28faae17074153491fca.pdf", "updated_at": "2026-10-05T19:53:38.601Z", "categories": [ "ptq", "training" ], "matched_terms": [ "int4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "T1USDcgiBK", "number": 22844, "title": "Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation", "abstract": "Modern video diffusion models achieve higher generation quality through scaling, but this also increases inference cost. Although many acceleration methods have been proposed, a central challenge is that the most effective acceleration strategy is highly instance-specific: a recipe that works well for one combination of model, hardware, and inference configuration often does not transfer to another. Different models vary in architecture, numerical sensitivity, and attention concentration patterns. Inference settings differ in spatial and temporal resolution and video duration, while hardware platforms differ in memory hierarchy, supported numerical formats, and kernel throughput. These factors create a large tuning space, making manual performance engineering costly. We present Sol Video Inference Engine, an agentic, native, training-free acceleration framework for video diffusion models. It organizes five broadly applicable techniques, cache, sparse attention, token pruning, quantization, and kernel fusion, into an agentic acceleration stack for instance-specific optimization. For a concrete deployment target defined by a model, hardware platform, and serving configuration, parallel skill agents optimize the implementation of each technique, an agent integrator composes them into a global acceleration stack, and a human validator provides feedback on generation quality. We instantiate this workflow on three video models with different sizes and architectures: 64B Cosmos3-Super, 33B MiniMax-H3, 22B LTX-2.3, and 2B SANA-Video. With little human effort, the full stack achieves more than 2x end-to-end acceleration while maintaining near-lossless VBench quality, demonstrating the effectiveness of the agent framework for video diffusion acceleration.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Agentic workflow", "Video generation", "Inference acceleration", "Kernel optimization" ], "forum_url": "https://openreview.net/forum?id=T1USDcgiBK", "pdf_url": "https://openreview.net/pdf/49975650618c874d93c5f6eb70643edbf2ccf82a.pdf", "updated_at": "2026-10-05T19:53:52.425Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Lz755URdm1", "number": 23007, "title": "Ghost Your System: Audit Fibers Hide Hijacks Behind Passing Checks", "abstract": "Deployment pipelines routinely accept non-identical model components produced by quantization, export, or compilation as long as they preserve the behavior checked during qualification. However, qualification observes only the declared component interface, whereas the deployed system later composes that component with downstream consumers. A malicious replacement must preserve every checked output on the same execution it redirects. For reusable components, this requires one fixed replacement that operates across inputs. We show that the check may discard distinctions that downstream consumers still use. The states it treats as equivalent form an audit fiber. We prove that qualification is unsafe whenever a component can reach two states in one fiber that induce different downstream behavior. For normalized representations, we derive an exact targeted attack within the fiber and compile it into FiberRouter, a fixed malicious replacement. Experiments across image and text representations and a Qwen model KV-cache handoff demonstrate reliable downstream hijacking while preserving the qualification that authorizes deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "AI Security", "Model Supply Chain", "Compositional Integrity" ], "forum_url": "https://openreview.net/forum?id=Lz755URdm1", "pdf_url": "https://openreview.net/pdf/ab97e684d03b43065f09b19de64acee4e5aadcca.pdf", "updated_at": "2026-10-05T19:54:18.679Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KZGndNHcs6", "number": 23093, "title": "No Man Is an Island: Overlapping Pruning for Near-Unstructured Quality at 2:4 Efficiency", "abstract": "Semi-structured 2:4 sparsity is the dominant fine-grained pruning pattern on modern ML\naccelerators, yet for large language models it degrades quality far more than unstructured pruning\nat the same 50\\% sparsity. We trace this gap to mask rigidity: 2:4 partitions weights into\ndisjoint four-element windows, so both selectors draw from the same four inputs, severely\nconstraining the feasible mask space. We propose Overlapping Pruning (OP), an\nalgorithm--hardware co-designed sparsity structure that overlaps neighboring selector windows,\nenlarging the mask space while preserving 2:4's selectors, 2-bit index, and datapath. At 4-, 8-,\nand 16-bit, OP-1D and OP-F2D cost at most $1.09\\times$ 2:4's core area, $0.94$-$1.04\\times$ its\ndelay, and $0.93$-$1.16\\times$ its power. OP-2D additionally yields transposable masks. Feasible\nmasks are the independent sets of a transversal matroid, so a saliency-ordered greedy pass with a\nper-candidate feasibility check is exact for every OP pattern; our kernels generate OP-F2D masks\nwithin $1.6\\times$ the 2:4 mask-generation cost. One-shot OP-F2D with Wanda++ increases perplexity\nby at most $10.4\\%$ over unstructured 50\\%, versus up to $59.8\\%$ for 2:4. OP-1D attains lower\nperplexity than 4:8 and OP-F2D stays within $1.1\\%$ of 8:16; $N{:}M$ patterns cost\n$1.2$-$4.2\\times$ 2:4's core area and $1.7$-$3.4\\times$ its delay. OP-F2D matches or improves on\nreported results from learned-mask and factorized 2:4 methods at a fraction of their pruning cost,\nwith no mask training or added parameters. By relaxing mask geometry alone, OP-F2D closes\n$75.8$-$88.8\\%$ of the unstructured-to-2:4 quality gap at 2:4's hardware cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Pruning", "Sparsity", "N:M Semi-Structured Pruning", "Large Language Models", "Efficient ML", "Inference Acceleration", "Combinatorial Optimization", "2:4" ], "forum_url": "https://openreview.net/forum?id=KZGndNHcs6", "pdf_url": "https://openreview.net/pdf/9f86f16c6e1f63e328b1e77ebc2ae4c6dcbcd52c.pdf", "updated_at": "2026-10-05T19:54:32.703Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bXpH3h8p52", "number": 23186, "title": "Observation-Aware QIM for Gaussian Splats", "abstract": "When one 3D Gaussian-splatting asset is released to several recipients, a leaked copy should\nreveal which release it came from. Releases share geometry exactly, so the identifier can live\nonly in the spherical-harmonic (SH) colour coefficients, where it must stay invisible yet survive\nre-export. Existing marks need trained decoders, and a direct write budgets its change in\ncoefficient distance although camera sensitivity spans a median 7.2 orders of magnitude, leaving\nmargins too thin for 8-bit export. We instead spend the budget in the renderer's own error metric,\nplacing the mark where cameras barely look. Observation-aware quantization index\nmodulation (QIM) solves each keyed quantization write in closed form in a metric built from\ncompositing weights, and a compositing inequality bounds the image error of the complete write on\nthe declared cameras, certifying each write before release. On fourteen scenes at matched\ntraining error, exact recovery after 8-bit export rises from 10/70 to 70/70 over Euclidean and\ndiagonal controls, at higher held-out error; at a matched 48-bit payload, every identifier\nsurvives half-pruning at a marking cost of $.0005$ dB, against $3.99$ dB for GaussianMarker.\nFloat16 coordinates, SH truncation, and a keyless recipient holding the declared cameras remain\nfailure modes. These results show that solving the write in the renderer's metric lets a\ntraining-free mark survive export, and they give native 3D assets a certified release identifier\nwith measured limits.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "3D Gaussian splatting watermarking", "quantization index modulation", "spherical harmonic coefficients", "certified distortion bound", "robustness to pruning and quantization", "asset provenance" ], "forum_url": "https://openreview.net/forum?id=bXpH3h8p52", "pdf_url": "https://openreview.net/pdf/4d9e8add1ea2d80f1534fd4f1ab2f5966d697b27.pdf", "updated_at": "2026-10-05T19:54:47.351Z", "categories": [ "theory" ], "matched_terms": [ "8-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LLnTvYzOq0", "number": 23285, "title": "DyMPQ: Dynamic Mixed-Precision Quantization for Efficient Diffusion Transformers", "abstract": "Diffusion transformers (DiTs) have demonstrated strong performance in image and video generation, but their large model sizes and iterative denoising incur substantial computational and memory costs.\nHowever, timestep-dependent activation outliers and channel-wise magnitude disparities hinder accurate 4-bit quantization.\nTo address these challenges, we propose Dynamic Channel-wise Mixed-Precision Quantization (DyMPQ) for DiTs. It combines a shared W4A4 main branch with stage-specific high-precision branches to absorb timestep-varying activation outliers.\nWe then apply trajectory-based channel reordering to reduce within-group magnitude disparities in the W4A4 main branch.\nBuilding on this mixed-precision representation, we further introduce cross-stage joint calibration to reduce weight quantization error and kernel fusion to enable efficient inference.\nExtensive experiments on PixArt-$\\Sigma$, Z-Image, and SANA-Video demonstrate that our method effectively preserves image and video generation quality under low-bit quantization.\nOn an NVIDIA RTX 4090, our method achieves up to $2.24\\times$ single-step speedup and reduces peak GPU memory by up to $61.8\\%$ on Z-Image compared with BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Transformers", "Post-Training Quantization", "Mixed-Precision Quantization", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=LLnTvYzOq0", "pdf_url": "https://openreview.net/pdf/19a4c1b97b1b92d9afd663ea073e3817d3753a90.pdf", "updated_at": "2026-10-05T19:55:09.271Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "low-bit", "mixed-precision", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FGVr95nHCv", "number": 23288, "title": "Projected-Distortion Guided Quantization", "abstract": "Post-training quantization (PTQ) reduces the memory and latency costs of large\nlanguage models (LLMs), but existing methods degrade sharply at 2--3 bits, in\npart because codebook fitting, outlier handling, and pruning are designed and\ntuned separately. We present Projected-Distortion Guided Quantization\n(PDGQ), which drives all three with a single activation-aware measure of output\ndistortion. PDGQ combines (1) a maximum-entropy codebook refined by\nactivation-weighted Lloyd iterations, (2) Projected-Distortion Outlier\nRestoration (PDOR), which restores weights whose distortion is large relative\nto a rate--distortion reference, on a scale shared across layers and\nbit-widths, and (3) Distortion-Consistent Pruning (DCP), which zeroes weights\nthat zero represents better than their codeword. On LLaMA and Qwen3 models,\nPDGQ matches recent PTQ methods in perplexity and zero-shot accuracy at 4 bits\nand outperforms them by a margin that grows as the bit budget shrinks, being\nlargest at nominal 2--3 bits, with a few narrow exceptions that we report.\nWith simple activation and KV-cache quantization, PDGQ also preserves\nDeepSeek-R1-Distill accuracy on GSM8K and MATH-500 under W4A4KV4 better than\nFlatQuant and QuaRot.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Low-Bit Quantization", "Model Compression", "Rate-Distortion Theory" ], "forum_url": "https://openreview.net/forum?id=FGVr95nHCv", "pdf_url": "https://openreview.net/pdf/59553368b37e79bbe9da3af2009b51bee57047e5.pdf", "updated_at": "2026-10-05T19:55:08.727Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HWi9uA3nEe", "number": 23305, "title": "Multi-Bitwidth Quantization for LLMs Using Additive Codebooks", "abstract": "As large language models (LLMs) are increasingly deployed across heterogeneous hardware with varying resource constraints, the ability to adaptively manage the performance-efficiency trade-off without retraining is critical. We propose Drop-by-Drop, a novel multi-bitwidth post-training quantization framework enabling inference-time precision control over LLM weights from a single quantized model. As theoretical motivation, we prove that Gaussian sources are successively refinable under a weighted mean squared error distortion motivated by LLM loss functions: successive refinement across distortion levels incurs no rate penalty relative to encoding at each level separately. Drop-by-Drop approximates this hierarchical structure in practice by incorporating Matryoshka-style supervision into additive codebook training, inducing an ordering in which codebook prefixes yield accurate partial reconstructions at each precision level. Furthermore, a block-Hadamard rotation brings the weight distribution closer to our Gaussian source assumption. The result is a single model that serves multiple bitwidths by dropping codebooks, reducing storage and quantization cost relative to the static per-bitwidth models. Across Qwen, LLaMA, Gemma, and Mistral, Drop-by-Drop achieves lower perplexity than state-of-the-art multi-bitwidth methods with competitive zero-shot accuracy, while our specialized kernels further reduce decoding latency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "multi bit width quantization", "adaptive inference", "successive refinement", "large language model compression", "post-training quantization" ], "forum_url": "https://openreview.net/forum?id=HWi9uA3nEe", "pdf_url": "https://openreview.net/pdf/7756612b491cc5019aab0d5000d7d1d762c71fed.pdf", "updated_at": "2026-10-05T19:55:11.295Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "bit width", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VqDpqib2hh", "number": 23318, "title": "MeshOp: Operator Learning for Adaptive Mesh Refinement in Parametric PDEs", "abstract": "Adaptive finite-element methods (AFEM) construct problem-dependent meshes through repeated solve--estimate--mark--refine cycles, but this sequential mesh-discovery process must be repeated across PDE instances with different input functions or physical parameters. We propose MeshOp, an operator-learning framework for adaptive mesh refinement across parametric PDE families. MeshOp predicts AFEM-derived refinement fields directly from PDE inputs, thereby reducing the cost of sequential mesh discovery. The predicted field is converted into a conforming hierarchical mesh by a dedicated score-to-mesh realization procedure, while residual-based AFEM can continue from this learned initialization when higher accuracy is required. Across the three benchmarks, the mean exact-level disagreement between quantized predictions and AFEM targets ranges from $11.4\\%$ to $27.4\\%$, yet the realized meshes attain finite-element errors comparable to those of fixed-cycle AFEM. This empirical downstream tolerance suggests that effective discretization does not require exact pointwise recovery of the AFEM refinement field. MeshOp reduces online time by $2.34\\times$ and $2.10\\times$ relative to reference-error-matched AFEM on Burgers and disk convection--diffusion, respectively, and by $3.06\\times$ relative to comparable-accuracy fixed-cycle AFEM on three-dimensional reaction--diffusion.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Operator Learning", "Adaptive Mesh Refinement", "Adaptive Finite Element Methods", "Parametric PDEs" ], "forum_url": "https://openreview.net/forum?id=VqDpqib2hh", "pdf_url": "https://openreview.net/pdf/2e4b8663fa46ec37731c3ccc4b02119c4f2a2655.pdf", "updated_at": "2026-10-05T19:55:13.991Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9OEOQLKqJj", "number": 23321, "title": "P4Q: Co-designing Token Pruning and Quantization for Vision-Language Model Acceleration", "abstract": "Vision language models have achieved strong performance across a wide range of multimodal applications, yet their substantial computational and memory costs hinder efficient deployment. Visual token pruning and post-training quantization reduce inference overhead along two complementary dimensions, namely sequence length and numerical precision. Existing workflows typically optimize these techniques independently or apply them sequentially. Their distinct optimization objectives leave critical interactions unaddressed and constrain the achievable compression performance. \nWe revisit these designs and present P4Q, a practical co-design framework that jointly optimizes visual token pruning and low-bit quantization for efficient VLM inference. First, P4Q introduces a quantization-aware visual token selection strategy before the LLM. It applies fake quantization to copies of the features produced by the projector and selects visual tokens using statistics computed from these fake-quantized features, thereby conditioning the selector’s feature-based decisions on simulated low-bit perturbations. Second, P4Q introduces a pruning-aware quantization calibration strategy. It uses the same selection strategy as pruning to calibrate the quantized model on the retained-token distribution, thereby aligning the calibration process with the pruned execution path used during deployment. By coupling these two components, P4Q achieves substantial inference speedups while maintaining comparable task performance, resulting in a better efficiency-accuracy trade-off than independently optimized pipelines. For instance, on LLaVA-NeXT, P4Q achieves an average end-to-end inference speedup of 2.8$\\times$ across eight distinct test sets, while retaining higher accuracy than prior compression and quantization methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Visual Token Pruning", "Post-Training Quantization", "Pruning-Quantization Co-design", "Efficient Multimodal Inference" ], "forum_url": "https://openreview.net/forum?id=9OEOQLKqJj", "pdf_url": "https://openreview.net/pdf/6e671d240bbb58cfd8ed5667c79d924ee0ffdd3c.pdf", "updated_at": "2026-10-05T19:55:14.057Z", "categories": [ "ptq", "qat", "hardware", "multimodal" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Z8rKEqGUNQ", "number": 23394, "title": "Stabilizing Logical Composition in Differentiable Logic Gate Networks", "abstract": "Differentiable logic gate networks (DLGNs) can be discretized into efficient hard logic circuits after training, but their intermediate relaxed Boolean features can be sensitive to input perturbations. This property is particularly relevant in DLGNs, where intermediate features serve as operands for subsequent logic gates and therefore directly participate in downstream logical composition. In this paper, we study convolutional DLGNs from the perspective of Boolean feature stability. We show that the bilinear interactions in relaxed logic gates induce feature-dependent local Jacobians, such that the local sensitivity of a gate varies with the values of its Boolean operands. Motivated by this analysis, we propose feature-level Jacobian regularization to stabilize selected intermediate relaxed Boolean features. We further introduce Binarization-Aware Jacobian regularization (BAJ), which assigns larger penalties to features near the binarization boundary, where perturbations are more likely to alter their Boolean interpretation. Experiments across image classification benchmarks show that the proposed regularizers improve classification performance and reduce the finite-perturbation sensitivity of intermediate Boolean features. Downstream analyses further show that feature stabilization affects subsequent logical computation. These results suggest that stabilizing intermediate Boolean operands is an effective approach to improving the stability of logical composition in convolutional DLGNs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Differentiable Logic Gate Network", "Robust Logical Composition", "Jacobian Regularization", "Representation Stability", "Boolean Feature Stability", "Differentiable Logic" ], "forum_url": "https://openreview.net/forum?id=Z8rKEqGUNQ", "pdf_url": "https://openreview.net/pdf/f28d8ce7527a8e84e26c3b334f5bbfb8747cc090.pdf", "updated_at": "2026-10-05T19:55:26.140Z", "categories": [], "matched_terms": [ "binarization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ka6SGUebh5", "number": 23406, "title": "ScaleSweep: Accurate NVFP4 Post-Training Quantization of LLMs via Block Scale Initialization", "abstract": "NVFP4 is a recently introduced hardware-supported FP4 format that improves the fidelity of 4-bit quantization through fine-grained block scales. However, existing NVFP4 scale initialization methods still rely on AbsMax initialization, which leaves a noticeable gap to the optimal solution.\nTo address this, we propose ScaleSweep, a simple and efficient scale optimization method that sweeps over block scale candidates and selects the candidate that minimizes a target objective. We further provide a theoretical analysis of NVFP4 quantization and derive both lower and upper bounds for the required sweep range under mean square error (MSE) and weighted mean square error (WMSE) between the original tensor and the quantized reconstructed tensor. The proposed bounds substantially reduce the sweep space while preserving the optimal candidate and maintaining end-to-end inference throughput close to the baseline.\nExperiments on Llama and Qwen models demonstrate that ScaleSweep consistently improves quantization performance over existing initialization methods and further narrows the gap to full precision. In particular, under aggressive end-to-end quantization of weights, activations, KV cache, and query states, ScaleSweep preserves more than 93\\% of the full-precision performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "nvfp4 quantization" ], "forum_url": "https://openreview.net/forum?id=Ka6SGUebh5", "pdf_url": "https://openreview.net/pdf/3238ea965c025cb18508e33b58343f8077b10994.pdf", "updated_at": "2026-10-05T19:55:27.480Z", "categories": [ "ptq", "training", "kv-cache", "hardware", "theory" ], "matched_terms": [ "4-bit", "fp4", "nvfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7lofmCNeVd", "number": 23424, "title": "Dense Checkpoints, Sparse Execution: Retraining-Free Acceleration of Query-Based Segmentation", "abstract": "Query-based segmentation repeatedly materializes dense query-pixel tensors during decoding and output processing, generating substantial memory traffic at high image resolutions. We use a frozen checkpoint's predicted masks and logits as an input-dependent execution plan to reduce this traffic. Mask-guided sparse cross-attention with a fixed budget of 128 feature cells per query preserves dense-level panoptic, instance, and semantic accuracy across compatible checkpoints on Cityscapes, COCO, and ADE20K, without retraining or checkpoint-specific tuning. Coarse intermediate mask prediction at the next attention stride avoids temporary stride-4 masks. Task-sufficient output processing evaluates masks only in regions that can affect the final output. For panoptic segmentation, we find that masks can influence pixel assignments even where they predict background, and derive conditions for safely skipping computation. Given identical FP32 logits from a released Mask2Former R50 checkpoint, output processing exactly reproduces the original panoptic maps, instance masks, and labels on all 500 Cityscapes validation images. On Mask2Former at native Cityscapes resolution, these transformations reduce FP32 DRAM traffic by 67% in the transformer decoder and over 80% in output processing. On an AGX Orin, output processing achieves a $1.8\\times$ panoptic speedup and a $4.2\\times$ instance speedup over optimized dense processing. With a PPM-FPN pixel decoder under mixed precision, complete-model inference achieves speedups of up to $1.31\\times$ for panoptic and $1.54\\times$ for instance segmentation over an optimized dense reference. Dense segmentation checkpoints can therefore be deployed more efficiently without retraining by targeting the inference paths that dominate memory traffic.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Query-based segmentation", "efficient inference", "sparse attention", "retraining-free acceleration", "high-resolution segmentation", "edge deployment" ], "forum_url": "https://openreview.net/forum?id=7lofmCNeVd", "pdf_url": "https://openreview.net/pdf/fdcdd5d9e0b616385a4dc79242de9dbe18b0c8af.pdf", "updated_at": "2026-10-05T19:55:30.943Z", "categories": [ "hardware" ], "matched_terms": [ "mixed precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VsQjZFCXQN", "number": 23430, "title": "TRPQuant: Token Rearrangement and Channel Pruning for Quantizing Large Vision-Language Models", "abstract": "High memory and computation costs prevent the practical edge deployment of vision-language models (VLMs). Post-training quantization (PTQ) offers a promising solution without expensive retraining. Given the visual and textual inputs in VLMs, existing methods typically adopt modality-level token partitioning and processing. However, they treat tokens uniformly within each modality, overlooking intra-modality activation distribution differences. This coarse-grained strategy leads to degraded accuracy and redundant computation under low-bit quantization. We propose TRPQuant, a new framework that addresses these limitations through Token Rearrangement and Channel Pruning, enabling fine-grained token grouping and computation reduction. Specifically, TRPQuant classifies tokens into Dense, Intermediate, and Sparse groups based on their activation patterns and applies tailored quantization strategies accordingly: dense tokens preserve full channels and go through the standard branch, sparse tokens are routed to the sparse branch with channel pruning to reduce computation, and intermediate tokens combine both mechanisms via dual-branch processing. Extensive experiments on mainstream VLMs demonstrate that TRPQuant achieves strong quantization accuracy. Notably, TRPQuant preserves 94.1% of FP16 performance under the challenging W3A3 setting (69.9 vs. 74.3) on Qwen2.5-VL-7B, pushing the frontier for ultra-low-bit quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Models", "Post-Training Quantization" ], "forum_url": "https://openreview.net/forum?id=VsQjZFCXQN", "pdf_url": "https://openreview.net/pdf/71f99283fb3e9358c2774e8a1d62ff1b712d1f16.pdf", "updated_at": "2026-10-05T19:55:31.217Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantizing", "w3a3" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aK7uF5MnlY", "number": 23431, "title": "Prediction Instability and Task Utility in Low-Precision Language Models", "abstract": "Selective precision and quantization repair are evaluated through changes in predictions, reconstruction error, and task accuracy. We examine how these measurements relate using a controlled audit of low-precision language models. A prespecified experiment evaluates two instruction-tuned models on 1,024 HellaSwag training examples across four precision paths and three execution layouts. Changing option batch size while preserving padding changes per-item precision-upgrade utility on 42 examples for Qwen and 14 for SmolLM2; corresponding full-model FP32 arithmetic controls show zero changes. Within fixed 256-example margin selections, opposing changes cancel, leaving aggregate shifts of -6 and +1 net-correct units. Subsequent SmolLM2 repair diagnostics compare four teacher-target objectives with the same rank-16 parameterization. Centered-score fitting reduces training error by 46.17%, while error on two evaluation panels increases by 15.24% and 8.16%. On the previously observed 1,024-question transfer panel, it yields three additional correct answers, compared with five for option-level distillation. An ARC replay further distinguishes offline fallback utility from executed predictions and cost. The measurements identify where prediction stability and reconstruction fidelity diverge from task utility, and provide a protocol for evaluating these quantities together.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "selective precision", "numerical reproducibility", "knowledge distillation", "language model evaluation" ], "forum_url": "https://openreview.net/forum?id=aK7uF5MnlY", "pdf_url": "https://openreview.net/pdf/3be58cee607735a04b9441150a8bdbef3a560518.pdf", "updated_at": "2026-10-05T19:55:31.477Z", "categories": [ "ptq" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dUDzh4Xgr0", "number": 23533, "title": "WMKD: Planning-Oriented World Model via Future-Guided Knowledge Distillation for End-to-End Autonomous Driving", "abstract": "End-to-end autonomous driving requires anticipating how the surrounding scene will evolve to make safe and effective planning decisions. World models provide this context by learning to predict future scene evolution, yet a fundamental question remains: \"What should a world model predict for planning?\"\nExisting approaches largely formulate world modeling as reconstructing\nfuture observations or scene representations, often without explicitly modeling how these targets should be constructed to support the current planning decision. To this end, we present World Model Knowledge Distillation (WMKD), a teacher--student framework in which a future-guided teacher constructs a planning-oriented prediction target and a student learns to predict it from sensor observations. The teacher encodes ground-truth future semantic evolution under joint supervision from future semantic reconstruction and ego planning, and quantizes the resulting scene representation into discrete codes. The student's Discrete World Model (DWM) predicts these codes to anticipate planning-relevant future knowledge without privileged inputs at inference. To incorporate the predicted representation into planning, the Residual Fusion Module (RFM) fuses it with the observed BEV and their residual, while Planning Response Distillation (PRD) transfers the teacher's planning responses. The Trajectory Repropagation Planner (TRP) further re-encodes refined trajectory geometry to guide successive refinement. Extensive experiments across NAVSIM and Bench2Drive validate WMKD, achieving 92.0 PDMS and 90.6 EPDMS on NAVSIM, as well as a Driving Score (DS) of 87.56 and a Success Rate (SR) of 72.27% on Bench2Drive.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "End-to-End autonomous driving", "world model", "knowledge distillation" ], "forum_url": "https://openreview.net/forum?id=dUDzh4Xgr0", "pdf_url": "https://openreview.net/pdf/02286614c5100030a15b72bc8c94900441c8810f.pdf", "updated_at": "2026-10-05T19:55:43.447Z", "categories": [ "ptq" ], "matched_terms": [ "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xtMGsGokzG", "number": 23574, "title": "Rotation Is Token-Blind: A Structural Criterion for Post-Training Quantization of Feed-Forward Geometry Transformers", "abstract": "Feed-forward 3D models use Transformers to predict camera poses, depth, and structure in one pass, avoiding iterative multi-view optimization. Long sequences and deep attention incur high compute and memory costs, motivating Post-Training Quantization (PTQ) for deployment. Existing PTQ methods mainly suppress channel-wise outliers by smoothing or orthogonal rotation. Geometry Transformers introduce a different challenge: heterogeneous token types share one activation quantization grid. A few special tokens may exert high leverage over global outputs and be updated by only a subset of blocks. Token-blind transforms, which apply the same transformation to every token, reduce overall quantization error but cannot remove differentiated degradation across token types. We propose QMapAnything, a structure-aware PTQ framework. We cast floating-point-preserving rotation and smoothing as a token-blind reparameterization group, and prove that they yield common-mode gains while leaving relative token SNR unchanged. We introduce a statistic $\\Gamma$ that combines a special token's quantization disadvantage and output leverage to predict whether differentiated bias correction is needed beyond rotation. Experiments on geometric benchmarks show that W4A8 closely matches full-precision performance. Under W8A8, peak memory falls by $2.2$-$3.1\\times$ versus FP32, with up to $3.1\\times$ speedup on real hardware. These results provide a principled and efficient approach to deploying feed-forward 3D reconstruction models in resource-constrained edges.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "model compression", "3D reconstruction", "feed-forward transformers", "geometry transformers", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=xtMGsGokzG", "pdf_url": "https://openreview.net/pdf/4e219ee90ba9eee89b8a0cea48d2f0bfb2facf00.pdf", "updated_at": "2026-10-05T19:55:49.820Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "ptq", "quantization", "w4a8", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Z7J62RTkst", "number": 23581, "title": "MUL-GPT: Timing-Structure-Aware Generative Optimization for Multiplier Circuits", "abstract": "Multiplier optimization is a challenging combinatorial optimization problem, as its design space grows rapidly with increasing operand bit width. In recent years, learning-driven methods have provided a promising approach to discovering high-speed and area-efficient circuit structures. However, existing methods typically optimize only a small set of global post-synthesis metrics, such as area and delay, without explicitly modeling the internal timing structure of multipliers or the mechanisms underlying critical path formation. Consequently, they struggle to accurately identify timing bottlenecks and effectively exploit internal timing information to guide structural optimization. To address these limitations, we propose MUL-GPT, an internal-timing-structure-aware generative framework for multiplier optimization. MUL-GPT is a generative pretrained Transformer (GPT)-style generative framework that addresses the internally coupled timing optimization problem of multipliers. It unifies constrained autoregressive compressor-tree generation, arrival-time-conditioned adder optimization, and static timing analysis (STA)-guided local interconnection rewiring, with internal timing information serving as the link across these stages. Experiments show that the best designs found by MUL-GPT consistently outperform state-of-the-art methods across various arithmetic-unit design tasks, reducing the area-delay product (ADP) by up to 10.57%. These improvements are preserved when the optimized arithmetic units are integrated into large-scale artificial-intelligence (AI)-accelerator circuits, where MUL-GPT achieves an ADP reduction of up to 7.90%, further demonstrating its scalability and practical applicability.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Computing Circuits Optimization", "Electronic Design Automation", "Datapath Optimization", "Reinforcement Learning" ], "forum_url": "https://openreview.net/forum?id=Z7J62RTkst", "pdf_url": "https://openreview.net/pdf/fa8d1106e19f7c98ef5894a6c40e7987d47698e7.pdf", "updated_at": "2026-10-05T19:55:50.946Z", "categories": [ "hardware" ], "matched_terms": [ "bit width" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PAquDumY4i", "number": 23611, "title": "MOSES: Geometric Preconditioning and Hybrid Quantization for Memory-Efficient LLM Optimization", "abstract": "Optimizer states are the largest static memory cost in large model training. We introduce MOSES, a memory-efficient optimizer that addresses this challenge through two core innovations. First, Geometry-Aware Adaptive Preconditioning integrates gradient orthogonalization with factorized second-moment estimation, leveraging the geometry of the gradient space to provide both spatial smoothing and temporal adaptivity with negligible memory overhead. Second, Distribution-Aware Hybrid 1/4/8-bit Quantization adaptively allocates precision to post-preconditioning momentum based on the spectral properties of EMA updates, enabling aggressive compression while preserving optimization quality. Together, these innovations reduce optimizer state memory by 12.3× compared to Adam. Across model scales from 125M to 6.7B parameters, Moses achieves a compelling memory–performance trade-off that existing memory-efficient methods struggle to match, particularly excelling in the small-learning-rate regime that is characteristic of large-scale production training. Comprehensive evaluations across pretraining and downstream tasks show that Moses achieves competitive optimization quality while significantly improving memory efficiency, thereby reducing the hardware requirements for large language model training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Memory Efficient Optimizer; Large Language Model Training" ], "forum_url": "https://openreview.net/forum?id=PAquDumY4i", "pdf_url": "https://openreview.net/pdf/6d555be32d60707b383c0c241ec618f87e99c8b2.pdf", "updated_at": "2026-10-05T19:55:55.644Z", "categories": [ "training", "hardware" ], "matched_terms": [ "8-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cYHReGptgT", "number": 23615, "title": "Consistency over Accuracy: Outlier-Bypass Layer-Adaptive Quantization for Codec-Based KV Cache Compression", "abstract": "Reusing key–value (KV) caches can eliminate repeated prefill computation in long-context large language model serving, but practical reuse requires the cache to be compressed without significantly altering the model response. Practical lossy KV cache compression raises two challenges: how to measure reuse fidelity and how to mitigate quantization distortion. For the former, aggregate task accuracy is insufficient to reliably characterize answer consistency before and after compression, as compression-induced perturbations may spuriously turn an incorrect answer into a correct one. To address this issue, we propose replacing accuracy with answer consistency as the metric for evaluating compression-induced distortion. For the latter, existing codec-based systems compress the dense representation effectively, yet sparse outliers are not adequately handled. To be specific, a few extreme values can consequently dominate the shared group scale, reducing the effective resolution of normal values and compromising subsequent related allocations. We refer to this effect as outlier-scale pollution. Accordingly, we propose outlier-bypass layer-adaptive quantization. The method detects group-wise extremes by gap prominence, stores them exactly in a separate sparse bypass. Then, the remaining values are quantized using layer-specific level where the corresponding thresholds are customized for each model through offline cumulative-distribution fitting (CDF) calibration. By the calibrated CDF, we derive two presets corresponding to distinct consistency–compression operating points. Across five models and three long-context benchmarks, the high-fidelity preset (Ours-HF) attains 82.33% consistency at 7.41× compression, simultaneously outperforming KVCodec and CacheGen by +1.96 percentage points in consistency with +5.26% and +26.24% in compression ratio, respectively. For workloads that prioritize storage efficiency, the high-compression preset (Ours-HC) further increases the compression ratio to 8.35×, a gain of 18.61% over KVCodec and 42.25% over CacheGen, at a modest consistency cost of 1.45 percentage points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "long-context LLM serving", "quantization", "outlier bypass", "layer-adaptive quantization", "video coding" ], "forum_url": "https://openreview.net/forum?id=cYHReGptgT", "pdf_url": "https://openreview.net/pdf/c6da2cfb6dfb5a22d4cb5eea3ad90b917d609506.pdf", "updated_at": "2026-10-05T19:55:56.075Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZYfgzNZN0r", "number": 23625, "title": "NADA: Unlocking Fast and Accurate NVFP4 PTQ through Distortion-Aware Optimization", "abstract": "NVFP4 provides a wide effective dynamic range while retaining hardware-native FP4 computation on NVIDIA Blackwell GPUs, enabling high quantization fidelity with efficient low-precision inference. However, accurate NVFP4 post-training quantization (PTQ) remains challenging, particularly for activations. Simply suppressing outliers does not guarantee lower activation quantization error, largely because the nonuniform E2M1 codebook induces uneven distortion. We identify that NVFP4 activation reconstruction is governed by three interdependent and sequentially coupled decisions at the tensor, channel, and block levels. This paper proposes an NVFP4 Activation Distortion-Aware (NADA) optimization framework for fast and accurate NVFP4 PTQ. NADA consists of tensor-level Distortion-Aware Activation Channel Clustering (DACC), channel-level Bounded Iterative Top-$k$ Rescaling (BITR), and block-level Analytical Block-Scale Optimization (ABSO). DACC and BITR are calibrated offline to reorganize activation channels and selectively rescale sensitive channels, respectively, while ABSO performs a search-free analytical scale update for dynamic activation blocks at runtime. Across eight models ranging from 135M to 70B, NADA achieves state-of-the-art accuracy while preserving high end-to-end inference efficiency. Code is available at https://anonymous.4open.science/r/NADA-5B92.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "NVFP4", "Weight-Activation Quantization", "Microscaling Data Formats", "Distortion-Aware Quantization", "Efficient LLM Inference" ], "forum_url": "https://openreview.net/forum?id=ZYfgzNZN0r", "pdf_url": "https://openreview.net/pdf/d1531c848c788cd9e1ba96bbca6348a495c1ab53.pdf", "updated_at": "2026-10-05T19:55:57.815Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "low-precision", "nvfp4", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "J09wbKGDLf", "number": 23665, "title": "Quantization Can Preserve Verification While Degrading Self-Generated Supervision", "abstract": "Can a quantized model retain its ability to select answers yet lose the learning value of its own solutions? We separate verification ranking, final-answer correctness and held-out learning gain. On Gemma 4 12B and Qwen3.6-35B-A3B, all nine registered 4- and 3-bit verifier contrasts are equivalent to full precision within ±0.025 AUROC, even as generation degrades. Crossing sample source with learner precision reveals that SmolLM3-3B's full-precision solutions at 38% correctness teach both learners more than NF4 solutions at 74% (by 1.9 and 3.6 accuracy points) on a holdout selected for partial solvability. On the 12B, approximately matching sample count and problem coverage leaves an 18.9-point source advantage for the quantized learner on this holdout and 4.6 points on MATH-500. Completed, mostly correct NF4 solutions also reduce the full-precision learner's accuracy. The penalty increases along the GGUF precision ladder on both learning subjects, with format-dependent severity, and no tested selector shows an established recovery. Reciprocal continuations separately localize the 12B's excess truncation to the continuing model for the tested prefixes. Under this math supervised fine-tuning recipe, verification robustness and final-answer correctness are insufficient summaries of useful self-generated supervision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "self-generated supervision", "self-verification", "synthetic data", "supervised fine-tuning" ], "forum_url": "https://openreview.net/forum?id=J09wbKGDLf", "pdf_url": "https://openreview.net/pdf/37bbf6ba03e32bde689930afe2bc1b33a0c1e360.pdf", "updated_at": "2026-10-05T19:56:06.661Z", "categories": [ "theory" ], "matched_terms": [ "3-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eTic76nLQW", "number": 23711, "title": "BitMoE: Post-Training Binarization for MoE-based Large Language Models", "abstract": "Large language models (LLMs) with Mixture-of-Experts (MoE) architectures have achieved remarkable progress, yet their substantial weight storage and memory traffic hinder practical deployment.\nWeight binarization offers extreme compression, but transferring dense-model methods to MoE faces three coupled challenges: decomposition--binarization mismatch, interaction-dependent residual allocation, and quantization-induced expert-shift.\nTo address these challenges, we propose \\textbf{BitMoE}, a post-training binarization framework coordinating shared representation, residual capacity, and routing stability.\nFirst, Routing-Coupled Binary Factorization alternately refines higher-precision shared factors and expert-specific binary factors to reduce the reconstruction error of routed expert aggregation.\nSecond, Conditional Expert Residual Allocation selects additional binary residual blocks by their conditional reconstruction gain per added storage bit after readapting the shared output factor, updating gains as corrections accumulate.\nThird, Margin-Aware Optimal Brain Routing uses boundary-aware curvature to preserve selected--unselected expert margins and mitigate expert-shift.\nBitMoE retains all experts and performs optimization offline under an explicit storage budget.\nExtensive experiments demonstrate that BitMoE consistently outperforms state-of-the-art binary baselines across multiple MoE models and benchmarks.\nFor instance, on Qwen3-30B-A3B, BitMoE delivers 54.2\\% lower perplexity and improves average zero-shot accuracy by 19.71 percentage points over the strongest binary baseline, alongside an over 2$\\times$ inference speedup over the BF16 model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Binarization", "PTQ", "MoE" ], "forum_url": "https://openreview.net/forum?id=eTic76nLQW", "pdf_url": "https://openreview.net/pdf/aefed59e7b64e1e8531760ee388b74f04361d54a.pdf", "updated_at": "2026-10-05T19:56:14.217Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "binarization", "binary", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pX7KzJPRpY", "number": 23735, "title": "QCRL: Quantization-Consistent Reinforcement Learning for LLM Reasoning", "abstract": "Obtaining a capable W4A4 reasoning model through reinforcement learning with verifiable rewards remains challenging because direct W4A4 optimization suffers from rounding-induced reward instability and rollout-optimization policy mismatch. We propose Quantization-Consistent Reinforcement Learning (QCRL), a dual-view framework that combines two strategies, Quantization-Consistent Rollout and Quantization-Weight Dampening, to improve the final W4A4 policy. Specifically, the rollout strategy collects W4A16 completions from round-to-nearest and stochastic-rounding views of shared latent weights and calibrates their rewards through cross-view comparisons and within-view support to guide W4A4 policy optimization. The weight-dampening strategy stabilizes weights near their quantization centers after parameter updates. On GSM8K with Qwen2.5-3B, QCRL achieves 78.32\\% pass@1 under W4A4 evaluation, surpassing BF16 reinforcement learning followed by W4A4 post-training quantization by 2.28 percentage points and direct W4A4 reinforcement learning by 4.25 percentage points. These results support reinforcement learning for stronger W4A4 reasoning models while retaining compatibility with W4A16 rollout.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Reinforcement Learning" ], "forum_url": "https://openreview.net/forum?id=pX7KzJPRpY", "pdf_url": "https://openreview.net/pdf/21356aa6f23c5f17c47084b6a9c3a8f08c581e58.pdf", "updated_at": "2026-10-05T19:56:18.987Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "quantization", "w4a16", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FvwT42QPDE", "number": 23859, "title": "The Misallocation of Precision in Quantized Language Models", "abstract": "Mixed-precision quantization must decide where to spend a fixed budget of full-precision weights, and the criteria in use are derived from activation statistics: GPTQ minimises layer-wise error under the Hessian $H{=}2XX^\\top$, ResQ keeps that covariance's highest-variance subspace at 8 bits. We test the premise underneath, that numerical sensitivity tracks functional importance, by taking its natural per-layer summary, $\\mathrm{Tr}(H_\\ell)/d$, as a cross-layer allocation rule. In Llama-3.1-8B it is not merely uncorrelated with functional importance but *inverted*: Spearman $\\rho{=}{-}0.497$ ($p{<}0.001$, $n{=}64$). The sub-layers it ranks highest sit in a functional dead zone; those it ranks lowest are load-bearing. The cost is measurable: at an identical 6/32 FP16 budget and identical attention/MLP composition, allocating by Hessian rank recovers $7\\times$ less of the damage 4-bit quantization does to ARC-Challenge ($7.7\\%$ vs. $53.8\\%$, $p{=}0.002$), and leaves the model significantly worse than FP16 where the functionally-chosen budget does not. We establish this by treating 4-bit NF4 quantization as a controlled perturbation probe, analogous to lesion studies in neuroscience: quantize globally, restore full precision to one sub-layer, measure task recovery. A 64-sub-layer sweep localises the effect to *named sub-layers* rather than to components as classes. Against a do-nothing null measured on the same examples (ARC exactly 0/468) with BH-FDR correction, L13–14 attention clear the null on GSM8K and MATH but on neither measure of ARC, while L6–7 MLP clear it on all three; the component-level contrast is itself not significant ($p{=}0.107$). Split-half selection shows the ranking beats a random equal-budget subset on ARC ($+10.4$pp, 95% CI [$+3.7$, $+18.5$]) while carrying 4.6–12.2pp of selection optimism. The inversion replicates on Mistral-7B and under GPTQ, and the Llama-3.1-8B-derived layer set transferred to Mistral-7B unchanged recovers $47.1\\%$ of its ARC damage against $2.0\\%$ for Mistral-7B's own Hessian-top layers, though offset by collateral damage elsewhere, so it transfers as a diagnostic rather than an allocator. A surgical model keeping 6 of 32 sub-layers at full precision reaches FP16 equivalence at 42% memory, yet neither Hessian sensitivity nor weight-space SQNR ($\\rho{=}{+}0.032$, ns) predicts which six.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "mixed-precision quantization", "post-training quantization", "mechanistic interpretability", "hessian sensitivity", "layer-wise sensitivity", "large language models", "model compression" ], "forum_url": "https://openreview.net/forum?id=FvwT42QPDE", "pdf_url": "https://openreview.net/pdf/a215bb67ee1266432b02c0d6491e9eb30d55d65c.pdf", "updated_at": "2026-10-05T19:56:44.025Z", "categories": [ "ptq" ], "matched_terms": [ "4-bit", "mixed-precision", "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "X09ivFsbtU", "number": 24058, "title": "Discrete Weight Updates for Extreme Low-Bit Training without Master Weights", "abstract": "Memory consumption has become a major bottleneck for training and deploying increasingly large models. While extreme low-bit models with binary or ternary weights remarkably reduce the memory required for inference, their training still commonly relies on full-precision master weights that are optimized in a contin- uous space and quantized for the forward pass. Removing such master weights is challenging, as discrete weights can only change through transitions between a finite set of states and thus cannot directly represent the small continuous updates made by conventional optimizers. In this paper, we propose Discrete Updates for Extreme Low-bit Training (DUET), a scalable framework that directly trains ex- treme low-bit weights without maintaining master weights. We first derive a tran- sition budget from the magnitude of a continuous optimization step, specifying the expected number of discrete weight transitions. Given this transition budget, we select the weight transitions that maximize the alignment of the resulting discrete update with the continuous step. We further show that optimizer states used for these transition decisions can be significantly compressed while retaining training quality. Our experiments demonstrate that DUET maintains competitive training quality over long training horizons and on large models, with training memory reductions that increase with model size and reach 12.0× at 235B parameters compared to conventional training with master weights", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Extreme Low-Bit Training", "Memory-Efficient Training", "Master Weight Free Training" ], "forum_url": "https://openreview.net/forum?id=X09ivFsbtU", "pdf_url": "https://openreview.net/pdf/10dad097e2364169bb36df0144234db5eddbc999.pdf", "updated_at": "2026-10-05T19:57:10.727Z", "categories": [ "training" ], "matched_terms": [ "binary", "low-bit", "quantized", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "prw4GsTDr0", "number": 24095, "title": "LLC: A Linear Low-Rank Codec for Joint Key-Value Compression Before Token Eviction", "abstract": "Token eviction controls KV cache growth but permanently discards states that may become relevant later. Retaining these states in compressed form requires a compact representation that remains useful to attention. We introduce LLC, a closed-form, training-free codec that jointly encodes post-RoPE keys and values across all heads into a shared low-rank code. Fixed linear maps enable direct attention over the codes without reconstructing full-width KV states. We deploy the codec in a two-tier cache: a base eviction policy manages exact residence, while a separately budgeted compressed tier retains demoted tokens with the lowest reconstruction errors. The joint representation improves reconstruction fidelity and yields stronger accuracy–memory trade-offs across models and eviction policies. Across RULER at 8k, 16k, and 32k, LLC approaches dense accuracy with less KV cache memory; at 32k, it stays within 1.5 points of dense across four models using 31–38% of the dense cache. On LongBench, LLC outperforms a prior codec in all eight matched-memory comparisons. Under CUDA Graph replay, an A100 implementation achieves $1.31$–$4.10\\times$ speedups over matched dense SDPA in single-token, per-layer attention microbenchmarks from 8k to 64k. INT8 quantization of the stored codes further halves their storage with negligible accuracy loss. Together, LLC mitigates irreversible eviction without sacrificing memory efficiency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "efficient LLM inference", "low-rank approximation", "long-context language models", "token eviction" ], "forum_url": "https://openreview.net/forum?id=prw4GsTDr0", "pdf_url": "https://openreview.net/pdf/c8cda41a189f9e38011ff47afae1ce350c7712b7.pdf", "updated_at": "2026-10-05T19:57:16.052Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CnG68VRdPZ", "number": 24197, "title": "SAFE-Align: Alignment-Conditioned Evidence Routing for Multimodal Manipulation Detection and Grounding", "abstract": "Multimodal manipulation detection requires reasoning over image--text pairs whose semantic correspondence can vary substantially across samples. Existing methods typically use multimodal fusion mechanisms whose aggregation is not explicitly conditioned on the sample-level image--text alignment regime. We challenge this assumption and propose SAFE-Align, an alignment-conditioned framework that treats image--text correspondence as a sample-level routing prior rather than an authenticity score. SAFE-Align combines a discrete alignment state, obtained by quantizing global image--text similarity and represented by a learned state embedding, with the original continuous similarity to characterize the correspondence regime, and uses this alignment prior to route complementary global-interaction, local-spatial, and text-only evidence. To preserve fine-grained manipulation cues, it further models bidirectional patch--token correspondence and captures local agreement and discrepancy cues for image and text grounding. On the official DGM4 split, SAFE-Align achieves 95.63% AUC, 88.74% accuracy, and 89.63% mAP, improving over ASAP by 1.25, 1.03, and 1.10 percentage points, respectively. Alignment-conditioned routing also improves AUC by 1.55 points over uniform fusion and by 1.97 points over a learned gate without the alignment prior. These results support explicitly conditioning multimodal evidence routing on cross-modal alignment rather than relying on a fixed fusion pattern.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multimodal Manipulation Detection", "Multimodal Grounding", "Cross-Modal Alignment", "Evidence Routing", "Vision-Language Models" ], "forum_url": "https://openreview.net/forum?id=CnG68VRdPZ", "pdf_url": "https://openreview.net/pdf/57d39db5a5cad8dafb122523171cc7768ff6ed72.pdf", "updated_at": "2026-10-05T19:57:33.836Z", "categories": [ "ptq", "reasoning", "multimodal" ], "matched_terms": [ "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xYATh2U73h", "number": 24335, "title": "ReSC-Q: Residual-Aware Reliable PTQ with Shrinkage-Controlled Compensation", "abstract": "Post-training quantization (PTQ) reduces the memory footprint and inference cost of LLMs under tight deployment constraints. Hessian-based PTQ methods such as GPTQ use estimated curvature information for error compensation, but compensation estimated from calibration data does not always lead to optimal reconstruction loss on unseen inputs.\nThis becomes especially problematic in ultra low-bit quantization.\nWe propose ReSC-Q, which controls uncertain cross-group curvature through shrinkage while explicitly accounting for the reconstruction loss remaining after compensation. \nReSC-Q supports practical group-wise quantization by separating the blockwise reconstruction objective into cross-group compensation and within-group quantization, applying shrinkage and residual-aware weighting to each, respectively.\nAcross seven LLMs evaluated at 2-bit, ReSC-Q improves average zero-shot accuracy by 1.09 percentage points over the strongest baseline for each model, reducing the mean accuracy gap to full precision by 23.2\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Post Training Quantization", "Weight-only Quantization" ], "forum_url": "https://openreview.net/forum?id=xYATh2U73h", "pdf_url": "https://openreview.net/pdf/e23cd109f3343ffda754def2f29176b55e1ce8b1.pdf", "updated_at": "2026-10-05T19:57:52.784Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jYLRArW7rw", "number": 24368, "title": "Not All Reconstruction Errors Matter: Score-Guided Block-wise Reconstruction for Low-Bit LLM Quantization", "abstract": "Block-wise reconstruction has been widely used for post-training quantization (PTQ) of large language models (LLMs), but conventional objectives penalize all block-output perturbation directions equally despite their different effects on the final predictive distribution. In this paper, we propose SABRE, a score-guided anisotropic reconstruction approach that accounts for these unequal predictive effects during block-wise reconstruction. To identify prediction-sensitive directions, SABRE approximates the Fisher curvature of the full-precision (FP) predictive distribution using rank-one outer products of sampled score vectors. Based on this approximation, SABRE assigns additional penalties to errors aligned with these directions while retaining isotropic reconstruction to control the overall perturbation magnitude. To ensure stable weighting across blocks and tokens, we further introduce gradient-norm normalization that bounds token-dependent directional weights, thereby reducing sensitivity to variations in score-vector scale. The FP-derived score vectors are computed once and cached, enabling efficient block-local reconstruction without repeated end-to-end evaluation. Because SABRE modifies only the reconstruction objective, it can be readily integrated into existing block-wise quantization methods without changing their optimization variables or inference-time computation. Extensive experiments across diverse quantization methods, LLM architectures, and low-bit settings demonstrate consistent perplexity improvements, substantial recovery under aggressive 2-bit quantization, and broad downstream gains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-training Quantization", "Block-wise Reconstruction", "Predictive Sensitivity", "Fisher Information" ], "forum_url": "https://openreview.net/forum?id=jYLRArW7rw", "pdf_url": "https://openreview.net/pdf/9e18fe2457618f5d5b5b244d5f20c305ab6102dc.pdf", "updated_at": "2026-10-05T19:57:56.819Z", "categories": [ "ptq", "training" ], "matched_terms": [ "2-bit", "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3VSa1Kx7HD", "number": 24488, "title": "What Must Quantization Preserve in Latent World Models?", "abstract": "What must quantization preserve in latent world models to retain their planning\nability? Existing work has primarily characterized sensitivity to precision\nconfiguration. To our knowledge, this is the first study to investigate latent\nworld-model quantization from the perspective of representation preservation\nand calibration-target design. We analyze how quantization-induced\nrepresentation errors propagate to goal-distance estimates and candidate\nrankings, and argue that calibration should preserve reusable latent\nrepresentations rather than merely match scalar planning costs. We derive a\nbound relating terminal relative-representation error to the second moments of\nlocal encoder and predictor reconstruction errors, providing a theoretical\nbasis for Module-MSE. Across LeWM, DINO-WM, and PLDM on four control tasks under\nsymmetric per-channel W4 quantization, Module-MSE reaches 79.33\\% average\nsuccess compared with 79.83\\% for full precision, while achieving the strongest\noverall planning-decision fidelity. It also maintains strong performance across\nthe tested learning rates and calibration-set sizes. These results show that\nsevere degradation under naive W4 quantization is not inevitable, establish\nrepresentation preservation as a practical calibration principle, and support\nModule-MSE as a default choice for latent world-model quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "latent world models", "post-training quantization", "model-based planning", "model compression" ], "forum_url": "https://openreview.net/forum?id=3VSa1Kx7HD", "pdf_url": "https://openreview.net/pdf/f969d6d88ab299396d710184f87f682ad6d0c13a.pdf", "updated_at": "2026-10-05T19:58:16.766Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NQMitBFYEQ", "number": 24489, "title": "TrajeQ: Trajectory-Preserving Flow Matching Quantization for Low-Bit Video Generation", "abstract": "While flow matching (FM) has emerged as a prominent paradigm for high-quality video generation, its significant computational overhead presents a major obstacle to practical implementation. Although post-training quantization (PTQ) is an established approach for efficient generation, existing PTQ methods primarily minimize reconstruction error in a magnitude-centric manner. This objective is misaligned with deterministic FM trajectories, where low-bit quantization introduces velocity-field orientation errors that can compound across integration steps and cause deviations from the full-precision trajectory. To address this issue, we introduce ***TrajeQ*** to preserve the geometric integrity of the velocity field and reduce trajectory drift under low-bit quantization. Our approach introduces *Curvature-Adaptive Quantization (CAQ)* to detect angular shifts across timesteps and mitigate geometric distortion via adaptive precision. We also integrate *Directional Bottleneck Preservation (DBP)*, a sensitivity-guided precision assignment scheme that retains only the most vulnerable layers in full precision. Experiments demonstrate that ***TrajeQ*** achieves state-of-the-art VBench performance and maintains stable video generation under extreme W4A4 and W3A6 quantization regimes where existing methods degrade severely. We provide the code in the supplementary material and additional video results at https://anonymous-orange.github.io/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Generation", "Flow-Matching", "Quantization" ], "forum_url": "https://openreview.net/forum?id=NQMitBFYEQ", "pdf_url": "https://openreview.net/pdf/a9d432eaa0d2c1c6f9b17877fd18f8ef046fe256.pdf", "updated_at": "2026-10-05T19:58:20.179Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "low-bit", "ptq", "quantization", "w3a6", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bjR44Lxl7y", "number": 24503, "title": "HAMoE: Heterogeneity-Aware Mixed-Precision Quantization for Mixture-of-Experts Models", "abstract": "Mixture-of-Experts (MoE) models face deployment challenges in resource-constrained devices due to their massive parameter scales. \nQuantization is recognized as one of the most effective approaches to address such challenges. \nIn this work, we conduct a systematic quantization study of MoE models and identify two critical heterogeneity insights that have not been considered in prior work: 1) inter-expert heterogeneity, i.e., distinct activation frequencies and processed token types across experts, and 2) intra-expert heterogeneity, i.e., differential quantization sensitivity of weights within a single expert. \nTo this end, we introduce HAMoE, a mixed-precision quantization framework that considers both heterogeneities. \nHAMoE assigns fine-grained mixed precision to inter-experts based on activation frequency and token type.\nIn addition, it proposes an adaptive column replacement mechanism, which replaces the most quantization-sensitive intra-expert weight columns with INT8-quantized versions, thereby effectively reducing quantization loss.\nEvaluation results show that HAMoE achieves competitive results against strong baselines under equivalent bit budgets, reducing WikiText-2 perplexity by 4.0\\% relative to GPTQ at 3.35-bit, recovering about 45\\% of the quantization loss. HAMoE also consistently outperforms MxMoE in both average accuracy and perplexity across all four models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "Model Quantization", "Mixed-Precision Quantization", "Large Language Models", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=bjR44Lxl7y", "pdf_url": "https://openreview.net/pdf/fc50590d98120e8285310ccb257e1901b45b920b.pdf", "updated_at": "2026-10-05T19:58:21.699Z", "categories": [ "hardware" ], "matched_terms": [ "int8", "mixed precision", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "f7RycDoIZM", "number": 24516, "title": "SAR4CAD: Structure-Aware Reinforcement Learning for Parametric CAD Reconstruction from Multi-View Drawings", "abstract": "Computer-Aided Design (CAD) models can be reconstructed from engineering drawings; however, existing methods are limited by the scarcity of well-annotated drawing datasets and an excessive reliance on a single sequence representation. Specifically, most sequence generators optimize only one serialization scheme, although neither sketch ordering nor the decomposition of certain features is necessarily unique. SAR4CAD is a two-stage framework that trains a multimodal large language model (MLLM) using multi-view renderings, scalable vector graphics (SVG) projection tokens, and concise textual descriptions. The SVG tokens encode visible and hidden edges, primitive types, and quantized coordinates extracted from the CAD model, thereby providing reliable geometric information for reconstruction. In the first stage, supervised fine-tuning teaches the model the required output syntax and establishes a stable generation policy. In the second stage, structure-aware group relative policy optimization (GRPO) evaluates sampled outputs after they have been parsed and executed. The associated reference-guided structural reward integrates CAD-kernel validity, structural-count similarity, sketch overlap computed via Hungarian matching, ordered extrusion-parameter similarity, and 3D Chamfer similarity. Experiments on a constructed parametric CAD dataset demonstrate that SAR4CAD consistently outperforms the evaluated baselines across all evaluation metrics. Code is available at \\textcolor{blue}{\\url{https://anonymous.4open.science/r/SAR4CAD-F64F}}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "CAD Reconstruction", "Reinforcement Learning", "Multimodal Large Language Models" ], "forum_url": "https://openreview.net/forum?id=f7RycDoIZM", "pdf_url": "https://openreview.net/pdf/90a740707a3310652a97175a795c8685d342938a.pdf", "updated_at": "2026-10-05T19:58:26.209Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6n5OgsGBuJ", "number": 24539, "title": "INT8-RL: Training--Inference Co-Optimization for Quantized Rollouts on Ascend", "abstract": "Autoregressive rollout generation is a major cost of reinforcement learning (RL) post-training for large language models. INT8 inference can reduce this cost on accelerators with efficient integer kernels, but quantizing the rollout policy changes the distribution from which training trajectories are sampled. Activation outliers further complicate W8A8 execution. We present \\irlMethod{}, a training--inference co-optimization framework for INT8 rollouts on Ascend. The framework combines quantization-aware training with jointly learned smoothing and weight scales, reproducible hash-based rounding, and selective precision fallback. These components adapt the policy to quantized computation while reducing avoidable numerical discrepancies between training and rollout. Experiments on openPangu-7B and Qwen3-30B MoE show substantial reward recovery over the corresponding INT8 baselines. On openPangu-7B with DAPO-Math-17K, the combined W8A8 configuration improves reward from 0.252 to 0.305, compared with 0.309 for BF16. Profiling shows a 19.7\\% reduction in per-token decode latency. A separate long-sequence ablation that disables training-side fake quantization reduces step time by 18.7\\%, indicating optimization headroom rather than the measured end-to-end gain of the complete QAT pipeline. These results highlight both the feasibility of INT8 rollouts and the need to control quantization error and training overhead jointly.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "RL", "INT8 Quantization; Quantization-Aware Training", "Low-Precision Rollouts", "Learnable Smoothing", "Training–Inference Alignment" ], "forum_url": "https://openreview.net/forum?id=6n5OgsGBuJ", "pdf_url": "https://openreview.net/pdf/16026f2a80d8e6e62266f8a8a6d1fcd44f4baaf0.pdf", "updated_at": "2026-10-05T19:58:28.233Z", "categories": [ "ptq", "qat", "training", "hardware" ], "matched_terms": [ "int8", "low-precision", "qat", "quantization", "quantized", "quantizing", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "87Zb0sXV0W", "number": 24557, "title": "Transform-Aligned Learned Features for Lossy Point Cloud Attribute Compression", "abstract": "Transform-based methods provide an effective framework for point cloud attribute compression by representing attributes as transform coefficients. Introducing learned spatial context into this framework requires mapping spatial representations to the transform domain, but this known basis change is often left for the network to learn implicitly. We propose Transform-Aligned Learned Features (TALF) by applying the attribute transform to learned spatial representations, explicitly aligning them with the coding targets. Our analysis shows that the resulting features exactly represent the first-order prediction term of a smooth nonlinear model, with a bounded Taylor remainder. We integrate TALF into a transform-based attribute codec with explicit coefficient prediction and conditional residual entropy modeling under a unified coefficient-domain rate--distortion objective, while retaining explicit quantization-step control. Extensive experiments across three benchmark datasets and multiple transform bases demonstrate that TALF improves rate--distortion performance over conventional and learned baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Point cloud attribute compression; Transform-aligned learned features; Learned entropy modeling; Rate–distortion optimization" ], "forum_url": "https://openreview.net/forum?id=87Zb0sXV0W", "pdf_url": "https://openreview.net/pdf/f24427c7c175baff85b951f4975f9282ad845e71.pdf", "updated_at": "2026-10-05T19:58:30.641Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NqKt2d3j1s", "number": 24590, "title": "LatchQ: Latching onto Hot-Cache Query Interactions for KV Cache Quantization", "abstract": "KV cache quantization reduces the memory footprint of large language model (LLM) inference, but minimizing reconstruction error alone may fail to preserve attention behavior. Although future queries are unavailable at quantization time, keys can interact with subsequent queries while residing in the full-precision hot cache. Motivated by this observation, we introduce ***LatchQ***, which uses these observed interactions to guide key clipping while preserving the baseline bit allocation. *QTrace* accumulates attention-weighted query energy for each key and channel, preserving differences in observed key usage and channel sensitivity. *QClip* leverages this evidence to select shared clipping ranges by minimizing interaction-weighted reconstruction error. Our method preserves the baseline storage format and cache schedule without extending hot-cache residency. Our analysis reveals 63\\% agreement between clipping range selections based on hot-cache observations and those based on both observed and future queries. Furthermore, ***LatchQ*** achieves the highest average LongBench scores among evaluated KV quantizers at 3.0 and 2.25 bits per dimension (bpd). It also outperforms the baseline quantizer on reasoning and code generation. Code is included in the supplementary material, and the project page is available at https://anonymous-latchq.github.io/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "KV Cache", "Quantization" ], "forum_url": "https://openreview.net/forum?id=NqKt2d3j1s", "pdf_url": "https://openreview.net/pdf/5e4f49606472e0cb250f755eab7cac1b44f587c9.pdf", "updated_at": "2026-10-05T19:58:35.830Z", "categories": [ "ptq", "kv-cache", "reasoning" ], "matched_terms": [ "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UDFkcoNTGQ", "number": 24653, "title": "RANDOMLLM: HOW LITTLE RANDOMNESS DOES A STRONG LOTTERY TICKET NEED?", "abstract": "Strong-ticket training freezes a network's weights at random values and learns only a mask over them.\nThe standard construction for language models samples an independent random weight for every projection, so the random source holds as many independent scalars as the model has connections.\nWe propose RandomLLM, which instead slices every random weight from a single rank-$r$ random pool and keeps a separate multivalued mask for every projection.\nWe find that the randomness of the random weight has little influence on the strong ticket that mask learning produces: at source rank $1$, two frozen vectors supply $24{,}576$ independent scalars, about $10^5$ times fewer than one per connection, and the resulting model reaches $16.00$ perplexity against $15.95$ and $38.12$ against $37.21$ mean zero-shot accuracy at 3B, while a controlled three-seed experiment finds no resolvable effect from either sharing the pool or constraining its rank.\nMoreover, instead of accelerating this paradigm with dedicated hardware; we give a general-purpose GPU implementation that avoids rebuilding the mask at every forward pass, reducing prefill and decode latency by $6.25\\times$ and $6.16\\times$ and training time by $1.60\\times$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Strong Lottery Ticket", "Low-bit inference" ], "forum_url": "https://openreview.net/forum?id=UDFkcoNTGQ", "pdf_url": "https://openreview.net/pdf/a21c176f364e0e229a191c4500b42f7bf3df1365.pdf", "updated_at": "2026-10-05T19:58:45.897Z", "categories": [ "hardware" ], "matched_terms": [ "low-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zlpasjGSoD", "number": 24783, "title": "Hierarchical Runtime Alignment for Safe Long-Context Agents", "abstract": "Long-context agents require safety controls that remain effective as model state,\nretrieved context, and cache state change over time. We introduce hierarchical run-\ntime alignment, a deployment-time framework that combines three coupled layers:\nL1 parametric safety adaptation, L2 contextual policy evidence, and L3 cache-level\nsafety preservation. Using matched vanilla/abliterated pairs from Qwen3-8B and\nLlama-3.2-3B, we show that runtime safety is a state-dependent calibration prob-\nlem rather than a single guardrail problem. Our experiments yield three findings.\nFirst, contextual policy substantially reduces harmful compliance in the tested\nabliterated models. Its false-refusal cost varies with model state and policy strength.\nSecond, the tested safety–utility operating point depends on model state. Third,\nmatched benign controls distinguish cache-allocation effects from uniform preci-\nsion reduction. Generic INT4 changes harmful and benign behavior together. The\nQwen allocation control changes which signals retain precision. Key comparisons\nuse three seeds, paired uncertainty, two independent semantic judges, and a prevali-\ndated fixed-subset LongSafetyBench evaluation. Across factorial sweeps, agent\ntasks, adaptive attacks, many-shot jailbreaks, RULER, and LongSafetyBench, the\nresults support joint, model-state-aware calibration rather than a universal all-on\nconfiguration.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Alignment", "Safety" ], "forum_url": "https://openreview.net/forum?id=zlpasjGSoD", "pdf_url": "https://openreview.net/pdf/83ecb373f91665e554b778b5252a032f5ae8791f.pdf", "updated_at": "2026-10-05T19:59:05.654Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "int4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KODwGYKY00", "number": 24887, "title": "Latest Exact Match Attention", "abstract": "We introduce latest exact match attention (LEMA), an attention variant for transformers where queries and keys are binarized and each query attends only to the latest exactly matching key. We prove that LEMA transformers with chain of thought can simulate word-RAMs, as was recently shown for the less restrictive rightmost hard attention. In contrast to prior hard attention variants, the restriction to exact matches enables an efficient converse direction: word-RAMs can simulate LEMA transformers at a cost per token independent of the context length. Together, these results yield a close correspondence between the two computational models in terms of both compute and memory. Beyond the theory, we propose a training method for LEMA transformers that handles their non-differentiable operations with a straight-through estimator for the binarization and a soft attention surrogate annealed towards LEMA. On a synthetic associative recall task, LEMA models trained this way use their growing state to store and recall a large number of associations, outperforming gated DeltaNet (GDN) with its fixed state size. As a first scaling test, we train LEMA language models with up to 834 million parameters. They match softmax transformers of around half their size in loss and, on repeated rare phrases and a needle-retrieval task, remain behind softmax transformers but recall across longer distances than GDN models of comparable size. Finally, we implement dictionary-based inference for LEMA transformers and show constant generation speed comparable to GDN despite their growing state, with the dictionaries residing in main memory rather than VRAM.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "attention", "transformer", "hard attention", "expressivity", "word-RAM", "chain of thought", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=KODwGYKY00", "pdf_url": "https://openreview.net/pdf/eaae78bf4dbccc0c78d7a59fad5b65743614b683.pdf", "updated_at": "2026-10-05T19:59:23.719Z", "categories": [ "reasoning" ], "matched_terms": [ "binarization", "binarized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MB117f2ICa", "number": 24930, "title": "Theoretically Scaled Quantization: A Data-Free Analytical Framework for LLMs", "abstract": "Large Language Models (LLMs) require efficient post-training quantization (PTQ) to reduce memory and computational costs while preserving model quality. Existing PTQ methods commonly rely on calibration data, activation statistics, Hessian information, or iterative optimization, introducing additional preprocessing overhead and scale storage. We propose Theoretically Scaled Quantization (TSQ), a fully data-free analytical PTQ framework that derives quantization parameters directly from pretrained weight statistics. We formulate the TSQ Distortion Objective, $R_N(k)$, an exact distribution-independent geometric distortion objective for finite multi-region quantization, and define the optimal geometric scaling factor through one-dimensional minimization over $k$. We further derive exact region-wise absolute-error bounds that decrease geometrically with region index and a region-independent relative-error bound of $k/(2q_{\\max})$ for all non-innermost regions. Under the empirically observed near-Gaussian structure of pretrained LLM weights, we derive the TSQ Governing Optimality Equation, $\\Phi_N(k,r)=0$, which analytically characterizes the optimal scaling factor using the normalized dynamic range $r=M/\\sigma$, eliminating calibration data, activation statistics, Hessian computation, and iterative optimization. Building on the regional distortion analysis, we derive a generalized TSQ distortion law that decomposes normalized distortion into the weight dynamic range, optimized geometric distortion, and quantization resolution. TSQ also employs a scale-efficient representation requiring only one per-row maximum and a single global scaling parameter, from which all thresholds and quantization step sizes are analytically derived, reducing scale storage from $O(mn/g)$ to $O(m)$. Experiments across OPT, Llama, and Qwen model families show that TSQ achieves competitive or lower perplexity than AWQ across the evaluated settings while requiring no calibration data, substantially less scale storage, and lower full-model quantization time. These results demonstrate that analytical geometric scaling can provide a data-free and scalable approach to LLM post-training quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Large Language Models", "Data-Free Methods", "Analytical Methods", "Geometric Scaling" ], "forum_url": "https://openreview.net/forum?id=MB117f2ICa", "pdf_url": "https://openreview.net/pdf/00f9ee83bb0a9257b31ca21ca4d706ce08b31a6a.pdf", "updated_at": "2026-10-05T19:59:29.607Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KCMR4XvFWM", "number": 24979, "title": "Arbitrary-Accuracy Neural Approximation with Optimal Neuron Count and Near-Optimal Bit Complexity", "abstract": "We study the minimum number of hidden neurons required for\narbitrary-accuracy approximation of multivariate Hölder-continuous functions on $[0,1]^d$ and the associated encoding complexity. For $d\\geq 2$, we construct a fixed, explicitly defined activation function for which a closed-form network with two hidden layers of widths $d$ and $1$ achieves arbitrary accuracy in the uniform norm. We prove that $d+1$ is the exact minimum total number of hidden neurons among standard feedforward networks with locally integrable activations and affine outputs. We further give a simpler construction using a single elementary activation that combines the floor and exponential functions. This construction requires three hidden layers of widths $d$, $1$, and $2$, only two neurons above the minimum. If a skip connection is allowed, widths $d$, $1$, and $1$ suffice. These\nconstructions use explicit grid addressing and integer encoding of quantized function values. For a bounded $\\alpha$-Hölder class, the target-dependent encoding requires $O(\\varepsilon^{-d/\\alpha}\\log(1/\\varepsilon))$ bits, matching the metric-entropy lower bound up to a logarithmic factor.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Universal approximation", "optimal neuron count", "bit complexity", "superexpressive activation functions", "Hölder-continuous functions." ], "forum_url": "https://openreview.net/forum?id=KCMR4XvFWM", "pdf_url": "https://openreview.net/pdf/11521dbd276f951d36e045dc37cceaa6b4951230.pdf", "updated_at": "2026-10-05T19:59:35.152Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "L1xiufYI8B", "number": 25196, "title": "QFast: Fast Quantized Inference for Robotics Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models have demonstrated strong performance in robotic manipulation by unifying visual perception, language understanding, and action prediction within transformer-based architectures. \nHowever, their deployment on resource-constrained edge devices faces critical challenges: inference latency and memory requirements that prevent real-time operation. \nExisting quantization methods prioritize accuracy preservation and memory compression while largely neglecting inference speed—a bottleneck for real-time robotic control requiring sub-100ms latency.\nWe present \\textbf{QFast}, a quantization inference system designed to accelerate VLA models on edge devices. \nQFast employs speed-oriented quantization that balances accuracy and latency through module-specific bit-width selection, per-token weight quantization with dynamic per-channel activation quantization, and selective preservation of precision-sensitive layers. \nTo exploit hardware capabilities, we implement optimized GEMM kernels tailored to diverse VLA matrix dimensions, enabling native low-bit computation on Tensor Cores.\nEvaluated on Pi0.5 and OpenVLA, QFast achieves 1.3$\\times$ end-to-end speedup with up to 3.6$\\times$ memory reduction while maintaining task performance on LIBERO benchmarks. \nQFast represents the first quantization system enabling hardware-accelerated low-bit inference for VLA models, making real-time edge deployment practical. \nCode is available at https://anonymous.4open.science/r/QFast-D157/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action (VLA) models", "Inference Accleration" ], "forum_url": "https://openreview.net/forum?id=L1xiufYI8B", "pdf_url": "https://openreview.net/pdf/64e0517dc0b8d7706265a4b612eb88409537ab04.pdf", "updated_at": "2026-10-05T20:00:07.331Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "bit-width", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mxGIkpABn7", "number": 25240, "title": "Focus Calibration for Ultra-Low-Bit All-Reduce in Bandwidth-Constrained Collaborative LLM Inference", "abstract": "In bandwidth-constrained environments, collaborative inference of large language models (LLMs) is increasingly bottlenecked by activation communication under tensor parallelism. Although low-bit activation compression can reduce communication overhead in distributed LLM inference, further bit-width reduction is approaching practical limits, as aggressive compression often incurs substantial accuracy degradation or requires costly retraining. To address this challenge, we propose Focus Calibration, a lightweight calibration strategy for retraining-free ultra-low-bit vector quantisation, which enables an index-based all-reduce mechanism that exchanges compact centroid indices instead of full-precision activation tensors. Experimental results show that our method substantially reduces activation communication overhead, achieving up to 38.8\\% and 12.3\\% reductions in end-to-end prefill and decoding latency, respectively, while maintaining accuracy comparable to existing quantisation-based approaches at ultra-low bit-widths of 2-2.5 bits. These results establish index-based all-reduce with Focus Calibration as a practical solution for communication-efficient LLM inference in bandwidth-constrained collaborative systems.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Communication path Quantisation", "Quantisation calibration", "Collabrative LLM inference" ], "forum_url": "https://openreview.net/forum?id=mxGIkpABn7", "pdf_url": "https://openreview.net/pdf/c461eb8f582a9b8b4fdd65757dd8eb40c0488d74.pdf", "updated_at": "2026-10-05T20:00:11.482Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "bit-width", "low bit", "low-bit", "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zbBSzTSQw8", "number": 25267, "title": "Beyond Action Deviation: Execution-Level Quantization Sensitivity for Vision-Language-Action Models", "abstract": "Post-training quantization (PTQ) offers a practical way to reduce the deployment cost of vision-language-action (VLA) models. When uniform low precision cannot preserve control fidelity, mixed-precision PTQ relies on sensitivity estimation to identify which layers should retain higher precision. Existing sensitivity criteria mainly measure internal numerical distortion or action-level deviation, overlooking how quantization perturbations propagate through robot execution and influence task performance. Consequently, mixed-precision quantization guided by these criteria can lead to substantial performance degradation, particularly when applied to the sensitive action generation modules. We identify the need for an intermediate perspective between action deviations and task outcomes that captures robot-dependent execution information. To this end, we introduce Execution-Level Quantization Sensitivity (EQS), a layer-wise sensitivity criterion for mixed-precision quantization of VLA action generators. It models how quantization-altered action commands translate into horizon-level motion discrepancy and evaluates the discrepancy's execution significance under the current robot configuration. Together, these components provide an execution-aware sensitivity signal for layer prioritization. Experiments across simulation benchmarks and real-robot manipulation tasks show that EQS exceeds the best-performing baseline by 0.44~8.06 percentage points in average task success, with gains reaching 33.82 percentage points over individual baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision language action model", "post training quantization", "robot execution" ], "forum_url": "https://openreview.net/forum?id=zbBSzTSQw8", "pdf_url": "https://openreview.net/pdf/59d3bca848c6c004d93ec83bc41f722239690365.pdf", "updated_at": "2026-10-05T20:00:14.946Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low precision", "mixed-precision", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ezYCsakKyn", "number": 25277, "title": "Quantize the Keys, Audit the Values: Retained-State Certificates for KV-Cache Compression", "abstract": "KV-cache compression aims to support long-context inference within limited memory while preserving output quality. How that memory is distributed matters because compression in different cache regions can affect the current attention output differently. Aggregate task scores assess overall quality but do not by themselves provide query-specific feedback for allocating memory within a compressed cache. We provide this feedback with a deterministic certificate that bounds attention error for each query, even after the original entries are discarded. It also gives the exact reduction in the bound that each increase in value precision would produce, which enables adaptive allocation while the cache stays compressed. This allocation lowers attention error by 30–67% under the same memory budget and reaches a calibrated value-error target with roughly two-thirds as much refinement memory as uniform allocation. In packed serving of two 7–8B models at 32K, adaptive allocation brings the next-token distribution 2.6–3.4 times closer to dense decoding in KL divergence than uniform allocation with the same memory. The design is also practical to serve. A fused kernel evaluates the certificate alongside attention with little additional arithmetic and reads the selected state directly from a packed cache. Its base representation uses about $0.26\\times$ dense KV storage and serves a 70B model at 128K on one GPU.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache compression", "attention error bounds", "adaptive memory allocation", "quantization", "long-context inference" ], "forum_url": "https://openreview.net/forum?id=ezYCsakKyn", "pdf_url": "https://openreview.net/pdf/040a78d0696b60384067cb170e4e891aa9cbd478.pdf", "updated_at": "2026-10-05T20:00:16.128Z", "categories": [ "ptq", "kv-cache", "hardware", "theory" ], "matched_terms": [ "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qD67x7kVIo", "number": 25427, "title": "Beyond Scalar Recovery: Residual Geometry in Quantized Vision-Language-Action Models", "abstract": "Post-quantization recovery in vision-language-action (VLA) models is typically summarized by scalar error or task-level metrics, which measure how much error remains but not how that error is organized across a multi-step action chunk. We characterize this missing structure through teacher-relative residual geometry in a controlled OpenVLA-OFT setting. Across 13 heterogeneous ternary quantization configurations, mean pairwise Spearman similarity increases from 0.037 before adaptation to 0.557 afterward, with a demonstration-bootstrap median gain of +0.520 (95% CI [0.472, 0.568]). This geometric reorganization is not explained by uniform error shrinkage, and scalar recovery magnitude can diverge substantially from final residual alignment. The shared-alignment trend persists across additional optimization seeds, matched ground-truth-only adaptation, and a distinct 3-bit quantizer. Matched full-precision controls further demonstrate that shared alignment does not require an initial quantization perturbation. Finally, a leave-one-out template formed directly from other configurations’ adapted residuals strongly predicts the final residual geometry of a held-out configuration, reaching mean Spearman correlation 0.678 and substantially outperforming both raw geometry and a leave-one-out recovery-field predictor. Shared chunk-position and action-dimension effects explain much of this predictability, while the complete template retains additional predictive structure in reconstruction error.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision-language-action models", "quantization", "post-quantization adaptation", "residual geometry", "action chunking", "robot learning", "model compression" ], "forum_url": "https://openreview.net/forum?id=qD67x7kVIo", "pdf_url": "https://openreview.net/pdf/614f02b69affb3c6fb2bfe183147bc3d555bc6e5.pdf", "updated_at": "2026-10-05T20:00:42.012Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "3-bit", "quantization", "quantized", "quantizer", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xoc8GGXykA", "number": 25451, "title": "MBVQ: Multi-Level Block-Wise Vector Quantization for Hardware Friendly Post-Training Quantization of LLMs", "abstract": "The rapid scaling of large language models (LLMs) has substantially increased their deployment cost. Post-training quantization (PTQ), especially vector quantization (VQ), focuses on model compression to below 2 bits per parameter, but usually overlooks the data movement bottleneck that dominates real-world inference latency. In this paper, we propose **M**ulti-level **B**lock-wise **V**ector **Q**uantization (MBVQ), a hardware-friendly VQ framework designed to address this issue. MBVQ introduces two key innovations: (1) capacity-constrained per-block codebooks for efficient memory residency during dequantization, and (2) multi-level residual quantization scheme with row-wise importance selection and per-block scaling to compensate for reconstruction error. Specifically, we first partition weight matrices into blocks to adapt large-scale codebooks to varying hardwares. Subsequently, we merge redundant adjacent blocks for quantizing residuals with selective codebooks and design codebook rescaling to adaptively adjust the block-wise distribution of values to simultaneously improve model accuracy and achieve substantial inference speedup. Experiments on a range of LLMs demonstrate that MBVQ consistently outperforms state-of-the-art low-bit quantization baselines under comparable compression ratios, delivering approximately $6.62\\times$ inference speedup, while maintaining competitive perplexity and zero-shot accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Large Language Models", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=xoc8GGXykA", "pdf_url": "https://openreview.net/pdf/4fe74af164fe4f20e0aa6669d871ef4c1062cd5b.pdf", "updated_at": "2026-10-05T20:00:46.853Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TlO7NHkqAg", "number": 25513, "title": "ACT-SNN: Activity-Constrained Two-Stage Conversion for Efficient Spiking Large Language Models", "abstract": "Large language models (LLMs) incur substantial computational and energy costs during inference, motivating their conversion into energy-efficient spiking neural networks (SNNs). \nAlthough existing LLM-to-SNN methods have achieved high conversion fidelity, this does not necessarily guarantee efficient inference: accurate discrete representations can require high spike activity and thus higher energy consumption, while layer-wise temporal accumulation leads to serial inference latency.\nIn the first stage, our Basis-Augmented Multi-Threshold Neuron (BA-MTN) jointly learns basis transformations and channel-wise thresholds under an explicit activity constraint, decoupling reconstruction optimization from spike activity control while enlarging the feasible space for a favorable reconstruction--activity trade-off.\nIn the second stage, the Basis-Augmented Symmetric Integrate-and-Fire (BA-SIF) neuron provides an equivalent temporal realization with causal step-by-step spike propagation, enabling pipelined cross-layer execution without layer-wise temporal accumulation; relaxation steps and differential operators further compensate for temporal propagation delays and nonlinear operations.\nExperiments across LLaMA-2, LLaMA-3, and Qwen2.5 models ranging from 7B to 70B parameters demonstrate that ACT-SNN achieves a strong accuracy--energy trade-off under both full-precision and 4-bit weight settings. In particular, under 4-bit weights, ACT-SNN attains energy ratios of $0.016$--$0.022$, reducing estimated inference energy by $49$--$64$% over representative W4A4 QANN baselines while maintaining competitive zero-shot accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Spiking Neural Networks", "ANN-to-SNN Conversion", "Spike Activity", "Energy-Efficient Inference", "Neuromorphic Computing" ], "forum_url": "https://openreview.net/forum?id=TlO7NHkqAg", "pdf_url": "https://openreview.net/pdf/dec9961ca5e64f8c58d7a7146e01b3aaaadd2670.pdf", "updated_at": "2026-10-05T20:00:58.179Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EpP44Vczot", "number": 25538, "title": "JoLT: Near-Lossless KV Cache Compression via Joint Rank-Bit Allocation", "abstract": "The key-value (KV) cache is the dominant memory bottleneck in long-context language model inference.\nExisting compression methods apply low-rank factorization or quantization independently, without jointly allocating rank and precision under a shared storage budget.\nWe introduce JoLT, a training-free compressor that treats grouped prefill caches as fourth-order tensors and applies partial Tucker decomposition along the token and feature modes, the two axes that carry low-rank structure, while leaving the head and layer modes intact.\nA rotated low-bit quantizer captures the truncation residual, and a single Lagrangian dual allocates per-group Tucker ranks and residual bit-widths under a global byte constraint.\nFlashJoLT replaces the exact token-mode SVD with a randomized approximation that matches JoLT within the free zone at a fraction of the compression cost, and a fused Triton decode kernel evaluates attention directly over the stored factors without materializing dense KV tensors.\nAcross five models from four architecture families, covering multi-head attention, grouped-query attention, and mixture-of-experts architecture, JoLT achieves 2 - 3x compression with less than 0.2% perplexity degradation, without retraining.\nOn RULER at 64K context with LLaMA-3.1-8B, retrieval accuracy remains near-lossless through 3x and declines by only 0.90 and 2.40pp at 4x and 5x, respectively.\nJoLT demonstrates that tensor-aware low-rank decomposition and quantized residuals, unified under a single storage budget, achieve near-lossless KV-cache compression across diverse model architectures without retraining.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache compression", "large language models", "Tucker decomposition", "low-rank approximation", "quantization", "Lagrangian relaxation", "rate-distortion", "long-context inference", "randomized SVD", "fused attention kernel" ], "forum_url": "https://openreview.net/forum?id=EpP44Vczot", "pdf_url": "https://openreview.net/pdf/c9ebaefa519d48bc591f30291933adf3a4e2ac6e.pdf", "updated_at": "2026-10-05T20:01:03.741Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EaBNBOUY4i", "number": 25555, "title": "PhotoFinish: Adaptive Precision for Exact Language Model Sampling", "abstract": "Sampling a token is a race with a single winner, yet output projection reads every weight at a fixed precision. PhotoFinish makes precision follow the race. A low-bit scan bounds token scores under fixed Gumbel perturbations, and an exact evaluation raises the threshold for remaining contenders. Only rows that can still win receive additional precision. Valid bounds certify each exclusion, preserving the same sampled token as a full scan. We characterize the minimum information required to determine the winner and analyze when another precision stage repays its cost. The analysis gives a tight worst-case guarantee for a reference access policy and explains why shrinking the candidate set need not reduce runtime. Across three large-vocabulary language models at small decoding batches, PhotoFinish outperforms the fastest of three evaluated full-scan samplers, including FlashSampling, with sampling-head speedups up to 2.61× and lower complete GPU decode-step latency during natural generation. Matched INT8 controls isolate the benefit of selective four-bit reads. PhotoFinish opens a path to exact sampling in which precision is allocated to the decision being made. Open-source code and reproduction materials accompany the submission.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "efficient inference", "exact sampling", "adaptive precision", "GPU kernels" ], "forum_url": "https://openreview.net/forum?id=EaBNBOUY4i", "pdf_url": "https://openreview.net/pdf/4c9dcafb1a0dcdb366c7e3dc76e0f1ea93cea3e5.pdf", "updated_at": "2026-10-05T20:01:07.262Z", "categories": [ "hardware" ], "matched_terms": [ "int8", "low-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fArqFprckh", "number": 25574, "title": "HEAT: Faster Fully Homomorphic Inference via Approximations-Weights Co-Adaptation", "abstract": "Fully homomorphic encryption (FHE) allows a server to run a language model directly on encrypted user prompts, but current approaches remain prohibitively slow. Ciphertexts natively support only addition, multiplication, and rotation, and multiplications may be composed only to a bounded depth before a costly bootstrapping operation is required to continue. Every nonlinearity must therefore be approximated by an iterative method; each iteration increasing the number of multiplications. A higher iteration count buys precision but exhausts the available depth more frequently and thus triggers more bootstraps, which dominate latency. \nWe introduce **Homomorphic Encryption-Aware Training (HEAT)**, a _fine-tuning_ method that makes the per-nonlinearity iteration counts _learnable_, enabling them and the model weights to co-adapt during training. HEAT optimizes iterations with respect to the task objective, allowing the model to adapt to approximation errors encountered during inference without architectural changes or retraining from scratch. We further relate iteration count to quantization bit width and bound, at fixed weights, the gap between our objective and quantization-aware training. On encrypted GPT-2 decoding, HEAT reduces iterations by $3.1\\times$, bootstraps by $1.6\\times$, and end-to-end latency by $1.4\\times$, while improving decode agreement over the calibrated encrypted baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Fully homomorphic encryption", "FHE", "Fine-tuning", "Encryption", "CKKS" ], "forum_url": "https://openreview.net/forum?id=fArqFprckh", "pdf_url": "https://openreview.net/pdf/e0f3c45c17492a47c96342442468f3b0e0aa478a.pdf", "updated_at": "2026-10-05T20:01:13.345Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "bit width", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TevsFAso12", "number": 25637, "title": "Compression-Aware Fine-Tuning: Adapting Language Model Weights to a Frozen Cache Bottleneck", "abstract": "The low-rank projection of the key-value (KV) cache is a natural strategy to reduce inference memory. However, compressing a model whose weights were never trained to tolerate the resulting information loss leads to severe quality collapse at moderate compression ratios. Compression-aware fine-tuning (CAFT) closes this gap by adapting the model weights to a frozen cache compressor rather than compressing weights that were never trained for the loss. At a 2x cache budget on Llama-3.2-1B, CAFT reduces WikiText-103 perplexity by 11.5x relative to the strongest post-hoc baseline (a fine-tuned checkpoint compressed under a basis freshly calibrated on its own activations). The advantage holds across six compression operators, three model scales (1B to 8B), and two training domains, and persists when stacked with 4-bit code quantization. At 2x and 3x compression, CAFT preserves single-needle retrieval and maintains a large advantage on LongBench where the post-hoc baseline collapses. Mechanistically, the gain is functional robustness: the network routes around the discarded dimensions rather than concentrating cache energy into the retained subspace, as shown by spectral probes and basis-transfer controls that falsify induced compressibility. Through this paper, we demonstrate that CAFT recovers the capability that low-rank KV cache compression destroys.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "compression-aware fine-tuning", "low-rank projection", "efficient LLM inference", "long-context inference", "straight-through estimator", "large language models", "memory-efficient inference", "model compression", "fine-tuning" ], "forum_url": "https://openreview.net/forum?id=TevsFAso12", "pdf_url": "https://openreview.net/pdf/f2afdb58f93b7695ed7f394112e38500efe0421b.pdf", "updated_at": "2026-10-05T20:01:24.803Z", "categories": [ "ptq", "kv-cache", "theory" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "o2wbM8RqOQ", "number": 25643, "title": "CroGen: Probing Spatial Control at Extreme Target Densities in Image Generation", "abstract": "Advanced image diffusion models generate realistic images and support flexible spatial control.\nHowever, we find that spatial correspondence deteriorates as the number of instances we aim to control increases, despite visually convincing outputs.\nWe term this phenomenon spatial constraint overload.\nIn this paper, we explore this phenomenon using dense crowds as a case study and introduce CroGen, a simple and elegant method to alleviate it.\nSpecifically, CroGen combines Proximity Field Conditioning (PFC) and Pixel-Packing Injection (PPI).\nOur PFC converts target locations into a quantized distance field, providing dense guidance while keeping individual locations explicit.\nOur PPI retains fine spatial details through pixel packing and injects the packed condition through a single linear projection.\nWe further introduce a dataset of 71K crowd images spanning diverse crowd densities, with point annotations and text prompts.\nComprehensive experiments on controllable generation and downstream dense crowd counting show that our CroGen improves spatial consistency without sacrificing image quality.\nIn particular, our method better maintains this consistency as the number of controlled instances grows, demonstrating its effectiveness in alleviating spatial constraint overload.\nThe dataset and model will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Dense Spatial Control; Controllable Image Generation; Diffusion Models; Spatial Condition Encoding; Crowd Generation" ], "forum_url": "https://openreview.net/forum?id=o2wbM8RqOQ", "pdf_url": "https://openreview.net/pdf/06418ac008390fa545fc6ed5cbe99b3bf9cd887a.pdf", "updated_at": "2026-10-05T20:01:25.719Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qUpUiOSDpP", "number": 25656, "title": "CONTRACT-ALIGNED QUANTIZATION FOR FSS- BASED SECURE TRANSFORMER INFERENCE", "abstract": "Function secret sharing (FSS) makes online two-party inference cheap, yet existing quantized Transformers for private inference are trained with at mostbitwidth-level awareness of the protocol---an idealized forward pass that doesnot match what the deployed protocol gates actually compute: FSS gates impose share-wise rounding with correlated truncation errors, public lookup-table encodings, and gate-specific input domains. We show that this mismatch systematically degrades accuracy. We therefore propose contract-aligned quantization: every nonlinear operator first publishes an explicit operator contract — ring, fixed-point scale, slot domain, LUT payload encoding, and truncation path — and the pruned, quantized model is then retrained with the exact integer semantics of the deployed gates in its forward pass. On SST-2, we retain 98.5\\% of the teacher's accuracy, Compared item by item with the values reported by SIGMA, Softmax achieves approximately 77\\% lower latency and approximately 51\\% lower communication, while GELU achieves approximately 55\\% lower latency and approximately 96\\% lower communication.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Privacy-Preserving Inference; Secure Two-Party Computation; Function Secret Sharing; Quant; Transformer Inference" ], "forum_url": "https://openreview.net/forum?id=qUpUiOSDpP", "pdf_url": "https://openreview.net/pdf/61ac7e5ff1c4c10a62da6fb8fb8bf326428050ba.pdf", "updated_at": "2026-10-05T20:01:27.030Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gRbFEs5deR", "number": 26198, "title": "Normalization-Free Residual Networks by Bounding Weights and Residuals", "abstract": "Normalization layers are widely used for stabilizing deep residual networks, yet they incur computational overhead and obscure training dynamics. We propose a stable, normalization-free residual block rooted in constrained nonlinear optimization that replaces traditional activation normalization with an inexpensive, element-wise clipping operation on the output of each residual branch. We extend this framework to weights, combining it with a parameterization that enables a single global constant to bound the magnitudes of both activations and weights. This parameterization additionally results in controllable and uniform effective learning rates across all layers. Compared to networks that do not use clipping, our approach yields activation and weight distributions that are remarkably smooth and free of outliers, enabling direct post-training quantization of weights to INT4 with only a minor loss degradation. Our experiments demonstrate that this approach successfully eliminates the need for both activation and weight normalization, offering a simple and principled foundation for training deep residual networks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "architecture research", "constrained optimization", "language modeling", "transformers", "quantization" ], "forum_url": "https://openreview.net/forum?id=gRbFEs5deR", "pdf_url": "https://openreview.net/pdf/d2842bbce50bf791d53c2435406f9f74a54f8410.pdf", "updated_at": "2026-10-05T20:02:55.868Z", "categories": [ "ptq" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bstvvfFV4y", "number": 26300, "title": "TerMeZO: Ternary Sparse Zeroth-Order Optimization for Fine-tuning BitNet Models at the Edge", "abstract": "Fine-tuning large language models (LLMs) with first-order optimizers requires a memory several times larger than that required for inference. Memory-efficient zeroth-order optimization (MeZO) sidesteps this cost by estimating gradients from forward passes only. However, for BitNet architectures, a family of LLMs with ternary {-1,0,1} weights and 8-bit activations, fine-tuning requires updating full-precision latent weights, and thus the memory footprint of MeZO no longer matches that of inference. A promising solution is to fine-tune only a subset of the latent weights, but existing sparse zeroth-order (ZO) methods either ignore the ternary structure or require first-order gradient information to build a sparse mask, which is at odds with the purpose of ZO fine-tuning. We propose TerMeZO, a sparse MeZO scheme that exploits the geometry of the ternary quantizer itself to identify the latent weights that are more likely to change values during fine-tuning, at no additional data or memory cost. Our convergence analysis shows that TerMeZO can converge faster than full-parameter MeZO, owing to its optimized reduction of the fine-tuning effective dimension. We run extensive experiments on BitNet models ranging from 1B to 3B parameters, spanning classification, instruction-following, and mathematical reasoning tasks. TerMeZO matches or exceeds the performance of full-parameter MeZO while substantially reducing the fine-tuning memory footprint.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models; BitNet; fine-tuning; zeroth-order optimization" ], "forum_url": "https://openreview.net/forum?id=bstvvfFV4y", "pdf_url": "https://openreview.net/pdf/db15a4aee717d40ae0b9d7e03c96393709e96379.pdf", "updated_at": "2026-10-05T20:03:10.659Z", "categories": [ "training", "reasoning", "theory" ], "matched_terms": [ "8-bit", "quantizer", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qtothsftO6", "number": 26309, "title": "Quantitative Generalization Bounds for Multi-Layer Transformers: A Few Long Strings can Beat Many Short Ones", "abstract": "Transformer-based language models exhibit an astonishing capacity to generalize to longer and more difficult problem instances than those seen in training. However, theoretical understanding of this phenomenon still lags behind. Unlike most machine learning models, the generalisation abilities of transformers inherently require considering two axes. On one hand, one must examine in-distribution statistical generalisation, which concerns the ability of a model to broaden a function from training samples to novel samples of the same distribution. On the other hand, given the sequential nature of language data, one must also consider length generalisation, which reflects whether a model trained on inputs of a given length can extrapolate to longer ones. These two notions have mostly been studied separately, leaving open how they interact. In this work, we give a joint treatment of both axes for fixed-precision transformers. We provide bounds for generalisation guarantees in terms of the parameter sizes of the involved models, as well as the quantization space and the alphabet size that supply the sequences. Finally, we illustrate how the requirements interact between number and length of training samples. In particular, we find that increasing training length can be exponentially more efficient than increasing the sample size. Experiments support predictions of the theory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "transformers", "learning theory", "length generalization" ], "forum_url": "https://openreview.net/forum?id=qtothsftO6", "pdf_url": "https://openreview.net/pdf/4f845a9a0cec2c1ebf0b34fa8d0a64081e3c7080.pdf", "updated_at": "2026-10-05T20:03:11.882Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aGfpzml2qT", "number": 26311, "title": "TORQUE: Optimizing What (not) to Quantize Before and After Rotation", "abstract": "Random rotations are an effective preprocessing step for quantization: regardless of the input, they make coordinate distributions approximately Gaussian, enabling the use of precomputed optimal codebooks. We introduce TrunQuant, a framework that improves on previous work using random rotations by jointly optimizing how many and which coordinates to preserve at high precision before and after rotation, under a fixed overall expected bit budget.\n\nIntuitively, before rotation, preserving large input coordinates at high precision can reduce overall error by preventing the rotation from spreading their values across many coordinates. Likewise, after rotation, preserving a small fraction of the largest-magnitude coordinates at high precision allows the remaining values to be quantized more accurately using codebooks optimized offline for the resulting truncated Gaussian distribution. \n\nWe demonstrate TrunQuant's effectiveness across various biased and unbiased quantizers, including both scalar and vector methods, and show that it improves worst-case accuracy and applications such as KV-cache compression and approximate nearest-neighbor search.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "random rotations", "vector quantization", "model compression", "approximate nearest-neighbor search" ], "forum_url": "https://openreview.net/forum?id=aGfpzml2qT", "pdf_url": "https://openreview.net/pdf/0addc92057c2d72f1af49a7bc7e24ecaba9ef772.pdf", "updated_at": "2026-10-05T20:03:12.635Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization", "quantize", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3PXn4E10OA", "number": 26362, "title": "Demystifying Manifold Constraints in LLM Pre-training", "abstract": "The recent success of matrix optimizers (e.g., Muon) suggests that specific normalization of momentum—such as orthogonalization and row-wise normalization—benefits both the stability and acceleration of LLM training. Consequently, several recent studies have suggested that weights should also be normalized, leading to a Riemannian optimization problem. While such constrained training frameworks demonstrate superior performance, the effects of explicitly constraining weights, and their interaction with existing stabilization mechanisms, remain less understood. To bridge this gap, we study manifold constrained training dynamics through activation scales, rotational dynamics, and the update-to-weight ratio. We propose a Riemannian spectral steepest descent optimizer called MACRO, alongside a radius selection principle to serve as our testbed. Our analysis and numerical experiments reveal that RMSNorm and manifold constraints serve overlapping roles, and that weight decay can be completely eliminated when manifold constraints are applied. By controlling the update-to-weight ratio, constrained training significantly alleviates update cancellation, empirically demonstrating that MACRO is robust to low-precision computation and competitive with existing algorithms for standard LLM pre-training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM pre-training", "manifold-constrained optimization", "training dynamics", "normalization", "weight decay", "low-precision training" ], "forum_url": "https://openreview.net/forum?id=3PXn4E10OA", "pdf_url": "https://openreview.net/pdf/653df46efe348d3afb8b2cfd5035ea6cfb29ad12.pdf", "updated_at": "2026-10-05T20:03:19.993Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZGeR4gkEZq", "number": 26502, "title": "AdaptQuant: Hardware-Aligned Per-Tile MX Format Selection for Post-Training Quantization on NVIDIA Blackwell", "abstract": "Post-Training Quantization cuts the cost of serving Large Language Models, and NVIDIA Blackwell's native support for microscaling (MX) formats pushes the speedup further. Yet uniform MXFP4 still incurs substantial quantization error, and existing fixes assign higher-precision formats to sensitive regions using calibration data: a cost repeated for every new model before it can be served at all. We propose AdaptQuant, which assigns an MX element format per $16\\times32$ tile, the granularity at which Blackwell's block-scaled MMA reads its operands, from a single statistic computed in one pass over the tile. The rule mapping this statistic to a format is derived once, offline, on a single reference model against NVFP4 and reused unchanged on every other model. Across 13 open-weight LLMs on a single RTX 5090, AdaptQuant raises mean perplexity over BF16 by only $5.0\\%$ on WikiText-2 and $3.9\\%$ on C4 while spending fewer effective weight bits on average than the strongest mixed-precision baseline and quantizing $2.9\\times$ faster, cutting stored weight memory by up to $3.08\\times$ over BF16 and reaching $1.81\\times$ BF16's prefill throughput.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "model acceleration", "model compression", "microscaling quantization" ], "forum_url": "https://openreview.net/forum?id=ZGeR4gkEZq", "pdf_url": "https://openreview.net/pdf/3608529edfd0caf56e0908a36546b47564b0f16f.pdf", "updated_at": "2026-10-05T20:03:39.268Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "mixed-precision", "mxfp4", "nvfp4", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "t1JWQsnZ6C", "number": 26510, "title": "Stable-LoRA+: Regularizing Low-Rank Adaptation for Robust Scalin", "abstract": "The efficiency of low-rank adaptation does not ensure stable training as language models increase in size. Settings that are effective at one scale may degrade accuracy or lead to divergence at another. We study robust scaling of adaptation for mathematical reasoning, evaluating training stability and predictive accuracy across model sizes under a fixed training configuration, separately from performance after tuning. We introduce Stable-LoRA⁺, a training method that regularizes the factorized adaptation process. Its premise is that scaling the adapter output alone does not fully constrain factor evolution. The method therefore controls factor growth during training while penalizing update components aligned with the pretrained weight's leading singular subspaces. The same regularization is explicitly implemented for 4-bit backbones. Across more than 130 training-and-evaluation runs on five Qwen backbones (0.5B–14B), the strict fixed-configuration recipe matches vanilla LoRA's macro-average GSM8K accuracy (69.55%). With protocol and seed selection, the broader best-protocol comparison reaches 70.14% and mean rank 1.50. At lr = 10⁻³, vanilla LoRA diverges on 3B/7B, losing at least 76 percentage points, while our shared 10⁻⁴ configuration remains stable across the evaluated backbones. A quantization audit also identifies a silently inactive anchor, whose correction recovers 1.36 points at 7B. These results motivate regularization of factorized updates as an approach to robust adaptation across the evaluated model scales.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Training Stability", "Low-Rank Adaptation", "Cross-Scale Robustness" ], "forum_url": "https://openreview.net/forum?id=t1JWQsnZ6C", "pdf_url": "https://openreview.net/pdf/d1658b05db203a533a774460dba34a43fe165080.pdf", "updated_at": "2026-10-05T20:03:40.198Z", "categories": [ "reasoning", "theory" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FIS8CBDssc", "number": 26608, "title": "SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication", "abstract": "Communication increasingly dominates the cost of Large Language Model (LLM) pre-training, especially under data-parallel and sharded training schemes, where gradient synchronization and parameter reconstruction overhead increase with model size and system scale.\nExisting communication-reduction methods either sparsify raw gradients, which can be unstable for modern Adam-style optimizers at high sparsity, or quantize communication, whose savings are fundamentally bounded by bit width. \nWe present SCAPE, a communication-efficient distributed optimizer for LLM training that exploits the stability of AdamS's first-moment to enable aggressive sparsification while preserving LLM quality. Instead of constructing masks from gradients, SCAPE derives them from first-moment-based statistics, partitions mask generation across workers to align with optimizer sharding, and delays mask usage by one step so that mask synchronization can overlap with computation. SCAPE also reconstructs the quantities required for second-moment updates from a single synchronized sparse buffer, avoiding an additional collective. We implement SCAPE in Megatron-LM and evaluate its convergence by pre-training GPT-345M on OpenWebText and Llama-500M on SlimPajama-6B using 32 NVIDIA GH200 GPUs on TACC Vista. In both models, SCAPE preserves training stability, validation loss, and downstream task accuracy under 90\\% and 99\\% sparsity. \nFor Llama-500M, SCAPE reduces end-to-end pre-training wall-clock time by up to 43.3\\% while maintaining model quality comparable to dense AdamW and AdamS. For Llama-1.8B, SCAPE achieves up to 3.20$\\times$ speedup per step compared to dense distributed AdamS.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Distributed Computing", "Communciation Compression", "Sparsity" ], "forum_url": "https://openreview.net/forum?id=FIS8CBDssc", "pdf_url": "https://openreview.net/pdf/e1e301956ea08212d5db772472d6c22be9d65056.pdf", "updated_at": "2026-10-05T20:03:51.968Z", "categories": [ "ptq", "training", "hardware", "theory" ], "matched_terms": [ "bit width", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2OJ187fKm6", "number": 26646, "title": "When Tools Fail Quietly, LLM Agents State Wrong Answers as Fact", "abstract": "Tool-using LLM agents sometimes state a wrong answer as fact. We ask which properties of a tool failure cause this, using QuietFail, a fault-injection testbed over a real filesystem, SQLite databases, an HTTP service and a shell, in which each of 30 multi-step tasks also runs without its fault. Across seven agents from five providers and 4,305 graded runs, what the failing call returns shapes the outcome. Access denials end most runs in open failure and other errors raise confident wrong answers by at most 10.3 points, while a tool that silently returns half the requested records raises them from 7.3% to 26.1%, and to 38.1% in the runs where it removed data. A notice that the result is truncated, added to otherwise unchanged output, cuts them by 17.2 points, and by 15.6 in same-window cells of five agents (task-level p = 0.008); in a concurrent three-agent collection the notice arm matches the fault-free rate. For Claude Haiku the notice must say something about this result, either that it is truncated or how much; a hedge or a bare flag does nothing, and false notices cost little. When the agent cannot re-fetch, frontier agents fail too: Claude Sonnet 5 states a wrong answer in 39.7% of runs and Claude Opus 5 in 23.8% (27.0% and 12.7% under a second judge); the notice stops most of this, and a cursor to the withheld rows restores the answer (Sonnet 5: 95.2% correct). An ungated frontier auditor catches 80% of the wrong answers when told to look for incompleteness and 58% when not, at low precision, and in the loop neither a targeted nor a generic check significantly helps. Of 87 audited tool surfaces, 63 withhold output by default; 8 silently cut a result and 26 return the top k without a total. Other graders find the same directions at smaller sizes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM agents", "tool use", "tool failure", "silent failure", "fault injection", "hallucination", "abstention", "AI oversight", "agent-computer interface", "benchmarks and testbeds" ], "forum_url": "https://openreview.net/forum?id=2OJ187fKm6", "pdf_url": "https://openreview.net/pdf/253aa35245d7d7a52974add086f80c30b7f978c3.pdf", "updated_at": "2026-10-05T20:03:54.749Z", "categories": [], "matched_terms": [ "low precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "N6bS03pkah", "number": 26686, "title": "Communicate the Change: Efficient Expert Parallelism via Activation Compression", "abstract": "Scaling has driven the success of large language models (LLMs), with sparse mixture-of-experts (MoE) architectures enabling greater parameter capacity without a proportional increase in computation. At large scales, MoE models rely on expert parallelism to distribute experts across devices. Each MoE layer requires two synchronous collective communication steps during inference: dispatching hidden states to the activated experts and combining the resulting expert outputs. As compute throughput has outpaced interconnect bandwidth, these synchronous collectives have become a major bottleneck to MoE scaling and efficiency. We address this challenge with EchoMoE, a lightweight architecture modification that retrofits existing MoE models, requiring about 8500 times fewer training tokens than full pretraining. Motivated by the observation that transformer blocks update the residual stream incrementally, we use a hidden state from an earlier layer as a reference and communicate only the compressed difference between it and the current state, rather than communicating each hidden state in full. EchoMoE periodically refreshes these reference states and transmits only the compressed differences synchronously, reducing the traffic on latency-critical paths without significant computational overhead. EchoMoE reduces inter-node communication volume by 46% and improve the throughput by 17.2%, while preserving average performance within 2 percentage points across 18 benchmarks. EchoMoE also tolerates quantization of the communicated representations without measurable quality loss. We release an efficient implementation for applying EchoMoE to pretrained MoE models, together with an inference integration for vLLM", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture of Experts", "Expert Parallelism", "Efficiency", "Inference" ], "forum_url": "https://openreview.net/forum?id=N6bS03pkah", "pdf_url": "https://openreview.net/pdf/b9cd54a951c9920d20eaedb165f3314603ab261e.pdf", "updated_at": "2026-10-05T20:03:59.484Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yCofSMItv9", "number": 26747, "title": "OSFP4: Joint Optimization of Diagonal Smoothing and Block Scales for NVFP4 Quantization", "abstract": "NVFP4 is an attractive datatype for large language model (LLM) inference, offering compact storage and native tensor-core acceleration, but preserving accuracy requires careful quantization.\nIn this work we develop a novel quantization scheme called Optimized Smoothing and Scaling for NVFP4 (OSFP4). For each linear projection it uses a diagonal smoothing matrix whose entries are optimized to minimize the squared matrix-product quantization error under NVFP4, taking into account the rounding procedure that is used (either round-to-nearest, or GPTQ-style successive interference cancellation). This requires performing joint optimization on the smoothing entries as well as the block scales, which is facilitated by analyzing a multiplicative-dither FP4 quantizer instead of the fixed deterministic one. Experiments show that OSFP4 achieves the highest average accuracy among the evaluated competitors in the corresponding quantization settings, while retaining approximately 94-97\\% of vendor NVFP4 prefill throughput on the measured workloads.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "NVFP4", "Dither", "Diagonal Smoothing" ], "forum_url": "https://openreview.net/forum?id=yCofSMItv9", "pdf_url": "https://openreview.net/pdf/71041b99f2b1f9a1c47d625ef4e24136c5106fbd.pdf", "updated_at": "2026-10-05T20:04:08.645Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "nvfp4", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yQzCqXYYxJ", "number": 26768, "title": "MeMix: Selective State Preservation for Streaming 3D Reconstruction", "abstract": "Streaming 3D reconstruction must absorb new observations while retaining historical context in bounded memory. This requires deciding not only how strongly to update a recurrent state, but which locations should temporarily avoid a write and when to reselect them. We present MeMix, a training-free module that allocates a fixed preservation budget using current-frame compatibility. At each frame, it scores candidate state locations against the observation, preserves the previous values at the highest-scoring locations, and leaves the remaining locations on the original update path. MeMix adds no learnable parameters and keeps recurrent memory constant in stream length. Matched interventions show that compatibility-guided selection and current-frame reselection improve long-horizon performance beyond gate binarization or a fixed protected subset. We evaluate MeMix across update rules sharing a pretrained CUT3R checkpoint. On 16 1,000-view 7-Scenes streams, MeMix reduces TTSA3R's accuracy and completeness errors by 29.9% and 30.7%. With the FILT3R Kalman update, it lowers 1,000-view ScanNet pose ATE by 17.8% across 65 scenes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "streaming 3D reconstruction", "recurrent state", "selective state preservation", "long-horizon reconstruction" ], "forum_url": "https://openreview.net/forum?id=yQzCqXYYxJ", "pdf_url": "https://openreview.net/pdf/6f35f46c272dfb770b828efc6b8e30a6e69e13c4.pdf", "updated_at": "2026-10-05T20:04:13.973Z", "categories": [ "ptq" ], "matched_terms": [ "binarization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KF0jzKebwH", "number": 26777, "title": "Compact Activation Storage for ReLU Feed-Forward Networks", "abstract": "Reducing memory use in large language model (LLM) training is an ongoing challenge. Activation checkpointing saves memory by recomputation, while lossy compression changes the stored activations. The reemergence of sparse $\\relu$ and $\\relu^2$ activations creates a third opportunity: losslessly compressing the activation cache itself. We present BitSparse, a GPU-resident lossless compression format for sparse activations. Compression and reconstruction run in custom Triton kernels fused into an autograd function, shrinking the activation cache while recovering it exactly for the backward pass. Bfloat16 (BF16) and FP8 activations are supported. Optional sign-bit packing further increases compression up to close to the theoretical limit. We analyze its compression ratio and compare it to the theoretical optimum. BitSparse lowers per-token VRAM by 31\\% and 30\\% Modded-NanoGPT and Nemotron-H respectively, with overheads of no more than 1\\% at practical sequence lengths. Additionally, a new preallocated buffer system avoids host synchronization during training when variable size tensors are allocated.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Memory Reduction", "Compression", "LLM training" ], "forum_url": "https://openreview.net/forum?id=KF0jzKebwH", "pdf_url": "https://openreview.net/pdf/c5045b47ef3bbad13f1f7214a5c901a3dd4be835.pdf", "updated_at": "2026-10-05T20:04:14.367Z", "categories": [ "ptq", "training", "hardware", "theory" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1AFucgFgEv", "number": 26781, "title": "SmoothGPTQ: Optimizing Diagonal Scaling for GPTQ Weight-Activation Quantization", "abstract": "Diagonal smoothing, as in SmoothQuant, is a useful tool for balancing the errors incurred from weights quantization and activations quantization. For weight quantization the error can be further reduced by accounting for the activations second-order statistics via GPTQ. Prior work have considered the optimization of diagonal smoothing parameters for round-to-nearest (RTN) quantization. In this work we develop a framework, termed SmoothGPTQ, for optimizing those parameters for GPTQ-quantized weights and RTN-quantized activations. We derive an approximate expression for the quantization mean-squared error MSE as a function of $S$. This expression is convex in log-scale and is therefore amenable to minimization, and its accuracy is established through experiments. Applying the developed SmoothGPTQ algorithm to various LLMs give state-of-the-art results for W4A4 INT quantization across various benchmarks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "GPTQ", "SmoothQuant" ], "forum_url": "https://openreview.net/forum?id=1AFucgFgEv", "pdf_url": "https://openreview.net/pdf/312f508bd42e4adcf6621ba50018c2aac18e9ad7.pdf", "updated_at": "2026-10-05T20:04:15.296Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "kk4xPkadff", "number": 26844, "title": "Shift-Accumulate Attention: Multiplier-Free Query-Key Products for Transformer Decoding", "abstract": "Power-of-two quantisation turns a multiplication into a bit shift, but it has\nso far been applied primarily to the post-softmax attention--value product. We\nreformulate the earlier and larger score product,\n$\\mathbf{S}=\\mathbf{Q}\\mathbf{K}^{T}$, by quantising the key cache to a signed\npower-of-two fixed-point representation, so that every scalar multiplication\nbecomes a sign flip, a shift, and an integer accumulation. We call the resulting\ndatapath ShiftAtten. A fixed-point headroom condition makes every shift a\nnon-negative left shift, rendering the accumulation exact with respect to the\nquantised operands. A mantissa-extended power-of-two family then trades\nadditional shift--add operations for accuracy across 4- to 8-bit\nrepresentations, while a shift-exact online softmax extends the same treatment\nto the attention--value product.\n\nFused CUDA kernels for a 1.1\\,B-parameter Llama decoder on an RTX 4090 reach\n$4.60\\times$ the throughput of FP16 scaled dot-product attention at a 32k\ncontext length with a $2.5\\times$ smaller KV cache. We separate the gain into\n$2.53\\times$ from the narrower representation and $1.81\\times$ from kernel\nengineering. On zero-shot reasoning, the 8-bit representation achieves\n$61.61$ percent versus $61.51$ percent for FP16, while at 4 bits it loses only\n$0.94$ percentage points compared with $2.68$ points for uniform INT4,\ndespite worse perplexity and tensor-level error.\n\nA cost model calibrated by kernel disassembly shows that the remaining\nbottleneck is instruction issue rather than memory bandwidth: current GPUs\nprovide a packed four-way integer multiply--accumulate but no corresponding\npacked shift primitive, and PTX scalar shift--accumulate is emulated. We\ntherefore propose DS4A, a packed four-way shift-dot-accumulate instruction.\nUnder the same kernel structure, DS4A is projected to reach $6.51\\times$ FP16\nthroughput, compared with $4.94\\times$ for an optimised INT8 MAC kernel,\nbecause the power-of-two representation moves 101 bytes per cached token\nversus 133 bytes for INT8.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Transformer inference", "attention", "power-of-two quantisation", "fixed-point arithmetic", "KV cache", "CUDA kernels", "hardware-software co-design." ], "forum_url": "https://openreview.net/forum?id=kk4xPkadff", "pdf_url": "https://openreview.net/pdf/c0f40cfdb841b2f38b1a9af296e2c984e622c45d.pdf", "updated_at": "2026-10-05T20:04:22.355Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "8-bit", "int4", "int8", "quantisation", "quantised", "quantising" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UrRfXOR2eF", "number": 26917, "title": "Quantization Everywhere with Rotated Fourier Surrogate", "abstract": "Quantization is essential for efficient neural network deployment, whereas its optimization overwhelmingly relies on the Straight-Through Estimator (STE). However, STE discards the structure of the quantization operator by replacing its derivative with an identity mapping, resulting in a mismatch between the forward and backward passes that can fundamentally limit the achievable quality of quantized models. To address this limitation, we introduce Rotated Fourier Surrogate (RFS), a unified framework that extends STE with a principled gradient estimator. RFS analytically reflects the structure of quantization while remaining stable and computationally efficient. We further provide rigorous theoretical guarantees for RFS, demonstrating well-conditioned gradients and favorable optimization properties. Moreover, RFS is plug-and-play and incurs negligible computational overhead with our optimized kernel. We extensively evaluate RFS across diverse quantization paradigms, spanning integer and floating-point quantization, weight and activation quantization, as well as quantization-aware training (QAT), quantization-aware distillation (QAD), and post-training quantization (PTQ). Our experiments cover a broad spectrum of architectures, including large language models, diffusion language models, vision transformers, and convolutional neural networks, as well as diverse applications including generative recommendation, image classification, pre-training, and reasoning post-training. Across these diverse settings, RFS consistently improves upon strong state-of-the-art quantization baselines, demonstrating its broad applicability across quantization paradigms, model architectures, and applications.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization" ], "forum_url": "https://openreview.net/forum?id=UrRfXOR2eF", "pdf_url": "https://openreview.net/pdf/e427d4b2a267302c0c25f9f08e7fc1b89cfdf516.pdf", "updated_at": "2026-10-05T20:04:31.565Z", "categories": [ "ptq", "qat", "training", "hardware", "reasoning", "diffusion", "theory" ], "matched_terms": [ "ptq", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ywrYebhyLF", "number": 26944, "title": "Scaling Laws for Quantized Neural Computation: Geometry, Synthesis, and Execution", "abstract": "Classical empirical neural scaling laws commonly treat scaling exponents as fitted constants, limiting mechanistic prediction before costly resource sweeps are executed. We develop a first-principles theory of quantized sequential neural computation that derives joint scaling surfaces directly from independently measurable structural profiles. First, we prove that marginal depth and library-size scaling rates cannot identify their joint law because depth dynamically reuses a finite operation library as a decoder. We resolve this bottleneck by introducing the task-quotiented shared-library reachable width. For a canonical spectral residual class, we solve the finite-resource minimax problem exactly, obtaining $\\Delta_a(D,M)=\\frac{1}{2D}\\sum_{i=1}^{2D}\\max{a_i,a_{M+1}}$, where $D$ is depth, $M$ is library cardinality, and $(a_i)$ is the task-visible operation spectrum. This formula shows that depth resolves a clipped Cesàro profile rather than a primitive power law. Monotone regular variation yields a sharp summability transition in which nonsummable spectra retain spectrum-dependent rates, summable spectra enter a universal $\\Theta(D^{-1})$ synthesis regime, and critical spectra are governed by an integrated slowly varying profile. Routed nonlinear residual adapters realize these phases, while task norm and loss geometry transform the observed exponent. Furthermore, execution semantics supplies a second structural law for arithmetic transport. Under fixed-horizon residual refinement, exact Green-operator identities show that full-state write-back has $\\Theta(D)$ arithmetic amplification via backward-propagator mass, whereas first-order error feedback remains $\\Theta(1)$ via adjoint semivariation. On a frozen pretrained ResNet-18 representation, a task-visible spectrum estimated from 5,000 calibration examples predicts all 49 registered depth–library errors on untouched test data with 1.02% median and 2.49% worst relative error without fitting a held-out exponent. Holding the ideal route fixed while altering only arithmetic semantics shifts the measured depth slope from 0.90 under write-back to 0.00 under error feedback, isolating a bit-width law of $2^{-1.01b}$. Consequently, neural scaling exponents emerge not as primitive constants, but as identifiable, predictable summaries of operation geometry, finite synthesis, task geometry, and execution-error transport.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "scaling laws", "quantization", "neural computation", "residual networks", "approximation theory", "neural scaling", "error feedback", "representation learning", "deep learning" ], "forum_url": "https://openreview.net/forum?id=ywrYebhyLF", "pdf_url": "https://openreview.net/pdf/818469aca3273d5e0c7fce20b3635ebf139bf299.pdf", "updated_at": "2026-10-05T20:04:33.980Z", "categories": [ "ptq" ], "matched_terms": [ "bit-width", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fQjddUg3tg", "number": 26946, "title": "Compute Less, Capture More: Improving Sparse Attention with Mixed Precision", "abstract": "Sparse attention reduces computation by selecting which query--key interactions to evaluate, but typically computes every selected interaction at the same numerical precision. We find that this is inefficient: importance remains highly uneven within the selected set, while useful interactions are still omitted. We introduce mixed-precision sparse attention, which converts this excess precision into broader coverage. It preserves high precision for important interactions, lowers the precision of less important retained interactions, and uses the saved compute to recover interactions that the original sparse method would otherwise skip. By reusing the geometry or importance ranking of existing sparse methods, it expands coverage without a new selection mechanism or additional compute. Our analysis shows that this exchange can reduce attention-output error because broader coverage better preserves dense attention distributions and its benefit can outweigh the error introduced by lower precision. We further develop a fused kernel that executes BF16, FP8, and FP4 interactions within a single attention operation with little overhead. Across video generation and long-context language understanding, mixed-precision sparse attention consistently improves the quality--compute frontier of existing sparse attention methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sparse attention", "mixed precision", "quantization", "efficient inference", "long-context language models", "video generation", "GPU kernel optimization" ], "forum_url": "https://openreview.net/forum?id=fQjddUg3tg", "pdf_url": "https://openreview.net/pdf/59f3031468df99a216496329ac0a698da09055db.pdf", "updated_at": "2026-10-05T20:04:35.597Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "fp8", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OJwZQtHiQF", "number": 26988, "title": "Croissattn: A Complete FP4 Attention Backward", "abstract": "FP4 training speeds up the linear layers of language models, but at long context\nthe cost shifts to attention. Four of attention’s six products sit in the backward\npass but existing FP4 recipes keep all or part of it in higher precision. CroissAttn\nextends native FP4 execution to both forward and all four backward attention\nproducts. It combines stochastic rounding of gradients and saved-factor recasts\nwith a consistent softmax correction, preserving the quantized forward’s straight-\nthrough gradient in conditional expectation. An error decomposition separates\nnoise propagated through the backward pass from new product noise. Query-row-\nowned accumulation and packed storage reduce the memory required by tiled\nexecution. On long-context workloads, native B200 attention kernels achieve\n1.5–2.2× backward speedups over BF16. Training from scratch reaches 10B\ntokens with validation gaps of 0.0439 nats at 500M parameters and 0.0451 at\n1.7B. Native Qwen3-8B continued training reaches a 0.056-nat gap. Under FP4\nserving, its checkpoint achieves lower validation loss than directly cast BF16-\ntrained weights. These results connect complete FP4 attention backward to training\nquality, deployment, and long-context execution.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "backward pass", "FP4 training", "Stochastic noise control" ], "forum_url": "https://openreview.net/forum?id=OJwZQtHiQF", "pdf_url": "https://openreview.net/pdf/4b88e9ece208fc90c8053a136df684038c16aa3d.pdf", "updated_at": "2026-10-05T20:04:39.448Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BLtMieC3Ds", "number": 27012, "title": "Behavioral Capacity Certificates for Quantized Language Models", "abstract": "Activation and key--value cache precision change what a quantized language model\ncomputes without altering its stored weights. Direct weight-code bounds, however,\nassign identical complexity to deployments that behave differently and charge\nseparately for weight codes that behave identically. Behavioral Capacity\nCertificates (BCC) charge for behavior using the aggregate prior mass of complete\nimplementations---weights, scales, activation and cache rules---that induce the\nsame bounded loss. When quantization merges implementations, this shared mass\nlowers the complexity penalty, and a break-even law determines when the saving\nsurvives the cost of validating it. BCC supports a three-step deployment\nworkflow, and our experiments verify each step. First, a forward-only screen\nshortlists per-layer bit-widths by how often candidate perturbations preserve the\nreference predictions, with quality comparable to Hessian-guided selection at\nlower preprocessing cost. Second, margin-certified cells identify weights that can be pruned or\nsign-flipped without changing the deployed behavior: every permitted\ncombination preserves all declared predictions, and on OLMoE-1B-7B and\nSmolLM2-1.7B, independent probes bound the probability that any permitted\ncombination changes a prediction on new text. Third, BCC bounds the population loss of the deployed model,\nnonvacuously for complete decoders and more tightly than the\ncompressed-code route. At equal cache memory, giving keys higher\nprecision than values yields lower NLL and higher prediction\nagreement on GPT-2, Qwen2.5, and SmolLM2, together with a tighter\ncomplexity bound in the GPT-2 audit.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Weight-Activation Quantization", "Generalization Certificates", "PAC-Bayes Bounds", "Mixed-Precision Allocation" ], "forum_url": "https://openreview.net/forum?id=BLtMieC3Ds", "pdf_url": "https://openreview.net/pdf/e3f4243c39cfc51e367e3d880d834e9297f5e9cf.pdf", "updated_at": "2026-10-05T20:04:43.125Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gxTTerbmox", "number": 27016, "title": "Statistical Rank Transport for Efficient Conformal Recalibration After Model Updates", "abstract": "Model updates invalidate cached conformal scores, while full recalibration reevaluates every calibration example with the updated model. We study arbitrary fixed updates through statistical rank transport. A small sample fits a monotone map from cached reference scores to updated scores; an independent residual sample supplies a maximum-residual radius; and a contamination-aware order-statistic argument localizes the updated threshold to a contiguous block that is rescored exactly. The maximum residual induces a distribution-free beta-binomial upper reference for the number of calibration violations, including discrete residual laws, so the contamination allowance is chosen directly from $(r_E,\\delta_V)$ rather than through an intermediate tolerance parameter. The same target-$k$ construction also yields a training-conditional guarantee after adding the calibration-order margin required by $\\delta_Q$. On Llama-3-8B backbone LoRA and ViT-B/16 quantization and pruning, the marginal-efficient setting uses roughly \\ensuremath{6-18\\%} of full-calibration forwards while stale thresholds materially undercover. Against a matched-budget PAC mini split, the PAC-tuned transport setting reduces the delivered target by roughly \\ensuremath{0.3-0.6} percentage points and mean set growth by about \\ensuremath{3} percentage points in the evaluated moderate-cache regime. The advantage is not monotone in cache size: at fixed $r_E$ it peaks before boundary-block cost overtakes the statistical benefit, while scaling $r_E$ with the cache preserves the favorable regime over a wider range of $n$.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "conformal prediction", "uncertainty quantification", "recalibration", "model updates", "prediction sets", "model compression", "pruning", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=gxTTerbmox", "pdf_url": "https://openreview.net/pdf/1307deeb524d2010e9453b2d66d3297e465a4e87.pdf", "updated_at": "2026-10-05T20:04:43.503Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pZ7ntwUEBz", "number": 27055, "title": "Quantization-Aware Preconditioning for Physics-Informed Neural Networks", "abstract": "Ill-conditioned physics-informed neural networks (PINNs) residual objectives can\nslow or destabilize optimization; residual preconditioning reshapes this geometry,\nwhile quantization reduces memory and arithmetic cost. These two design choices\nare usually treated independently, even though the same residual transform acts on\nboth the physical error and the perturbations induced by low-bit weights and activa-\ntions. We introduce a quantization-aware framework for residual preconditioning in\nphysics-informed neural networks, in which the preconditioner is designed jointly\nfrom the physical tangent geometry and measured quantization-induced residual\nstatistics. The resulting quantization-aware preconditioning (QAP) framework\ncombines isotropic and covariance-aware regularization with scalable low-rank\nestimates of the quantization geometry. Our analysis separates optimization con-\ntraction from quantization injection and shows why a preconditioner that is effective\nin full precision can become ineffective after quantization. Across ten-seed end-\nto-end quantization-aware training (QAT) experiments, QAP reduces deployed\nsolution error by up to 2.07× on heterogeneous one-dimensional diffusion and\nby 1.45× on a 1000:1 two-dimensional checkerboard-diffusion problem relative\nto no preconditioning; in the two-dimensional setting, the stabilized directional\nvariant also outperforms isotropic QAP. These results identify quantization-induced\nresidual geometry as a preconditioner-design variable. On the evaluated diffusion\nproblems, QAP improves deployed solution accuracy over quantized PINNs trained\nwithout preconditioning, at the same weight and activation bitwidths", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Aware Preconditioning; Physics-Informed Neural Networks; Low-Bit Neural Networks; Neural PDE Solvers; Quantization-Aware Training" ], "forum_url": "https://openreview.net/forum?id=pZ7ntwUEBz", "pdf_url": "https://openreview.net/pdf/040f4301774ab4c05891d6e965d128bd97a70840.pdf", "updated_at": "2026-10-05T20:04:48.753Z", "categories": [ "qat", "diffusion" ], "matched_terms": [ "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jvTgvKYRDe", "number": 27578, "title": "Counterfactual Fragility Certificates for Tabular Classifications: Guarantees and Limits", "abstract": "A high-confidence tabular prediction need not survive unreliable input fields. We study protocol-relative fragility: loss of support for the original prediction under a declared finite collection of field failures. Redundant fields can mask one-step effects, while partial degradation can differ from deletion. Counterfactual Fragility Certificates (CFC) combine a sequential removal path, recomputed after every step, with graded tests of every group. An original-class margin records collapse even after the predicted class changes. A reconstructable record exposes the tests, decisions, and witnesses behind a fixed fragility dominance score (FDS). We distinguish record verification from worst-case robustness, causal recourse, and prediction correctness. A conditional greedy bound and counterexamples delineate generic search limits; a model-specific affine solver certifies minimum removal depth. We evaluate a released implementation on seven OpenML datasets, eight classifier families, and three seeds, with paired row-level outputs and reference checks. On the in-family endpoint, CFC-FDS achieves 0.7743 AUROC versus 0.7741 for prespecified, query-matched random search with the same three-term score; the paired difference is +0.0001 (95% interval [-0.0054, 0.0063]). On withheld donor/quantization mechanisms, the difference is -0.0040 ([-0.0089, 0.0004]). These comparisons do not establish a consistent advantage for adaptive removal over query-matched random search across both endpoints. A separate five-dataset restoration study measures 2.36 net errors averted per 100 cases at 10% review using a validation-trained utility ranker, versus 2.61 for raw confidence.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "tabular data", "robustness certificates", "counterfactual audit", "model fragility", "selective prediction", "trustworthy machine learning" ], "forum_url": "https://openreview.net/forum?id=jvTgvKYRDe", "pdf_url": "https://openreview.net/pdf/66148acec11b339afda39c5af785b396249f052f.pdf", "updated_at": "2026-10-05T20:06:05.269Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pyB4n82RVJ", "number": 27604, "title": "PowerStep: Memory-Efficient Adaptive Optimization via $\\ell_p$-Norm Steepest Descent", "abstract": "Adaptive optimizers such as Adam are standard for training Transformers, but storing gradient first and second moments incurs substantial memory overhead. We introduce PowerStep, a memory-efficient optimizer that achieves coordinate-wise adaptivity without storing second-moment statistics. Motivated by $\\ell_p$-norm steepest descent, PowerStep applies a signed-power transform directly to one momentum buffer. We establish a finite-horizon stationarity bound for exact, unregularized updates, with an $O(1/\\sqrt{T})$ term and a noise-dependent residual. Experiments on Transformers from 124M to 235B parameters show competitive validation quality while halving $\\texttt{fp32}$ optimizer-state memory. Combined with uniform $\\texttt{int8}$ quantization, PowerStep remains numerically stable and reduces optimizer-state memory by $\\sim8\\times$ compared to $\\texttt{fp32}$ AdamW. PowerStep thus provides a simple, memory-efficient alternative for large-scale training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Adam", "AdamW", "adaptive optimization", "l_p norm" ], "forum_url": "https://openreview.net/forum?id=pyB4n82RVJ", "pdf_url": "https://openreview.net/pdf/325cd2cd1af6306a7f387a6c5363dabe5a1054b9.pdf", "updated_at": "2026-10-05T20:06:09.540Z", "categories": [ "training" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2lYGsmAKvN", "number": 27655, "title": "SALTQ: Saliency-Allocated Low-Bit Trainability for Quantized LLM Fine-Tuning", "abstract": "Fine-tuning for deployment below 4 bits should directly produce a uniform low-bit weight-only checkpoint of integer codes and per-group quantization parameters, without an adapter branch or FP16 columns. Existing approaches fall short in different ways: adapter-based methods require post-merge re-quantization, mergeable variants freeze integer codes, weight-updating QAT computes the full weight gradient, and selective methods keep salient columns in FP16. To address this gap, we propose SALTQ, which allocates trainability by saliency within the deployed representation. A second-order analysis of the task loss shows that the columns where rounding costs the most loss are also those where a fine-tuning update moves the output the most, namely the few input columns with the largest activation energy, so that repairing the quantization error and adapting to the task can share one set of trainable columns. SALTQ makes these salient columns fully trainable at the deployment bit-width. In every other quantization group it trains only the zero-point, whose gradient pools within the group and is therefore almost free, whereas training the scales or the integer codes would require the full weight gradient. A channel permutation gathers the salient columns into complete groups; it is shared within contiguous segments of layers rather than across the whole network, which leaves one to three residual-stream reorderings at inference. Controlled comparisons show that the salient columns need to be trained but not in FP16: training them at the deployment bit-width retains nearly all of the gain of training them in FP16, with a gap of at most 0.8 points. At INT2 on Llama-2-7B and Qwen2.5-32B, SALTQ outperforms the compared mergeable and mixed-precision baselines by at least 0.7 points on the Commonsense-170k average and 3.8 points on GSM8K. At INT3, it reaches the level of the FP16 QLoRA reference on commonsense reasoning and GSM8K. It trains in about the time of QLoRA and in less than a third of the time of LR-QAT, which updates all weights and reaches a similar accuracy. The code will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-aware training; low-bit quantization; parameter-efficient fine-tuning; large language models; weight-only quantization; saliency; channel permutation" ], "forum_url": "https://openreview.net/forum?id=2lYGsmAKvN", "pdf_url": "https://openreview.net/pdf/a2015097cea2ac6dd65c43500384b5a28bbd967f.pdf", "updated_at": "2026-10-05T20:06:17.623Z", "categories": [ "ptq", "qat", "training", "hardware", "reasoning" ], "matched_terms": [ "bit-width", "int2", "int3", "low-bit", "mixed-precision", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Q4Irv7k5Ue", "number": 27762, "title": "Mixtures of Neural Operators Reduce Active Complexity in Operator Learning", "abstract": "Operator-learning systems are not governed solely by total parameter count; for one query, the relevant bottleneck can be the model that must be loaded and evaluated. We study this distinction for classical neural operators on compact Sobolev subsets through a constructive comparison between routed mixtures of neural operators (MoNOs) and a fixed single-neural-operator construction. A MoNO routes each input function through a tree to one expert. Our main theorem shows that every scalar uniformly continuous nonlinear operator with bounded output Sobolev radius on the approximation set admits a MoNO approximation whose active expert has smaller depth and width scaling, and no larger rank scaling, than the analyzed single-neural-operator construction; for Lipschitz targets these expert quantities are bounded by $\\mathcal{O}(\\varepsilon^{-1})$. A coefficient-quantization tree also gives polynomial routing work for Lipschitz targets under an explicit coefficient-access model, while storage remains exponential. For bounded $C^{1,1}$ functionals, affine experts improve the exponential storage power over constant lookup under target-independent routing. We also prove a quantitative universal approximation theorem for the underlying neural-operator architecture, with explicit dependence on compact-set diameter and modulus of continuity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture of Experts", "Neural Operators", "Distributed Computing" ], "forum_url": "https://openreview.net/forum?id=Q4Irv7k5Ue", "pdf_url": "https://openreview.net/pdf/dc72db9e1380c7fdd34ca38d4145a244320c3b2b.pdf", "updated_at": "2026-10-05T20:06:35.352Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "O045bZKD4s", "number": 27836, "title": "LoReS: Low-Resolution Selection for Faster-and-Better Diffusion Generation", "abstract": "Diffusion inference pursues two opposed goals: making one trajectory cheaper, and spending more to select a better one. What divides them is the price of evidence — selection pays only if a candidate can be judged for far less than it costs to produce — so we move selection into the cheap stage of dynamic-resolution generation. Two measurements license this: at a fixed prompt the initial noise alone moves quality across almost the whole attainable range, for generation and editing alike, and a completed low-resolution trajectory agrees with the full-resolution image its handoff produces in content and ranking. LoReS (Low-Resolution noise-trajectory Selection) batches K low-resolution trajectories, ranks previews with one or several standardized verifiers, and refines a top-T pool rather than the argmax, re-choosing the winner on full-resolution evidence refinement exposes for free. Pooling pays because the cheap ranking is informative and cheap to widen: at K=8 its top-1 holds the true best for 32.5% of prompts against 12.5% for random, and T=2,4 give 53.5% and 76.5%. On FLUX.1-dev/DrawBench200, LoReS is the only training-free accelerator we test that is faster and better than the 50-step reference, by a wide margin: at reference latency ImageReward rises 32% with all eight metrics up, where every baseline falls below the reference on ImageReward. LoReS+FreqCa holds seven of eight up at 6.24×, pure LoReS six at 9.16×, and a 0.260× point lifts all eight; at the quality end K=80, T=8 search overtakes Best-of-16 at 2.4× lower latency. Editing repeats on two judges: 7.50→7.70 at 6.26× (Qwen-Image-Edit), 6.21→6.92 at 3.52× (FLUX.1-Kontext). LoReS composes with caching, quantization, and distillation; code will be released.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Acceleration", "Noise-Trajectory Selection", "Dynamic-Resolution Generation" ], "forum_url": "https://openreview.net/forum?id=O045bZKD4s", "pdf_url": "https://openreview.net/pdf/905d8c68d71524256b7bd18bec058c3296c5a289.pdf", "updated_at": "2026-10-05T20:06:51.477Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RlWsMbYAy8", "number": 27904, "title": "TempoQuant: Timestep-Aware Calibration for Post-Training Quantization of Masked Diffusion Language Models", "abstract": "Post-training quantization (PTQ) of masked diffusion language models (MDLMs) is challenging because their activation states evolve throughout the denoising trajectory. We find that temporal activation variation alone does not reliably indicate quantization difficulty; instead, target-bit quantization vulnerability is localized across denoising states, channels, and operations. Based on this observation, we propose TempoQuant, a trajectory-aware PTQ framework that directly measures channel-wise computation distortion under joint weight--activation quantization. TempoQuant aggregates distortion across denoising states into trajectory-level vulnerability and uses it to guide quantizer adaptation, allocating greater adaptation freedom to vulnerable channels. To account for accumulated quantization errors, Transformer blocks are progressively calibrated on quantized-prefix inputs. All trajectory-dependent information is consolidated into static quantization parameters, requiring no timestep- or state-dependent quantizers at inference. Experiments on LLaDA-8B-Instruct demonstrate strong performance under both W8A8 and W4A4 quantization, with particularly substantial gains in the challenging W4A4 setting. The code is available at https://anonymous.4open.science/r/TempoQuant-1DC7/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Diffusion Language Models" ], "forum_url": "https://openreview.net/forum?id=RlWsMbYAy8", "pdf_url": "https://openreview.net/pdf/02f5f397729a323589a059d42010f8aa75ea3a72.pdf", "updated_at": "2026-10-05T20:07:00.529Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "ptq", "quantization", "quantized", "quantizer", "quantizers", "w4a4", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "69osbWvnYJ", "number": 27950, "title": "Capability Scaling-Down Laws for LLM Compression", "abstract": "LLM compression reduces inference costs and memory requirements, but selecting a method and configuration remains largely empirical because comparable resource reductions can produce different capability losses. We present a systematic investigation of capability scaling-down laws for LLM compression across pruning, quantization, and distillation. Our framework measures capability loss in mathematics, code generation, and question answering, and relates these measurements to model size, training stage, compression settings, data availability, and training exposure. We develop simple predictive relations and evaluate their accuracy, measurement efficiency, and generalization to unseen configurations and model states. The results identify response patterns that can be shared across capabilities and differences that require separate treatment. Sharing the density response across pruning levels halves the configuration measurements needed to fit a pruning predictor: on new Pythia states, on pre-registered OLMo-2 test states and under Wanda pruning, the compact relation matches a regression fitted with all measurements on math and code to within 0.020 nats per token, with coefficients refitted for each setting. Controlled distillation experiments show that the cost of heavy data reuse recurs across question-answering distributions, while the net benefit depends on the evaluation distribution. We further evaluate the decision value of these predictions by comparing numerical selection with configuration medians and fixed method priorities. Independent evaluations across two model families show that selection captures most of the available cross-method benefit for question answering within the tested candidate sets, where a fixed method priority attains the same regret, with smaller opportunities for mathematics and code. These results clarify the predictive scope of capability scaling-down laws and their use in compression method selection. Our code is publicly available at: https://anonymous.4open.science/r/scaling_down_law .", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Scaling-down law", "LLM compression", "LLM capability" ], "forum_url": "https://openreview.net/forum?id=69osbWvnYJ", "pdf_url": "https://openreview.net/pdf/4f586207d599c389839fbee28acaf71a00686616.pdf", "updated_at": "2026-10-05T20:07:06.401Z", "categories": [ "reasoning", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aS9kIpVJSW", "number": 28085, "title": "IO-SVD: Input-Output Whitened SVD for Adaptive-Rank LLM Compression", "abstract": "Large language models deliver strong performance across language and reasoning tasks, but their storage and compute costs remain major barriers to deployment in resource-constrained and latency-sensitive settings. SVD-based post-training compression offers a hardware-agnostic way to reduce model size and improve inference efficiency through low-rank factorization. However, existing methods often rely on input-only whitening spaces, homogeneous rank allocation, or loss-agnostic allocation heuristics, limiting their ability to preserve model quality under aggressive compression. We propose Input-Output Whitened SVD (IO-SVD), a post-training compression method that forms a KL-aware double-sided whitening space for model weights. Using a second-order expansion of the KL loss over the top-$K$ token probabilities, IO-SVD constructs an output-side metric that captures predictive sensitivity, while input whitening captures activation statistics. We further introduce an efficient heterogeneous rank-allocation strategy that scores whitened singular components using first-order calibration loss estimates and prunes the least sensitive components under a global budget. Inspired by prior work that combines SVD truncation with quantization, we improve hybrid SVD-quantization compression through loss-aware remapping, which selects low-rank factor rows for 8-bit quantization based on the predicted loss change incurred by quantizing them. Across LLM and VLM families, we show that double-sided whitening preserves the model's predictive output distribution more closely than input-only whitening at matched ranks. This improvement is also reflected in stronger free-form generation and higher speculative-decoding acceptance, while our inference analysis characterizes the memory and throughput benefits of low-rank compression across different serving settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model compression", "low-rank approximation", "singular value decomposition", "post-training compression", "double-sided whitening", "adaptive rank allocation", "loss-aware compression", "quantization" ], "forum_url": "https://openreview.net/forum?id=aS9kIpVJSW", "pdf_url": "https://openreview.net/pdf/d601f4a942b78eebf758f21cc3120874b3e4d279.pdf", "updated_at": "2026-10-05T20:07:30.166Z", "categories": [ "ptq", "hardware", "reasoning", "multimodal" ], "matched_terms": [ "8-bit", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "edcnWg2WAV", "number": 28098, "title": "RTAQ: Reasoning Termination-Aware Quantization for Mixture-of-Experts Architecture", "abstract": "Extreme low-bit post-training quantization is key to serving large-scale mixture-of-experts (MoE) language models, yet reasoning models quantized to around two bits often collapse on long-horizon reasoning tasks. We show that this collapse stems not from a loss of knowledge but from a failure to stop reasoning properly. When a 2-bit or near-1-bit reasoning model emits the end-of-thinking token (), it terminates in every case we observe and answers nearly as accurately as its full-precision counterpart; most errors arise from reasoning traces that never close. We trace this termination decision to a small subset of routed experts (approximately 2\\% of all routed experts) that are consistently selected at the end-of-thinking position but are largely overlooked by standard sensitivity metrics. Building on this finding, we propose RTAQ (Reasoning-Termination-Aware Quantization), a weight-only quantization pipeline with three stages: (1) constructing a bank of activation-weighted, DP-optimal scalar codebooks for every expert and bit-width; (2) performing termination-aware bit allocation, which fixes termination experts at higher precision within the bit budget and assigns the remaining bits through knapsack optimization; and (3) applying grid-preserving recalibration, which reassigns codes using GPTQ-style error compensation and rescales expert outputs in under one second per expert. At an average bit-width of 2.0 for routed experts, RTAQ preserves over 95% of the full-precision performance on long-reasoning benchmarks such as AIME.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Extreme Quantization", "Reasoning", "Mixture-of-Experts", "Post Training Quantization(PTQ)" ], "forum_url": "https://openreview.net/forum?id=edcnWg2WAV", "pdf_url": "https://openreview.net/pdf/54930f93c93b6081060291dec55cf5c152eb5d5c.pdf", "updated_at": "2026-10-05T20:07:31.866Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "1-bit", "2-bit", "bit-width", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "abstract", "summary": "RTAQ 在其观测中把极低位推理失败定位为思考轨迹未能正常结束，并识别约占 routed experts 2% 的终止相关专家。方法构造 activation-weighted 动态规划码本，为终止专家保留较高精度，通过背包分配其余专家位宽，再进行保持量化网格的误差补偿与重校准。", "contribution": "在 routed experts 平均 2.0 bit 的设置下，摘要报告在 AIME 等长推理基准保留超过 95% 的全精度性能，并称 expert 输出重缩放低于每专家一秒。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。超过 95% 指保留全精度基线的性能比例，不是 AIME 的绝对准确率；终止机制结论限定于作者观测，平均位宽也仅针对 routed experts。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "MoE reasoning 模型 routed expert 权重及终止专家的混合位宽分配", "bit_width": "routed experts 平均 2.0 bit；分析也涉及近 1 bit", "evidence": [ { "location": "官方摘要", "claim": "约 2% 的 routed experts 与 end-of-thinking 位置持续相关。" }, { "location": "官方摘要", "claim": "routed expert 平均 2.0 bit 时，在 AIME 等长推理基准保留超过 95% 全精度性能。" } ] }, { "id": "lgfSKz6HSS", "number": 28168, "title": "Never Evict Always Taper: A Static Logarithmic Bit-Width Schedule for KV Cache Quantization", "abstract": "Long-context inference relies on an ever-growing key-value (KV) cache, making precision allocation a key challenge for memory-efficient generation. We study how to distribute precision across cached tokens while preserving access to older information, without requiring online attention tracking. Under an exponential distortion surrogate and a prescribed precision floor, reverse water-filling yields an optimal allocation in which precision depends logarithmically on token importance; with power-law age weighting, this naturally produces a precision schedule that decreases with token age. A complementary perturbation analysis establishes a sufficient precision floor for preserving the highest-scoring attention key. Motivated by these results, we propose TaperKV, a token-preserving KV cache with a static, age-indexed integer-precision schedule that is compatible with a broad range of quantizers. Experiments on retrieval, language modeling, and mathematical reasoning show that TaperKV substantially reduces cache storage while maintaining strong task performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "long-context inference", "memory-efficient inference", "quantization" ], "forum_url": "https://openreview.net/forum?id=lgfSKz6HSS", "pdf_url": "https://openreview.net/pdf/2121788c3ef770ae3704a015c53283a44d436424.pdf", "updated_at": "2026-10-05T20:07:40.408Z", "categories": [ "kv-cache", "reasoning" ], "matched_terms": [ "bit-width", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7CxAFHyO1S", "number": 28216, "title": "Beyond Exact Equivalence: Rethinking Quantization Preprocessing", "abstract": "Most quantization preprocessing methods search for transformations that exactly preserve the pretrained model’s full-precision function. However, the model that matters is the one after quantization: why should optimization remain confined to equivalence before rounding? We challenge this restriction and formulate quantization preprocessing as optimization over a broader space of representations, where exact equivalence is only one possible constraint. Our framework accommodates flexible offline preprocessing, nonlinear parameter transformations, and direct weight adaptation while keeping deployment constraints explicit. We instantiate this perspective with an equivalence-preserving structured initialization followed by global quantization-aware distillation, allowing the full-precision parameters to deviate from the original model to better preserve its behavior after quantization. Across multiple model families, this relaxation consistently enhances post-quantization language-modeling quality compared to matched post-training quantization baselines, without requiring preprocessing-specific online transformations at deployment. These results establish a broader principle for low-bit compression: exact equivalence can provide a safe starting point, but should not define the space in which the quantized model is optimized.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model quantization", "quantization preprocessing", "low-bit compression", "quantization-aware distillation" ], "forum_url": "https://openreview.net/forum?id=7CxAFHyO1S", "pdf_url": "https://openreview.net/pdf/ec614e1dab9f19c553c96ef0e7b68c24e2710b5d.pdf", "updated_at": "2026-10-05T20:07:50.067Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BidbgdJ7Vv", "number": 28226, "title": "When Timing Matters: Interleaved Cross-Block Post-Training Quantization", "abstract": "Block-wise post-training quantization is making low bit-width compression of large language models increasingly feasible, especially under strict memory budgets. However, PTQ still lacks the global optimization view of quantization-aware training. While most recent PTQ work focuses on block-level improvements, cross-block variants use a moving window to capture inter-block dependencies, offering a step toward global behavior. We go further and propose Interleaved Cross-Block Quantization (ICBQ), a scheduling modification that interleaves short cross-block refinements with block-wise quantization so that errors are corrected before they propagate further, offering a new route toward more global optimization. The interleaved schedule naturally creates boundary overlap, which we analyze separately as a secondary effect. In the reported experiments, ICBQ reduces ternary-quantization perplexity relative to a matched Sequential CBQ baseline, yields finite perplexity in configurations where the baseline has severe degradation, and also transfers to 3-bit and 2-bit GPTQ.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "large language models", "weight-only quantization", "low-bit inference" ], "forum_url": "https://openreview.net/forum?id=BidbgdJ7Vv", "pdf_url": "https://openreview.net/pdf/95809cbd723d4cdeb52b4edd1b81c6898da942d2.pdf", "updated_at": "2026-10-05T20:07:50.657Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "2-bit", "3-bit", "low bit", "low-bit", "ptq", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZsDGfzVnKt", "number": 28417, "title": "ViLo-TM: Visual-Logic Mapping from Deep CNNs to Tsetlin Machines", "abstract": "Tsetlin Machines (TMs) provide explicit conjunctive rules and gradient-free adaptation, yet learning expressive visual features jointly with their discrete clauses remains difficult. We introduce ViLo-TM, a family of vision TMs that jointly learns CNN representations and visually traceable Boolean rules. Its learned spatial-evidence-to-literal (SE2L) interface maps pooled convolutional channel responses to ternary predicate states and paired literals, while continuous relaxations of literal inclusion and conjunction enable end-to-end optimization. Quantization-aware training (QAT) and export yield an integer CNN encoder and a Boolean TM in which each fired clause exposes its literal conditions and signed vote, linked to the source channel maps. The Boolean TM supports bitwise inference and multiplication-free learning through standard TM feedback (Type I/II) with the encoder and thresholds fixed. On CIFAR-10, the deployed Boolean TM exceeds the strongest compared vision-TM baseline (90.15% vs. 82.8%; initializer selected on test accuracy). On CUB-200-2011, Boolean conversion retains most of the QAT accuracy (75.92% $\\rightarrow$ 75.01%, mean of three runs), and the deployed TM exactly matches the reference Boolean clause outputs and class scores on all test images (5,794).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Tsetlin Machine", "Neuro-Symbolic Learning", "Interpretable Computer Vision", "Visual Concept Learning", "Differentiable Logic", "Discrete Inference" ], "forum_url": "https://openreview.net/forum?id=ZsDGfzVnKt", "pdf_url": "https://openreview.net/pdf/d2b5b1d67407240a7ef5be7bcc9e03fbc45492df.pdf", "updated_at": "2026-10-05T20:08:24.506Z", "categories": [ "qat", "training" ], "matched_terms": [ "qat", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LR8wVOqE28", "number": 28432, "title": "Quantized Evolution Strategies: High-precision Fine-tuning of Quantized LLMs at Low-precision Cost", "abstract": "Post-Training Quantization (PTQ) is essential for deploying Large Language Models (LLMs) on memory-constrained devices, yet it renders models static and difficult to fine-tune. Standard fine-tuning paradigms, including Reinforcement Learning (RL), fundamentally rely on backpropagation and high-precision weights to compute gradients. Thus they cannot be used on quantized models, where the parameter space is discrete and non-differentiable. While Evolution Strategies (ES) offer a backpropagation-free alternative, optimization of the quantized parameters can still fail due to vanishing or inaccurate gradient estimation. This paper introduces Quantized Evolution Strategies (QES), an optimization paradigm that performs full-parameter fine-tuning directly in the quantized space. QES is based on two innovations: (1) it integrates accumulated error feedback to preserve high-precision weight updating signals, and (2) it utilizes a stateless seed replay to reduce memory usage to low-precision inference levels. QES significantly outperforms the state-of-the-art zeroth-order fine-tuning method on a variety of tasks, making direct fine-tuning for quantized models possible. It therefore opens up the possibility for scaling up LLMs entirely in the quantized space.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "evolution strategies", "large language models", "quantization", "fine-tuning" ], "forum_url": "https://openreview.net/forum?id=LR8wVOqE28", "pdf_url": "https://openreview.net/pdf/793332f6a246b6ad1dbab4b2fdb5728804e7480c.pdf", "updated_at": "2026-10-05T20:08:26.295Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-precision", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mGHHGkYKLK", "number": 28483, "title": "PRISM: A Geometric Risk Bound for Decomposing Drift into Scale, Shape, and Head", "abstract": "A single base LLM now comes with dozens of post-training variants, quantized, LoRA-adapted, or distilled, and each has to be checked before release.\nExisting evaluations provide only a partial picture: benchmark scores and likelihood screens say that a variant has degraded, similarity scores such as CKA and SVCCA say how its features moved, and nothing connects the two.\nWe connect them with one structural fact and one design choice: the prediction head is linear, so feature geometry reaches the loss, and we compare the two feature sets through an orthogonal map, which leaves the geometry being measured unchanged.\nFrom these we derive PRISM, a closed-form upper bound on the cross-entropy risk gap between a target model and a proxy variant, and prove that it splits exactly into three measurable axes: scale, shape, and head.\nEach axis names a failure mode and where to intervene: low-bit quantization distorts shape and, at the lowest bit-widths, also inflates activation scale; quantizing the output embedding inflates the head term, which dominates the bound at high bit-widths.\nBecause the shape term is differentiable, the same geometry becomes a regularizer that curbs catastrophic forgetting more than experience replay.\nAcross two model families and five benchmarks, PRISM ranks quantized and fine-tuned variants from a single forward pass at mean Spearman above $0.8$, and eight reference sequences already recover the ranking that the risk gap itself needs over a hundred to reach.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "representation geometry", "representational similarity", "Procrustes analysis", "generalization and risk bounds", "large language models", "post-training quantization", "parameter-efficient fine-tuning", "model diagnostics", "interpretability" ], "forum_url": "https://openreview.net/forum?id=mGHHGkYKLK", "pdf_url": "https://openreview.net/pdf/88f70f95185d7be9704f1c3282eeb7bfaccbc74e.pdf", "updated_at": "2026-10-05T20:08:34.483Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Nx6znWy7IF", "number": 28542, "title": "Beyond Joint Refinement: Bounded Residual Flows for Multi-Modal Trajectory Forecasting", "abstract": "Multi-future trajectory prediction is a safety-critical bottleneck between perception and planning: an over-confident wrong-mode prediction can steer a planner into an unsafe manoeuvre. The dominant paradigm appends a refinement module to a strong backbone and trains the two jointly from random initialisation — leaving the refined output without an a-priori bound and letting the refiner lean on a learned gate that masks a poorly-behaved refiner on clean data and fails wherever its protection lapses: out-of-distribution input, closed-loop recursion, or low-bit quantisation. We therefore argue that a refiner should be judged by whether its raw correction is valid without the gate, and we construct one that is. BReF (Bounded Residual Flow) is a plug-and-play module that attaches to any frozen multi-modal predictor and refines its trajectories by integrating a learned drift field along a latent “refinement” time, kept separate from the physical trajectory time. A smoothness penalty in this latent time turns the refiner into a contraction, so its raw correction stays small and physically plausible on its own and the gate becomes near-redundant. Because the refiner is identity-initialised on the frozen backbone, the backbone acts as a verifiable performance floor that jointly-trained refiners cannot give. On nuScenes, BReF improves a strong backbone among twenty recent baselines on every multi-mode metric and on single-mode minADE while preserving the full mode set; the decisive gate-forced diagnostic — removing the gate at inference — separates refiners that are indistinguishable on displacement error by nearly three orders of magnitude, tracing the advantage to the latent-time contraction.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multi-Modal Trajectory Forecasting", "Bounded Residual Flows", "Frozen-Backbone Refinement" ], "forum_url": "https://openreview.net/forum?id=Nx6znWy7IF", "pdf_url": "https://openreview.net/pdf/87f936f38df46e6367c524da7a50c15d4dd13c06.pdf", "updated_at": "2026-10-05T20:08:42.116Z", "categories": [ "multimodal" ], "matched_terms": [ "low-bit", "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IVUGHkPXHT", "number": 28730, "title": "SPARQ: A Subspace-Preserving Mixed Precision LMO Optimizer", "abstract": "Optimizer-state memory remains a major bottleneck in large language model\ntraining. Existing low-bit state-compression methods do not explicitly account\nfor the geometric sensitivity of linear-minimization-oracle (LMO) updates.\nThis limitation is particularly important for Scion, whose spectral updates\ndepend on the singular subspaces of persistent exponential-moving-average\n(EMA) states. We study Subspace-Preserving Adaptive Representation\nQuantization (SPARQ), a Scion-oriented mixed-precision state-compression\nmethod. SPARQ stores a more accurate low-rank component of each matrix-shaped\nEMA state and compresses the residual more aggressively, while applying the\nspectral LMO to matrix-shaped parameters and the root-mean-square (RMS) LMO to\none-dimensional parameters. Our analysis shows that bounded reconstruction error\npreserves the decaying terms of the full-precision convergence bound while\nintroducing an explicit error floor. Experiments with models up to 3B parameters show that \nSPARQ remains close to full-precision Scion, improves over uniform\n4-bit integer (INT4) state storage, and reduces Scion optimizer-state memory by\nup to $6.9\\times$. By reducing optimizer-state memory while preserving training\nquality, SPARQ can facilitate the training of larger models under\nmemory-constrained settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-precision training", "training efficiency" ], "forum_url": "https://openreview.net/forum?id=IVUGHkPXHT", "pdf_url": "https://openreview.net/pdf/002b7b4a639aa69c49557f8ce7140bab38f85628.pdf", "updated_at": "2026-10-05T20:09:09.752Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "4-bit", "int4", "low-bit", "low-precision", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "agxyBt81WX", "number": 28825, "title": "Neural Language Dynamics: Short-Horizon Stability and Dissipation of Perturbations", "abstract": "We study neural language dynamics: treating the text rollout as a dynamic system, quantized by a pretrained embedding model one token at a time. We formulate this inference-time process as a discrete input–state–output system in which tokens drive an internal Markov state, read out by the embedding model. Our benchmark spans 15 model configurations, 7 languages, parallel sentence-scale FLORES+, and document-scale FineWeb/FineWeb2 trajectories. Local geometry separates sharply by readout: last-token trajectories retain large turning and approximately constant token-scale speed, whereas pooled trajectories slow with prefix length. A predictive gate accepts a reduced or delay coordinate on 267 of 280 shards. Accepted causal coordinates are moderately predictable and usually contract rapidly under a fitted output-space surrogate; pooled coordinates are far more predictable but evolve near the unit timescale, while Chinese causal trajectories form a slower, more metastable intermediate regime. Paired interventions on English FineWeb across three Qwen3-Base sizes expose the central stability distinction. Under a shared suffix, plausible and random one-token edits contract with half-lives of roughly 1.5–5.5 tokens. Once the branches generate freely, their deviation instead returns to the ordinary sampling-noise scale, so shared-input contraction does not imply synchronization under free generation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language models", "Embedding models", "Dynamic system" ], "forum_url": "https://openreview.net/forum?id=agxyBt81WX", "pdf_url": "https://openreview.net/pdf/b328b665ead6f30c9926b83c0f159deb05f1d79d.pdf", "updated_at": "2026-10-05T20:09:22.781Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "G3Bjxzn3w6", "number": 28959, "title": "Gradient Norms, Not Attention Mass: Cheap Proxies for Layer-Wise KV-Cache Bit Allocation", "abstract": "Layer-wise mixed-precision KV-cache quantization needs a per-layer sensitivity signal to decide where to spend bits. Search-based methods such as KVTuner measure damage for every (layer, bit-width) pair on calibration data; proxy-based methods replace the search with a cheap statistic. We evaluate three single-pass proxies—attention mass on outlier positions, key-quantization MSE, and the L2 norm of gradients with respect to the K/V projections—inside one matched harness on Phi-4-mini-instruct (3.8B) and Llama-3.2-1B, against our own re-implementation of the KVTuner sensitivity search as the published baseline. The gradient proxy is the only signal whose layer ranking correlates with measured 2-bit damage (Spearman ρ = 0.60 and 0.76 on the two backbones, vs. ≤0.21 for the others), and its allocations beat both activation-based proxies and the search-based allocation at the aggressive 3.25-bit budget, and beat the search on Phi and match it on Llama at 3.5 bits—with the advantage growing as the budget tightens—at two to three orders of magnitude lower calibration cost. This confirms, head-to-head and at matched allocation algorithm, recent rate-distortion analyses that claim gradient sensitivity dominates activation-based heuristics. We further document a failure mode of greedy water-filling on measured damage: when the candidate bit set extends above the budget (e.g. 8-bit candidates at a 4.0-bit budget), the allocator promotes insensitive layers to high precision and pushes sensitive layers into super-linearly damaging 2-bit territory, producing allocations worse than uniform quantization. A restricted-candidate control isolates the failure to the candidate-set width rather than the damage estimates themselves.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache", "quantization", "bit allocation", "KVTuner", "gradient sensitivity", "mixed precision" ], "forum_url": "https://openreview.net/forum?id=G3Bjxzn3w6", "pdf_url": "https://openreview.net/pdf/e7eee51290ba50e27221b4a3d3893f5eea46a561.pdf", "updated_at": "2026-10-05T20:09:45.323Z", "categories": [ "ptq", "training", "kv-cache" ], "matched_terms": [ "2-bit", "8-bit", "bit-width", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xiuvspOsro", "number": 28996, "title": "Where Precision Matters: Selection-Regret Learning for Quantized SAM 2", "abstract": "Segment Anything Model 2 (SAM 2) generates multiple mask candidates for each prompt and selects one using its predicted-IoU head, but existing quantization objectives evaluate candidates independently and do not protect this selection process. We propose a mixed-precision quantization method that jointly learns the weights and a bit width for every layer, together with a teacher-relative selection-regret objective that penalizes only the additional mask quality forfeited by the quantized selector compared with its full-precision teacher. A controlled precision sweep over video propagation units shows that hard selection accuracy degrades significantly already at eight bits, even when candidate quality remains within half a percent of full precision, and that the resulting damage increases with propagation depth. Independently, precision learned using only the segmentation loss concentrates around the same predicted-IoU interface, identifying it as a critical quantization bottleneck. At an average precision of 4.79 bits, our method achieves 71.1 SA-V J&F and 0.743 COCO mIoU on SAM 2.1 Base+, whereas two mixed-precision baselines at the same budget preserve COCO accuracy but fall below 2.4 on video. At a matched bit budget, the regret objective raises candidate-set and selected-mask quality, each with a paired interval excluding zero.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixed-precision quantization", "quantization-aware training", "vision foundation models", "video object segmentation", "mask candidate selection", "teacher-relative selection regret" ], "forum_url": "https://openreview.net/forum?id=xiuvspOsro", "pdf_url": "https://openreview.net/pdf/281408b7e4865d7d874a2b154e33d92b63f5327a.pdf", "updated_at": "2026-10-05T20:09:49.401Z", "categories": [ "qat" ], "matched_terms": [ "bit width", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "np710wGfpZ", "number": 29210, "title": "Rep-Q: Replaceability-Aware Mixed-Precision Quantization for MoE LLMs", "abstract": "Mixture-of-Experts Large Language Models (MoE-LLMs) can expand model capacity while controlling computational costs, yet their massive expert parameters still introduce substantial memory overhead. Mixed-precision quantization mitigates this issue by assigning differentiated bit widths to distinct experts. Existing approaches mostly assess expert importance based on activation frequency or loss sensitivity but largely overlook expert replaceability, i.e., whether the functionality of one expert can be fulfilled by others.\nHighly replaceable experts encode knowledge already covered by other experts, but experts with low replaceability require higher precision. \nAccordingly, this paper proposes Rep-Q, a mixed-precision quantization framework that jointly models expert replaceability and importance for quantization bit-width adjustment. \n(1) To translate the directly unmeasurable replaceability among experts into a computable metric, we design a similarity score to quantify similarities between expert output representations, establishing a quantitative foundation for modeling information dependencies across experts. \n(2) Leveraging these similarity scores, we construct a unified expert uniqueness metric by aggregating representation similarities between each expert and all remaining experts. This metric identifies single-expert replaceability from pairwise expert correlations and quantifies the irreplaceability of each expert’s information, enabling the transition from relationship-level measurements to individual-level evaluations. \n(3) Subject to quantization budget constraints, we jointly model expert quantization sensitivity and uniqueness scores. This allocation ensures experts with high importance and low replaceability are assigned higher precision.\nEvaluations across multiple datasets on Mixtral 8×7B (pure routing experts) and DeepSeek-V2-Lite (with shared experts) demonstrate that under an identical quantization budget, Rep-Q yields consistent improvements in perplexity and downstream task accuracy over existing expert-level mixed-precision quantization baselines, and exhibits strong cross-architecture generalization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture of Experts", "LLM", "Quantization", "mixed-precision quantization" ], "forum_url": "https://openreview.net/forum?id=np710wGfpZ", "pdf_url": "https://openreview.net/pdf/9f159f62b78c98325f8e554a014d7f7f7a436178.pdf", "updated_at": "2026-10-05T20:10:20.185Z", "categories": [ "theory" ], "matched_terms": [ "bit-width", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mJmUiHJyf9", "number": 29246, "title": "Beyond Modality Importance: Token Routing Attention for Multimodal LLM Quantization", "abstract": "Post-training quantization (PTQ) has become a key technique for reducing the storage and inference costs of multimodal large language models (MLLMs). Existing modality-aware PTQ methods learn independent activation smoothing scales for text and vision modalities, yet still rely on modality labels as the unit of quantization decisions, overlooking fine-grained variations in activation magnitudes across tokens within the same modality. In this paper, we propose Token Routing Attention Quantization (TRAQ), a PTQ framework that treats modality-specific activation smoothing scales as anchors and performs token-level scale routing through compatibility-based attention. Our core insight is that the optimal activation smoothing scale depends not only on global modality activation distributions but also on token-specific activation magnitudes. To this end, TRAQ treats modality-specific activation smoothing scales as prior anchors and employs a token-level attention mechanism that evaluates the compatibility between each token and each anchor, thereby assigning each token an activation smoothing scale aligned with its own activation magnitudes. On this basis, TRAQ computes layer-wise text and vision reconstruction errors and converts them into dynamic weights for the modality-specific reconstruction losses. The resulting weighted objective assigns greater emphasis to the modality with larger quantization error and is minimized to update the scale anchors. Extensive experiments show that TRAQ consistently outperforms competing PTQ methods, with gains of up to 7.70 points, while introducing no additional learnable modules and maintaining comparable inference efficiency to existing modality-aware PTQ methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multimodal Large Language Models", "Post-Training Quantization", "Token Routing Attention", "Token-Adaptive Quantization", "Activation Smoothing" ], "forum_url": "https://openreview.net/forum?id=mJmUiHJyf9", "pdf_url": "https://openreview.net/pdf/dd2659f442d07b8ff3e2b7ab04f9830ea9bdcda5.pdf", "updated_at": "2026-10-05T20:10:24.254Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pPunBTKTY3", "number": 29302, "title": "ByteBudget: Optimizing Weight–KV precision allocation for CPU inference", "abstract": "Autoregressive decoding on CPUs is memory-bandwidth-bound; generating each token requires streaming model weights, whose traffic grows with the context length. However, the quantization methods typically optimize weights or KV caches separately, and a fixed precision policy cannot account for changes in context length, hardware characteristics, or kernel efficiency. We introduce ByteBudget, a hardware-and-context-aware planner that converts a decode-latency target into a calibrated effective-traffic budget and jointly allocates precision across weight blocks and a KV cache to minimize estimated quality loss. The framework consists of three components: a format-aware CPU cost model, modular quality-loss curves supplied by off-the-shelf quantizers, and a discrete budgeted optimizer. For practical deployment, we use an efficient marginal-gain planner to produce mixed-precision weight and KV cache configurations, with a dynamic-programming oracle and lagrangian lower bound to audit solution quality. Since KV cache traffic grows with context length while weight traffic remains roughly constant, \\textsc{ByteBudget} adapts precision allocation changing deployment conditions. Our framework integrates the quantization methods with CPU-based LLM decoding kernels and allocates precision across layers to satisfy deployment-specific latency constraints, rather than a uniform bit-width.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization", "Weight-KV Precision Allocation", "CPU Infrerence for LLM", "CPU-Based LLM Decoding", "Bit-width optimization" ], "forum_url": "https://openreview.net/forum?id=pPunBTKTY3", "pdf_url": "https://openreview.net/pdf/c238e7863e0f2a2c6a7e0d6ff26ecc7daccb6b9a.pdf", "updated_at": "2026-10-05T20:10:33.905Z", "categories": [ "ptq", "training", "kv-cache", "hardware" ], "matched_terms": [ "bit-width", "mixed-precision", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JVIYRDANKn", "number": 29314, "title": "SubRot: Signed Gradient Subspace Calibration for VLM Rotation Quantization", "abstract": "Post-training quantization has emerged as an important technique for reducing the deployment cost of vision-language models (VLMs). Unlike conventional large language models, VLMs jointly process visual and textual tokens, which differ substantially in their distributions, quantities, and semantics. Existing VLM quantization methods primarily characterize such heterogeneity at either the modality or token level. However, the number, semantic roles, and visual content of tokens vary considerably across samples, making it difficult to establish stable token-level correspondences across samples. In contrast, the channel space provides a shared coordinate system across samples, making it a more natural domain in which to accumulate stable task-sensitive structures. In this work, we employ the empirical Fisher matrix as a approximation to the local curvature of the autoregressive loss with respect to intermediate activations, and perform eigendecomposition to identify a gradient-based subspace along the channel dimension. This subspace exhibits three desirable properties: cross-sample stability, clear sensitivity separation, and consistent effects along certain directions. Based on these observations, we employ a Fisher-based curvature term to suppress excessive quantization errors along sensitive directions, while using a signed first-order term to align unavoidable quantization perturbations with loss-decreasing directions. Consequently, our method goes beyond merely reducing the overall magnitude of quantization errors, further ensuring that errors within the task-sensitive subspace are both magnitude-controlled and directionally favorable. Extensive experiments demonstrate the effectiveness of our method for low-bit rotation quantization of vision-language models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Models", "Post-Training Quantization" ], "forum_url": "https://openreview.net/forum?id=JVIYRDANKn", "pdf_url": "https://openreview.net/pdf/1dfa7a68901b23859d7ef8eb46c11e11a450523a.pdf", "updated_at": "2026-10-05T20:10:34.823Z", "categories": [ "ptq", "training", "hardware", "multimodal" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Hu4rFmLvX9", "number": 29431, "title": "TetraBack: Identifying and Correcting Optimization Bias in Microscaled 4-bit Training", "abstract": "Microscaled 4-bit weight and activation quantization (W4A4) \naccelerates LLM inference, but maintaining accuracy at such \nlow precision remains challenging.\nActivation quantization is a major source of this degradation,\nyet its effects on training dynamics remain underexplored.\nIn this work, we show that this accuracy gap stems partly from \noptimization bias in the backward approximation, \nrather than solely from the limitations of low-bit representation.\nWe consistently observe sustained growth in activation magnitudes\nand pre-FFN RMSNorm weights during standard MXFP4 W4A4\nquantization-aware training (QAT), a pattern that impairs\noptimization and is absent when activations remain unquantized.\nWe trace this behavior to the conventional \nstraight-through estimator (STE) used for activation quantization, \nwhich ignores the gradient contribution from each group's scale \nand biases gradients reaching the preceding RMSNorm.\nWe propose **TetraBack** to restore this missing gradient contribution \nwithout modifying forward computation.\nOur method curbs the abnormal growth of activation magnitudes and RMSNorm weights, \nconsistently improves performance across 3B--94B MoE models, \nand reduces the loss gap induced by activation quantization by up to **41%**.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Aware Training", "W4A4 Quantization", "MXFP4", "Microscaling", "Activation Quantization" ], "forum_url": "https://openreview.net/forum?id=Hu4rFmLvX9", "pdf_url": "https://openreview.net/pdf/0c9543e6158fb7efa8e9774d4d6305fb23d67e61.pdf", "updated_at": "2026-10-05T20:10:54.969Z", "categories": [ "qat", "training" ], "matched_terms": [ "4-bit", "low precision", "low-bit", "mxfp4", "qat", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xoueBlu031", "number": 29470, "title": "SyRA-DiT: Synergistic Reconstruction-Aware Rotation for W4A4 Diffusion Transformers", "abstract": "Post-training quantization (PTQ) enables low-precision deployment of pretrained\nDiffusion Transformers without retraining, yet aggressive W4A4 quantization\nremains challenging. Coarsely quantized weights are reused across denoising\ntimesteps, while activation distributions vary over time, making layer reconstruction error difficult to minimize consistently throughout the denoising process. Group-wise quantization improves local range adaptation, but but its quantization fidelity remains highly sensitive to the within-group distribution. To this end, we propose SyRA-DiT, a reconstruction-driven PTQ framework that learns an orthogonal rotation to jointly reorganize weight and activation quantization groups across denoising timesteps, directly minimizing W4A4 layer reconstruction error on a limited calibration set via Cayley parameterization and straight-through estimation. Furthermore, SyRA-DiT uses the same per-timestep incoherence weights to construct the initialization covariance, the W4A4 reconstruction objective, and the GPTQ Hessian. The rotation is optimized to minimize the relative W4A4 reconstruction error, after which GPTQ quantizes the rotated weights using a Hessian constructed under the same temporal weighting. We evaluate SyRA-DiT on image, text-to-image, and video diffusion models under 4-bit weight and activation quantization, where it achieves state-of-the-art visual quality and quantitative performance. Notably, on ImageNet at $256{\\times}256$ resolution with DiT-XL/2, SyRA-DiT remains competitive with previously reported W4A8 results across diverse sampling and guidance settings, demonstrating that PTQ remains effective even in the challenging W4A4 regime.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Diffusion Transformer", "learnable rotation", "reconstruction" ], "forum_url": "https://openreview.net/forum?id=xoueBlu031", "pdf_url": "https://openreview.net/pdf/593c78e0484ff0d5b342ec23b14c6588b327de56.pdf", "updated_at": "2026-10-05T20:11:00.078Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "low-precision", "ptq", "quantization", "quantized", "quantizes", "w4a4", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Za7lxmQR73", "number": 29478, "title": "Adaptive Precision Allocation Can Starve Itself: Consequence-Aware Allocation for Bounded World Models", "abstract": "A recurrent world model operating under a hard per-transition representation\nbudget must decide where to spend its available precision. We show that\nadaptive precision allocation can starve itself: when source-referred\nimportance is repeatedly measured through the precision assignment it is used\nto update, low precision attenuates the very signal needed for a group to\nrecover rate. We formalize this reference-point failure through\nmulti-step predictive consequence---how strongly current representation\nerror propagates into future latent predictions under the same action\ncontinuation---and show that the resulting feedback admits locally absorbing\nstarvation fixed points. Measuring all groups from a common\nallocation-independent reference removes the assignment-dependent gain.\nBuilding on this analysis, we introduce CAP-WM, which predicts\nstate-dependent multi-step consequence and converts it into an exact recurrent\nprecision allocation by reverse water filling at a no-bypass interface. A\nlinear--Gaussian characterization recovers the same groupwise allocation rule\nunder a standardized group-block-isotropic specialization. In trained models,\nadaptive source-referred refresh reaches $\\leq4$ active groups from ordinary\ninitializations with probability $0.74$--$0.91$, versus $0.06$--$0.11$ for\nmemory-referred refresh. Under a directly countable serialized interface,\nCAP-WM reaches $90\\%$ retention of the no-channel DreamerV3 reference at\n$4$ bits/coordinate, while the tested adaptive controls that reach this\ncriterion require $6$; at the tight nominal rate, CAP-WM improves DMC-GB2 IQM\nby $+88\\,[55,119]$ over Uniform and $+24\\,[8,40]$ over One-step consequence.\nThe additional multi-step benefit is conditional rather than universal,\nresolving when precision is restrictive and predictive dependence extends\nbeyond one latent transition. More broadly, the results expose a design hazard\nfor adaptive resource allocation: an assignment can become self-reinforcing\nwhen it attenuates the importance signal used to revise itself.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "world models", "adaptive resource allocation", "representation learning", "rate-distortion", "predictive consequence", "mixed-precision representation", "model-based reinforcement learning" ], "forum_url": "https://openreview.net/forum?id=Za7lxmQR73", "pdf_url": "https://openreview.net/pdf/a11051a17dd934ace4e41a72939c7d329062e0b8.pdf", "updated_at": "2026-10-05T20:11:00.809Z", "categories": [], "matched_terms": [ "low precision", "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "J28Jfjqb4P", "number": 29724, "title": "AB-Sparse: Sparse Attention with Adaptive Block Size for Accurate and Efficient Long-Context Inference", "abstract": "Block sparse attention makes long-context decoding affordable by partitioning the KV cache into fixed-size blocks and loading only the blocks a query needs, and it has become the most deployable form of sparse attention because it leaves paged KV cache management untouched.\nIts weakness is accuracy: at a 4% KV budget, state-of-the-art methods trail full attention by up to 10 points on standard long context benchmarks.\nWe trace a large part of this gap to a single design choice, the use of one block size for every attention head.\nMeasuring attention recall head by head, we find that heads differ widely in their sensitivity to block granularity: some keep near-perfect recall, others lose over 90% of it when blocks grow from 16 to 32 tokens, and no single size serves both.\nWe present AB-Sparse, a training-free algorithm--system co-design that gives each head its own block size.\nA lightweight one-time calibration assigns block sizes that transfer across domains; INT4 per-channel asymmetric quantization of block centroids, which only rank blocks and never enter the attention output, keeps centroid memory below that of the uniform baseline; and custom GPU kernels batch heads with heterogeneous block counts without padding and map variable-size blocks onto standard paged attention.\nOn Llama-3.1-8B, Qwen3-8B, and Qwen3-32B, AB-Sparse raises the accuracy of Quest and ArkVale by 3.0--3.7 points on RULER and 1.9--2.6 points on LongBench on average (up to 5.4 points), recovering 27--85% of the gap to full attention, with comparable or higher decoding throughput.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language model", "long context inference", "attention" ], "forum_url": "https://openreview.net/forum?id=J28Jfjqb4P", "pdf_url": "https://openreview.net/pdf/8fd479084418916b8dce524eb92fc8693ae5efef.pdf", "updated_at": "2026-10-05T20:11:39.906Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "juv4oFeB0T", "number": 29739, "title": "GAP: Generation-Aligned Paired Fidelity Evaluation for GGUF-Quantized VLMs", "abstract": "Reliably distinguishing small fidelity differences between quantized candidates is challenging during quantization development. We introduce GAP, a Generation-Aligned Paired fidelity evaluation protocol for GGUF-quantized vision-language models. GAP generates image-conditioned responses with the full-precision reference, teacher-forces all candidates on these shared responses, and applies item-level paired tests to their fidelity divergence (e.g., KLD) from the reference to detect small fidelity differences with quantified uncertainty. A pilot-based procedure selects mode-specific token caps to reduce per-item evaluation cost, enabling larger samples and greater statistical power within a fixed budget. Across 126 quantized checkpoints near the Q4 regime, six benchmarks, and 15 model-mode configurations, the median paired-to-unpaired SNR ratio is 3.7 in thinking mode and 2.0 in non-thinking mode for candidate pairs with relative fidelity gaps below 5%; pairing reduces the estimated median sample size per candidate required for 80% power by approximately 15.5× relative to an independent-item unpaired design for such pairs in thinking mode, whereas close non-thinking pairs still need an estimated 11,000 items, beyond our budget. Fidelity rankings can also differ between thinking and non-thinking modes, while responses generated by a different reference model can change verdicts, including statistically significant reversals. These findings support evaluating candidates in their intended inference mode using responses from their corresponding reference model. We release the associated artifacts on Hugging Face Hub: https://huggingface.co/datasets/Anonymous778862/anony778862cr.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "evaluation", "benchmark", "quantization", "vision language models" ], "forum_url": "https://openreview.net/forum?id=juv4oFeB0T", "pdf_url": "https://openreview.net/pdf/4de280da97a4f089d8404ec7236aa92c1b78e922.pdf", "updated_at": "2026-10-05T20:11:42.155Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hP1yMU1rEj", "number": 29780, "title": "Ternova: Adaptive Low-Rank Post-Training Ternarization for Large Language Models", "abstract": "Large language models (LLMs) have achieved remarkable capabilities, yet their massive memory footprint and inference cost hinder efficient deployment. Ternary post-training quantization (PTQ) offers an attractive solution by combining compact storage with richer representation capacity than binary quantization. However, existing ternary PTQ methods remain challenged by limited ternary representation capacity, heterogeneous structural sensitivity across Transformer modules, and the mismatch between local reconstruction and model-level objectives. To address these challenges, we propose Ternova, an adaptive low-rank post-training ternarization framework for LLMs. Ternova is built on three core components: 1) Hessian-Aware Annealed Ternary Factorization (HATF) constructs high-fidelity low-rank ternary representations through sensitivity-aware factorization and annealed discrete refinement; 2) Sensitivity-Reweighted Mixed-Rank Allocation (SRMA) adaptively distributes rank capacity according to structural irreplaceability, task sensitivity, and storage cost; and 3) Task-Aligned Ternary Refinement (TATR) further aligns the compressed model with task-level behavior through global scale calibration and sparse topology refinement. Ternova preserves a compact ternary deployment form without additional dense high-precision branches. Extensive experiments across diverse LLM families demonstrate that Ternova\nachieves competitive or superior performance over existing low-bit PTQ methods\nat only 1.59 BPW while retaining competitive reasoning capabilities. On Qwen3.5-27B, Ternova achieves 74.62\\% average zero-shot accuracy, approaching the 75.97\\% BF16 baseline. Moreover, Ternova achieves roughly 3$\\times$ inference speedup and up to 80.6\\% memory savings over BF16, demonstrating a strong balance among accuracy, compression, and practical efficiency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "ternarization", "LLMs" ], "forum_url": "https://openreview.net/forum?id=hP1yMU1rEj", "pdf_url": "https://openreview.net/pdf/b2bf5deb1802c327f091a6c7a033ccf11a903a58.pdf", "updated_at": "2026-10-05T20:11:48.494Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "binary", "low-bit", "ptq", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AhbiH9yiFo", "number": 29815, "title": "Blaze: Block-Level Adaptive KV Quantization Guided by Attention", "abstract": "The KV cache is the memory bottleneck of long-context inference and grows with sequence length. Attention concentrates on a few positions, yet compression methods treat every token alike or evict tokens irreversibly. We present Blaze, a block-level adaptive KV cache quantization scheme that derives the precision of each block from the attention its positions receive. During prefill, Blaze marks a position closed when its attention stays below a threshold that decays with layer depth, since deeper layers attend to fewer positions, and it requires agreement across consecutive layers, so single-layer fluctuations do not drive the decision. Blocks are stored in INT8 or INT4 according to how many of their positions closed, blocks that carry long-distance positions keep the higher precision, and every decision is fixed during prefill, so decoding reads a lookup table and adds $O(1)$ work per step. Across five models, Blaze compresses the KV cache by 70--71\\% at 4K and 8K prefixes at a perplexity within 1.6\\% of FP16, and it stays within 2 points of FP16 on eight downstream tasks and hard four-needle retrieval, where uniform INT4 falls 6 to 14 points behind. Blaze lowers the memory barrier for long-context inference by assigning precision to blocks instead of discarding tokens.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "adaptive quantization", "cross-layer attention voting", "long-context inference", "large language models" ], "forum_url": "https://openreview.net/forum?id=AhbiH9yiFo", "pdf_url": "https://openreview.net/pdf/c498165bf72f54cce2aa737740fa4fc23b415aa6.pdf", "updated_at": "2026-10-05T20:11:53.496Z", "categories": [ "kv-cache" ], "matched_terms": [ "int4", "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IwvIHWajmN", "number": 29852, "title": "REVQ: Rotation-Enhanced Product Vector Quantization with Low-Rank Residual Compensation", "abstract": "Joint low-bit quantization of weight and activation could substantially reduce the storage and computational costs of Transformer-based architectures and convolutional neural networks (CNNs). However, existing methods cannot simultaneously minimize the layer-wise local quantization errors caused by joint misalignment between weight and activation distributions and address accumulation of errors across layers. In this paper, we formulate the cross-layer and local-layer errors for joint weight and activation quantization, and consequently, propose a rotation-based two-stage framework named REVQ for low-bit quantization to simultaneously address the errors. Specifically, we develop learnable groupwise rotation transformations parametrized by the Lie algebra of the special orthogonal group SO(g) to learn the coordinate system of parameter (weight and activation) space for achieving product vector quantization and reducing local-layer quantization error. Moreover, we design cross-layer low-rank residual compensation to bridge rotation-based quantization in adjacent linear layers with learned low-rank compensation via Haar wavelets for minimizing cross-layer quantization error. Extensive experiments demonstrate that REVQ achieves state-of-the-art performance in various tasks of image classification on ImageNet and CIFAR-100, object detection and instance segmentation on COCO, and language modeling on Wikipedia using a wide range of Transformer-based architectures and CNNs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model quantization", "quantization-aware training", "vector quantization", "learnable rotations", "low-rank residual compensation" ], "forum_url": "https://openreview.net/forum?id=IwvIHWajmN", "pdf_url": "https://openreview.net/pdf/68f7ab3cfdeafac125c89276cb629f22606e7042.pdf", "updated_at": "2026-10-05T20:11:59.548Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6EUV1IdU9C", "number": 29883, "title": "FROM PREDICTABLE DISCREPANCY TO RESIDUAL VARIANCE: CORRECTING QUANTIZED DDIM SAMPLING", "abstract": "Post-training quantization (PTQ) reduces diffusion-model inference cost, but denoiser-output discrepancies perturb subsequent sampling states. We introduce State-Conditioned Predictive Error Correction (SPEC), a two-stage output corrector for a fixed quantized denoiser under DDIM-family sampling. At sparse refresh steps, its temporal branch learns an output-matching offset for the quantized denoiser's time condition from full-precision supervision on the same latent state; the DDIM grid and transition coefficients remain nominal. A full-resolution branch is then trained on trajectories recollected with this adjustment active to predict the remaining same-state output discrepancy. On held-out trajectories, the Predictable Error Ratio (PER) shows that SPEC reduces the summed squared same-state output discrepancy between the quantized and full-precision denoisers by 82.7\\%--88.1\\% across the evaluated bitwidths. For stochastic DDIM, Student-\\(t\\)-estimated variance reallocation (\\(t\\)-VR) models the remaining post-SPEC residual variance and reallocates the nominal stochastic variance between its estimated contribution and the explicit Gaussian innovation. With 100-step stochastic DDIM on CIFAR-10 at W4A8, FID falls from 6.84 for Q-Diffusion to 5.83 for SPEC and 5.66 with \\(t\\)-VR.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "Post-Training Quantization", "Low-Bit Quantization", "Model Compression", "Efficient Deep Learning", "Quantization Error Correction", "Generative Models", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=6EUV1IdU9C", "pdf_url": "https://openreview.net/pdf/fca0a9e54995d2de9e8e37861bec5bda536a0235.pdf", "updated_at": "2026-10-05T20:12:04.359Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RhoXdIJKzZ", "number": 30130, "title": "QuantaSpike: Short-Window Spike-Driven Quantization for Large Language Models", "abstract": "Large language models (LLMs) achieve strong performance across many tasks but rely on dense multiply-accumulate (MAC) operations during inference, resulting in high energy cost.\nSpiking neural networks (SNNs) offer an event-driven alternative in which synaptic integration uses lightweight accumulation.\nHowever, spike-driven LLM inference remains difficult because outlier-heavy activations typically require long firing windows or auxiliary non-spiking paths.\nWe propose QuantaSpike, a short-window spike-driven quantization framework for LLMs built around Logarithmic Ternary Integrate-and-Fire (LTIF) neurons.\nLTIF uses ternary events with power-of-two membrane-response quanta, improving the information represented by each firing step while retaining shift-ACC-compatible computation. \nQuantaSpike combines this neuron with group-adaptive gain and selective outlier admission: normal values use residual LTIF steps, whereas admitted outliers receive one additional onset spike before entering the same residual dynamics.\nAcross OPT and Llama-2, QuantaSpike achieves state-of-the-art or competitive perplexity and zero-shot accuracy among spike-driven LLM quantization methods. It also transfers to newer dense LLMs, remaining close to the FP16 reference on Llama-3-8B and Qwen3-8B under the same four-step firing window.\nAnalytical linear-energy projections show that QuantaSpike reduces the energy of one linear transformation by about $80.0\\%$ on OPT models and $67.1\\%$ on Llama-2 models relative to SpikeQuant, providing an accurate and energy-efficient spike-driven path for LLM inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Large Language Models", "Activation Quantization", "Energy Efficient" ], "forum_url": "https://openreview.net/forum?id=RhoXdIJKzZ", "pdf_url": "https://openreview.net/pdf/e78f0ab6f218f6bfa2ded28bd48c5993a6357aa5.pdf", "updated_at": "2026-10-05T20:12:42.399Z", "categories": [], "matched_terms": [ "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VDdXI0OSRx", "number": 30134, "title": "SparRotQuant: Patterned Sparsity-Aware Rotation-Based Quantization for LLM Compression", "abstract": "Low-bit quantization and semi-structured N:M sparsity are two hardware-friendly techniques for efficient large language model deployment. \nHowever, their direct combination suffers from a structural conflict: rotation-based quantization smooths weight distributions to improve low-bit representations, whereas N:M sparsity relies on strong local contrast within each weight group to determine a stable keep/prune pattern. Consequently, applying N:M pruning after rotation produces unstable sparse masks and significant quality degradation. \nWe propose SparRotQuant, a sparsity-aware rotation-based quantization framework that incorporates the N:M sparsity constraint into calibration, aligning the learned sparse support with the rotated low-bit representation used for deployment. \nTo guide this process, we introduce Probability-aware Balanced Group Separation (PBGS), a group-level criterion that combines retention probability with Hessian-aware sensitivity to encourage each weight group to form a reliable sparse pattern before hard projection. \nWith frozen pretrained weights, SparRotQuant learns differentiable sparse masks during calibration while adapting rotations and quantization parameters to the resulting sparse low-bit representation. \nExperiments across multiple model families show that SparRotQuant maintains strong model quality under W4A4KV16 and W4A4KV4 while preserving the efficiency benefits of sparse low-bit execution.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "GPU acceleration", "hardware-aware optimization" ], "forum_url": "https://openreview.net/forum?id=VDdXI0OSRx", "pdf_url": "https://openreview.net/pdf/d1c67fd5c8b5b3f015fcab9c66de09a8b21a7e4b.pdf", "updated_at": "2026-10-05T20:12:42.924Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization", "w4a4kv16", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SycfZ4UmJD", "number": 30225, "title": "BTT: Quantizing Diffusion Language Models via Block-wise Token Transformations", "abstract": "Low-bit activation quantization remains challenging for diffusion\nlanguage models (dLLMs), whose activations vary across token positions\nand denoising states. Existing channel transformations reshape feature\ncoordinates within tokens, leaving the token dimension available for\ncomplementary transformations. We propose \\textbf{Block-wise Token\nTransformation (BTT)} for linear-input activation quantization in dLLMs.\nBTT applies randomized Hadamard transforms within blocks of\nsimultaneously processed tokens before activation quantization and\nrestores token coordinates after the linear projection. This\nconstruction preserves the linear function in exact arithmetic\nwithout quantization, while redistributing activations presented to\nthe low-bit quantizer. The block-wise design enables efficient\ncomputation and composes naturally with fixed channel rotations.\nBecause token transformations do not benefit all operators equally,\nwe select them independently for each linear operator by minimizing\noutput reconstruction loss on shared calibration inputs with fixed\nquantized weights, retaining the identity when no candidate improves\nreconstruction. Experiments show that adding BTT to QuaRot and\nDuQuant improves downstream accuracy under W4A4 quantization.\nThese results demonstrate the potential of token-dimension\ntransformations as a complement to channel transformations for\nlow-bit dLLM quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Language Models", "Post-Training Quantization", "Low-Bit Quantization", "Token Mixing", "Orthogonal Transformations" ], "forum_url": "https://openreview.net/forum?id=SycfZ4UmJD", "pdf_url": "https://openreview.net/pdf/79ff2781577af52ac3c379d675845686c9e82397.pdf", "updated_at": "2026-10-05T20:12:57.254Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizer", "quantizing", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XBZnOHUDxC", "number": 30268, "title": "HayateQuant: More Geometry per Bit with Folded Rotations for KV Cache Quantization", "abstract": "Compressing the key--value (KV) cache can substantially reduce memory traffic during language-model decoding. However, improved quantization geometry is useful only if it can be realized with low encoding and serving overhead. We investigate whether short-block quantization can improve a data-free cache codec while preserving a compact, fixed-rate representation. We introduce HayateQuant (HQ), which replaces independent scalar assignments with compact joint codebooks. A folded rotation promotes uniform within-block directions, while precomputed candidate lists keep online assignment efficient. To account for the distinct effects of key and value quantization errors, HQ further employs role-specific scaling and key and value bit allocation. Across diverse model families, HayateQuant achieves a favorable quality-compression tradeoff, particularly at moderate rates. Its optimized packed serving implementation also improves decoding throughput over the existing baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "kv cashe quantization" ], "forum_url": "https://openreview.net/forum?id=XBZnOHUDxC", "pdf_url": "https://openreview.net/pdf/20cca174e4d76499ae78bd1e851bc8d4e6f3a656.pdf", "updated_at": "2026-10-05T20:13:02.496Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZCJ8N4y9BC", "number": 30279, "title": "PSG: Preserving Activation Structure for Low-Bit Image Restoration", "abstract": "Post-training quantization (PTQ) is a practical route to deploying image restoration models, yet restoration fidelity degrades sharply at ultra-low bit widths. A key challenge is channel-wise outliers, which are particularly pronounced in Mamba-based models and substantially inflate the activation range. Suppressing these extremes can tighten the range, but is not a reliable remedy, as even limited removal can severely degrade reconstruction quality. We show that these outliers are not random spikes, but exhibit strong persistence: their dominant direction remains stable across inputs, and even within each input, the response along this direction is effectively shared across tokens. We term this shared directional component the pedestal. Crucially, the pedestal not only inflates the range, but also defines a direction highly sensitive to quantization error, making it a structure that should be preserved rather than suppressed. Building on this, we propose PSG, a PTQ framework that combines Pedestal Shifting and GRILL. Pedestal Shifting keeps the pedestal outside the quantizer while reducing the activation range, and GRILL calibrates quantization bounds for the residual activations and weights through closed-form updates. Across diverse image restoration tasks and architectures, PSG consistently outperforms existing methods, delivering up to $3.73$ dB higher PSNR at 2 bits, $4.33\\times$ smaller models, and $6.13\\times$ fewer bit-operations, while deploying as standard TensorRT engines with no additional inference overhead. Our code is available at: https://anonymous.4open.science/r/PSG-AD60.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training Quantization", "Image Restoration" ], "forum_url": "https://openreview.net/forum?id=ZCJ8N4y9BC", "pdf_url": "https://openreview.net/pdf/9bbf78f1e62420514108cc9e2d59041f586cc7a6.pdf", "updated_at": "2026-10-05T20:13:05.761Z", "categories": [ "ptq" ], "matched_terms": [ "low bit", "low-bit", "ptq", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "V1BxlV2FAC", "number": 30291, "title": "From Jump Degradation to Warped-Space Quantization", "abstract": "The growing parameter scale of large language models imposes substantial memory and computational costs, making weight quantization an important approach\nto efficient deployment. Quantization represents continuous weights with a limited set of discrete levels, and its effectiveness depends increasingly on how this\nlimited precision is allocated as the bit width decreases. Existing methods improve\nthis allocation through optimized rounding, scaling, codebooks, or transforms,\nbut they are commonly designed using static statistics collected from weights or\ncalibration data. During model execution, however, quantization error evolves\ndynamically: fixed weight residuals interact with changing activations that may\nalready contain perturbations propagated from upstream layers. To characterize\nthis process, we progressively replace full-precision weights with their quantized\ncounterparts and compare the resulting intermediate activations with those of the\noriginal model. We observe jump degradation, in which activation error rises\nabruptly at a small number of layers and continues to amplify downstream. Further analysis identifies the mismatch between equally spaced quantization levels\nand non-uniform weight distributions as one controllable source of the residual involved in this process. We therefore propose Warped-Space Quantization (WSQ),\nwhich maps weights into a flatter space for uniform sampling and applies the inverse map to construct distribution-matched non-uniform levels. By allocating\nmore of the limited precision to high-density regions, WSQ reduces the inputindependent weight residual before it interacts with changing activations. Experiments across multiple models and benchmarks show that WSQ consistently\nimproves quantized-model quality and outperforms the compared baselines under\ndirectly comparable low-bit quantization settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "ML: Efficient", "Edge", "Green & Hardware-aware ML" ], "forum_url": "https://openreview.net/forum?id=V1BxlV2FAC", "pdf_url": "https://openreview.net/pdf/141bcb5a1752df63a044c57984adc4b45e88b7a6.pdf", "updated_at": "2026-10-05T20:13:07.468Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "bit width", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pBIUl16cPA", "number": 30295, "title": "Disaggregated Quantization: Specializing LLM Prefill and Decode", "abstract": "The prefill and decode phases of LLM inference reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce memory traffic during generation. We propose \"disaggregated quantization\" (DQ), which specializes computational formats, weights and storage placement to both of these phases. Starting with format disaggregation, removing activation quantization specifically on decode improves accuracy on decode-heavy tasks without increasing inference cost. Moving to full disaggregation, training separate compute-native prefill weights accelerates prompt processing relative to weight-only inference while exceeding weight-only accuracy on both decode-heavy and prefill-heavy tasks for 2-3-bit quantization. To accommodate the extra prefill checkpoint on a single device, offloaded disaggregated prefill (ODP) streams its weights from SSD, amortizing loading over prompt length. Using released Qwen3.8-27B GGUF decoders, training an auxiliary NVFP4 prefill model in a fully-disaggregated fashion improves 1-bit accuracy by 32.5 points on MMLU-Pro and 35.3 on MMMU-Pro, without modifying the decode checkpoint. Combined with ODP, it delivers a 1.78x prefill speedup over the weight-only baseline at 8K prompt length in llama.cpp at no device memory overhead. Additional PTQ validation includes models with up to 2.8T parameters.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "llm", "compression", "disaggregated", "quantization", "inference", "serving", "gpu", "cuda", "blackwell", "prefill", "decode", "vllm", "distillation", "offloading" ], "forum_url": "https://openreview.net/forum?id=pBIUl16cPA", "pdf_url": "https://openreview.net/pdf/eb63c86fa03b7f9f6b7043f958d5964631858dae.pdf", "updated_at": "2026-10-05T20:13:08.786Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "1-bit", "3-bit", "low-precision", "nvfp4", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DnAsidFUmk", "number": 30302, "title": "Quantize Once, Adapt in Place: Zeroth-Order Optimization with Detectable Signals", "abstract": "Quantization reduces the memory required to store and deploy large language\n models, while zeroth-order (ZO) optimization avoids storing activations for\n backpropagation. Combining these two advantages still leaves a key question: \n which parameters should be optimized? We \n therefore study a simple constraint: keep the packed weights and quantization \n metadata fixed, and fine-tune only the floating-point model tensors retained \n in the checkpoint. We collect these tensors into the Quantization-Preserved Subspace (QPS) \n and propose Quantization-Preserved Zeroth-order Optimization (QPZO), which applies\n two-point ZO updates only in QPS. This enables adaptation without adding\n trainable modules or dequantizing the low-bit backbone for optimization.\n We support QPZO with theoretical analysis and empirical evidence. Specifically, \n (1) we show that an updateable output head retained in QPS yields a nonzero projected \n gradient whenever its aggregate gradient is nonzero, establishing the presence of \n task-relevant signal; and (2) because ZO accesses this signal only through finite forward-loss \n differences, we introduce Signal-to-Error Ratio (SER) to compare directional task-signal energy with \n measurement-error energy. Under the stated error bound, we prove that SER above one \n is the sharp threshold for guaranteeing positive expected alignment, \n characterizing when finite forward queries robustly preserve a useful task direction.\n Our theoretical analysis and extensive experiments demonstrate that QPZO offers a simple \n approach that substantially reduces GPU memory usage while maintaining highly competitive \n performance and stability.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantized Large Language Models", "Zeroth-Order Optimization", "Low-Bit Quantization", "Memory-Efficient Fine-Tuning" ], "forum_url": "https://openreview.net/forum?id=DnAsidFUmk", "pdf_url": "https://openreview.net/pdf/cb42d0bf243d6a89ae951d1e248bfa0e60f92ce4.pdf", "updated_at": "2026-10-05T20:13:09.969Z", "categories": [ "training", "hardware", "theory" ], "matched_terms": [ "low-bit", "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "byJQ7BNyPd", "number": 30335, "title": "Easy-Kuant: Diagnosing and Mitigating KV-Cache Quantization Collapse with Selective Layer Protection", "abstract": "Serving large language models increasingly relies on quantizing the key–value (KV) cache to hardware-native low-bit formats such as NVFP4 and FP8, widely reported to be near-lossless at four bits. On the Qwen2.5 family and its derivatives the same one-line switch is catastrophic: in vLLM, Qwen2.5-7B-Instruct falls from 90.2% to 11.0% on GSM8K under NVFP4 and to 2.2% under FP8, and vision-language checkpoints fall to 0%. Rather than tuning against cache reconstruction error, as prior methods do, we ask *where* the error becomes a failure. Replaying each layer's attention from clean activations shows that cache error is flat across depth while attention-output error varies by an order of magnitude, enters almost entirely through keys, and concentrates in one to four layers; two-directional interventions show that these layers are sufficient both to cause and to repair the collapse. The diagnosis becomes Easy-Kuant: a few forward passes yield one statistic per layer, a fixed threshold turns it into the list of layers to keep in BF16, and stock vLLM applies the list through an existing argument—no kernel, no search, no gradients. On the native serving path Easy-Kuant returns NVFP4 to 90–96% of BF16 accuracy on the 3B and 7B checkpoints and FP8 to 98–100%, transfers unchanged to long-context and multimodal checkpoints, matches software baselines that cannot run in vLLM, and costs 0.94× the per-token KV processing of full NVFP4 at 85% of its cache capacity. An optional quantization-aware stage, trained on the model's own rollouts and guided by the same layer list, adds up to 3.7 further points where a gap remains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "efficient LLM inference" ], "forum_url": "https://openreview.net/forum?id=byJQ7BNyPd", "pdf_url": "https://openreview.net/pdf/9380d9d9cd5fd96f402adad182c4c7d20fd4c5a1.pdf", "updated_at": "2026-10-05T20:13:22.169Z", "categories": [ "ptq", "qat", "training", "kv-cache", "hardware", "multimodal" ], "matched_terms": [ "fp8", "low-bit", "nvfp4", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NFUiPrlPj0", "number": 30388, "title": "The Price of Waiting: Budgeted Online Routing for Latency-Tiered LLM Services", "abstract": "Interactive large language model (LLM) applications must trade off response quality, latency, and the cost of premium compute. A routing platform with a limited budget must decide which LLM to query and when to pay for faster delivery, all while learning from feedback that may be arbitrarily delayed or censored. We formulate this challenge as an online routing problem with immediate and delayed service modes under a hard premium budget. For a given LLM, both modes share the same response-quality distribution, but delayed service postpones feedback and discounts the response value. To address this, we bound the remaining value of any pending request using its current age and a known discount kernel. By combining these envelopes with completed observations, we construct confidence intervals that accommodate arbitrary reward--delay dependence without requiring known delay distributions or bounded delays. We propose \\texttt{DC-LADA} for cost-penalized objectives and \\texttt{EQF-LADA} for hard-budget constraints. The latter integrates optimistic allocation with quantized updates and deterministic local credit. Under stationary outcomes and deterministic costs, \\texttt{EQF-LADA} strictly satisfies the budget on every realized sequence and achieves a regret bound that disentangles standard learning error, uncertainty from premium observations, and uncertainty in pending values. For a finite action space, summable pending-value envelopes guarantee sublinear regret without relying on raw-delay bounds. Trace-replay experiments on MMLU and HellaSwag demonstrate that \\texttt{EQF-LADA} consistently outperforms evaluated baselines in mean discounted reward across various budget and discount settings while strictly adhering to spending constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Routing", "Budgeted Bandit", "LLM" ], "forum_url": "https://openreview.net/forum?id=NFUiPrlPj0", "pdf_url": "https://openreview.net/pdf/1807dd80c8ed600f6599c479248be71be4192a6b.pdf", "updated_at": "2026-10-05T20:13:29.263Z", "categories": [ "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "N6KiGSOYfX", "number": 30406, "title": "Rethinking Memory-Efficient Optimizers for LLM Pre-Training", "abstract": "A growing family of memory-efficient optimizers reduces state storage through gradient projection, moment factorization, and shared statistics. Yet their practical value depends on when optimizer states constrain training and whether structural compression offers an advantage over simple state quantization. We examine these questions through a broad benchmark of memory-efficient methods and strong conventional optimizers under a common language-model pretraining setup. Our evaluation covers 257M and 500M dense models and extends to mixture-of-experts architectures, varying compute and state precision independently. We investigate how compression strength affects training quality and compare the approaches in terms of peak memory, validation loss at fixed token budgets, and time to quality. In the main 500M comparison, Muon with FP8 states achieves lower validation loss and lower peak training memory than the evaluated structural methods with FP32 states under both BF16 and FP8 computation. Quantizing the states of structural methods remains feasible, but yields limited further memory savings over quantized-state Muon without closing the quality gap. These findings support quantizing a strong optimizer as a practical starting point for memory-constrained pretraining and make competitive quantized baselines essential to assessing the benefits of structural compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "optimizer state compression", "optimizer state quantization", "memory-efficient optimization", "low-precision training", "FP8 training", "large language model pretraining", "matrix-aware optimizers" ], "forum_url": "https://openreview.net/forum?id=N6KiGSOYfX", "pdf_url": "https://openreview.net/pdf/50df875d99d583964ff94aa95e4f6c7adde47d70.pdf", "updated_at": "2026-10-05T20:13:31.475Z", "categories": [ "training" ], "matched_terms": [ "fp8", "low-precision", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sN2eYUoWvb", "number": 30435, "title": "Are Scaling Laws Universal in Spiking Large Language Models?", "abstract": "The scaling law is a foundational property of large language models (LLMs), dictating predictable performance scaling in foundation models. Spiking LLMs are often presented as neuromorphic analogues of quantized models which convert continuous-valued LLM activations into discrete temporal spike trains. However, whether Spiking LLMs genuinely exhibit consistent scaling laws remains an open and unresolved question. In this paper, we demonstrate that the validity of scaling laws in Spiking LLMs depends fundamentally on whether the system is analyzed under a static conversion regime or a temporal dynamical regime.\nUnder the static \\emph{accumulate-then-fire (ATF)} regime, where synaptic input is fully accumulated prior to firing, conversion acts effectively as a static quantizer. The excess loss exhibits a predictable representation-limited scaling law governed by network scale $N$ and spike-code capacity $L(T, \\phi)$. In contrast, when operating in the temporal \\emph{dynamical regime} via \\emph{online charge-and-fire (OCF)}, which represents the asynchronous parallel mode native to neuromorphic hardware, scaling laws become strictly \\emph{conditional}. In this dynamical regime, small time windows $T$ and coarse threshold levels let temporal unevenness dominate, causing excess loss to deviate from power-law scaling against $N$ or $L$. Power-law scaling is only restored when $T$ and code capacity exceed critical dynamic thresholds. Our findings reveal that scaling claims for SLLMs cannot treat them simply as bio-inspired quantized LLMs, but must explicitly specify the dynamical regime and the $(T,m)$ condition under OCF.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking Neural Networks", "Scaling law", "Language model" ], "forum_url": "https://openreview.net/forum?id=sN2eYUoWvb", "pdf_url": "https://openreview.net/pdf/5143f9426536a1ee0e42f7130c2034c924d2df8f.pdf", "updated_at": "2026-10-05T20:13:35.505Z", "categories": [ "hardware" ], "matched_terms": [ "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fDCtw2lywD", "number": 30503, "title": "The Privacy Side Effect of Cheap Inference", "abstract": "Reducing the numerical precision of model weights is common in deployment because it lowers computational cost. This reduction also selectively lowers a model’s ability to reproduce the training data word-for-word, by several bits, before it substantially affects general capability. This difference is termed the selectivity gap. The gap is measured across a 29-fold range in parameter count and is replicated in a second model family, an error-correcting production quantizer, and text-to-image diffusion, where the gap is more pronounced. Although the gap narrows as models grow, it stays above 24.2 % points of retention across all sizes tested. The content destroyed is not predicted by data duplication, although a magnitude filter, such as quantization, should spare duplicated content. Instead, the sharpness of the loss surface around a memorized sequence predicts its sensitivity to numerical rounding. This extends prior observations linking memorization to sharp loss regions by showing that the same sharpness also predicts which memorized content is lost under quantization. In diffusion models, the selectivity gap persists, but weight-space sharpness is no longer predictive. Instead, the gap is associated with the stability of image generation to perturbations in the starting noise. Finally, decomposing a realistic deployment recipe shows that the 8-bit activation quantizer, rather than the 4-bit weight quantizer, is primarily responsible for the loss of training-image reproduction. Numerical precision is therefore both an efficiency-related deployment choice and a privacy-relevant configuration that can substantially affect a model’s ability to reproduce its training data.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "memorization", "training data extraction", "diffusion models" ], "forum_url": "https://openreview.net/forum?id=fDCtw2lywD", "pdf_url": "https://openreview.net/pdf/5ddf86d8a5a78a9a6faf27e905586f492bdb08f0.pdf", "updated_at": "2026-10-05T20:13:46.555Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "8-bit", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "seKHg9ZiYL", "number": 30750, "title": "ByteKV: Reusable KV Cache Compression via Bit Allocation and Model Adaptation", "abstract": "The key–value (KV) cache accelerates autoregressive decoding in large language models by avoiding redundant context computation, but storing long contexts incurs substantial memory overhead. Reusing a compressed cache across questions requires preserving context information within a tight byte budget before downstream queries are revealed. Here we present ByteKV, a framework combining existing query-blind selection rules with bit allocation and model adaptation. Lowering KV precision enables ByteKV to retain more context positions per byte, while a shared low-rank adapter (LoRA) adapts the model to the sparse, quantized cache through self-distillation from a frozen teacher using a full 16-bit KV cache (all context positions). Across four selection rules evaluated on Llama and Qwen models, 4-bit ByteKV outperforms matched frozen 16-bit baselines on RULER-13, delivering absolute accuracy gains of 17.7 to 58.1 percentage points at a nominal 3% budget relative to full 16-bit KV storage. Integrated with custom fused packed-KV attention and vLLM memory management, ByteKV delivers 1.91× the peak output throughput of full-cache serving with the same adapted weights under concurrent serving. Our code is available at https://anonymous.4open.science/r/ByteKV.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "KV cache compression", "efficient inference", "quantization", "low-rank adaptation", "KV cache reuse", "query-agnostic eviction", "query-blind eviction", "low-bit quantization", "LoRA", "self-distillation", "inference memory budget", "RULER" ], "forum_url": "https://openreview.net/forum?id=seKHg9ZiYL", "pdf_url": "https://openreview.net/pdf/3ab6ff66031e2de2e5b94e6f217900bfceec34bc.pdf", "updated_at": "2026-10-05T20:14:27.598Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "4-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ng9ggtZX6D", "number": 30778, "title": "Beyond Reconstruction Loss in Post-Training Quantization: Balanced Fitting for Large Vision-Language Models", "abstract": "Post-training quantization (PTQ) enables efficient deployment of large vision-language models (LVLMs), but is typically calibrated on a small set while expected to generalize across diverse downstream tasks. Although recent PTQ methods for LVLMs incorporate sensitivity signals, they still minimize reconstruction loss with respect to the full-precision model, potentially over-preserving FP behavior and calibration-specific bias. Rather than treating quantization solely as an error to be minimized, we observe that it can also provide beneficial regularization for certain layers and modalities. Motivated by this observation, we propose Balanced Fitting, a quantization effect-based framework that balances precision and regularization beyond reconstruction-based optimization. By measuring layer- and component-wise quantization effects for weights, vision activations, and text activations, Balanced Fitting combines fine-grained fitting for sensitive components with coarser fitting to exploit potential regularization benefits. Experiments on multiple LVLMs show that our method consistently outperforms prior PTQ approaches under both weight-only and weight-activation quantization, while lower reconstruction loss does not reliably translate into better downstream performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Model Compression", "Vision-Language Models" ], "forum_url": "https://openreview.net/forum?id=ng9ggtZX6D", "pdf_url": "https://openreview.net/pdf/4b575e5e9eecce105c5ccb9a6f74283622ca48eb.pdf", "updated_at": "2026-10-05T20:14:35.207Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CzwnhuTaBw", "number": 30809, "title": "SCOPE: From Analytical Scale Oracles to Candidate Scale Search for NVFP4 Quantized Training", "abstract": "Recent hardware support for microscaled FP4 formats has made NVFP4 a compelling target for reducing the cost of large language model pre-training.\nHowever, existing NVFP4 training recipes still show a noticeable quality gap to BF16, primarily due to forward weight-activation quantization under the sparse FP4 codebook and shared block scale.\nWe propose **SCOPE**, a forward-centered NVFP4 quantization framework that combines an analytical scale oracle with a deployable candidate search. \nSCOPE first introduces an analytical per-block scale oracle based on FP4 rounding breakpoints to characterize the reconstruction improvement recoverable by adaptive scale selection. \nFor deployment, SCOPE replaces online breakpoint enumeration with a compact fixed-candidate search, preserving data-dependent scale selection while keeping the online path simple. \nBackward gradient-to-quantization-noise diagnostics further show that lightweight SR/Had-SR estimators remain well above the theoretically motivated diagnostic reference in our setting, allowing us to use a low-overhead backward path and focus the main optimization effort on forward scale selection. \nOn NanoChat pre-training up to 1.9B parameters and 38B tokens, SCOPE reduces the BF16 BPB gap to 1.01\\%, a relative gap reduction of \\(～30\\%\\) over Quartet-II and \\(～47\\%\\) over the NVIDIA NVFP4 recipe. In a matched 500M experiment, SCOPE reduces the measured complete-recipe training time by 16.5\\% relative to Quartet-II while achieving better validation quality. Additional PTQ and pre-training experiments show transfer across dense LLMs, MoE models, and vision-language settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization; NVFP4 training" ], "forum_url": "https://openreview.net/forum?id=CzwnhuTaBw", "pdf_url": "https://openreview.net/pdf/9d2bdc6593e93a205a7294fa614a8f69c7d14060.pdf", "updated_at": "2026-10-05T20:14:41.847Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "nvfp4", "ptq", "quantization", "quantized" ], "review_status": "abstract", "summary": "SCOPE 将 NVFP4 训练的改进重点放在前向 block scale 选择，先根据 FP4 舍入分界点构造解析尺度 oracle，衡量可恢复的重建收益。实际在线路径用固定小候选集合替代分界点枚举，同时采用摘要所述低开销反向估计路径。", "contribution": "在最高 1.9B 参数、38B token 的 NanoChat 预训练中，与 BF16 的 BPB 差距降至 1.01%，该差距相对 Quartet-II 和 NVIDIA NVFP4 配方缩小约 30% 和 47%；匹配的 500M 实验中完整训练配方耗时相对 Quartet-II 减少 16.5%。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。1.01% 是相对 BF16 的 BPB 差距，不是 BPB 数值；16.5% 耗时结果来自特定 500M 配对实验，不能外推到所有模型规模。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "NVFP4 预训练的前向权重/激活与共享 block scale", "bit_width": "NVFP4（前向 FP4；反向各张量位宽摘要未逐项列出）", "evidence": [ { "location": "官方摘要", "claim": "最高 1.9B 参数、38B token 的 NanoChat 预训练，BF16 BPB 差距为 1.01%。" }, { "location": "官方摘要", "claim": "匹配 500M 实验的完整配方训练时间比 Quartet-II 少 16.5%，验证质量更好。" } ] }, { "id": "M331HDrzRn", "number": 30825, "title": "Certified Feasibility for Inexact Riemannian Frank--Wolfe on Strongly Convex Sets", "abstract": "We study constrained optimization on Riemannian manifolds where every iterate must remain feasible despite numerical approximation of the updates. Riemannian Frank--Wolfe preserves feasibility by moving along geodesics within a geodesically\nconvex constraint set. Numerical approximation and finite-precision storage, however, can destroy\nthis guarantee even when the output remains a valid manifold point. For numerical maps, we derive\nan acceptance rule combining a geometric tolerance from strong convexity of the feasible set with\na quadratic descent tolerance; given feasible oracle targets, valid error bounds, and the stated\noracle-accuracy and step-schedule conditions, accepted iterates remain feasible and retain the\nstandard sublinear bound for smooth geodesically convex objectives. We then derive an intrinsic\ntangent-cone margin and, under smooth logarithms, an exact-oracle scaling inequality with a\nmultiplicative conditioning factor. We use this inequality to transfer known faster rates when\nthe gradient norm is bounded away from zero. We derive sharp rotation-ball formulas that make the\ngeometric tolerances explicit. Finally, for a fixed-sub-step Euler integrator in a rotation chart,\nwe prove that mantissa precision logarithmic in the inverse target accuracy suffices, given an\nexact feasible oracle, the stated primitive-error model, and sufficient available precision.\nRotation experiments show that the acceptance rule maintains feasibility and recovers reference\naccuracy under low-precision storage through accurate fallback.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Riemannian optimization", "Frank-Wolfe methods", "geodesically strongly convex sets", "inexact optimization", "finite-precision analysis" ], "forum_url": "https://openreview.net/forum?id=M331HDrzRn", "pdf_url": "https://openreview.net/pdf/0da7e82cfd4f8c8dacd6ed698223504498cb7686.pdf", "updated_at": "2026-10-05T20:14:44.850Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WANnDuFpxf", "number": 30931, "title": "Stout: Structured Order-Aware Quantization Training for Large Language Models", "abstract": "Layerwise quantization-aware training (QAT) reduces the training cost of large language models by optimizing one layer at a time. However, a fixed input-to-output order assigns adaptation by depth, without considering which layers can\nbest recover from errors introduced by other layers. It can therefore leave layers with limited adaptation ability to handle errors after more capable layers have been frozen. We introduce Oats, Order-aware Adaptation Score, which estimates a layer’s potential recovery from other layers’ quantization errors. Our local analysis explains this recovery, and experiments show that Oats predicts it more effectively than own-layer sensitivity. Based on Oats, we propose Stout, Structured Order-Aware Quantization Training, a two-stage framework for efficient QAT. The first stage computes Oats through shared measurements and trains layers with the model’s output loss in ascending score order, preserving stronger adaptation opportunities for later. The second stage trains scales with fixed quantized weights using ScaleKernel, a fused backward kernel that avoids dense weight-gradient intermediates in global memory. On Qwen3.6-35B-A3B at 2bit, Stout improves LiveCodeBench pass@1 and the BFCL aggregate score by 0.87% and 2.68% over the development-selected ordering baseline. Compared with a matched compiled or fused baseline, ScaleKernel lowers second-stage training time and peak alloated GPU memory by 0.85% and 14.47%, respectively. Our code is available at https://anonymous.4open.science/status/stout-613D", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large+Language+Model", "Weight+Quantization", "Quantization-aware+Training" ], "forum_url": "https://openreview.net/forum?id=WANnDuFpxf", "pdf_url": "https://openreview.net/pdf/e4daef193f43e7de5cd9c029a505f5898a4713ae.pdf", "updated_at": "2026-10-05T20:15:01.975Z", "categories": [ "qat", "training", "hardware" ], "matched_terms": [ "2bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cWWd9pZpe0", "number": 31011, "title": "HD-Fuse: Single-Pass Hallucination Detection via Confidence Reshaping with Conformal Likelihood-Ratio Fusion under Logprob-Only APIs", "abstract": "Hallucination detection under API logprob-only contracts is practically important but constrained: production LLM APIs expose only final-token logprob distributions, while hidden states, attention weights, and repeated sampling are unavailable or expensive. We present a systematic empirical study of single-pass, logprob-only hallucination detection for short-form QA ($\\le 16$ tokens) and identify a validated recipe that outperforms individual signals: extract a 5-dimensional confidence vector, model the test as a Gaussian likelihood ratio (LDA/QDA), fuse complementary likelihood-ratio members (including thermodynamic trajectory statistics of the top-$K$ distribution) via a calibrated cross-validated rank ensemble, and attach a split-conformal $p$-value with a finite-sample false-positive-rate guarantee, at zero extra passes. On real public benchmarks the recipe attains the best cross-cell mean AUC at one pass versus five; we confirm the signal and false-positive control on a production API (DeepSeek), so the logprob-only claim does not rest on local quantised weights. Ablation attributes the gain to signal decorrelation. Scope and prerequisites (short answers; a per-(model, benchmark) calibration fold) are stated up front, so the contribution is a validated deployment recipe plus an interpretable diagnostic framework; longer-form generation is future work.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Hallucination detection", "Large language models", "Log‑probability", "Conformal prediction", "Likelihood‑ratio fusion" ], "forum_url": "https://openreview.net/forum?id=cWWd9pZpe0", "pdf_url": "https://openreview.net/pdf/7aa5e8541fcc1cd2d89cca4fb36f9f9de5c88fd8.pdf", "updated_at": "2026-10-05T20:15:15.703Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantised" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yPW4aTeUq4", "number": 31036, "title": "GraphPTQ: Post-Training Quantization for Graph Models, Memory Savings and Generalization", "abstract": "Post-training quantization (PTQ) is essentially unexplored for graph models. Yet graph foundation models (GFMs) are scaling rapidly, straining memory budgets, while high-stakes deployment demands robustness to distribution shift. We address this with GraphPTQ, the first unified evaluation of how PTQ affects generalization in message passing models, graph transformers, and pretrained GFMs, featuring hardware-measured GPU peak memory. We evaluate eight models and eleven modes, spanning both weight-dominated and activation-dominated memory regimes across covariate shifts and zero-shot tasks. Our findings reveal clear memory-generalization tradeoffs: weight-only INT8 is the safe default for significant memory reduction, while under INT4 weights or activation quantization, out-of-distribution (OOD) accuracy responses become bidirectional and architecture-dependent. We further show that the in-distribution loss curvature along the rounding direction successfully predicts this OOD risk, with the predictive ordering strengthening as bit-width drops.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-Training Quantization", "Out-of-Distribution Generalization", "Zero-Shot Transfer", "Graph Foundation Models", "Graph Transformers", "GNNs", "Hessian Analysis" ], "forum_url": "https://openreview.net/forum?id=yPW4aTeUq4", "pdf_url": "https://openreview.net/pdf/7c73340a43d1a85b53e41bd8b92915919758de64.pdf", "updated_at": "2026-10-05T20:15:20.345Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "bit-width", "int4", "int8", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ETM5cVneOg", "number": 31091, "title": "Where Massive Activations Come From: Causal Evidence from MLP Geometry in Large Language Models", "abstract": "Large language models (LLMs) commonly exhibit massive activations (MAs) across Transformer architectures. MAs are sparse hidden-state coordinates whose magnitudes greatly exceed typical activations and concentrate at specific tokens and dimensions. These extreme values complicate quantization, compression, and efficient inference, yet their generation mechanism remains unclear. Existing studies mainly associate MAs with attention sinks or mitigate them as numerical outliers. However, these approaches do not explain how MAs are generated inside the network. In particular, the causal connection between token-level triggers, layer-wise sources, and weight-space geometry remains unresolved. We address this gap through a systematic study of 26 language models combining component ablations, singular value decomposition, and causal interventions. Our results show that MAs primarily arise from MLP down-projections, where function-token representations align with load-bearing right-singular directions that amplify and write extreme activations, while attention mainly regulates this pathway. These findings establish MAs as structured geometric events and provide a causal account of their emergence across architectures.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Massive Activations", "Large Language Models", "Mechanistic Interpretability", "MLP Geometry", "Causal Intervention", "Singular Value Decomposition." ], "forum_url": "https://openreview.net/forum?id=ETM5cVneOg", "pdf_url": "https://openreview.net/pdf/cf08cb31f1ef2c8d2e10a95bfed7d54b262dcc3e.pdf", "updated_at": "2026-10-05T20:15:28.377Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Mha5JdZBfU", "number": 31130, "title": "Real-time Streaming Video Stylization via Rectified Style Flow", "abstract": "Diffusion-based image stylization provides strong appearance control but is\ntoo slow and temporally unstable for streaming video. Directly training a\ncausal video generator, on the other hand, entangles style, motion, and\nlong-horizon autoregression and makes every new style expensive to support.\nWe introduce RSF, a staged framework for real-time streaming video\nstylization. Its core is rectified style flow: rather than generating\nfrom pure noise, the model transports a partially corrupted source\ndistribution at noise level s to a target-style distribution. We first\nlearn modular style LoRAs on paired source-target video clips. We then freeze a\nbank of these LoRAs and train the full video DiT with teacher ODE pairs and\ncausal forcing, separating reusable autoregressive dynamics from\nstyle-specific parameters. The resulting causal teacher is distilled to a\none-step student using distribution matching distillation, followed by generated-history rollout\nfine-tuning to reduce exposure bias. Finally, a lightweight recovery procedure\nadapts previously unseen style LoRAs without repeating the complete video\ntraining pipeline. At 832x480 resolution, optimized mixed-FP8\ninference sustains 102.1 FPS with 78.4 ms input-to-output latency on a single\nNVIDIA H100 and 36.3 FPS with 220 ms latency on a single NVIDIA\nRTX 5090. The H100 deployment requires only 4.81 GiB of steady-state resident\nGPU memory. We evaluate target-style text alignment, source preservation, and\ntemporal stability on a 50-video benchmark under five styles against four\npublic video stylization systems; the strictly paired CLIP-T comparison uses\n47 complete videos. We will open-source all training code, model weights,\nand prepared datasets.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "generative model", "style transfer", "video generation" ], "forum_url": "https://openreview.net/forum?id=Mha5JdZBfU", "pdf_url": "https://openreview.net/pdf/3ef21144bbab2cc8bb9aa7a7ab2d65b417bc8cc2.pdf", "updated_at": "2026-10-05T20:15:40.787Z", "categories": [ "training", "hardware", "diffusion" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2FM86b3HVZ", "number": 31234, "title": "Spiking Learned ISTA for Inverse Problems with Finite-Time Convergence Analysis", "abstract": "Deep unfolding offers an interpretable framework for learned sparse recovery. Learned Iterative Shrinkage-Thresholding Algorithm (LISTA) unfolds sparse-recovery iterations into a trainable reconstruction network. However, it represents the recovered coefficients as real-valued activations processed by dense multiply-accumulates. By contrast, a spiking implementation can represent them through sparse signed spike trains and replace dense coefficient multiplications with event-triggered accumulations. We therefore propose Spiking LISTA (S-LISTA), which replaces LISTA soft-thresholding with a Signed Soft-Threshold Integrate-and-Fire (SST-IF) unit that uses paired non-leaky integrate-and-fire neurons for the positive and negative branches and coordinate- and stage-specific scales to balance firing-rate resolution and saturation. We prove a finite-time approximation bound for SST-IF, with error proportional to the scale divided by the number of timesteps, and propagate this error through a fixed-depth unfolded network. The resulting bound shows that the spiking error of S-LISTA decreases as the number of timesteps increases. We evaluate S-LISTA on synthetic support recovery, natural-image compressive sensing, compressed-sensing MRI, and low-precision deployment. S-LISTA preserves the support-recovery behavior of its non-spiking counterpart, and after fine-tuning, it closely matches the corresponding backbone on Set11, BSD68, and MRI while remaining competitive with unfolded baselines. Under low-precision deployment, S-LISTA remains stable under reduced weight precision and avoids the sharp failures observed in its non-spiking counterpart under aggressive post-training quantization. Its measured spike activity further suggests that event-driven affine execution can reduce synaptic arithmetic.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "spiking neural networks", "inverse problems", "compressed sensing", "iterative shrinkage-thresholding algorithms", "deep unfolding" ], "forum_url": "https://openreview.net/forum?id=2FM86b3HVZ", "pdf_url": "https://openreview.net/pdf/83d21e29b9e6d872914af958a0844864d2f725b7.pdf", "updated_at": "2026-10-05T20:15:53.349Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bhD9D7FTiM", "number": 31362, "title": "iKVC: Inference-Time KV-Cache Correction for Quantized LLMs in Long-Horizon Tasks", "abstract": "Post-training quantization (PTQ) reduces the memory footprint of large language models (LLMs) and can improve inference throughput. \nHowever, the resulting quantization errors can significantly degrade accuracy, creating a fundamental trade-off between the efficiency gained from quantization and the accuracy retained from the full-precision model, particularly in long-horizon generation.\nPrior work mitigates this degradation by preserving the KV cache of selected prompt tokens at full precision, thereby reducing errors introduced during prefill. \nThis strategy, however, does not address errors introduced during decoding, as quantized weights continuously perturb the key and value representations of newly generated tokens, and these errors accumulate in the KV cache, altering subsequent predictions. \nWe empirically observe that this error accumulation becomes non-negligible as a large number of tokens are generated, and we theoretically derive that the accumulated KV-cache errors can cause the attention output to deviate from the values computed by the original full-precision model.\nMotivated by this finding, we propose inference-time KV-cache Correction (iKVC), a training-free inference method that periodically corrects the KV cache for both prompt and generated tokens before decoding errors substantially accumulate. \nThe correction interval is configurable, allowing iKVC to explicitly control the accuracy--throughput trade-off. \nAcross experiments, iKVC consistently improves the trade-off over PTQ alone in long-horizon tasks. \nOur results identify KV-cache error accumulation as an important, previously underexplored factor in achieving accurate and efficient long-horizon tasks with quantized LLMs, suggesting that managing such errors during inference can complement PTQ techniques.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Quantization", "KV cache" ], "forum_url": "https://openreview.net/forum?id=bhD9D7FTiM", "pdf_url": "https://openreview.net/pdf/dc521d84b438494b9a5492163b702aa5a2ddfbad.pdf", "updated_at": "2026-10-05T20:16:11.536Z", "categories": [ "ptq", "kv-cache", "hardware", "theory" ], "matched_terms": [ "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aL4SasjCZg", "number": 31487, "title": "PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models", "abstract": "Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed–accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PACEVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PACEVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation. The $\\text{\\href{https:\\/\\/iclr-pacevla.github.io\\/iclr-pacevla\\/}{project page}}$ is available.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-language-action models", "Robot manipulation" ], "forum_url": "https://openreview.net/forum?id=aL4SasjCZg", "pdf_url": "https://openreview.net/pdf/04836e8808aa6bd599af1304aef3eddee319543d.pdf", "updated_at": "2026-10-05T20:16:30.554Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "quantizability", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "kD59Z4R6Kf", "number": 31503, "title": "Diagnosing Poisoning Effects in Adapted LLM Guardrails: Beyond Fixed-Threshold Evaluation", "abstract": "LLM guardrails are increasingly adapted with newly collected feedback, but evaluating an updated guard at its old deployment threshold mixes two effects: ordinary adaptation can move the operating point, while poisoned feedback can additionally degrade ranking. We ask how to distinguish poisoning-induced degradation from adaptation-induced operating-point drift.\n\nWe first conduct controlled INJECTION→SAFE label-flip experiments on encoder guardrails. DistilBERT-CE and DeBERTa-GCE, three training seeds, and 1–10% corruption reveal dose-dependent behavior whose magnitude and ranking effects vary by architecture, seed, and evaluation distribution. Frozen-threshold blocking can change dramatically even when threshold-independent ranking changes little. A locked DeBERTa evaluation and a BIPIA-derived external evaluation further show that restoring a measured operating point on one distribution does not establish transfer to another.\n\nWe then study an online 4-bit QLoRA guard with clean and 10%-poisoned updates, a frozen Mistral downstream model, HarmBench, and XSTest. Both clean and poisoned adaptation move fixed-threshold operating points toward severe over-refusal, so the fixed threshold alone does not isolate poisoning.\n\nPrompt-level CPU-only reanalysis separates these mechanisms. Under the all-HarmBench ranking definition, clean-adapted AUROC ranges from 0.82 to 0.98, whereas poisoned-adapted AUROC ranges from 0.09 to 0.48; the Judge-positive analysis likewise yields paired poisoning losses in all three seeds. At the exact shared BR≈20% point available for seed 29, Clean HR is 28.02% and Poisoned HR is 52.59%, a +24.57 percentage-point gap (95% CI 18.97–30.17). Broader empirical budget frontiers are worse in 11 of 12 estimable seed–budget cells. These results motivate reporting fixed and adjusted operating points together with ranking and matched-utility analyses, rather than treating a fixed threshold as a poisoning-specific measure.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM guardrails; safety evaluation; data poisoning; model adaptation; operating-point drift; ranking degradation; safety-utility trade-off" ], "forum_url": "https://openreview.net/forum?id=kD59Z4R6Kf", "pdf_url": "https://openreview.net/pdf/1cc8260d4da3599bf3e2505711e833c4531546b0.pdf", "updated_at": "2026-10-05T20:16:32.563Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "k20Y0kwuLu", "number": 31573, "title": "Task-Free Visual-Token Repair for Frozen Multimodal LLMs", "abstract": "Multimodal large language models (MLLMs) rely on an internal interface that\nmaps visual representations into the language model. When these representations\nare altered by quantization, compression, pruning, or other efficiency\nconstraints, recovering downstream utility would ideally require optimizing them\ndirectly against the MLLM's downstream objective. However, this requires\ntask-specific supervision and repeated forward/backward passes through a large\nMLLM, making end-to-end adaptation expensive and often impractical.\nWe instead study task-free recovery at the visual-token interface. We introduce\na lightweight repairer trained only from paired clean and corrupted visual\nrepresentations, while keeping the visual encoder, multimodal projector, and\nlanguage model frozen. Training requires neither task labels, prompts, language\nlosses, nor LLM forward passes. Controlled perturbation experiments show that\npointwise token fidelity is incomplete: representations with similar\nreconstruction error can induce substantially different downstream behavior,\ndepending on their structure and direction.\nWe evaluate this approach under practical corruption across Honeybee,\nLLaVA-v1.5-7B, and Qwen2.5-VL-7B. Under severe static 2-bit visual-token\nquantization, task-free repair recovers 81.8\\%, 72.5\\%, and 52.9\\% of the POPE\ndegradation on Honeybee, LLaVA-v1.5, and Qwen2.5-VL, respectively, while\nobjectives that preserve richer representation properties can outperform\nstandard pointwise reconstruction. Learned image compression at approximately $0.155$ bpp provides a second,\nupstream corruption regime: repair recovers downstream utility across all three\narchitectures, while robust and structure-aware objectives outperform MSE in\nmatched settings. Overall, our results show that lightweight, task-free\nvisual-token repair provides a practical alternative to end-to-end MLLM\nadaptation and helps identify which representation properties should be\npreserved under different forms of visual corruption.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multimodal Large Language Models", "Visual Token Repair", "Representation Corruption", "Task-Free Adaptation", "Visual Token Quantization" ], "forum_url": "https://openreview.net/forum?id=k20Y0kwuLu", "pdf_url": "https://openreview.net/pdf/fdb8b1a09419c9cacd2a3e115e95f5e80710d081.pdf", "updated_at": "2026-10-05T20:16:42.472Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "2-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zpXMAB7b1x", "number": 31581, "title": "SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models produce remarkably capable robot behavior, but their memory footprint makes deployment on resource-constrained robotic platforms challenging. Weight quantization is a natural remedy, yet post-training quantization (PTQ) recipes inherited from large language models degrade sharply when applied to VLAs at ultra-low precision. Two limitations become particularly important: existing low-bit PTQ methods typically assess quantization error at the layer or projection level without modeling the downstream operation through which it propagates, and allocate precision without directly measuring how reducing a layer’s precision changes the policy action. We introduce SACQ, an ultra-low-bit weight-only vector-codebook quantization framework for VLAs. SACQ combines (1) structure-aware codebook fitting, which fits each layer using a role-specific quadratic metric derived from the operation its projection feeds; (2) action-aware mixed-bit allocation, which assigns precision according to decoded-action degradation as individual layers are compressed; and (3) two-stage post-quantization refinement, which first reassigns codeword indices with second-order error compensation and then calibrates codebooks and row-group scales against the decoded action. On GR00T N1.7, SACQ retains 93.85% average success across all four LIBERO suites at approximately 1.5 bits per language/action weight, only 0.25 percentage points below the 94.10% BF16 reference, while reducing resident model memory from 5.677 GB to 1.175 GB, a 4.83× reduction. On SimplER Google Robot, at approximately 2 bits per language/action weight, SACQ achieves 63.76% and 52.27% success under Visual Matching and Variant Aggregation, respectively, compared with 58.90% and 53.45% for BF16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action Models", "Post-Training Quantization", "Codebook Quantization", "Model Compression", "Robotics" ], "forum_url": "https://openreview.net/forum?id=zpXMAB7b1x", "pdf_url": "https://openreview.net/pdf/0157df3cdb2750964f016ba4527a46b8e33bffe3.pdf", "updated_at": "2026-10-05T20:16:43.214Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low precision", "low-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WTnhd7ACZw", "number": 31592, "title": "Risk Aware Compositional Evidence Learning for Heterogeneous Biomedical Observations", "abstract": "High dimensional continuous observations provide rich local information for representation learning in biomedical and scientific data. Recent advances in patch based modeling, structured encoding, and feature selection have improved the ability to capture local dependencies. However, existing methods typically aggregate local observations into a single deterministic representation or retain one preferred subset, making it difficult to preserve multiple useful evidence combinations when local information is redundant, heterogeneous, and unequally reliable. Our key idea is to learn an input conditioned, risk aware distribution over multiple evidence compositions rather than searching for a single optimal representation. Based on this idea, we propose Quantized Structured Flow Representation (QSFR). Specifically, QSFR maps continuously varying local observations into reusable discrete evidence units, constructs alternative evidence compositions through flow based sequential sampling, and models composition level utility as a distribution. The estimated utility and utility dispersion jointly reshape probability mass toward informative and reliable evidence states. We evaluate QSFR across multiple heterogeneous cohorts spanning different biological sources and acquisition conditions. The results show consistently favorable representation quality across multiple evaluation criteria. Mechanism analyses reveal improved local stability, expanded compositional support, and lower risk evidence organization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Compositional Evidence Learning", "Structured Representation Learning", "Heterogeneous Continuous Observations" ], "forum_url": "https://openreview.net/forum?id=WTnhd7ACZw", "pdf_url": "https://openreview.net/pdf/2a3f05c3af494584d99bb73b3e9574ee78e112be.pdf", "updated_at": "2026-10-05T20:16:45.576Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ug2ONLMU65", "number": 31643, "title": "Optimization Is Part of the Codec: Training Compressible Implicit Neural Video Representations", "abstract": "Implicit neural representations (INRs) compress a video by fitting a network to it and then pruning and quantizing the network into a bitstream. This pipeline implicitly assumes that improvements in the dense representation translate into improvements after compression. We show that this assumption can fail: as dense fitting improves reconstruction quality, compressed quality eventually declines. The decline stems from a spectral drift of weight energy toward weakly excited input directions, which lowers a sensitivity factor shared by pruning and quantization. A local spectral analysis of the optimizer indicates that weight decay sets where the weights are driven, while the learning rate sets how far. Without weight decay, a larger learning rate deepens the drift, whereas with weight decay it reverses the drift. Building on this, we propose SpecFit (Spectrally-Guided Dense Fitting), a training strategy that pairs moderate weight decay with a large learning rate to reverse the drift. SpecFit leaves the architecture, pruning criterion, and quantizer unchanged, making it a drop-in strategy for NeRV-based video compression. Relative to the standard training of each backbone, it reduces the PSNR BD-rate by 23--28\\% on average across three NeRV backbones, and it remains effective under a post-training quantization pipeline. Within the standard three-stage pipeline, SpecFit achieves rate--distortion performance competitive with NVRC without NVRC's learned entropy-modeling framework. This simpler decoder achieves a 65.8$\\times$ speedup in model-decoding time and increases full decoding throughput from 3.08 to 19.75 FPS. Our findings reveal that optimization in video INRs is not merely a procedure for improving dense reconstruction, but a key mechanism that determines how compressible the learned representation ultimately becomes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video Compression、implicit neural representation" ], "forum_url": "https://openreview.net/forum?id=ug2ONLMU65", "pdf_url": "https://openreview.net/pdf/db601f51989c214ae28269908f6c46138f93f243.pdf", "updated_at": "2026-10-05T20:16:54.235Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "quantization", "quantizer", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "e1LpPIjhQW", "number": 31757, "title": "HeadGuard: Selective Head Protection for Low-Bit VLM KV-Cache Quantization", "abstract": "Low-bit key–value (KV) cache quantization saves storage but can sharply degrade vision–language model (VLM) accuracy. We introduce HeadGuard, a composable head-protection method that augments a base KV-cache quantizer with a fixed high-precision mask. Image-sensitivity and output-sensitivity scores select physical KV heads offline, with approximately 1/8 protected in the main experiments; their image keys and optionally values remain in bfloat16 (BF16), while the base quantizes unprotected image entries. Across eight VLMs, three base quantizers, and eight benchmarks (six discriminative and two generative), HeadGuard recovers a substantial fraction of lost accuracy on weaker quantizers, with the strongest gains for Qwen and InternVL. At 2 bits, the six-task discriminative mean over eight models rises from 0.436 to 0.580 on the weakest base; protection can also improve generated answers and caption fidelity to BF16 outputs. Mean accuracy gains persist across all three quantizers with both tested calibration datasets. Keys-only protection retains substantial recovery at lower modeled storage cost. Evaluated through simulated quantization, HeadGuard offers a composable way to improve low-bit VLM accuracy without replacing the underlying quantizer.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision-language models", "KV-cache quantization", "mixed-precision quantization", "attention-head selection", "sensitivity analysis", "multimodal inference" ], "forum_url": "https://openreview.net/forum?id=e1LpPIjhQW", "pdf_url": "https://openreview.net/pdf/ebd3636379d2f2ccdaad7cf4b21d588b9d78d3fe.pdf", "updated_at": "2026-10-05T20:17:13.174Z", "categories": [ "kv-cache", "multimodal" ], "matched_terms": [ "low-bit", "mixed-precision", "quantization", "quantizer", "quantizers", "quantizes" ], "review_status": "reviewed", "summary": "HeadGuard 为基础 KV 量化器增加一个固定的高精度头掩码。它对图像贡献和整个头输出设置不同的标量门，分别用正确答案概率与提示文本损失的梯度评估敏感性；将共享 query heads 的分数平均到物理 KV heads 后，标准化相加并选出 Top-K。主实验用 100 条带标签的 POPE 校准、约 1/8 的头预算，掩码跨任务复用；无需训练模型权重，但离线校准需要梯度。", "contribution": "原稿比较八种 VLM、三个简化基础量化器及六项判别和两项生成任务，显示保护收益取决于模型与基线质量。输出敏感性解释了大部分恢复，图像敏感性的额外收益并不统一；仅保护 K 可以在较低的估算缓存成本下保留大部分恢复。", "limitations": "原稿明确采用假量化，未实现打包低比特缓存；存储为模型估算，没有端到端峰值显存或速度证据。KIVI-style/AKVQ-style 是同一评估框架下的简化算术，不是完整原方法或原内核。图像尺寸受限，文本 KV 保持 BF16，未测长文本或多轮生成；所有模型的保护收益并不一致。指标采用自定义 TextVQA 和对 BF16 的 caption 相似度，不能直接当作标准生成质量成绩。未运行补充代码或独立复现实验。", "review_basis": "official_pdf", "official_pdf_verified": true, "pdf_sha256": "e72017727b1edf0f25a89e7aafa3d3d68a373422fc641421ebb6981467cdfd63", "reviewed_pdf_url": "https://openreview.net/pdf/ebd3636379d2f2ccdaad7cf4b21d588b9d78d3fe.pdf", "reviewed_source_updated_at": "2026-10-05T20:17:13.174Z", "reviewed_at": "2026-10-09T05:47:09Z", "quantization_target": "视觉语言模型的图像 token KV；文本 KV 和模型权重保持 BF16。", "bit_width": "主实验图像 KV 为 2/3-bit，约 1/8 物理 KV heads 保留 BF16；另有 2–4-bit 和 1/16–1/4 头预算消融。", "evidence": [ { "location": "PDF p.2–3，§3–4", "claim": "双门敏感性、物理 KV 头分数组合；主校准 100 条 POPE，Top-K 约 N_KV/8；文本 KV 与模型权重保持 BF16。" }, { "location": "PDF p.5，Tables 1–2", "claim": "Qwen2.5-VL 2-bit min–max 0.360→0.747；Gemma-3 0.372→0.367；min–max/KIVI-style 平均增益 14.37/0.85 个百分点。" }, { "location": "PDF p.6，Table 3 与 §5.4", "claim": "K-only min–max 0.434→0.580；估算 K+V/K-only 存储 3.97/3.11 bits/element。" }, { "location": "PDF p.4、p.14，§4、Table 10", "claim": "TextVQA 为自定义 scorer；Qwen2.5-VL 的 2-bit min–max K+V 从0.015到0.687；caption指标对比 BF16 输出。" }, { "location": "PDF p.8，Scope and limitations", "claim": "模拟量化与存储估算；端到端内存/速度、长文本、多轮生成尚未评估。" } ], "results": [ "Table 1–2：2-bit min–max 的八模型六任务平均准确率约从 0.436 到 0.580，未舍入均值计算的提升为 14.37 个百分点；Qwen2.5-VL 从 0.360 到 0.747，而 Gemma-3 从 0.372 到 0.367。", "Table 2：同样保护预算下，2-bit KIVI-style 增益仅 0.85 个百分点；保护后的弱 min–max 平均准确率仍低于不保护的 KIVI-style。图像分数加入 output-only 后，2-bit min–max 额外增加 1.03 个百分点；3-bit 并未优于 output-only。", "Table 3 与 §5.4：仅保护 K 的 2-bit min–max 均值从 0.434 到 0.580。假设 128 维等长 K/V、1/8 头预算且每个量化向量含 32-bit 元数据，K+V 与 K-only 估算为 3.97 与 3.11 bits/element，后者约少 22% 图像缓存存储。", "Table 10：2-bit min–max、K+V 保护将 Qwen2.5-VL 的自定义 TextVQA 得分从 0.015 到 0.687；这不是官方 VQA scorer 的成绩。Caption ROUGE-L 的参考是 BF16 生成结果，不是人工描述。" ], "reading_basis": "已阅读本次下载的 39 页官方原稿正文，核对 Tables 1–3、相关生成表、POPE bootstrap 与预算/校准附录；关键 Tables 1–3 另用 PDF 原页图像核对。只报告原稿作者实验，未独立复现。" }, { "id": "XZU7rhdDAo", "number": 31786, "title": "Ties Worth Keeping in Dueling Submodular Minimization", "abstract": "Many learning systems must choose a set but observe only which of two candidates is preferred. For submodular objectives, existing analyses combine the cost of reconstructing a Lovasz subgradient with optimization. Under sigmoid feedback, we characterize the worst-case tradeoff between expected comparison count and mean squared error for reconstructing a canonical subgradient. This cost depends on which coordinates of the iterate are equal, with quadratic dimension dependence when all coordinates differ and linear dependence when all coincide. A Shapley identity yields an unbiased estimator. For regular, locally unbiased estimators, an electrical network argument gives a matching structural lower bound that covers arbitrary adaptive comparisons. Rounding the iterate creates ties at controlled bias, yielding a family of algorithms that improves the optimization bound over both the unquantized method and prior work at intermediate budgets. Experiments support the predicted dependence on tie structure and the optimization benefits of quantization. Across two objective families, one fitted constant describes the measured product of query cost and mean squared error within 7%; controlled optimization experiments reduce median error by up to a factor of 19 relative to the unquantized method.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "submodular minimization", "pairwise comparison feedback", "query complexity", "stochastic optimization", "Shapley estimation" ], "forum_url": "https://openreview.net/forum?id=XZU7rhdDAo", "pdf_url": "https://openreview.net/pdf/36064419f1ad6801568b280e18c793fc6e93bd16.pdf", "updated_at": "2026-10-05T20:17:18.156Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ShjsDbBcCq", "number": 31806, "title": "GapSQ: Closing the Pruning-Induced Gap for Joint Sparsification and Quantization of LLMs", "abstract": "Post-training sparsification and quantization can be combined to compress large language models, but pruning changes the weight distribution seen by the quantizer in a systematic way. By removing many near-zero weights, it creates a pruning-induced gap between the retained negative and positive branches, leading to under-utilization of the quantization range and available levels. Motivated by this observation, we introduce GapSQ, a training-free method for aggressive joint compression. We primarily study 50\\% sparsity with 3- or 4-bit quantization, where GapSQ narrows the pruning-induced gap through group-wise inward shifting before quantization. To avoid storing per-weight branch information, GapSQ enforces quantized-space separability, allowing the shift to be reversed using only lightweight group-level metadata. Across Llama and Qwen models, multiple sparsity patterns, and several pruning--quantization backbones, GapSQ consistently improves joint-compression baselines, with larger gains observed in more aggressive low-bit regimes. For example, on Llama 3-8B with 3-bit quantization and 50\\% sparsity, GapSQ improves average zero-shot accuracy from 57.2\\% to 62.5\\% while reducing WikiText-2 perplexity from 30.52 to 11.72 relative to SparseGPT+GPTQ. These results show that explicitly adapting post-pruning weight distributions is an effective way to improve low-bit joint compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLMs", "model compression", "post-training", "training-free", "pruning", "quantization" ], "forum_url": "https://openreview.net/forum?id=ShjsDbBcCq", "pdf_url": "https://openreview.net/pdf/f5b2d063c2cccd4e15f691ca49234217866b4d7c.pdf", "updated_at": "2026-10-05T20:17:22.792Z", "categories": [ "ptq" ], "matched_terms": [ "3-bit", "4-bit", "low-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NMPncFF2We", "number": 31827, "title": "Vodka: Vector quantization-aware training via Differentiable Kernel Approximation", "abstract": "The low bits quantization of Large Language Models (LLMs) is essential for lowering deployment costs and enabling inference with limited resources.\nAs vector quantization (VQ) offers higher expressivity than scalar quantization (SQ) and\nquantization-aware training (QAT) demonstrates a higher upper-bound than post-training quantization (PTQ),\ncombining VQ with QAT shows better potential for low-bit quantization.\nHowever, in face of the optimization and efficiency problems, current VQ-QAT methods either constrain the expressivity, or require extensive search, which motivates us to explore the tradeoff between expressivity and efficiency.\nTherefore, we propose \\textbf{Vodka}, a novel VQ-QAT algorithm via differentiable kernel approximation.\nWe identify the key expressivity gap as the mapping function: current VQ-QAT restricts this map to be affine, whereas VQ allows it to be arbitrary.\nTo bridge the gap, Vodka constructs the \\textbf{Codebook Spectrum} with kernel by ordering the curvature, with two endpoint functions provably the affine and arbitrary, which stays search-free and admits an exact codebook gradient at fixed indices, serving as implicit codebook to approach the VQ controllably.\nIn contrast to prior work with a fixed initialized codebook, we optimize the implicit codebook jointly with the weights, so that both adapt to each other.\nWe also propose \\textbf{Shakernel}, a fused GPU kernel that implements Vodka without heavy intermediate tensors, reducing training cost and memory.\nEmpirical evaluations on Qwen3.6-35B-A3B and Llama3.1-8B on 7 benchmarks demonstrate that Vodka outperforms state-of-the-art (SOTA) low-bit methods while maintaining efficiency.\nNotably, Vodka improves over the strongest baseline by 1.31\\% on IFEval and 0.51\\% on BFCL, demonstrating the excellent performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Vector Quantization", "Quantization-Aware Training" ], "forum_url": "https://openreview.net/forum?id=NMPncFF2We", "pdf_url": "https://openreview.net/pdf/5ff4c01b9a1f31a8829dd5897d53a9b396adc096.pdf", "updated_at": "2026-10-05T20:17:27.512Z", "categories": [ "ptq", "qat", "training", "hardware" ], "matched_terms": [ "low-bit", "ptq", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "m8YL42U0m0", "number": 31910, "title": "Two-pass Trajectory Calibration for Compensation-based Post-training Quantization", "abstract": "In post-training quantization (PTQ), learned per-row codebooks are typically fitted once on the weights where an error compensation engine such as GPTQ or LDLQ starts from. The engine, however, does not round those weights: after each column is rounded, the remaining columns are updated to compensate for its rounding error. By the time a later column is rounded, its values have therefore already been shifted by the compensation from every column rounded before it. We call this sequence of shifted values the layer's trajectory. We propose 2-Pass Trajectory Calibration (2PTC), a drop-in correction for the mismatch between the static weights a codebook is fit on and the shifted trajectory the engine actually rounds: run the engine once to record the trajectory, refit the codebook on it, and rerun the same engine on the same starting weights with the refitted grid. 2PTC stores a single codebook of the same memory size and leaves inference unchanged; it only costs one additional offline pass. Across six open LLMs and two compensation engines, on top of QEP inter-layer correction, per-channel 2PTC lowers perplexity at every bit-width with GPTQ and at 2 and 3 bits with LDLQ. The gains are largest at 2 bits: WikiText-2 perplexity decreases in 36 of 36 paired runs, by 15.5% on average with GPTQ and 5.7% with LDLQ, and average zero-shot accuracy improves by up to 9.4 points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "LLM compression", "model compression", "non-uniform quantization" ], "forum_url": "https://openreview.net/forum?id=m8YL42U0m0", "pdf_url": "https://openreview.net/pdf/71d196665b8a776618e277539a6c0227ee0fea32.pdf", "updated_at": "2026-10-05T20:17:41.713Z", "categories": [ "ptq" ], "matched_terms": [ "bit-width", "ptq", "quantization" ], "review_status": "abstract", "summary": "2PTC 指出，GPTQ/LDLQ 后续列实际被舍入的是已经过误差补偿的权重，未必符合最初拟合码本时的静态分布。它先运行一遍补偿引擎记录权重轨迹，再按轨迹重拟合码本，最后从同一起始权重重新量化；部署仍只保存同等大小的一套码本。", "contribution": "在六个开放 LLM、两种补偿引擎并叠加 QEP 层间修正的设置中，2-bit WikiText-2 困惑度在 36/36 个配对实验中下降，GPTQ 和 LDLQ 的平均降幅为 15.5% 和 5.7%；平均零样本准确率最高提高 9.4 个百分点。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。提升是在 QEP 与所述补偿引擎组合下报告的；额外成本是一遍离线运行，摘要没有报告它带来的服务加速。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "补偿式 PTQ 中的 LLM 权重及逐行非均匀码本", "bit_width": "明确报告 2 bit 和 3 bit；其他测试位宽摘要未列出", "evidence": [ { "location": "官方摘要", "claim": "2-bit 的 36/36 个配对实验降低 WikiText-2 困惑度，平均降幅为 GPTQ 15.5%、LDLQ 5.7%。" }, { "location": "官方摘要", "claim": "只增加一遍离线运行，推理及码本存储大小保持原方案。" } ] }, { "id": "xuj4skIhiZ", "number": 31954, "title": "Auditing Dynamic Sensitivity-Selective Quantization For Low-Precision LLM Reinforcement Learning", "abstract": "Reinforcement learning has become a critical post-training phase for large language models, where the rollout stage dominates the computational and memory overhead. Low-precision rollout reduces the cost of reinforcement learning for large language models but introduces a rollout-training mismatch that can destabilize optimization and even leads to collapse. Selective high-precision protection mitigates this mismatch, but the modules that benefit most from such protection can change as the policy evolves. Consequently, a protection set determined through one-time calibration may become suboptimal during training. We formulate selective protection as a time-varying allocation problem under a fixed high-precision budget, periodically assessing the sensitivity of each module to low-precision quantization and reconfiguring the protection strategy based on their current benefit-to-cost ratios. This allocation mechanism complements Quantization-Aware Training (QAT): while QAT adapts weights to low-precision computation, our method adapts precision assignments to the evolving module sensitivities. Experimental results demonstrate that combining dynamic allocation with QAT effectively controls the rollout-training mismatch throughout the evaluation window, whereas using QAT alone or QAT with static protection exhibits earlier signs of instability.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning", "Low Precision", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=xuj4skIhiZ", "pdf_url": "https://openreview.net/pdf/6eb0d38ba42b21c1988d012f2378e34d557a0304.pdf", "updated_at": "2026-10-05T20:17:47.063Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "low precision", "low-precision", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OHXQcSAqfG", "number": 32059, "title": "DecoupleGen: Intent-Aware CoT Layout Planning and Controllable Text Rendering for Video Covers", "abstract": "Video covers are often a viewer's first encounter with a video, yet their typography must be composed over a content-bearing frame whose people, objects, actions, and native screen content remain recognizable. We study this progressive design problem with DecoupleGen, which connects frame selection, intent-aware layout planning, and background-conditioned text rendering through an explicit text--quadrilateral--color interface. The frame selector identifies a relevant and visually usable background. The planner first produces a natural-language rationale for title hierarchy, placement, and contrast, then predicts structured text regions and quantized colors. A diffusion renderer converts that plan into rasterized spatial and color conditions and uses localized OCR-feature supervision to strengthen character fidelity. We introduce \\textbf{Cover-100K}, containing 99,781 aligned examples of recovered clean backgrounds, completed covers, strings, geometry, colors, and generated design rationales. On Cover-600 with supplied clean backgrounds, DecoupleGen improves text accuracy and distributional image quality over released editors, while human raters favor its planned layouts over PosterLlama. Component studies support the contributions of cover-domain supervision and structured rendering. Together, the dataset and structured interface provide a practical path from video content and design intent to readable covers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion model" ], "forum_url": "https://openreview.net/forum?id=OHXQcSAqfG", "pdf_url": "https://openreview.net/pdf/c5ddb47bcab5d50c2d5d360c7ba7ec2da2e17082.pdf", "updated_at": "2026-10-05T20:18:01.868Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Hyo9CqELcu", "number": 32184, "title": "Owen-Value-Guided Mixed-Precision Quantization for Mixture-of-Experts Models", "abstract": "Mixture-of-experts (MoE) models activate only a few experts per token, yet low-latency inference can require the entire expert pool in memory. At ultra-low average precision, mixed-precision quantization must decide which experts retain scarce higher-bit slots. Individual sensitivity scores can miss the joint effects of co-routed experts. We propose Owen-MoE, which values experts in an availability game while accounting for an output-similarity partition. Two-level Monte Carlo sampling estimates Owen values in O(TK) evaluations per layer for T trials over K experts. A scale-invariant Influence Ratio converts these values into a static {2,4}-bit map that can protect locally influential experts within lower-value coalitions. We compare rankings under identical per-layer 4-bit counts and AutoRound settings. Across eight zero-shot benchmarks on Qwen3-30B-A3B, three-seed means at 2.75--2.125 average expert bits, Owen-MoE exceeds Router-norm by 3.57--7.02 points and PMQ by 4.86--10.56 points on average. It also exceeds measured fractional Uniform by 2.78--13.31 points at matched expert-weight payload. Equal-budget controls show an additional gain over flat Shapley of 1.78 points at 2.50 bits and 2.58 points at 2.25 bits. In 96 selected high-coactivation expert pairs where the rankings disagree, Owen's choice reduces round-to-nearest quantization MSE in 78.1% of cases, with a median reduction of 17.6%. Mixtral and Qwen1.5-MoE extend the evaluation to different expert counts and routing configurations. We will release the complete code upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "mixed-precision quantization", "model compression", "post-training quantization", "Owen value", "cooperative game theory", "expert importance" ], "forum_url": "https://openreview.net/forum?id=Hyo9CqELcu", "pdf_url": "https://openreview.net/pdf/169cd3c798805b395f1d4328b41791ab16babc97.pdf", "updated_at": "2026-10-05T20:18:21.158Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UNvzDOYA37", "number": 32266, "title": "Edit the Model You Deploy: Quantization Weakens Knowledge Edits Through Their Conditioning", "abstract": "Knowledge edits and unlearning updates are computed on full-precision weights, but the served model is quantized, which can weaken or undo them. For unlearning, the explanation given is that an update smaller than the quantization step is rounded away. We show that this explanation does not carry over to closed-form (locate-and-edit) edits. By a classical dither argument, rounding on a fixed, evenly spaced grid without clipping is unbiased on an update of any size when the weights' bin positions are uniform and independent of it, and on four models single edits below the bin width keep at least 99% of their margin, on average, when only the edited matrix is quantized. The damage comes instead from the model the edit was solved against: quantized layers below the edit move the key it fires at, and quantized layers above change the read-out its value was fitted to. Quantizing everything except the edited matrix reproduces 90–101% of the damage to a single MEMIT edit of one layer (33–91% of it generic at int3), on seven models (124M–8B), two datasets and weight-only round-to-nearest, NF4, AWQ and served GPTQ and AWQ at 3 and 4 bits. At the served int4, an edit loses 4–16% of its margin, and at 3 bits 20–56%. Solving the same closed form inside the quantized model returns 90–100% of the 3-bit single-edit loss on four of five models (72–96% for released MEMIT), 77–104% under 3-bit GPTQ and AWQ, and 72–95% on held-out paraphrases; on Qwen2.5-3B, whose symmetric int3 model is itself broken, it buys efficacy at the cost of locality. Where the dither's assumptions fail, the explanation holds: gradient unlearning at a small learning rate, stored in bf16, changes mainly near-zero weights, whose bin positions are not uniform, and int4 rounding delivers 5–29% of such an update for seven of eight methods on a 1B model. All code and result files will be public at https://github.com/xxx/xxx upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "knowledge editing", "model editing", "post-training quantization", "memit", "key vectors", "unlearning", "llm deployment" ], "forum_url": "https://openreview.net/forum?id=UNvzDOYA37", "pdf_url": "https://openreview.net/pdf/057fac824e982f977c1448b4ebc35b900e312d88.pdf", "updated_at": "2026-10-05T20:18:33.260Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "3-bit", "int3", "int4", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EnCAzPC8fN", "number": 32307, "title": "Adaptive Computation with Conformal Certification", "abstract": "Scientific imaging data involving medical (e.g., 3D radiographic, microscopic images) and astronomical sensors pose different challenges from natural images. \nOne key challenge is that these datasets are often orders of magnitude larger than natural images. \nEfficient training and inference on such data is challenging due to the increased hardware requirements, often needing bespoke solutions. \nMoreover, the relevant objects such as tumors and astronomical phenomena are sparsely distributed with respect to the background information. \nStandard efficiency techniques like quantization and pruning apply uniform scaling of all regions, potentially compromising the precision/fidelity of predictions. \nThus, efficient methods that reduce computational requirements without losing resolution for relevant parts of the data and with uncertainty estimates on predictions are essential.\nWe present \\textit{Conformal Adaptation} with risk guarantees for uncertainty-aware sparse computations. Our formalism is general, spanning dense pixel-level classification and regression tasks where hierarchical models that predict at low resolution, certify the signal using conformal risk control, and refine the certified predictions. \nWe benchmark our method on multiple tasks such as 3D medical image segmentation, 2D galaxy flux prediction, and 1D calcium spike rate prediction, achieving $2.5$-$72\\times$ wall-clock speedup across the various tasks, including on commodity CPUs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Deep Learning", "Conformal Prediction", "Computation Optimization" ], "forum_url": "https://openreview.net/forum?id=EnCAzPC8fN", "pdf_url": "https://openreview.net/pdf/22f665e3f4daf96b7290e825bc3cd5caf4a3cec8.pdf", "updated_at": "2026-10-05T20:18:42.668Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W3jcTOFu5O", "number": 32348, "title": "Measuring and Reducing the Train--Inference Mismatch in Discrete Diffusion Models", "abstract": "In discrete diffusion, we train a denoiser model on states from a noising process (i.e., on corrupted data). At inference time, the model has to reverse this process, iteratively denoising data, and is thus exposed to states created by its own samples. This creates a train–inference mismatch: the discrepancy between the distribution over states presented to the model during training and inference. We formalize this mismatch as the Mirror Gap. We measure this gap using three measures of discrepancy: the discriminability of training and inference states, the distance between their representation centroids, and the divergence between their quantized representation distributions. Together, these measurements give us a window into how the mismatch behaves at different resolutions and across the diffusion process. Our analyses reveal the train–inference mismatch has a low-dimensional component in representation space; they also show correlations between discriminability scores early in sampling and final sample quality. These results motivate two strategies to reduce this gap: steering generations in representation space, and resampling states based on measured gap. Experimentally, we show these strategies improve the generation compute–quality curve: at the same quality (measured by MAUVE or FID), our methods show sampling-efficiency gains exceeding 8× over strong baselines while requiring no denoiser retraining, fine-tuning, or distillation; this is true in both text and image generation, and across noising processes and samplers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "discrete diffusion models", "train-inference mismatch", "generative modeling", "sampling", "exposure bias", "inference-time scaling", "representation steering" ], "forum_url": "https://openreview.net/forum?id=W3jcTOFu5O", "pdf_url": "https://openreview.net/pdf/fbff393f77b35475384d1f4a489bdb0bc15dfef4.pdf", "updated_at": "2026-10-05T20:18:52.163Z", "categories": [ "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qZ9ZyaD9zO", "number": 32402, "title": "MINE: Entropy-coded Memory-efficient Inference for Large Language Models", "abstract": "We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as BF16 and under additional assumptions. Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments. At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference. It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost. We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency. MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget. On Llama 3.1 8B Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\\times$ and 1.48$\\times$ over the fastest baseline on an RTX 4060 Ti and an RTX 5090, at effective bitrates of 2.24 and 3.43 bits per weight. MINE is available as open source.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Entropy coding", "Asymmetric Numeral Systems (ANS)", "LLM inference", "Quantization", "Compression", "Low-latency inference", "Consumer GPUs" ], "forum_url": "https://openreview.net/forum?id=qZ9ZyaD9zO", "pdf_url": "https://openreview.net/pdf/38382ce8e8c01f08d9bd62f0fb6bf82562cab9da.pdf", "updated_at": "2026-10-05T20:19:01.194Z", "categories": [ "hardware" ], "matched_terms": [ "1-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "u7THv0qTCi", "number": 32471, "title": "BEYOND LOCAL RECONSTRUCTION: GEOMETRY- PRESERVING OUTPUT ALIGNMENT FOR BINARY LLM QUANTIZATION", "abstract": "Recent advances in post-training quantization for large language models (LLMs) have demonstrated that low-bit quantization methods can maintain most of the original model performance. Despite such successes, 1-bit quantization remains particularly challenging. A common strategy in 1-bit quantization is to determine binary weights by matching full-precision parameters, following a weight-driven criterion. However, this objective is not directly aligned with the quantized model's objective, which is to preserve the model's output behavior under the impact of quantization. A natural alternative is to adopt output-driven criteria that minimize discrepancies in model outputs using calibration data. Surprisingly, naive output-driven approaches often perform even worse in the 1-bit regime.\n In this paper, we show that this failure arises from two fundamental issues: error accumulation across layers and, more critically, \\emph{anisotropic distortion} of the representation space. \n Based on this observation, we introduce a geometry-preserving output-alignment framework for 1-bit LLM quantization. First, we formulate output alignment as a \\emph{constrained optimization} problem that improves reconstruction while explicitly controlling changes in token similarity, preventing reconstruction gains from being achieved at the expense of representation geometry. Second, we propose a \\emph{geometry-aware salient-mask selection} strategy that identifies quantization configurations according to their ability to preserve token relations, rather than relying solely on reconstruction error.\n These components enable effective output-driven optimization while retaining the structural information lost by naive alignment. Extensive experiments across multiple LLM families and downstream tasks demonstrate consistent improvements over existing 1-bit PTQ methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "binary quantization", "post training quantization", "large language model" ], "forum_url": "https://openreview.net/forum?id=u7THv0qTCi", "pdf_url": "https://openreview.net/pdf/4130b831579a4039918cd4a8ccdbebbe21cc1e87.pdf", "updated_at": "2026-10-05T20:19:15.286Z", "categories": [ "ptq" ], "matched_terms": [ "1-bit", "binary", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ccVPJNt8JD", "number": 32486, "title": "Hadamizer: Spatio-Temporally Adaptive Hadamard Rotation for Low-Precision Distributed Optimization", "abstract": "Large Transformer models have become increasingly vital across numerous applications, yet training them is fundamentally bottlenecked on high-bandwidth memory (HBM) with 32-bit optimizer states occupying a dominant share. Quantization is a key technique to alleviate this pressure, but standard coordinate-wise quantization suffers from severe clipping errors caused by heavy-tailed outliers that widen the quantization range. While orthogonal rotations effectively disperse outliers, mixing coordinates across distributed systems makes collective communication prohibitively expensive. \n\nHere, we propose HADAMIZER, an optimizer-native orthogonal rotation framework designed for distributed training on accelerator\nmeshes. Spatially, it aligns rotation blocks with local shard boundaries to avoid rotation-induced collectives. Temporally, HADAMIZER adapts the randomization of Hadamard transformations to tensor dynamics: target-adaptive keying employs static keys for persistent optimizer states and dynamic keys for stochastic gradients, preserving preconditioning with minimal quantization errors. Evaluated on TPUv5e meshes, HADAMIZER achieves near-full-precision-equivalent results on various pre-training and fine-tuning tasks, while delivering up to 9.58% higher end-to-end training throughput over the current rotation-based training frameworks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Quantization", "Distributed Training", "Hadamard Rotation", "Optimization" ], "forum_url": "https://openreview.net/forum?id=ccVPJNt8JD", "pdf_url": "https://openreview.net/pdf/251e7e70f88eefd5a0148af98b8bcef9b2dc7771.pdf", "updated_at": "2026-10-05T20:19:17.887Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qEBQy8q7Wd", "number": 32519, "title": "MoE-GS: π-Gated Mixture-of-Experts for Content-Adaptive 2D Gaussian Image Representation", "abstract": "Explicit Gaussian image representations support direct spatial rendering, but their quality--storage trade-off depends on the precision of every stored attribute. We investigate the interaction between shared appearance, quantization, and truncated rendering in MoE-GS. Each Gaussian combines a small shared color table through continuous mixture coordinates, while a learned positive amplitude participates in both top-$K$ selection and normalized blending. We formulate the storage break-even condition, characterize the color and selection constraints, and optimize the representation through its quantized inference path. A self-contained bitstream makes the reported quality reproducible from the stored bytes alone. Our evaluation distinguishes 16-bit uniform quantization from IEEE FP16, combines fixed-budget joint fitting with fixed-geometry conditional controls, and measures component interventions without changing geometry. The results delimit when appearance sharing is useful and when its quantization error outweighs its storage savings. Region queries retain the global tile partition; their cost includes candidate discovery rather than only the final rasterization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "2DGS", "Image Compression", "MoE" ], "forum_url": "https://openreview.net/forum?id=qEBQy8q7Wd", "pdf_url": "https://openreview.net/pdf/e353f3360b88d5b2fa71722b957caa692ab179b6.pdf", "updated_at": "2026-10-05T20:19:22.868Z", "categories": [], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "t5Rr4kcINa", "number": 32613, "title": "SQuAT: Spherical Lattice Vector Quantization-Aware Training", "abstract": "Vector quantization (VQ), in particular lattice-based methods, offer strong rate-distortion trade-offs for low-bit large language models (LLMs), yet their behavior under quantization-aware training (VQ‑QAT) remains poorly understood. We scale lattice VQ‑QAT to billion-parameter LLMs and multi-billion-token training regimes, and introduce a shape–gain VQ‑QAT approach, SQuAT, that decomposes weights into direction and magnitude: normalized directions are quantized using an E8-derived spherical codebook, while gains are quantized independently with low-bit scalar quantizers.\nWe analyze the interaction of shape-gain parameterizations and straight-through estimators (STE) in the context of VQ optimization, and show that factored STEs induce tangent-space updates for directions, yielding an implicit form of spherical (Riemannian) optimization.\nTo enable scalability, we develop a distributed exact nearest-neighbor assignment scheme, in which workers quantize disjoint weight blocks and communicate only discrete indices, reducing VQ-QAT training time by 12.8x and allowing us to scale it to 10B tokens.\nAcross instruction-tuned LLaMA-3.2-1B/3B, Qwen3-4B, and base Qwen3-1.7B, SQuAT consistently outperforms standard VQ‑QAT STE under identical quantizers and initialization, and achieves SOTA VQ-QAT accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Quantization-Aware Training", "Vector Quantization", "Lattice Quantization", "LLMs" ], "forum_url": "https://openreview.net/forum?id=t5Rr4kcINa", "pdf_url": "https://openreview.net/pdf/c292c22d135c54f45a8746415f05d8d8472aff2c.pdf", "updated_at": "2026-10-05T20:19:33.502Z", "categories": [ "qat", "training", "theory" ], "matched_terms": [ "low-bit", "qat", "quantization", "quantize", "quantized", "quantizers" ], "review_status": "abstract", "summary": "SQuAT 把权重拆成方向与幅度，方向使用 E8 衍生球面码本，幅度使用独立的低位标量量化。作者分析分解式 STE 对方向产生的切空间更新，并用分布式精确最近邻分配让不同 worker 量化互不重叠的权重块，仅传递离散索引。", "contribution": "摘要报告 VQ-QAT 训练时间减少至原方案约 1/12.8，并扩展到 100 亿 token；在所列 LLaMA-3.2、Qwen3 模型上，以相同量化器和初始化比较，持续优于标准 VQ-QAT STE。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。12.8 倍是训练时间改进，不能当作推理速度；摘要未列出具体位宽、worker 数或硬件配置。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "LLM 权重方向与幅度的 shape–gain lattice VQ-QAT", "bit_width": "摘要未注明", "evidence": [ { "location": "官方摘要", "claim": "分布式精确分配将 VQ-QAT 训练加速 12.8 倍，并支持 100 亿 token 的训练规模。" }, { "location": "官方摘要", "claim": "在相同量化器与初始化下优于标准 VQ-QAT STE，未提供具体准确率数值。" } ] }, { "id": "znwlEJDy1v", "number": 32630, "title": "Single Thread, Real Time: Heterogeneous Quantization for CPU Speech Recognition", "abstract": "LM-based speech recognition models are among the most accurate ASR systems today, but they typically require GPUs for real-time inference, so private audio leaves the device and latency is tied to the network. We move them onto device CPUs and propose a memory-traffic model that assigns a quantized format per module. The transformer is weight-bound, so its weights are ternarized; the acoustic stack is activation-bound, so it runs entirely in INT8, which shrinks activation memory 6.2x against FP32. Progressive fake-quantization keeps training from diverging, and AVX2 and NEON intrinsics accelerate both formats. Trained and deployed this way, a 1.5B recognition model is up to 4.32x faster on a single thread than FP16 and 2.9x smaller, and transcribes faster than real time on one thread of either an Apple M4 or an Intel Core i7-13700, for a 0.43% increase in recognition error across ten corpora relative to the FP16 baseline. The allocation transfers to other recognition and generation models with comparable gains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "on-device inference", "CPU inference", "speech recognition", "text-to-speech", "mixed-precision", "quantization-aware training", "ternary weights", "INT8", "SIMD kernels" ], "forum_url": "https://openreview.net/forum?id=znwlEJDy1v", "pdf_url": "https://openreview.net/pdf/7a9704bc3586ea2f108263e0ee4735d07f6406ec.pdf", "updated_at": "2026-10-05T20:19:35.233Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "int8", "mixed-precision", "quantization", "quantized", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qKZewFbAXh", "number": 32718, "title": "IMC-CLINIC: Coupled Loss-Informed Newton Iterations for Clipping in Analog In-Memory Computing", "abstract": "Analog in-memory computing (IMC) offers a promising path toward energy-efficient large language model (LLM) inference by executing matrix multiplications (MatMul) directly within memory arrays in the analog domain. Its efficiency, however, comes with an additional source of error: limited-precision analog-to-digital converters (ADCs) quantize the accumulated analog partial sums and introduce output-side error, which is a different problem from conventional activation and weight quantization at the MatMul inputs. Clipping can mitigate both operand and ADC quantization errors by reducing their dynamic ranges, but the optimal clipping factors must jointly balance activation rounding and clipping, weight rounding and clipping, and ADC quantization. Existing clipping methods, designed for digital quantization, do not explicitly optimize this coupled IMC error. Common approaches rely on costly search-based calibration, leading to suboptimal accuracy and long calibration time. We introduce **IMC-CLINIC** (**C**oupled **L**oss-**I**nformed **N**ewton **I**terations for **C**lipping), a clipping calibration framework built around an analytical surrogate for IMC MatMul output error. The loss surrogate models operand quantization, accumulated clipping-induced bias, and ADC quantization jointly, allowing its gradient and approximate curvature to be evaluated directly from only a small calibration set. **IMC-CLINIC** then jointly optimizes activation and weight clipping factors using a safeguarded Newton-type method. We show that **IMC-CLINIC** reduces analog-IMC MatMul output error by effectively balancing operand and ADC quantization errors. Across multiple models and datasets, it improves average zero-shot accuracy by 6.5–11.5 percentage points over the grid search baseline while reducing calibration time by 10.0×–12.1×. Moreover, its analytical surrogate closely tracks empirical IMC output error, while its optimizer is fast and certified within 1% of the global optimum under the loss objective across all projections on two representative models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "In-memory Computing", "LLM Quantization", "Clipping", "ADC Quantization" ], "forum_url": "https://openreview.net/forum?id=qKZewFbAXh", "pdf_url": "https://openreview.net/pdf/db638501ce69fae623ac24eebed2fb0540e343fc.pdf", "updated_at": "2026-10-05T20:19:47.995Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sEwLB21Bgz", "number": 32765, "title": "DISTRIQUANT: BEYOND RECONSTRUCTION FOR ROBUST BINARY LLMS", "abstract": "Post-training quantization (PTQ) for large language models (LLMs) is commonly formulated as a reconstruction problem, where quantized weights are optimized to reproduce full-precision behavior on a small calibration set. We show that, in the 1-bit regime, reconstruction plays a more consequential role: it implicitly determines \\emph{which representation directions survive quantization}. Since reconstruction error places greater emphasis on directions with high variance on the calibration distribution, binary PTQ preferentially preserves calibration-dominant structure while weak directions receive substantially less protection. Crucially, directions that are weak on the calibration distribution are not necessarily unimportant elsewhere. We find that many become relatively more prominent on unseen domains, yet are disproportionately suppressed after binarization. To quantify this effect, we introduce a directional survivor rate that measures how much of each full-precision representation direction remains after quantization. Motivated by this finding, we propose \\textbf{DistriQuant}, a spectrum-balanced approach to salient-mask selection for binary PTQ. DistriQuant uses \\emph{Covariance-Balanced Alignment} to remove the variance imbalance among calibration directions before comparing full-precision and quantized representations, preventing dominant calibration directions from overwhelming the allocation of limited binary capacity. The method requires only the standard calibration set, leaves the underlying binary representation unchanged, and introduces no additional inference-time computation. Extensive experiments across different architectures and datasets demonstrate the effectiveness of our proposed method.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "binary quantization", "post training quantization", "large language model" ], "forum_url": "https://openreview.net/forum?id=sEwLB21Bgz", "pdf_url": "https://openreview.net/pdf/423d86557674953f13251ee6ba0b3616b9acc1e1.pdf", "updated_at": "2026-10-05T20:19:54.200Z", "categories": [ "ptq" ], "matched_terms": [ "1-bit", "binarization", "binary", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zHmxUzb4kv", "number": 32772, "title": "Decentralized Personalized Federated Learning with Compressed Messages via Finite-Time Quantized Coordination", "abstract": "Decentralized Federated Learning (DFL) enables privacy-preserving collaborative training without a central server. However, it faces two major challenges: statistical data heterogeneity, which degrades global model performance on local tasks, and limited communication bandwidth, which hinders collaboration and degrades model accuracy. \nTo address these issues, we propose Q-pFedMe, a decentralized personalized FL algorithm with quantized communication. Q-pFedMe decouples local personalized model optimization from global reference learning using Moreau envelopes. Furthermore, it embeds fixed-point quantization into finite-time averaging, making all nodes reach exact agreement on a common quantized aggregate. We theoretically prove that Q-pFedMe converges linearly to a quantization-dependent neighborhood of the optimal solution for strongly convex objectives, and has an $\\mathcal{O}(1/R)$ optimization term with an explicit error floor for smooth non-convex objectives. We also derive a finite-time upper bound on the number of communication steps required by the quantized average consensus subroutine at each global round. Experiments on Synthetic, MNIST, and CIFAR-10 show that Q-pFedMe achieves accuracy comparable to centralized unquantized pFedMe. Additional CIFAR-100 experiments on two directed topologies demonstrate the highest personalized accuracy and lowest per-transmission payload among the evaluated decentralized baselines. Our code is available at https://anonymous.4open.science/r/N8HgpH8UoWHorEt/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Decentralized Federated Learning; Personalized Federated Learning; Quantized Communication; Finite-Time Consensus; Directed Networks" ], "forum_url": "https://openreview.net/forum?id=zHmxUzb4kv", "pdf_url": "https://openreview.net/pdf/fb4675e1de80fb26374a11e5192e79d82c94c220.pdf", "updated_at": "2026-10-05T20:19:55.292Z", "categories": [ "training", "theory" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IOSIv5EbWj", "number": 32941, "title": "InnoComm: Learning Decision-Relevant Latent Messages for Multi-Agent Language Models", "abstract": "When evidence is distributed across large language model agents, a communication\ninterface must preserve the information needed for receiver decisions under a\nreal byte budget. We introduce InnoComm, a sender-local latent codec that maps\nlocally contextualized evidence into 128-dimensional slots, serializes them as\nINT8 packets, and delivers them to a set-consistent language-model receiver.\nThe codec is trained to improve receiver decisions through task supervision,\nfrozen-base behavior matching, and packet-receiver co-adaptation rather than\nfull activation reconstruction. On two annotation-derived candidate-decision\ntasks, the uncompressed interface exhibits stable content utility across three\nreceiver seeds on Development, with a separate component confirmation on a\nheld-out FEVEROUS population. In single-receiver-seed Development experiments,\na fixed-reader four-slot packet compresses the full FP16 hidden-bank payload by\napproximately $1{,}054\\times$ to a mean of 1,428.872 application-layer bytes per\nrequest while lowering NLL by 0.044406 relative to the full base. Under a\ntighter 146-byte-per-sender budget, co-adapting the shared codec and memory\nadapter reduces NLL by\n$0.042625$ (95\\% CI $[0.019817, 0.063856]$) relative to a matched codec-only\ncontinuation, while preserving content effects, base utility, and tested\npermutation consistency. A strong adapted text interface remains ahead by\n$0.044275$ NLL at the same mean byte budget. Together, these results show that\nco-adapting the shared packet and receiver interface improves compact latent\ncommunication under an actual wire budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Latent Communication", "Multi-Agent Language Models", "Decision-Relevant Representations", "Learned Compression", "Byte-Constrained Inference" ], "forum_url": "https://openreview.net/forum?id=IOSIv5EbWj", "pdf_url": "https://openreview.net/pdf/7050c808efe6c66f20d44fbc7c46953a49164e34.pdf", "updated_at": "2026-10-05T20:20:18.659Z", "categories": [ "ptq" ], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dbB7Ni8kVL", "number": 32981, "title": "PPCC: Price-Preserving Calibration Coresets for Post-Training Quantization", "abstract": "Calibration coreset selection is essential for efficient post-training quantization (PTQ) of large language models, as limited calibration data must support reliable layer-wise choices of bitwidths, quantization granularity, and pre-quantization transformations.\nHowever, existing criteria based on data diversity or activation coverage do not directly characterize how calibration errors affect these decisions.\nTo address this, we propose Price-Preserving Calibration Coresets (\\ours{}), a framework that reinterprets calibration selection through the preservation of pairwise quantization price gaps.\nThis perspective decomposes price-gap distortion into distribution shift, finite-pool estimation, and coreset compression, where the compression term, which coreset selection controls, motivates a price-weighted moment-matching objective.\nBuilding on this foundation, inspired by traditional herding algorithms, we introduce an efficient greedy algorithm with a correction term derived by minimizing the expected terminal discrepancy.\nTheoretically, we establish finite-pool concentration bounds and prove that the resulting price-weighted squared moment discrepancy is no larger than its expectation under uniform random selection at the same budget.\nEmpirical results show that \\ours{} recovers the full-pool price minimizer at all three tested budgets on Llama-3.2-1B and improves average downstream task scores in most configuration selection settings, with perplexity reductions of up to $35.2\\%$ over the best baseline on Llama-3.1-8B.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "calibration selection" ], "forum_url": "https://openreview.net/forum?id=dbB7Ni8kVL", "pdf_url": "https://openreview.net/pdf/060ed1f05d7200cee0e67a36c05e56d7c9fb70ad.pdf", "updated_at": "2026-10-05T20:20:26.062Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OIqo7a2tPO", "number": 33079, "title": "AngularQ: Geometry-Guided Gradients for Low-Bit Large Language Models", "abstract": "Training large language models (LLMs) with one- to three-bit weights and activations offers major gains in efficiency, but preserving accuracy under extreme quantization remains a central challenge. The difficulty lies not only in representing discrete values but also in learning through them, since hard quantizers eliminate the derivatives required for gradient-based training. We introduce AngularQ, a mathematical framework for quantization-aware training that leverages the averaging effect of random embeddings to obtain a smooth population formulation of discrete low-bit computation. This formulation yields geometry-aware gradient surrogates that retain information from the quantized forward pass while providing an unbiased surrogate for the corresponding backward gradient. We characterize how different embedding constructions affect approximation accuracy and gradient quality, using these insights to guide a unified, computationally efficient approach to binary and multibit training. Beyond providing a comprehensive theoretical foundation for learning through quantization, this framework translates into substantial empirical gains in extreme-low-bit language modeling. Across language models from 45M to 1.7B parameters on FineWeb- Edu, AngularQ delivers its largest gains when both weights and activations are binary: 55.3% and 49.4% lower perplexity than matched-update QuEST at 0.6B and 1.7B, respectively. Two- and three-bit configurations retain substantial gains of 7.7–9.3%. Three paired training seeds at 140M consistently reproduce the binary advantage. At equal code-storage budgets, multibit quantization also outperforms binary projection expansion across the evaluated scales. These results establish angular geometry as a principled and effective foundation for learning extreme-low-bit language models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Bit Large Language Models", "Quantization-Aware Training", "Surrogate Gradients", "Angular Geometry", "Binary Embeddings" ], "forum_url": "https://openreview.net/forum?id=OIqo7a2tPO", "pdf_url": "https://openreview.net/pdf/a91da90d1864186a2d41d86986fa39bdde66d938.pdf", "updated_at": "2026-10-05T20:20:41.621Z", "categories": [ "qat", "training", "reasoning", "theory" ], "matched_terms": [ "binary", "low-bit", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rS9IUWWuiX", "number": 33096, "title": "Block-GTQ: RoPE-Block Rate Allocation for KV-Cache Quantization", "abstract": "Larger models and longer contexts make the KV cache a memory and bandwidth\nbottleneck. We introduce \\emph{Block-GTQ}, a framework that uses the\ntwo-dimensional frequency blocks of rotary position embeddings (RoPE) to\nallocate cached-key precision. A blockwise logit-error bound motivates a\nlabel-free query--key energy score, which we combine with codec-dependent\nrate penalties. Under equal-cost budgets and diminishing marginal gains,\nlargest-marginal-gain allocation exactly minimizes the separable integer\nproxy. Our TurboQuant-MSE (TQ-MSE) instantiation groups equal-rate blocks\nand supports attention over packed caches without materializing a decoded\nfp16 KV cache. Across ten model/path configurations, K-only diagnostics\nreduce layer-averaged logit mean absolute error by about $32$--$80\\%$\nrelative to uniform TQ-MSE, improving all $367$ layers at both 2- and 3-bit\naverage logical key rates. At K2V2 (2 average logical bits per K/V\ncoordinate) on Llama-3.1-8B-Instruct, Block-GTQ raises the six-task\nneedle-in-a-haystack (NIAH) average from $70.6$ to $97.4$ and LongBench-EN\nfrom $33.7$ to $44.9$; K2V2 gains also hold on the tested 32B and\n70B models. On DeepSeek-R1-Distill-Qwen-7B at K3V2, without fp16 sink or\nrecent-token buffers, it achieves $49.6/38.8\\%$ avg@8 on AIME 2024/2025 versus fp16’s $54.2/37.9\\%$,\nwhere uniform TQ-MSE scores zero on both. On Qwen2.5-3B-Instruct, packed\nK3V3 serving on one H800 reduces resident KV footprint by a factor of\n$3.24$ and retains perplexity close to fp16 at 128K, while uniform TQ-MSE\ndegrades severely. At a 1M-token context, Block-GTQ reduces decode latency by 8.0\\%\nrelative to FP16 FlashAttention-2.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache quantization", "efficient LLM inference", "long-context language models", "mixed-precision bit allocation", "rotary position embeddings (RoPE)" ], "forum_url": "https://openreview.net/forum?id=rS9IUWWuiX", "pdf_url": "https://openreview.net/pdf/1697b6d115851ebd6a8d38acade8433cb2d72608.pdf", "updated_at": "2026-10-05T20:20:43.125Z", "categories": [ "kv-cache", "hardware", "theory" ], "matched_terms": [ "3-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "x94ZH5nOgP", "number": 33172, "title": "Topology Matters: Evaluating Topological Fidelity in CAD Generation", "abstract": "Geometric similarity and execution validity do not guarantee structural correctness in Computer-Aided Design (CAD): small geometric perturbations can change a solid's connected components, tunnels, or cavities. \nWe find that such changes can arise during representation preprocessing alone: among 7,590 DeepCAD references reconstructed from both original CAD programs and quantized vectors, 762 (10.0\\%) have different global topology. \nMotivated by this gap, we study topology fidelity across five neural CAD generators, using the Betti signature of each final solid as a diagnostic. \nUnder standard generation, exact signature fidelity ranges from 78.2\\% to 88.9\\%, and error directions differ across generation paradigms: text-conditioned models tend to omit existing structures, while the evaluated reconstruction-based models show more balanced or over-realized errors.\nFidelity is also substantially lower on construction-novel patterns, with a gap persisting after standardization over measured topological complexity and construction length.\nFinally, even when a model reproduces the baseline topology, it rarely realizes explicit topology-changing instructions, with strict success rates of 0--1.7\\% across the evaluated text-conditioned models. \nThese findings show that aggregate topology fidelity alone is insufficient to characterize CAD generators: models can differ substantially in their failure modes, sensitivity to construction patterns, and ability to modify existing topology.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Computer-Aided Design", "CAD Generation", "Topological Fidelity", "Model Evaluation" ], "forum_url": "https://openreview.net/forum?id=x94ZH5nOgP", "pdf_url": "https://openreview.net/pdf/141dfa88ccce322b297c0427c4064cdff9a4b6ee.pdf", "updated_at": "2026-10-05T20:20:52.984Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1ED9UU1BP1", "number": 33194, "title": "Linger and Lose: Knowledge Collapse in Low-Bit Language Models", "abstract": "Training language models with ternary weights is commonly judged by loss and downstream accuracy, which record only a modest cost relative to full precision. We show that these metrics can conceal a much larger failure. We instead measure knowledge capacity, the factual bits stored per parameter, on synthetic biographies with known information content. We train GPT-2-style models from scratch with 2.5M to 50M parameters at five precisions. Under the standard cosine schedule, ternary models retain as little as 6% of an identically trained fp16 model's capacity. The deficit widens with model size while perplexity rises by only 1.4 to 1.6 times. Measured throughout training, these models first acquire capacity and then lose most of it. We identify this knowledge collapse as a learning-rate dwell instability. Held near $1$--$2\\times10^{-4}$ with no decay, a pre-collapse model collapses within a few hundred exposures, and returning to a safe rate does not restore capacity. We then locate the collapse in the output head. It happens at a value the model can never predict. The weights there grow unchecked, while every other prediction the model makes is unchanged. We find that making the value predictable removes the collapse, regardless of which attribute carries it. A warmup-stable-decay schedule with a 10\\% cooldown increases ternary capacity by 2.7 times at 25M and 4.3 times at 50M. Gains are larger at lower precision. Cutting the output head's learning rate prevents failure when training from scratch. The post-training quantization methods we tested recover no measurable capacity below 4 bits. Our findings suggest judging low-precision training by retained capacity during training rather than final loss.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-bit quantization", "Quantization Aware Training", "knowledge capacity" ], "forum_url": "https://openreview.net/forum?id=1ED9UU1BP1", "pdf_url": "https://openreview.net/pdf/fcac033e8748e373888ace415d7b2266ea0b4085.pdf", "updated_at": "2026-10-05T20:20:56.588Z", "categories": [ "ptq", "qat", "training" ], "matched_terms": [ "low-bit", "low-precision", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YMkHmZjJ6B", "number": 33220, "title": "Optimizer Lineage: The Pretraining Optimizer Leaves a Causal Spectral Fingerprint on Base-Model Weights", "abstract": "A base model is usually treated as a black box that downstream users prune, quantize, LoRA-adapt, or merge—as if these operations depend only on its architecture and final weights. We show that its weight spectrum also carries a durable, causal imprint of a choice fixed long before any user touches the checkpoint: the optimizer it was pretrained with. Adaptive optimizers such as Adam apply per-coordinate, effectively low-rank updates, whereas spectral optimizers such as Muon orthogonalize each update via a matrix-sign (msign) step, spreading energy across the full singular spectrum. Our single load-bearing result is a causal spectral fingerprint (H1). Under a clean-room control that pretrains 124M GPT pairs varying only the optimizer—architecture, data, initialization seed, and step count all held fixed— Muon raises the per-layer stable rank of hidden matrices by 6–14× (from ~17 to ~121) and lightens their spectral tail relative to Adam. The effect is unusually robust: it holds consistently across every layer, replicates across two independent seeds, and reproduces at two scales (124M and 350M). It is causal rather than a loss artifact—under an iso-loss control that trains Adam to matched (or lower) loss the gap persists and grows, at both scales (a 350M pair keeps a ~12–25× stable-rank gap, p=7e-15)—so it tracks update geometry, not training progress. The fingerprint has a concrete deployment consequence: it sets a checkpoint’s compression budget. Muon-lineage weights are 5–7× less low-rank-truncatable and ~29× less amenable to 4-bit quantization than Adam-lineage weights (again robust to the iso-loss control), so the pretraining optimizer—not just the architecture and final weights—determines how cheaply a base model can be compressed. Optimizer lineage is thus a measurable, causal, and deployment-relevant property that deserves to travel with a checkpoint as first-class metadata.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "optimizer", "Muon", "Adam", "weight spectrum", "singular values", "stable rank", "model compression", "quantization", "low-rank", "science of deep learning" ], "forum_url": "https://openreview.net/forum?id=YMkHmZjJ6B", "pdf_url": "https://openreview.net/pdf/8efb370291694ef58e024affc87d571c925f6b1e.pdf", "updated_at": "2026-10-05T20:21:00.674Z", "categories": [ "training", "hardware" ], "matched_terms": [ "4-bit", "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fmZUFWjHbE", "number": 33226, "title": "A Trail from Low-Precision Sampling: Diagnosing Super-Diffusive FP32--BF16 Drift in Diffusion Models", "abstract": "Low-precision inference is now the default deployment path for diffusion image generators: BF16 is standard and FP8 is spreading fast. The rounding error this introduces is usually treated as small, zero-mean noise that averages out over the tens of denoising steps of a sampler. We show it does not average out. Running SDXL from an identical seed in FP32 and BF16, we find that the same-seed latent trajectories do not wander apart like independent noise: their per-step L2 divergence grows super-diffusively, with a log–log growth exponent of 1.04 (diffusive random-walk null = 0.5) and a locatable takeoff around step 25 of 50. This is a robust same-seed phenomenon (n=64 pairs), and we pin down its cause with an unusually thorough elimination: five controls that rule out the tempting statistical explanations one at a time. It is not a mean bias (sign-bias statistic |b| ≈ 0.012). It is not the error’s step-to-step directional correlation: although consecutive error vectors are nearly collinear (cos(e_t, e_{t+1}) = 0.971), that merely rides the trajectory’s own velocity autocorrelation (0.979), and a magnitude-matched but uncorrelated Gaussian perturbation injected afresh at every step reproduces the super-diffusion (exponent 1.01 ≈ 1.04). And it is not a fixed-magnitude dynamical amplification: the same continual injection at constant per-step magnitude is merely diffusive (0.43), while a one-time kick is damped (0.19). What remains is a growing-magnitude positive feedback: the per-step rounding error scales with the accumulating separation, so as the same-seed trajectories diverge the error grows and drives them further apart—faster than √t, independent of the error’s mean or correlation. Crucially, we do not stop at inferring this feedback but measure it directly: regressing the per-step divergence increment on the current divergence is decisively positive (Spearman 0.89, positive for all 64 of 64 prompts). We package these measurements as DriftProbe, a training-free diagnostic—usable on any UNet or DiT block—that reports the per-step sign bias, the same-seed divergence curve and its onset t*, and the injection controls that isolate the mechanism. We treat the drift as a lens, not a defect to correct, and lay out how t* could serve as an early signal for downstream generation defects. Results are measured on SDXL/BF16 on a single 48 GB GPU (n=64 same-seed pairs; n=16 per injection control); the FP8 arm, the FLUX backbone, and the defect-prediction evaluation are scoped as future work.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion models", "low precision", "bfloat16", "numerical error", "sampling", "SDXL", "super-diffusion", "inference", "interpretability" ], "forum_url": "https://openreview.net/forum?id=fmZUFWjHbE", "pdf_url": "https://openreview.net/pdf/9fdd6acc24ec238a67d2026c44c054153f3c8418.pdf", "updated_at": "2026-10-05T20:21:01.643Z", "categories": [ "ptq", "training", "hardware", "diffusion" ], "matched_terms": [ "fp8", "low precision", "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "C8L2nuzZpB", "number": 33227, "title": "Precision-Aware Optimal Newton--Schulz Iterations for the Matrix Sign", "abstract": "The Muon optimizer orthogonalizes each gradient matrix by running a short quintic Newton–Schulz (NS) iteration that drives its singular values to one—an approximate matrix sign function, msign(M) = UV^T. Recent work (Polar Express) shows that the per-step coefficients of this iteration should not be hand-tuned but solved for: a per-step minimax (Remez) fit yields provably fastest convergence in exact arithmetic. But Muon actually runs the iteration in bfloat16, and no prior analysis characterizes these learned iterations under low precision. We give a precision-aware characterization of the coefficient design space, deliberately scoped to the idealized scalar reduction of the iteration—the singular-value band [1/κ, 1] with per-scalar rounding—and bounded honestly against real gradients and real training. We study the sign iteration as an optimal–structured–hybrid family—P1, a free per-step minimax fit (Polar Express); P2, the same minimax under a boundedness constraint; and P3, a hybrid that runs free steps early and bounded steps late—and measure error-versus-step curves in fp32 and bf16 across κ ∈ {10, 30, 100}. Three findings. (i) On the idealized band, the fp32 per-step-optimal P1 crushes fixed Muon at a matched budget (1e-8 vs. 0.318 at κ = 10; 7.8e-6 vs. 0.318 at κ = 100). (ii) On that same idealized band, bf16 is fragile at high conditioning: at κ = 100 the aggressive P1 coefficients make the bf16 scalar iteration diverge (terminal 4×10^18; P2 overflows to +∞), while the hybrid P3 stays finite (bf16 floor 0.141, beating fixed’s 0.328); we give a first-order model of this floor. (iii) This bf16-specificity is a feature of the idealized band, not of real gradients. Running the true matrix iteration on real 124M gradient matrices (fp32-accumulate tensor cores), the optimal P1 coefficients are fragile in both precisions—blowing up on 8/8 matrices in bf16 and 7/8 in fp32, so the clean bf16-only divergence appears for just 1/8—while fixed Muon stays stable (residual ≈ 0.32). And the deployed iteration is bf16-safe end-to-end: a nanoGPT-124M premise test trains equally well in bf16 and fp32 (a null, −0.038 nats). Our contribution is thus a precision-aware map of the coefficient design space in the idealized setting, honestly bounded: the optimal-coefficient instability does not clearly transfer to real gradients bf16-specifically, and today’s Muon is already bf16-safe—not a fix for a real deployment problem. Extending precision-aware optimal iterations to general spectral functions is left to future work.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Muon", "Newton-Schulz iteration", "matrix sign", "polar decomposition", "low precision", "bfloat16", "minimax approximation", "optimization" ], "forum_url": "https://openreview.net/forum?id=C8L2nuzZpB", "pdf_url": "https://openreview.net/pdf/d7e5dbcf0d9f94e880b2952c56d7e5c3cac2e889.pdf", "updated_at": "2026-10-05T20:21:01.699Z", "categories": [ "ptq", "training", "hardware", "theory" ], "matched_terms": [ "low precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dqbqIE4Ujm", "number": 33230, "title": "RECAP: Refine Quantized LLM Checkpoints, Don't Replace Them", "abstract": "When a deployed quantized large language model requires higher accuracy, transmitting a complete replacement checkpoint repeats information already available at the client. We formulate successive model refinement as encoding an incremental correction with the cached checkpoint as shared side information. The central challenge is to recover model quality with an incremental correction substantially smaller than a replacement checkpoint. We introduce RECAP, which combines two correction representations with different communication costs. An activation-subspace code transmits coefficients in a basis reconstructed at both endpoints from the cached model and public calibration data, with water-filling allocating bits by activation-weighted coefficient energy; a complement code transmits low-rank factors to correct the remaining residual. Our analysis establishes the shared basis's minimax optimality, derives optimal coefficients and approximate rate allocation, and quantifies the complement's benefit. We ask whether a correction can deliver at least the quality of the next-higher-bit replacement checkpoint while transmitting fewer bytes. From four 1.58-bit checkpoints of 7B models, RECAP reaches the perplexity of same-family 2-bit replacements with 2--19\\% of their bytes; at a fixed budget, it also achieves lower perplexity than conventional reconstruction baselines at both 7B and 13B scale. These findings show that even a checkpoint with poor standalone performance can remain valuable side information for communication-efficient model refinement. Code is available \\href{https://anonymous.4open.science/r/RECAP-7F3A/README.md}{here}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "successive model refinement", "post-training quantization", "LLMs", "edge LLMs", "low rank decomposition" ], "forum_url": "https://openreview.net/forum?id=dqbqIE4Ujm", "pdf_url": "https://openreview.net/pdf/7e52ec729936c93edc73179eeb57bc12249dfa63.pdf", "updated_at": "2026-10-05T20:21:02.188Z", "categories": [ "ptq" ], "matched_terms": [ "1.58-bit", "2-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "e3f4YZDEyV", "number": 33286, "title": "Best Arm Identification in Generalized Linear Bandits with 1-Bit Feedback", "abstract": "In many real-world applications such as IoT sensor networks, digital\nmarketing, and clinical trials, decision-making often operates under\nstrict 1-bit communication constraints and binary feedback (e.g.,\nsuccess/failure, click/no-click). Generalized Linear Models (GLMs) have\nbeen shown to be more efficient than linear models for binary data and\nare used effectively in contextual settings where decisions are driven\nby a multitude of high-dimensional variables. However, the existing\nliterature on Best Arm Identification (BAI) in GLMs has focused almost\nexclusively on full-precision rewards, leaving a critical gap in\ncommunication-constrained settings. In this paper, we study Best Arm\nIdentification in Generalized Linear Models with 1-bit observations,\nwhere the latent reward follows an exponential-family distribution but\nthe learner observes only thresholded binary feedback. To the best of\nour knowledge, this is the first work to formulate and solve BAI in the\n1-bit GLM setting. Our algorithm constructs confidence sets directly\nfor the pairwise reward gaps between arms, formulated in terms of the\nfeature matrices $V$ and an $m$-dimensional parameter vector $\\theta$.\nUnder standard regularity assumptions, we prove that the proposed\nalgorithm is $(\\epsilon, \\delta)$-PAC.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "generalized linear bandits", "best-arm identification", "1-bit feedback", "Fisher Information Matrix", "exponential family" ], "forum_url": "https://openreview.net/forum?id=e3f4YZDEyV", "pdf_url": "https://openreview.net/pdf/db48dd6e13efb78dd0889e9cc4e42357ea0a5c6b.pdf", "updated_at": "2026-10-05T20:21:08.471Z", "categories": [], "matched_terms": [ "1-bit", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lJj9QgYW2w", "number": 33371, "title": "ESPTQ: Post-Training Ternarization of Weights and KV Cache for Eliminating Per-Element Multiplications", "abstract": "Reading the weights and KV cache dominates large language model inference. Ternary quantization to {-1, 0, +1} shrinks both and removes per-element multipliers: each weight product becomes a sign flip or skip, so a signed adder tree suffices. Ternary weights alone, however, leave multipliers in attention's $QK^\\top$ and $PV$, whose operands are both run-time floating-point tensors.\n\nWe present ESPTQ (element-sparse-plane ternary post-training quantization), which quantizes the weights and KV cache into one ternary storage format, removing per-element multipliers from every matrix product inside the Transformer layers except the rotation of layer inputs. By design, ESPTQ caps the residual multiplication ratio at the weight-side budget of two multiplications per 128-element reduction ($2/128 = 1.5625\\%$). For weights, after a learnable butterfly rotation within each 128-column group, a rate-distortion criterion selects, at a continuously variable second-plane assignment ratio $\\rho$, which elements receive a sparse second ternary plane. For the KV cache, static tables replace per-token offsets and the V-side scaling factor, meeting the same budget.\n\nWe compare methods on effective bits per weight, serializing each in its own storage format with losslessly compressed discrete codes, so scaling factors, rotation angles and other auxiliary metadata count. On LLaMA-2-7B, ESPTQ matches TWLA in perplexity and average zero-shot accuracy at less than half the effective bits, and at comparable bits lowers perplexity by 36% against PT2-LLM. Ternarizing the KV cache as well shrinks it to about one fifth of its FP16 size, at less cost in perplexity and zero-shot accuracy than ternarizing the weights.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "ternary quantization", "KV cache quantization", "large language models", "model compression", "multiplication-free inference", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=lJj9QgYW2w", "pdf_url": "https://openreview.net/pdf/b9e8bcc76cb6d8b9e39e6c7ba6825d580849223a.pdf", "updated_at": "2026-10-05T20:21:18.968Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization", "quantizes", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nVAjpjQ5y1", "number": 33474, "title": "Configuration-Valued Expert Importance for Mixed-Precision Quantization of Multimodal Mixture-of-Experts", "abstract": "Sparse activation does not eliminate the expert-weight storage cost of\nmultimodal Mixture-of-Experts (MoE) models, motivating mixed-precision\nquantization under memory constraints. However, existing importance-based\napproaches often rely on isolated expert or projection scores, overlooking\nthe context dependence of expert importance, the natural hierarchy between\nexperts and their internal projections, and the coupled quantization effects\nwithin each expert. We propose Config-Owen, a hierarchical game-theoretic\nframework for mixed-precision MoE quantization. Config-Owen formulates\nexpert-level precision allocation as a cooperative game, uses Owen-style\nattribution to preserve the expert--projection hierarchy, and incorporates\nwithin-expert Harsanyi interactions to capture coupled quantization effects,\nyielding configuration-valued utilities for precision allocation. We\nefficiently estimate these utilities and optimize the resulting precision\nassignment under a global memory budget. Across four multimodal MoE models\nand Mixtral-8$\\times$7B, Config-Owen consistently improves low-bit performance\npreservation; on DeepSeek-VL2-S at 2.58 bits, it retains 98.39\\% of the FP16\nfive-task average score. Code will be released upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts; Mixed-Precision Quantization; Post-Training Quantization; Multimodal Large Language Models; Model Compression;" ], "forum_url": "https://openreview.net/forum?id=nVAjpjQ5y1", "pdf_url": "https://openreview.net/pdf/cc31880531559fda8451bcc54dc2b7e7c336b48b.pdf", "updated_at": "2026-10-05T20:21:31.473Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "low-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8bCJW8E9CL", "number": 33506, "title": "ReplaySSM: Cache SSM Inputs, Not State", "abstract": "State-space models (SSMs) maintain a fixed-size recurrent state, but reading and writing it at every decoding step can bottleneck hybrid models. This eager state management causes large state I/O with low arithmetic intensity, leaves speculative decoding no cheap rollback, and feeds state quantization error into subsequent recurrent updates. We introduce ReplaySSM, which unrolls the recurrence to represent the state as a checkpoint and a buffer of recent SSM inputs. This separates output computation from checkpoint updates. By updating the checkpoint periodically, ReplaySSM avoids state writes on most steps and roughly halves the dominant state memory traffic. Keeping recent inputs also makes speculative rollback a buffer-pointer operation and enables output-only verification without materializing intermediate states. The same separation allows a low-precision checkpoint to serve output computation on most steps while a higher-precision checkpoint serves recurrent updates, so rounding error in the low-precision copy never enters later updates. We implement ReplaySSM in vLLM for Mamba-2 and Gated DeltaNet. At serving batch size 256, it accelerates standard decoding by $1.20$–$1.47\\times$ over a tuned vLLM baseline across four hybrid models, and delivers about $1.8\\times$ the throughput of vLLM's speculative decoding on the two 120B-class models. Separate state precisions also reduce output KL divergence and downstream accuracy losses compared with a shared low-precision checkpoint.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "State Space Models", "Hybrid Models", "LLM Inference", "Linear Attention", "Speculative Decoding", "Quantization" ], "forum_url": "https://openreview.net/forum?id=8bCJW8E9CL", "pdf_url": "https://openreview.net/pdf/d990d1cf468df84ea77cd74411303bd31aba4b6f.pdf", "updated_at": "2026-10-05T20:21:34.718Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gpqH98THyv", "number": 33525, "title": "Hardware-Aware Language Model Quantization For A Many-Core AI Inference Accelerator", "abstract": "Edge accelerators for ML/AI workloads are designed around trade-offs between peak performance, memory capacity and bandwidth, power, and cost, with vendor economics driven by volume rather than per-unit margin. Because deployed silicon cannot be refreshed on the cadence of model releases, gains must increasingly come from the algorithmic side: pushing out the Pareto frontier of accuracy, memory footprint, and throughput within the constraints of hardware already in the field. Given the extensive work on post-training quantization, we choose it as the primary compression lever and treat it as a full inference-stack optimization problem, jointly optimizing accuracy, storage, and throughput rather than minimizing quantization error alone. We account for hardware features such as supported data formats, computational units, and the instruction set architecture, alongside structural constraints imposed by the runtime and compiler. Beyond the formats natively supported by the hardware, we treat the format design space itself — element and scale datatypes, group size, and hierarchical scales — as part of the optimization process, covering standardized block formats such as MXFP and NVFP as well as non-standard constructions. Algorithmically, we evaluate mixed-precision allocation and codebook-based quantization across dense and mixture-of-experts models from 4B to 35B parameters. We share with the wider community insights on what works well on paper and what actually improves the Pareto frontier across accuracy, memory footprint, and throughput for models deployed on edge silicon. We share with the wider community insights and trade-offs on what works well on paper and what actually improves the Pareto frontier across accuracy, memory footprint, and throughput for a model deployed on edge silicon. We demonstrate our hardware-aware quantization methodology on a many-core edge accelerator with over a thousand RISC-V cores, achieving 34\\%, 67\\%, and 24\\% decode speedups over the llama.cpp K-quant of equal storage at 4.3$\\times$, 3.7$\\times$ and 3.3$\\times$ weight-memory reduction over BF16, within one point of its mean accuracy, on Qwen3.5-35B-A3B, Qwen3.5-4B, and Llama-3.1-8B respectively.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "mixed-precision quantization", "hardware-aware quantization", "edge accelerators", "mixture-of-experts", "llama.cpp" ], "forum_url": "https://openreview.net/forum?id=gpqH98THyv", "pdf_url": "https://openreview.net/pdf/28e21f3fdabed4b728276fdd4f845df66722b3ad.pdf", "updated_at": "2026-10-05T20:21:38.014Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gF07Eukddi", "number": 33554, "title": "BlockFAMQ: Realized-Residual Curvature for Complete-Checkpoint Mamba Quantization", "abstract": "Mixed-precision quantization must decide which realized rounding residuals receive scarce high-precision bytes, but diagonal sensitivity scores discard the signed coordinate interactions that distinguish deterministic candidates. We introduce BlockFAMQ, a matrix-free allocator that scores squared per-token gradient–residual products and solves the resulting integer-byte multiple-choice problem under a complete-checkpoint budget. Across Mamba-1 130M, 370M, 1.4B, and 2.8B, the mean BlockFAMQ-minus-diagonal WikiText-2 NLL gap contracts from −0.218 at 2.5 effective bits to −0.026 at 4.0 bits and −0.009 at 4.5 bits. At 3.5 bits, BlockFAMQ improves six-task accuracy by 2.08 points and recovers 65.7–80.1% of the diagonal quantization gap; an activation-weighted reconstruction baseline narrows but does not remove the advantage. At 3.0 bits, direct candidate KL improves on BlockFAMQ by 0.038 nats across 1.4B and 2.8B, while a feasibility-anchored block-to-KL shortlist recovers 0.033 nats at 1.62× the block-scoring cost on average. Precision shifts toward x_proj and in_proj, whose realized residuals have the largest cross-term tails, and the block effect is larger on Mamba-1 1.4B than on matched-scale Pythia-1.4B. The direction persists on Mamba-2 2.7B and Falcon-Mamba 7B; packed W4A8 composition retains 96.8–97.7% of Quamba2 decode throughput while lowering NLL. These results support realized block curvature as a budget-dependent allocation correction, with direct KL becoming more useful when aggressive quantization leaves the local quadratic regime.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "mixed-precision quantization", "state-space models", "Mamba", "empirical Fisher information", "model compression" ], "forum_url": "https://openreview.net/forum?id=gF07Eukddi", "pdf_url": "https://openreview.net/pdf/2f877655a959f34846a5c3902bf83abaed3bbded.pdf", "updated_at": "2026-10-05T20:21:43.355Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "mixed-precision", "quantization", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NVocDpVffz", "number": 33608, "title": "What Survives Extreme Compression? Vision-Language Embedding Alignment at Microcontroller Scale", "abstract": "We ask what remains of CLIP’s image - text embedding geometry when ViT-B/32’s 87.8 M-parameter image tower is distilled into a 724 K-parameter microcontroller encoder whose class prototypes the teacher’s text encoder computes offline. Distilling on CC3M, we find a sharp capacity - granularity boundary: the student keeps 39% of teacher accuracy on COCO with classes weighted equally, 12% on ImageNetV2, 5 - 9% on fine-grained sets and near chance on two robotics sets, a boundary that tracks pretraining coverage. It spans ≈93 effective dimensions where the teacher needs ≈277, and a stronger teacher adds up to 3.2 points at the deployed dimension at no device cost. One coordinate of the teacher is its prototypes’ shared direction: CLIP ViT-B/32’s coordinate 133 carries 79% of the mean prototype’s squared norm and enters the prefix only at 256 dimensions, where it adds ≈6 per-image COCO points, all on “person”, while class-balanced accuracy falls; projecting it out lifts class-balanced 256d above 128d on COCO and ImageNetV2. Alignment is also teacher -specific: prototypes from any other text encoder score at chance, but a closed-form map fitted on class names alone recovers 64-71% of accuracy offline. The INT8 student runs at 2.5 FPS on an ESP32-S3 and fits an STM32H7 in 1,085 KB flash (464 KB arena, emulated); its 128 px input costs eleven points of top-1 at ten COCO classes (37.9% vs. 48.8%).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "vision-language models", "knowledge distillation", "model compression", "Matryoshka representation learning", "embedding geometry", "rogue dimensions", "zero-shot classification", "on-device inference", "anisotropy / rogue dimensions", "TinyML", "microcontrollers" ], "forum_url": "https://openreview.net/forum?id=NVocDpVffz", "pdf_url": "https://openreview.net/pdf/4cfd768e7c3073a9a09e9a47cc14e1820f16fc83.pdf", "updated_at": "2026-10-05T20:21:51.419Z", "categories": [ "multimodal" ], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fluIqGRwZa", "number": 33808, "title": "Compression order in spiking neural networks: normalization and model construction", "abstract": "Compression-order comparisons depend on how models are constructed and calibrated. We study channel pruning, unfolded weight-only quantization and threshold scaling in spiking neural networks (SNNs), without intervening training. Matched-state controls distinguish batch-normalization (BN) history from model construction: aggressive clipping without rounding reproduces large history-dependent gaps, including in ReLU networks, and common recalibration yields identical logits. Separately, a prespecified twelve-run CIFAR-100 ResNet-SNN comparison at two time steps gives a mean absolute accuracy gap of 0.17 percentage points (pp) on reused test images, despite 6.0\\% prediction disagreement under the specified construction. Both orders satisfy the prespecified operational 5-pp budget on evaluated validation and test sets, with 10.08\\% fewer convolution multiply--accumulate operations. Descriptive extensions include matched T=4 retraining, additional image and event-frame datasets, and Transformer MLPs. Calibration improves both-order budget passage in the complete ImageNet-100 comparison; full reconstruction changes one finite-split qualification decision that final-only recalibration preserves. In a descriptive comparison of twelve T=2 ResNets at $\\alpha$=1.02, the mean between-order relative synaptic-operation range contracts from 12.48\\% to 0.031\\% after common recalibration, with all orders remaining budget-feasible. Retrospective candidate policies on reused validation splits quantify compression choices and test-budget violations. These observations distinguish normalization history, model behavior, budget feasibility and activity-dependent work in compression-order evaluation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Interpreting Compression-Order Effects in Spiking Neural Networks" ], "forum_url": "https://openreview.net/forum?id=fluIqGRwZa", "pdf_url": "https://openreview.net/pdf/28133c8c4aa4ccfbfd1e661ee1b7dd1eb19edf0f.pdf", "updated_at": "2026-10-05T20:22:15.773Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "t9m2mBoay8", "number": 33877, "title": "Joint Branch-Space Transform Coding for Diffusion Activation Quantization with Classifier-Free Guidance", "abstract": "Post-training quantization for diffusion models increasingly exploits timestep, feature, and layer structure. While recent work has begun incorporating CFG structure into diffusion quantization, activation quantization still operates independently across conditional and unconditional coordinates, leaving cross-activation structure unexploited. We show that matched CFG activations form a strongly correlated two-dimensional source and that, under a fixed bit budget, the choice of branch coding basis materially affects quantization fidelity. Motivated by this observation, we introduce $\\textbf{branch-space transform coding}$, which rotates matched CFG branches via an offline derived $2\\times2$ orthogonal matrix, requiring minimal modifications to model parameters or the quantization pipeline. We further derive the $\\textbf{Guidance-Correlation Branch Transform (GCBT)}$, which jointly incorporates the CFG guidance direction and cross-branch second moments. Under an equal-rate quantization-noise surrogate, GCBT admits a closed-form per-layer solution without gradient optimization or angle search. Applied on top of existing diffusion PTQ methods, GCBT yields statistically significant fidelity gains in most evaluated comparisons with no statistically significant degradation, while leaving the underlying host quantization pipeline unchanged.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion models", "classifier-free guidance", "quantization" ], "forum_url": "https://openreview.net/forum?id=t9m2mBoay8", "pdf_url": "https://openreview.net/pdf/481aea53d7653ff9db455099458519b09f8efaab.pdf", "updated_at": "2026-10-05T20:22:22.593Z", "categories": [ "ptq", "training", "diffusion" ], "matched_terms": [ "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UubS9jd0J4", "number": 33938, "title": "Sensitivity-Guided Mixed-Precision Quantization for Large Language Models", "abstract": "Post-training quantization (PTQ) is essential for deploying large language models (LLMs) on resource-constrained hardware, but its impact on specific capabilities is poorly characterized. We present a capacity-controlled study of how weight-only quantization affects multi-step reasoning versus factual recall and commonsense inference, across six models (0.5B–8B; Qwen2.5, Qwen3, and Phi-3.5 families). We introduce the Reasoning Fragility Ratio (RFR), a diagnostic metric that quantifies capability-asymmetric degradation, and SGMPQ, a mixed-precision scheme that assigns 8-bit precision to the layers ranked most quantization-sensitive by a perturbation probe, with a capability-conditioned ratio as the ranking signal. Our evaluation uses real benchmarks (GSM8K, ARC-Challenge, HellaSwag, held-out perplexity) with Wilson confidence intervals and exact McNemar tests, and compares SGMPQ 8/4 (≈5.2 average bits) against capacity-matched controls: uniform W5/W6, a random mixed-precision control at the identical budget (replicated over five allocation seeds), and—decisively—two capability-agnostic sensitivity rankings at the same budget, plus GPTQ and AWQ at matched budgets. Three findings emerge. First, the quantizer dominates where it matters: GPTQ beats RTN by up to 63 accuracy points exactly where RTN breaks (W4 on the 3B base and Coder-3B models), rescuing Qwen2.5-Coder-3B, which RTN destroys outright, while the two quantizers converge on robust models at W5. Second, on robust models the bit budget masks placement: at equal ≈5.2 bits, sensitivity-guided placement is statistically indistinguishable from random placement on the two small instruct models, worse on the 3B base model, and better on the two largest models—though with five allocation seeds the permutation design (sign-flip floor p=0.0625) cannot reach conventional significance even if the advantage is real on every seed. Third, placement matters exactly where uniform quantization fails: on Coder-3B, both perturbation-probe-guided allocations tested preserve the model (59.4–64.2% GSM8K) where all five independently drawn random allocations collapse (0.4–2.4%; uniform W5 also rescues it), and the capability ratio itself is beneficial on one model (8B: +6.0 pp) but actively harmful on another (3B: -33.4 pp vs. the agnostic ranking). We release per-item results, sensitivity maps, and the full evaluation stack, and conclude that sensitivity-guided mixed precision is a targeted repair tool for quantization-brittle models—but no ranking signal we tested (capability ratio or agnostic) transfers safely across models, so any allocation must be validated against random-placement controls on the target model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model quantization", "mixed-precision", "reasoning", "model compression", "large language models" ], "forum_url": "https://openreview.net/forum?id=UubS9jd0J4", "pdf_url": "https://openreview.net/pdf/b075e68682d1db26a1ef37c5e4b9c016dee6206f.pdf", "updated_at": "2026-10-05T20:22:28.509Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "8-bit", "mixed precision", "mixed-precision", "ptq", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GPya82gZy7", "number": 34007, "title": "Whose Word Survives Compression?", "abstract": "Post-training quantization can preserve answer accuracy while changing the computation that selects between retrieved context and parametric memory. We test causal transfer of source-arbitration interventions under GPTQ INT4 with JuICE, a dual-run attention-head intervention, on Llama-3.1-8B and Gemma-2-9B base. A question-aware matcher audit reveals that apparent transfer can be a same-entity wrap artifact rather than a transferred causal object. On Llama, the original-matcher (un-audited) context-K10 transfer (+15 percentage points (pp), 95% CI [+8, +23]) is one such wrap: the audited number is +1 [+0, +3] (null), with 17 of 17 original INT4 flips naming the same context entity. The locked Llama parametric pair (clean K2) remains detected at n=100 on both precisions under gold-PAR-only counting (−8 [−14, −3] on FP16; −7 [−12, −2] on INT4), confirmed head-specific by a matched random-K2 control. On Gemma, identified context-suppression H− remains detected versus a locked random-K control but is attenuated on INT4 (inclusive −19 [−27, −11] vs. −1 [−5, +2] on FP16; −11 [−18, −4] vs. −1 [−5, +3] on INT4); context-K10 enlarge is a small real effect (+7 [+2, +12] / +5 [+1, +9]) but not INT4-specific, and the original-matcher (un-audited) raw peak L27_H6 is dead as a wrap artifact. Conversely, Gemma's clean parametric pair attenuates to an in-cell transfer null on INT4 despite a detected FP16 effect—the mirror image of Llama's result. A same-sign ∆ is therefore not evidence of transfer without question-aware matching, entity confirmation, and a matched random-head control.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mechanistic interpretability", "model compression", "post-training quantization", "GPTQ", "knowledge conflicts", "source arbitration", "attention heads", "causal intervention" ], "forum_url": "https://openreview.net/forum?id=GPya82gZy7", "pdf_url": "https://openreview.net/pdf/bb3e3775f060e91c66eff7547ea65c138c699a41.pdf", "updated_at": "2026-10-05T20:22:38.796Z", "categories": [ "ptq" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5aH1G4WwDg", "number": 34068, "title": "Curvature-Corrected Activation Gradients for Binarized Neural Networks", "abstract": "Binarized Neural Network (BNNs) training typically relies on the straight-through estimator (STE) as a practical approximation of the derivative of the quantization layer applied to the model weights and activations. \nWe identify that the activation gradients in standard STE are biased and that this bias can be decomposed into two components: one related directly to the STE and another related to applying the classical chain rule through a discontinuous quantizer, which propagates the tangent of the downstream loss at the selected quantization level.\nIn contrast, the distributional derivative of the population loss is governed by the finite loss jump between adjacent levels.\nA second-order expansion of this jump yields a gradient-correction term involving the quantization residual and the diagonal curvature of the downstream loss.\nWe turn this result into a practical backward rule for activation quantizers by approximating the curvature component with an empirical-Fisher proxy and apply the proposed correction during the final 10\\% of training.\nWe evaluate our method on binarized Llama-style language models trained with recent state-of-the-art methods, including QuEST and CAGE, and observe a 1.0--3.1\\% perplexity reduction across model sizes up to 1.6B parameters.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "aware", "training", "language", "model", "curvature", "correction", "gradient", "activation" ], "forum_url": "https://openreview.net/forum?id=5aH1G4WwDg", "pdf_url": "https://openreview.net/pdf/a581ae4abcdef40d668768212c4b970283a05ca9.pdf", "updated_at": "2026-10-05T20:22:50.490Z", "categories": [ "qat", "training" ], "matched_terms": [ "binarized", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cNmQQbumsi", "number": 34072, "title": "Quantization for Reasoning Preservation in Small LLMs", "abstract": "Large Language Models (LLMs) have demonstrated exceptional capabilities in textual inference, but their practical deployment is constrained by high computation and memory overhead. To address this, we propose LLM-QRP (Quantization for Reasoning Preservation), a layer-aware, mixed-precision quantization framework designed to reduce model memory footprint while preserving multi-step reasoning capabilities. Unlike standard uniform post-training quantization methods that treat all parameters equally, LLM-QRP identifies and protects critical \"thinking\" sub-components by analyzing layer-wise Self-Logits Evolution Decoding (SLED) and entropy dynamics across the residual stream. These signals are unified via Principal Component Analysis (PCA) into a data-driven reasoning importance score. We then formulate precision allocation as a 0-1 Multiple-Choice Knapsack Problem solved via Integer Linear Programming (ILP) to assign higher bit-widths to high-sensitivity components under strict VRAM targets. Experiments on sub-500M parameter models across GSM8K, TruthfulQA, and MMLU demonstrate that LLM-QRP achieves an optimal Pareto frontier between model compression and task accuracy, maintaining near-baseline reasoning performance while outperforming uniform 8-bit baselines and alternative quantization methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "mixed-precision quantization", "large language models", "model compression", "reasoning preservation", "chain-of-thought", "post-training quantization", "edge deployment" ], "forum_url": "https://openreview.net/forum?id=cNmQQbumsi", "pdf_url": "https://openreview.net/pdf/9a42a9864e63bccf1f64a4fb1b43393e657d549b.pdf", "updated_at": "2026-10-05T20:22:51.248Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "8-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jnPNPh46Zk", "number": 34346, "title": "MaskAhead: Unifying KV Cache Selection and Eviction in Block Diffusion Language Models", "abstract": "Block diffusion language models decode one block at a time against an exact key-value cache of the prefix, which grows with the prompt and with the model's own long generations.\nA small cache requires two decisions: which prefix entries the current block should read, and which entries later blocks may need. Selection is temporary. Eviction permanently removes entries from storage.\nSparse-read methods cut repeated cache reads but keep the full prefix resident. Eviction reduces storage, but the decision is irreversible: entries removed now cannot serve later blocks. Block diffusion exposes masked positions in upcoming blocks before generation, so a no-write probe can obtain model-derived queries for retention without producing draft tokens.\nMaskAhead assigns the model's mask queries to the decisions they can inform: current-block queries rank temporary reads, while a no-write probe of upcoming blocks ranks the entries retained until the next eviction. Both decisions use an output-sensitive score that accounts for values. The surviving cache can also be stored at four bits.\nOn MATH-500, Q-MaskAhead reduces peak KV by 75\\%, with a controlled accuracy degradation. For long-prompt QA, question-conditioned prefill reduces peak KV by 86.7\\% on average across HotpotQA, NarrativeQA, and MuSiQue.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language models", "KV cache", "quantization", "selection", "inference" ], "forum_url": "https://openreview.net/forum?id=jnPNPh46Zk", "pdf_url": "https://openreview.net/pdf/3be1e0da3d15ba8ed195fac1d1f1c23ff435bf62.pdf", "updated_at": "2026-10-05T20:23:29.069Z", "categories": [ "kv-cache", "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JcQlSCOUlH", "number": 34396, "title": "Looped Latent Attention: Cross-Loop Key-Value Compression For Looped Transformers", "abstract": "Looped, weight-tied Transformers reduce parameters by reusing a single block, but decoding still stores a\nseparate key/value (KV) cache for every recurrence step. We show this loop-indexed cache is highly structured:\nfor a fixed token, layer and head, the per-loop K/V vectors trace a short low-rank trajectory across loops, while\nthe head and layer axes stay much flatter. We introduce Looped Latent Attention (\\lla{}), a post-training cache\ncodec that stores compact K and V latents and reconstructs loop-specific vectors only when attention reads them.\nThe codec is initialized from the SVD of teacher activations and refined by logit and attention-output\ndistillation; an \\lla{}-2D variant folds heads into one latent for the extreme-compression regime. At matched\ncache budget, per-head \\lla{} beats head-axis MLA, cross-layer sharing, KV quantization and zero-parameter\nfinal-loop reuse; the last result shows that an unadapted endpoint cannot replace the loop trajectory. The advantage\nholds on Ouro-2.6B-Thinking and transfers to Huginn-3.5B, where an SVD codec is near-lossless to $32\\times$ on\ndecoder-independent fidelity. The same full-RoPE codecs that produce the accuracy results also serve: with a\nfused latent-read kernel, \\lla{} decodes end to end on one B200 at $1.2\\times$ the teacher's fixed-batch throughput\nand $2.6$--$5.9\\times$ its peak throughput while caching $4$--$21\\times$ less, with decoder-independent quality\nunchanged. Serving through the streaming latent store exposes a train/serve mismatch, which we trace to drift of the\nstored K/V of generated tokens and remove by fitting the codec under the served semantics with an exact 64-token\nwindow: served GSM8K rises from $0.31$ to $0.67$ at $3.8\\times$ on Ouro-1.4B (oracle re-forward $0.79$) and from\n$0.41$ to $0.84$ on Ouro-2.6B, matching the oracle, with the same fused kernel. An absorbed variant, LLA-D, reads latents without reconstruction and reaches $21\\times$ the teacher's peak throughput. On\nstudent-generated prefixes, on-policy refinement raises MATH-500 at $4\\times$ from 0.43 to 0.64--0.66 across two\nruns and reduces no-answer generations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "looped transformers", "latent attention", "KV cache reduction" ], "forum_url": "https://openreview.net/forum?id=JcQlSCOUlH", "pdf_url": "https://openreview.net/pdf/b686bbee80b35af3cd44e1f495f456f62a5a25e9.pdf", "updated_at": "2026-10-05T20:23:37.235Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "aJHDXSNKV8", "number": 34505, "title": "Decodable But Not Detachable: Training Data Granularity Determines Parametric Modularity in Large Language Models", "abstract": "Do large language models contain domain-specific parametric shells:\nconcentrated, causally necessary neuron populations whose removal selectively\ndegrades a target domain while sparing others? We apply a uniform causal\nmethodology across two domain granularities, three model families (1.5B to 7B\nparameters), and eight domains. At the academic subject level, zero neurons\nexceed 60% domain selectivity across 939,008 combined FFN neurons and\ncausal damage matrices are flat, despite domain identity being linearly decodable\nabove 85% accuracy. At the language and modality level, 0.65–1.14% of neurons\nexceed 60% selectivity, damage matrices are near-perfectly diagonal (ratios up\nto 595:1), and shell neuron sets are essentially disjoint (IoU < 0.003). Masking\ncode-selective neurons reduces mathematical reasoning accuracy by 16–24\npercentage points across all models; chain-of-thought prompting recovers up\nto 75% of this damage, revealing that the code shell carries both an answer-\nstructuring function and an irreducible reasoning function. Masking Spanish or\nChinese neurons leaves mathematical reasoning at or below random. Shell causal\nstrength increases monotonically with model scale, shells are spatially interleaved\nin a pattern that precludes group-level selective quantization, and amplification of\nshell neurons degrades performance monotonically, confirming that shells operate\nat a precisely calibrated functional point. Parametric shells form where and only\nwhere training data was modular at the token level.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "parametric modularity", "language-specific neurons", "domain selectivity", "causal intervention", "multilingual representations", "model interpretability", "FFN neurons", "knowledge localization", "training data distribution", "chain-of-thought reasoning", "model compression", "granularity", "neuron masking" ], "forum_url": "https://openreview.net/forum?id=aJHDXSNKV8", "pdf_url": "https://openreview.net/pdf/0434baf912bb2cebd684425bee69a9ca69f68d56.pdf", "updated_at": "2026-10-05T20:23:50.005Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sCiu5Miaxx", "number": 34520, "title": "Learning Perturbation Robust Policies for LLM Agents with Stable Optimization", "abstract": "Reinforcement learning (RL) has become an effective post-training paradigm for long-horizon large language model (LLM) agents. However, we find that the resulting policies can be sensitive to various policy perturbations, such as hidden-state noise, pruning, and quantization. In this work, we study how to improve perturbation robustness during policy optimization. We first introduce the notion of a perturbation robust policy and analyze conditions under which perturbed policy updates preserve stable monotonic improvement. Based on this analysis, we introduce Stable Perturbation-Robust Policy Optimization (SPrPO), which applies adaptive and sensitivity-aware perturbations during RL training. We evaluate SPrPO on ALFWorld and WebShop and conduct systematic experiments across multiple perturbation types and scales, showing improved perturbation robustness while maintaining stable policy optimization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Agents; Perturbation Robustness; Reinforcement Learning" ], "forum_url": "https://openreview.net/forum?id=sCiu5Miaxx", "pdf_url": "https://openreview.net/pdf/a41b057bb3cd263fdf5f418b764e7cdd4af65d2f.pdf", "updated_at": "2026-10-05T20:23:52.246Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "khyAiOYkPl", "number": 34546, "title": "WaterKron and FlipFlop Hessian: Information-Theoretically Grounded Quantization with Kronecker-Factored Hessians", "abstract": "How should a Kronecker-factored Hessian approximation be chosen for post-training quantization?\nWe address this question through WaterKron, which combines two-sided GPTQ with row- and column-dependent waterfilling scales and entropy coding.\nWe derive its high-rate distortion with respect to the full Hessian using an explicit Kronecker-Hessian mismatch factor $\\Phi$.\nThis factor quantifies the asymptotic distortion penalty due to the Kronecker Hessian approximation and provides a criterion for selecting the factors optimally.\nMinimizing $\\Phi$ leads to a Gaussian covariance-fitting problem with classical flip-flop updates.\nWe thus give a rate-distortion justification for using the resulting FlipFlop Hessian in quantization.\nWe evaluate it empirically, finding that the FlipFlop Hessian consistently improves KL divergence and perplexity over input-only, marginal, and Frobenius-based Hessian choices.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "gptq", "watersic", "kronecker factorization", "rate-distortion" ], "forum_url": "https://openreview.net/forum?id=khyAiOYkPl", "pdf_url": "https://openreview.net/pdf/4a2adb59fe12e2b5df934dbcf8f9e83c83fc64b3.pdf", "updated_at": "2026-10-05T20:23:55.926Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "O0eayhNGip", "number": 34571, "title": "FastKron: Efficient Quantization with Kronecker-Factored Hessians", "abstract": "We accelerate a family of algorithms for neural network quantization which utilize a two-sided version of the GPTQ/LDLQ algorithm.\nStandard GPTQ-style adaptive rounding uses one-sided correlation information derived from input activations.\nA natural two-sided extension can additionally capture correlations across output channels.\nIt utilizes a general Kronecker-factored approximation of the weight matrix curvature.\nThis approach has been used in BoA and YAQA.\nBut making a concrete algorithmic implementation of this two-sided GPTQ variant is nontrivial.\nBoA uses a large number of sequential steps, while YAQA improves over the sequential depth but still has quartic total cost.\nWe introduce FastKron, an efficient algorithmic implementation that combines anti-diagonal parallelism with a recursive divide-and-conquer construction.\nFor an $m\\times n$ weight matrix, FastKron uses $O(m+n)$ sequential steps while reducing the total work from $O(m^2n^2)$ to $O(mn(m+n))$.\nThus, it matches the cubic scaling of GPTQ while exploiting richer curvature information.\nMoreover, FastKron is modular with respect to both the base quantizer and the Hessian estimator. \nWe also provide practical benchmarks, consider a range of Hessian approximations that FastKron can be used with, and provide an efficient technique to compute these Hessians.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "gptq", "yaqa", "kronecker", "divide and conquer" ], "forum_url": "https://openreview.net/forum?id=O0eayhNGip", "pdf_url": "https://openreview.net/pdf/c22fd6460aac05a44d8d3701e9f40d3a28d9000d.pdf", "updated_at": "2026-10-05T20:23:59.561Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wD2AKOEkZo", "number": 34623, "title": "Accelerating SO(2)-Equivariant Materials Models through Low-Precision Emulation", "abstract": "Rotationally equivariant architectures improve the physical fidelity of machine-learned (ML) universal materials models based on graph neural networks (GNNs). However, their scalability is limited by the use of non-trivial, high-precision tensor operations during both training and inference. Adopting the reduced-precision formats of matrix multiplication units (MMUs), e.g., FP8 instead of FP32, provides massive increases in throughput, up to 30× (60×) on NVIDIA Hopper (Blackwell) GPUs, but may lead to drastic loss of accuracy. Here, we introduce tailored ‘emulated’ low-precision operations under the integer Ozaki scheme to accelerate the ‘SO(2)-equivariant’ processing layers of GNNs, which dominate the runtime of (1) many state-of-the-art ML interatomic potentials (MLIPs) and (2) most universal ML Hamiltonian (MLH) models. Our framework, FlashSO(2), outperforms TF32, the most accurate native low-precision MMU format, in throughput, accuracy, and memory footprint at once. At the SO(2) layer level, it reaches 40× (7×) the throughput of MALOQ’s (UMA’s) reference implementation, an MLH (MLIP) model, sustaining up to 236 (232) effective Tflop/s on an NVIDIA GH200, 3.5× the FP32 peak. At the model level, FlashSO(2) helps train MALOQ 3.9× faster than its pure-FP32 version while reducing peak memory by 58% at no loss in accuracy. In molecular dynamics simulations with UMA, the speedup factor is 3.3×, and the peak memory is reduced by 51%. FlashSO(2) thus pushes the speed-accuracy Pareto frontier of equivariant materials models beyond the native low-precision formats and substantially lowers the resources required to exploit their scaling laws.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "precision emulation", "mixed-precision training", "tensor cores", "GPU kernels", "kernel fusion", "equivariant graph neural networks", "machine learning Hamiltonians", "machine learning interatomic potentials" ], "forum_url": "https://openreview.net/forum?id=wD2AKOEkZo", "pdf_url": "https://openreview.net/pdf/a6388ea63ccafc7159c0485a613254e4eb80f32d.pdf", "updated_at": "2026-10-05T20:24:07.175Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp8", "low-precision", "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QRa6eTsdwr", "number": 34673, "title": "ACCOUNTING FOR THE UNSEEN: FRAGILITY-AWARE RECALIBRATION OF SEMANTIC UNCERTAINTY FOR HALLUCINATION DETECTION", "abstract": "Despite their impressive ability to generate natural and coherent text, large language models (LLMs) can produce plausible but fabricated responses. The same prompt, when repeatedly applied, may yield different answers, undermining the confidence one may place in an individual generation. The common approach of detecting such confabulations by observing multiple generations to estimate the uncertainty over their semantic content must necessarily rely on a finite sample of all generations that the model may produce. Such estimates can be fragile when rare or singleton answers dominate observed semantic support. To address this, we propose \\underbar{S}emantic \\underbar{U}ncertainty \\underbar{R}ecalibration accounting for \\underbar{F}ragility (SURF), a framework that corrects semantic uncertainty estimates using the \\emph{probability} of observed singletons as a proxy for unobserved semantic support. SURF incorporates this fragility proxy into an entropy-based optimization that recalibrates token-sequence probabilities according to their sensitivity, yielding uncertainty estimates that are more robust against finite observations. We also provide an estimate of the \\emph{number} of semantically distinct answers as an actionable signal for when additional observations or human intervention may be warranted. When evaluated across 112 different settings (involving four datasets, seven model variants, and different generation lengths and model quantization levels), SURF shows consistent improvement in AUROC and headroom gain (HRG), and maintains stable AURAC performance, which demonstrates its reliability even under resource constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models (LLMs)", "Hallucination Detection", "Semantic Uncertainty" ], "forum_url": "https://openreview.net/forum?id=QRa6eTsdwr", "pdf_url": "https://openreview.net/pdf/d9be0b2ed341162db9c483573e2d5576f27463b5.pdf", "updated_at": "2026-10-05T20:24:15.806Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vgOEJYwlsZ", "number": 34682, "title": "Decoupling Coordinate Permutations from Low-Dimensional Quantiles in LLMs", "abstract": "Post-training weight modeling and quantization schemes predominantly treat neural network tensors as monolithic coordinate arrays, simultaneously optimizing value discretization and positional assignment. In this work, we investigate a fundamental question: How much information truly resides within the scalar values of a Large Language Model (LLM) weight tensor when strictly decoupled from coordinates? We formalize an orthogonal decomposition of any weight matrix W ∈ Rm×n into a sorted value spectrum s = sort(vec(W)) and a spatial permutation π ∈ SN . Stripping away positional entropy reveals an unexpected empirical regularity: the position-agnostic value spectrum of trained LLM tensors is universally governed by an ultra-low-dimensional continuum. Across diverse open-weight architectures (Meta-Llama-3.2, Qwen-3, Gemma-4) and across every internal layer type, an unconstrained 5-term Metalog quantile distribution fits the empirical spectrum with near-zero error (MAE < 0.0002, relative MSE < 0.002). This closed-form parameterization decisively outperforms classical Gaussian, extreme-value, and generalized quantile baselines. Crucially, our findings expose an extreme information-theoretic asymmetry: while the value distribution collapses to an O(1) description length of just 5 scalars (160 bits), spatial permutations retain an irreducible combinatorial entropy of ⌈log2(N !)⌉ ∼ O(N log N ) bits. We demonstrate that learned model capacity resides almost entirely within spatial permutations rather than value variance, establishing that future model compression paradigms must pivot toward structured permutation coding.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Lineage Verification", "Weight-Space Forensics", "Metalog Distributions", "Quantile Representations", "Data-Free Auditing", "Open-Weight LLMs", "Parameter-Space Invariants", "Model Governance" ], "forum_url": "https://openreview.net/forum?id=vgOEJYwlsZ", "pdf_url": "https://openreview.net/pdf/f589991c8083616fd5f2014f0506831ac87812cb.pdf", "updated_at": "2026-10-05T20:24:14.730Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rVXjEd0dmE", "number": 34712, "title": "GIM: Evaluating models via tasks that integrate multiple cognitive domains", "abstract": "As LLM benchmarks saturate, the evaluation community has pursued two strategies to increase difficulty: escalating knowledge demands (GPQA, HLE) or removing knowledge entirely in favor of abstract reasoning (ARC-AGI). The first conflates memorization with capability; the second divorces reasoning from the practical contexts in which it matters. We take a different approach. The \\textbf{Grounded Integration Measure (GIM)} is a benchmark of 820 original problems (615 public, 205 private) where difficulty comes from \\emph{integration}; individual problems require coordinating multiple cognitive operations (constraint satisfaction, state tracking, epistemic vigilance, audience calibration) over broadly accessible knowledge, so that reasoning stays grounded in realistic tasks without being gated on specialized expertise. Each problem is an original expert-authored composition, majority with rubric-decomposed scoring (median 6 independently judged criteria). A balanced public--private split provides a built-in contamination diagnostic. We calibrate a judge-aware continuous response 2-parameter logistic (2PL) IRT model across 53 test-configurations (unique model × thinking-level pairs) and five calibrated judges, using 203{,}800 epoch-averaged prompt--judge cells derived from >1M raw judge-scored observations, producing robust ability estimates that correctly order test-configurations even when raw accuracy is distorted by errors, missing data, or judge leniency differences. Using this framework, we present a comprehensive leaderboard spanning 22 models and 47 reporting test-configurations, and conduct what is to our knowledge the most extensive published study of how test-time compute trades off against model capability on a fixed benchmark: 11 models swept across 35 test-configurations. We observe that within-family configuration choices, such as thinking budget and quantization, matter as much as model selection, and increasing thinking tokens has diminishing marginal returns. We release the evaluation framework, calibrated IRT parameters, and all public problems.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "benchmark", "item response theory", "rubric-based evaluation", "test-time compute" ], "forum_url": "https://openreview.net/forum?id=rVXjEd0dmE", "pdf_url": "https://openreview.net/pdf/4439d1eb6f7f626301e4060a20c22b90572e3dd0.pdf", "updated_at": "2026-10-05T20:24:18.971Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "e84SGSPW8T", "number": 34771, "title": "Latent-Action-Guided Video-Language Feature Learning for Surgical Instrument–Tissue Interaction Recognition", "abstract": "Recognizing instrument--tissue interactions is essential for\ncontext-aware surgical AI. Vision-language models offer a natural way to\ninject semantic structure into surgical representations by aligning video\nfeatures with textual action descriptions. However, pretrained encoders may\nlack spatial coherence, while global semantic alignment does not ensure precise\nspatial and temporal representations. By analyzing frame-to-frame feature\nchanges, we find that semantic alignment increases their dimensionality, but\nlarger increases do not necessarily improve recognition; encoders also differ\nin how strongly dominant changes localize to interaction regions.\nMotivated by these findings, we introduce LAViF, which compresses\nframe-to-frame changes into latent actions and predicts next-frame features\nduring end-to-end video--language alignment. Without additional spatial or\nmotion annotations, LAViF improves the interaction grounding of leading\nfeature changes and temporal-direction sensitivity in our evaluated settings.\nWe further characterize how action capacity and prediction strength affect\nrecognition across encoders and triplet components. Using image encoders\nwithout large-scale video pretraining, LAViF achieves competitive recognition\nwith faster inference and smaller INT4 accuracy drops than V-JEPA2/2.1,\nsupporting its deployment potential.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Surgical action recognition", "Video representation learning", "Vision-language learning" ], "forum_url": "https://openreview.net/forum?id=e84SGSPW8T", "pdf_url": "https://openreview.net/pdf/8d5ae5c917bd8783dc55c9702f5d916043e229f1.pdf", "updated_at": "2026-10-05T20:24:25.124Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "int4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qy0AJWUh1L", "number": 34843, "title": "ZO-Act: Efficient Zeroth-Order Fine-Tuning via One-Shot Activation-Informed Low-Rank Subspaces", "abstract": "Zeroth-order (ZO) optimization enables fine-tuning large language models when backpropagation is unavailable or memory-prohibitive, but existing methods typically perturb the full parameter space or rely on structured low-dimensional perturbations, still leading to noisy gradient estimates and limited optimization flexibility. We propose ZO-Act, an activation-informed ZO fine-tuning method that uses input activations to define a fixed low-rank parameterization. For each linear layer, ZO-Act computes a small activation basis once at initialization and optimizes only lightweight coefficient matrices using forward-only loss evaluations. This explicit coefficient-space parameterization reduces the effective perturbation dimension, enables direct use of standard momentum-based optimizers such as Adam, and naturally supports quantized LLM fine-tuning by keeping low-bit weights frozen. Our analysis characterizes a dimension-alignment trade-off: reducing the perturbation dimension improves ZO gradient estimation, while effective subspace optimization requires preserving sufficient gradient energy. We further show empirically that activation-informed subspaces capture substantially more gradient energy than random and weight-based alternatives. Across Llama-3-8B, OPT-13B, Qwen2.5-7B, Qwen2.5-32B, and INT4 Llama-3-8B, ZO-Act consistently achieves competitive performance against strong ZO baselines across language understanding, question answering, commonsense reasoning, and mathematical reasoning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Fine-tuning", "Zeroth-order optimization", "Activation-informed subspaces", "Low-rank adaptation", "Forward-only training" ], "forum_url": "https://openreview.net/forum?id=qy0AJWUh1L", "pdf_url": "https://openreview.net/pdf/a8ee6b1d2bea424029503f7b6960c01ae14d35da.pdf", "updated_at": "2026-10-05T20:24:31.588Z", "categories": [ "training", "reasoning" ], "matched_terms": [ "int4", "low-bit", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NsCmxAUHnb", "number": 34961, "title": "LampAttention: Look-Ahead Mixed-Precision FlashAttention for Dedicated Accelerators", "abstract": "While most attention logits can be computed in low precision without degrading numerical stability, current attention kernels fail to exploit this phenomenon.\nWe introduce a novel hardware-algorithm co-design in the form of mixed-precision FlashAttention.\nOur method accumulates key-query products and evaluates their exponentials in 8-bit formats, then adaptively identifies sensitive sub-blocks and recomputes them in 16-bit formats.\nWe propose the specifications for a dedicated accelerator capable of executing this pipeline efficiently.\nSimulated experiments with Qwen3 and Gemma 3 show that rerouting a selective minority of sub-blocks to high precision is sufficient to recover the baseline model performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "FlashAttention", "numerical stability", "mixed precision", "transformer inference" ], "forum_url": "https://openreview.net/forum?id=NsCmxAUHnb", "pdf_url": "https://openreview.net/pdf/61f79e57971d61e85eccf51208fba41465ee1bd3.pdf", "updated_at": "2026-10-05T20:24:50.929Z", "categories": [ "hardware" ], "matched_terms": [ "8-bit", "low precision", "mixed precision", "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tHh8TREObl", "number": 34982, "title": "Second Order Delta Space Search (SODSS),", "abstract": "Which weights of a trained network can be pruned or quantised is usu-\nally decided from the loss landscape at convergence, using gradients on\ncalibration data or extra Hessian-vector products. We show that the train-\ning trajectory already contains this information and that it can be read\nfrom checkpoints alone. Along a gradient-descent path the second differ-\nence of the parameters satisfies an exact identity, at =−η\nHtvt−1, with\na path-averaged Hessian. We characterise what this gives: the curvature\nthe optimiser experiences along its own path (a Rayleigh quotient), glob-\nally and per block, with a bias under minibatch noise bounded by 1/(η∆)\nthat vanishes with the checkpoint stride ∆; and what it does not: the\nHessian diagonal per parameter, which we show is not identifiable. Second-\nOrder Delta-Space Saliency (SODSS) builds on this. It detects second-\norder events in parameter, representation and token-embedding space with\na scale-free statistic (the second difference of log-speed), segments training\ninto temporal regimes, and scores each weight by a regime-weighted path-\nFisher saliency θ2\nN,j t wtv2\nt,j , which we prove is an Optimal-Brain-Damage\nsaliency with the empirical Fisher averaged over the path and which reduces\nto Synaptic Intelligence without the θ2 factor. The score needs no gradients,\ndata or Hessian products and is computed in one streaming pass with O(d)\nmemory. On SGD-trained networks (an MLP and a CNN) one-shot global\npruning with SODSS is within a point of magnitude pruning up to 95%\nsparsity, beats gradient-based Fisher and SNIP criteria throughout, and\nis clearly best at 98%; the event detector recovers learning-rate schedule\nchanges and representational phase changes without being told the sched-\nule. Under AdamW the raw score fails, because the premise vt ∝gt does\nnot hold; undoing the preconditioner with the second-moment state that\ncheckpoints already contain restores it, and the corrected score then beats\nmagnitude pruning on an AdamW-trained Transformer at every sparsity.\nWe also report what does not work: per-parameter second-order statistics\ndo not help, and regime weighting has no measurable effect.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "knowledge distillation", "model compression", "neural architecture search", "training dynamics", "second-order optimization", "Hessian-based pruning", "quantization" ], "forum_url": "https://openreview.net/forum?id=tHh8TREObl", "pdf_url": "https://openreview.net/pdf/c0f6b22a4e342a38dc5ab425c3ef3027de668f57.pdf", "updated_at": "2026-10-05T20:24:51.583Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "quantised", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2w6E5ddPeX", "number": 34984, "title": "Intrinsic Information Rates of Summaries in Neural Posterior Estimation", "abstract": "Neural posterior estimation (NPE) often encodes observations $Y$ into a fixed-dimensional summary $S$, with the neural posterior conditioning only on $S$. Any such summary imposes a finite information budget on inference, whereas exact preservation requires infinite information whenever posteriors vary continuously with the data. The budget typically emerges from architecture and training rather than an explicit measurement or choice. Better performance may therefore reflect better encoding or simply a larger budget. \nIn this paper, we make the budget explicit and ask two questions: how large must it be, and how best is it spent? To answer them, we model the summary as a (possibly stochastic) channel from $Y$ to $S$. For the first question, we show that this budget, measured by the mutual information $I(Y;S)$, grows unboundedly as the posterior concentrates: in regular $d$-dimensional models, bounded posterior distortion requires $\\tfrac d2\\log n-O(1)$ nats, even when a fixed-dimensional sufficient statistic exists. More generally, the required budget is set by the distribution of the random posterior, and we bound it from above by quantizing posteriors and from below by converses that weight each posterior state by its probability. For the second question, we show that a finite budget is best spent stochastically. Under a common-support condition on the simulator, which covers the continuous observation spaces typical of simulators, a deterministic summary necessarily loses something: at every positive budget too small for exact preservation, some stochastic summary achieves _strictly_ lower distortion than every deterministic summary and every time-shared mixture of deterministic summaries within the same budget. Motivated by these results, we insert a binary bottleneck into NPE, measure each trained summary's rate, and compare summaries at matched rate, which\nremoves the budget as a confounder in architecture comparisons. Across four simulator benchmarks and several encoder families, stochastic bottlenecks achieve favorable rate--loss tradeoffs, and architecture rankings change with the budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "neural posterior estimation", "bayesian computation" ], "forum_url": "https://openreview.net/forum?id=2w6E5ddPeX", "pdf_url": "https://openreview.net/pdf/edee27d676a45a5d94fca58f7f76fba511e9773a.pdf", "updated_at": "2026-10-05T20:24:51.747Z", "categories": [], "matched_terms": [ "binary", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "U3q0Ag8hN4", "number": 34997, "title": "SoftWater: The Rate-Distortion Geometry of Softmax Quantization", "abstract": "Post-training quantization methods typically optimize weighted mean-squared error (WMSE) on linear-layer outputs, yet the final softmax layer is fundamentally different: its relevant distortion is the change in the predictive distribution. \nWe formulate softmax-layer quantization as a rate-distortion problem under KL divergence and show that, to second order, its geometry is class-aware, coupling feature covariance with class-specific softmax curvature. \nThis distinction matters because output heads are often left in high precision even though, in modern language models, they can account for 15--30\\% of all parameters, and over half of the stored bytes once the body is quantized.\nWe introduce \\emph{SoftWater}, a practical quantizer derived from this KL geometry. \nA separable approximation reduces the otherwise prohibitive joint class-feature weighting to a single feature-side factorization together with inexpensive class-dependent rescaling, allowing the head to be quantized using statistics collected in one calibration pass. \nThe resulting allocation assigns precision according to the output class frequency discounted by its variance across inputs, which is especially important in highly imbalanced output spaces such as LLM vocabularies.\nAcross five language models from 1B to 32B parameters, SoftWater outperforms the released WaterSIC quantizer, near-optimal under WMSE, at matched head rates on 59 of 60 evaluation points, reducing head-induced KL by 6.5--8.3$\\times$ at 2 bits despite using none of WaterSIC's additional refinements. \nBecause the class-side statistic comes from calibration data, matching calibration to the deployment domain gives the lowest KL on that domain throughout. \nOn Llama-3.2-1B-Instruct with quantized bodies, where the tied head is also the embedding, a 2-bit head removes 45--60\\% of stored bytes for a 2.9--3.7\\% perplexity increase and a 4-bit head is near-lossless.\nThese results show that treating the softmax layer according to its own distortion geometry can make aggressive end-to-end model quantization practical.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "LLM", "softmax" ], "forum_url": "https://openreview.net/forum?id=U3q0Ag8hN4", "pdf_url": "https://openreview.net/pdf/bd209505f23f7240c5ff9e3a3bea515d00435623.pdf", "updated_at": "2026-10-05T20:24:53.909Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "4-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "arzRXC3814", "number": 35023, "title": "Robo-GPC: Compliant Robotic Control with Generative Pretrained Controllers", "abstract": "Humanoid robots operating in human environments must interact safely and robustly with both humans and their surroundings. They must remain stable under disturbances, respond compliantly to external forces, and recover naturally from falls. We introduce Robotic Generative Pretrained Controller (Robo-GPC), a discrete generative controller that learns motor skills from large-scale human motion data and adapts to downstream whole-body control tasks through parameter-efficient fine-tuning. Robo-GPC represents motor skills in a compact Finite Scalar Quantization (FSQ) latent space and models their distribution with an autoregressive transformer decoder. By reusing the learned skill distribution during task adaptation, Robo-GPC produces natural behaviors while supporting a wide range of downstream control tasks, providing a general framework for safer humanoid deployment. Deployed on a Unitree G1 humanoid, Robo-GPC enables robust whole-body control while responding naturally to pushes and pulls and recovering from large disturbances and falls. The supplementary videos are present in https://anonymous.4open.science/r/robogpc/index.html.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Humanoid Whole-Body Control", "Generative Model" ], "forum_url": "https://openreview.net/forum?id=arzRXC3814", "pdf_url": "https://openreview.net/pdf/793fbe9360e692d088d535fbac6ac4f08589fa20.pdf", "updated_at": "2026-10-05T20:24:59.505Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9a4dEL5xo8", "number": 35033, "title": "LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization", "abstract": "Recent LLMs increasingly adopt hybrid designs that replace standard attention with linear attention, such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA). Although they compress the context into a fixed-size recurrent state and substantially reduce the cost of long-context processing, repeatedly reading and updating that state remains a major inference bottleneck. Quantization offers a natural way to reduce this cost, but can significantly degrade model quality, due to the accumulation of rounding errors and the presence of outlier rows and columns in the state. To address these challenges, we propose LeapQuant, a training-free method that achieves near-lossless performance under 8-bit recurrent-state quantization. First, to mitigate error accumulation, we propose per-window quantization, which leaps over a window of tokens and quantizes the state only once at its end. Within a window, outputs are computed from the fixed low-bit state together with high-precision buffered updates. Second, to reduce the error introduced by each quantization, LeapQuant retains the state's largest outliers as a few high-precision compensator tokens, which share the update path of real tokens. We then smooth the remaining residual before quantization to further reduce the error. Comprehensive experiments across the Qwen, Kimi, and GLM model families show that LeapQuant substantially reduces memory and compute costs during inference. With accuracy comparable to the FP32 baseline, it achieves average speedups of 2.05--3.70$\\times$ at the kernel level and 1.47$\\times$ for end-to-end inference on NVIDIA B200, RTX PRO 6000, and RTX 5090 GPUs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Linear Attention", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=9a4dEL5xo8", "pdf_url": "https://openreview.net/pdf/47764dd74e8c0d3174457d2f388647f50de41581.pdf", "updated_at": "2026-10-05T20:24:59.938Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "8-bit", "low-bit", "quantization", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "l86eXLKuHS", "number": 35068, "title": "Does Quantization Preserve How Language Models Answer? Causal-Equivariant Quantization for Symbol-Binding Circuits", "abstract": "Post-training quantization (PTQ) is typically optimized to preserve weights and activations while maintaining perplexity, or task accuracy. However, these objectives do not reveal whether quantization preserves the causal computation underlying model decisions. Recent studies have localized answer selection and symbol binding in formatted multiple-choice question answering (MCQA) to a sparse set of attention components, making it possible to directly examine their preservation under quantization. To preserve the causal mechanisms underlying formatted MCQA after quantization, we propose Causal-Equivariant Quantization (CEQ), a structured mixed-precision framework consisting of three components: counterfactual causal probing, causal-regret-based precision allocation, and causal-equivariant calibration. First, counterfactual causal probing identifies model components whose causal behavior is most affected by low-bit quantization across different answer formats. CEQ measures these changes using causal-effect distortion (CED) and format-equivariance error (FEE). Second, causal-regret-based precision allocation uses this information to assign higher precision to causally important and quantization-sensitive components under a global bit budget. Third, causal-equivariant calibration (CEQ-Cal) further reduces the discrepancy between the quantized and full-precision models while maintaining consistent behavior across answer formats by preserving output distributions, answer margins, causal effects, and format equivariance. We evaluated CEQ on a novel controlled symbol-binding dataset constructed specifically for this work and on five public MCQA datasets. On the controlled symbol-binding benchmark, CEQ achieves 83.40 ± 14.21\\% worst-format accuracy at 2.72 effective matrix bits versus 68.00 ± 20.64\\% for uniform 2-bit RTN. Across the five public MCQA datasets, CEQ consistently improves the worst-format accuracy over the matched TaCQ baseline, achieving an average gain of 9.86\\% points. The effectiveness of CEQ further generalizes across model scales and architecture families, achieving 76.55\\% MMLU accuracy on Qwen2.5-32B-Instruct at 2.12 bits and a 67.95\\% zero-shot average on Llama-3-8B at 4.08 bits. These results show that preserving prediction quality under quantization does not necessarily preserve the underlying causal computation, and motivate mechanism preservation as a distinct objective for low-bit language-model compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model Quantization; Mechanistic Interpretability; Post-Training Quantization; Mixed-Precision Quantization; Model Compression" ], "forum_url": "https://openreview.net/forum?id=l86eXLKuHS", "pdf_url": "https://openreview.net/pdf/3d73e212b2b52188c32e37493276fdfa92200ccf.pdf", "updated_at": "2026-10-05T20:25:03.934Z", "categories": [ "ptq" ], "matched_terms": [ "2-bit", "low-bit", "mixed-precision", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Dhg6TEC1WG", "number": 35075, "title": "QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction", "abstract": "As large language model inference shifts toward lower precision, post-training quantization (PTQ) becomes increasingly brittle, making quantization-aware training (QAT) essential for preserving model quality. However, QAT has a structural mismatch: gradient updates are applied to latent full-precision weights, while the loss and gradients are computed on lossy reconstructions of those weights. This mismatch can lead to suboptimal training trajectories and a higher loss floor. Second-order PTQ methods address a similar problem by minimizing loss-aware reconstruction error, but applying such expensive reconstruction repeatedly during QAT as the weights evolve is impractical. We introduce QUASAR, a QAT method that brings lightweight, loss-aware reconstruction into the training loop. At each training step, QUASAR reconstructs the latent weights by searching over a small set of clipping ranges and fitting dequantization parameters through saliency-weighted least squares. We use an exponential moving average of squared gradients as the per-parameter saliency signal. Our theoretical analysis shows that optimizing QUASAR's reconstruction objective tightens both the convergence and final-loss bounds of QAT. We evaluate QUASAR across four model families and across INT4, INT3, INT2, and NVFP4 quantization formats. QUASAR consistently achieves lower training and evaluation loss than competitive QAT methods and outperforms QAT and PTQ baselines on downstream benchmarks. At INT2, QUASAR improves average accuracy over the best QAT baseline by 13.3 points with quantization-aware distillation and by 10.9 points with QAT on mathematical reasoning data. Notably, after distillation on only $\\sim$600M tokens, QUASAR's INT4 Gemma-4 E4B checkpoint outperforms the corresponding QAT checkpoint released by Google, with 66\\% lower KL divergence and 1.8 points higher average accuracy.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization-aware training", "large language models" ], "forum_url": "https://openreview.net/forum?id=Dhg6TEC1WG", "pdf_url": "https://openreview.net/pdf/f5f11332a35b3670e993235598fd898cb8dfedf0.pdf", "updated_at": "2026-10-05T20:25:04.535Z", "categories": [ "ptq", "qat", "training", "reasoning", "theory" ], "matched_terms": [ "int2", "int3", "int4", "nvfp4", "ptq", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bvQrd1Nndz", "number": 35239, "title": "QSpan: Training-Aware Low-Rank Quantization Error Compensation for Fully Quantized LoRA Fine-tuning", "abstract": "Parameter-efficient fine-tuning (PEFT) reduces adaptation costs, and quantized PEFT further improves efficiency by primarily quantizing the frozen backbone weights.\nHowever, most methods retain activations, gradients, and the LoRA branch in high precision.\nFully quantizing both the backbone and LoRA branch reduces these remaining costs but substantially degrades adaptation, as the frozen backbone cannot correct quantization errors and rank-constrained adapters have limited compensation capacity.\nExisting methods prioritize weight-quantization errors without considering their effects on the low-precision forward computation and training objective.\nWe introduce **QSpan**, a training-aware framework that determines which quantization-error directions each adapter should span and how rank should be distributed across layers.\nQSpan prioritizes directions of the weight-quantization error according to their exposure through quantized activations and sensitivity to quantized gradients.\nIt then initializes adapters along the most consequential directions through a closed-form solution and allocates a fixed total rank budget toward layers with large fractions of uncompensated training-weighted quantization error.\nAcross language reasoning, personalized language modeling, and text-to-image personalization, QSpan consistently outperforms fully quantized LoRA and representative quantization-aware baselines under matched NVFP4 precision and adaptation budgets.\nThese results demonstrate that aligning limited adapter capacity with training-relevant quantization errors enables effective fully quantized PEFT.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LoRA", "Quantization", "Fully Quantized Fine-tuning" ], "forum_url": "https://openreview.net/forum?id=bvQrd1Nndz", "pdf_url": "https://openreview.net/pdf/f2a3c5a94c539ead295a56da5c2ac0637abce185.pdf", "updated_at": "2026-10-05T20:25:32.335Z", "categories": [ "qat", "training", "reasoning" ], "matched_terms": [ "low-precision", "nvfp4", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SndB6lpD5M", "number": 35288, "title": "Deep Psychovisual Image Representations", "abstract": "Psychovisual models suggest human vision decouples low-level feature extraction from higher cognition by first forming intermediate abstractions. In contrast, deep learning-based vision models routinely extract and aggregate features using homogeneous stacks of spatial layers, rendering their decision-making processes opaque. In this paper, we propose Deep Visual Coding, a learned frequency-domain representation inspired by 1990s image codes that quantised perceptually salient frequencies, which together with complex-valued image representations produces psychovisual-style abstractions. This approach enables the first psychovisual-based deep learning framework, utilizing data-driven spectral filters that learn to encode task-relevant semantic structures within distinct frequency sub-bands. Salience analyses reveal that our psychovisual models extract highly interpretable object parts compared to the amorphous regions produced by regular Convolutional Neural Networks (CNNs). Furthermore, we find that our models are less depth dependent than CNNs for model scaling, since our complex-valued representations and learned abstractions subsume the role of the deep spatial layers. Together, these findings demonstrate that psychovisual coding provides a promising path toward more efficient and transparent vision models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "psychovisual", "visual codes", "frequency domain", "Fourier transform", "representation learning", "deep learning", "vision", "human vision" ], "forum_url": "https://openreview.net/forum?id=SndB6lpD5M", "pdf_url": "https://openreview.net/pdf/d04202ebfeb756195ef8bbe3071235dc6487deeb.pdf", "updated_at": "2026-10-05T20:25:39.558Z", "categories": [], "matched_terms": [ "quantised" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pnTndxkrC8", "number": 35293, "title": "Be Your Own Teacher: Steering Protein Language Models via Unsupervised Reward Optimization", "abstract": "Protein language models (PLMs) have emerged as powerful tools for controllable biomolecular design, yet their post-training adaptation typically relies on costly wet-lab validation or curated preference datasets. To overcome this supervision bottleneck, we introduce unsupervised reward optimization of PLMs, a comprehensive framework for steerable protein generation without ground-truth labels. Our key insight is that task-agnostic rewards, which combine intrinsic model uncertainty with extrinsic semantic consistency informed by protein representation models, exhibit strong correlation with controllability measures across base models and temperature regimes. Building upon this discovery, we propose two offline algorithms: Soft Reward Optimization (SRO) and Binarized Reward Optimization (BRO), which effectively maximize the classical RLHF objective induced by these proxy rewards. Extensive experiments on compositional out-of-distribution prompts demonstrate that both methods significantly outperform competitive baselines (DPO, KTO), while approaching oracle performance across multiple sampling temperatures, model scales and protein families. Moreover, PLMs fine-tuned with unsupervised rewards can achieve consistently higher coverage compared to their base model in pass@k evaluations. By enabling self-improvement of PLMs through their own generated experience, our framework provides a scalable pathway toward controllable biomolecular design in settings where labeled preferences or experimental feedback are scarce or unavailable.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Protein Language Models", "Generative Models", "Post-Training", "Reinforcement Learning" ], "forum_url": "https://openreview.net/forum?id=pnTndxkrC8", "pdf_url": "https://openreview.net/pdf/5fa0dda5ed379aa7078d11c7b6f45267df4fec19.pdf", "updated_at": "2026-10-05T20:25:38.747Z", "categories": [ "ptq" ], "matched_terms": [ "binarized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "0ptZNRyVjx", "number": 35339, "title": "Signs Beat Floats: Low-Rank Double-Binary Adaptation for Storage-Efficient Deployment", "abstract": "Task-specific LoRA adapters can be stored and exchanged independently of a shared frozen language model, but their floating-point factors remain a substantial deployment cost. We introduce LoRDBA, a low-rank double-binary adapter that represents both factors with binary signs and lightweight channel-wise scales. A post-training factorization initializes the signs, after which our default QAT Freeze procedure adapts only the scales; QAT Full additionally updates latent carriers. We provide a conditional finite-sample reconstruction bound for a sign-noise model. Across three-seed math-reasoning evaluations, LoRDBA achieves favorable compact accuracy-storage trade-offs on LLaMA-2-7B and 13B, with model-dependent margins on Mistral-7B. At approximately $9.54$ MiB logical payload on 7B, Freeze improves GSM8K/Minerva means by $6.24/2.24$ points over compact fp16 LoRA with equal total dataset passes. On 13B, it exceeds an SVD rank-2 control initialized from the same source and given the same two-epoch recovery by $3.61/3.34$ points at approximately $14.93$ MiB. Separately, packed execution reduces resident adapter memory by $6.90\\times$, and adapter-local recomputation brings Freeze training to the fp16-LoRA peak-memory level at higher step time. These findings support storage-efficient adapters produced on server GPUs, not inference speedup or demonstrated on-device training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "adapter compression", "parameter-efficient fine-tuning", "low-rank adaptation", "binary quantization", "large language models", "storage-efficient deployment" ], "forum_url": "https://openreview.net/forum?id=0ptZNRyVjx", "pdf_url": "https://openreview.net/pdf/9792ecf6968aeacc901184b802c35365a4df54aa.pdf", "updated_at": "2026-10-05T20:25:45.559Z", "categories": [ "ptq", "qat", "hardware", "reasoning" ], "matched_terms": [ "binary", "qat", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9LUInzhKZc", "number": 35390, "title": "Forgetting to Forget: How Rounding Bias Alters Memory in Gated DeltaNet", "abstract": "Repeatedly rounding a language model's recurrent state can change what it\nremembers during decoding. Lower-precision caches save memory, but may\ndiscard small updates each time the state is stored. We prove that\nround-to-nearest (RTN) permanently halts fixed-gate pure decay below an exact\nformat-dependent threshold. The measured active Gated DeltaNet heads do\nnot freeze: writes and erasures keep them moving, while rounding error accumulates\nacross the full update. Using Qwen3.5-0.8B as our main example, we connect\nthis arithmetic to state error and next-token prediction. In teacher-forced\nevaluation on 72 documents, storing selected heads in fp32 removes 75.4\\%\nof the excess negative log-likelihood caused by bf16 round-to-nearest\nrelative to an fp32-state control. In a separate pooled comparison,\nfull-update stochastic rounding removes 91.4\\% of the RTN excess\nwithout enlarging the bf16 cache.\nAcross five architectures, a post-hoc gate diagnostic flags at-risk\nbf16 heads; separate replays on fixed inputs show substantial state errors\nin selected small-decay heads. Free-running Qwen3.5-4B also shows changes in\ngeneration length under a fixed decoding protocol. The results motivate\ntesting recurrent-state precision over long decoding trajectories, not\nonly long-input prefill. They do not establish equal serving cost at equal", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Gated DeltaNet; recurrent language models; low-precision inference; rounding error; stochastic rounding; recurrent-state caches; long-context modeling" ], "forum_url": "https://openreview.net/forum?id=9LUInzhKZc", "pdf_url": "https://openreview.net/pdf/e49d48274eeefd917681e7109c4fe0d343084eb6.pdf", "updated_at": "2026-10-05T20:25:50.680Z", "categories": [ "ptq" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cBuI0SFIZL", "number": 35441, "title": "TEMPO: Temporal and Positional Precision for Diffusion Language Model Agents", "abstract": "Diffusion large language model (dLLM) agents recompute a fixed-length generation region at each denoising step. Reserved positions beyond the emitted call account for most of this computation. Removing them reduces accuracy because they provide context for uncommitted predictions, requiring a reduction in computation per position. Quantisation reduces this cost, but existing quantisers use a fixed configuration across steps and positions, irrespective of the decoding state defined by committed tokens and current predictions. We identify three phenomena that enable precision to adapt to this state. Filler-first ordering places end-of-text and padding commitments before content commitments, allowing early steps to run at low precision. Margin persistence correlates logit gaps across steps, indicating the risk of quantisation-induced prediction changes before each pass. Slack tolerance permits low-precision updates of the reserved region with limited changes in accuracy. These phenomena motivate TEMPO (Temporal and Positional Orchestration of Precision), a training-free hierarchical precision schedule. A step-wise gate selects low-precision steps using cached gaps. On other steps, position-wise routing retains full precision only for the prompt and estimated content region. Across four agent benchmarks and three backbones, TEMPO achieves 84\\% mean call fidelity, compared with 76\\% for full quantisation and 66\\% for Learn2PD. Adding TEMPO to parallel decoding with boundary fill yields a further $1.20$ to $1.45\\times$ speedup on Spider and API-Bank. The full composition achieves up to $98.8\\times$ at 128 denoising steps. Our code is openly accessible to support future research.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language models", "efficient inference", "quantization", "dynamic mixed precision", "parallel decoding" ], "forum_url": "https://openreview.net/forum?id=cBuI0SFIZL", "pdf_url": "https://openreview.net/pdf/83bb23aff6b2aa710360d9fa695f8ab494eab41d.pdf", "updated_at": "2026-10-05T20:25:56.664Z", "categories": [ "diffusion" ], "matched_terms": [ "low precision", "low-precision", "mixed precision", "quantisation", "quantisers", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "2j2PCoIKdi", "number": 35497, "title": "Towards Dense Capability Recovery of Ternary Quantized LLMs", "abstract": "Ternary quantization offers extreme compression, but the resulting perturbation is substantially more severe than conventional low-bit quantization. The key challenge is not merely reducing one-shot projection error, but finding a ternary representation that remains efficiently recoverable through subsequent training. We demonstrate that under matched knowledge distillation (KD), quantizers with vastly different initial errors converge to remarkably similar functional endpoints. Motivated by this, we introduce a calibration-free, energy-preserving quantizer as a stable starting point designed specifically to facilitate subsequent training.\nHowever, same-size KD restores broad post-trained capability while leaving substantial and highly nonuniform performance gaps across tasks. To address this, we first exploit test-time sampling to expose substantial residual headroom on several reasoning and coding tasks. We then apply quantization-aware reinforcement learning (RL) directly to the ternary policy, converting part of this headroom into stronger single-shot performance.\nWe evaluate our approach across several model sizes: Qwen3-1.7B, 4B, and 8B. The first math-RL stage substantially improves single-shot performance, while broader gains transfer across knowledge, reasoning, instruction following, mathematics, code, and the unoptimized thinking mode. At 8B, the final 2.03-GB model achieves 99.3\\% mean benchmark-wise retention relative to BF16 in non-thinking mode and 94.1\\% in thinking mode, while decoding at 77.9 tokens/s on an M5 Pro. Together, these results support a recovery-centric view of extreme quantization: preserve a recoverable ternary representation, restore the model distribution, and then optimize the deployed ternary policy directly.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "ternary quantization", "low-bit LLM compression", "knowledge distillation", "calibration-free quantization", "post-training reinforcement learning" ], "forum_url": "https://openreview.net/forum?id=2j2PCoIKdi", "pdf_url": "https://openreview.net/pdf/e53daf3c48b969a56fc5cc8d6f4943683969d6f6.pdf", "updated_at": "2026-10-05T20:26:04.939Z", "categories": [ "ptq", "qat", "reasoning" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizer", "quantizers", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SvI6SwedKB", "number": 35521, "title": "Deeper Is Not Better for Quantized CLIP: Early Readout Recovers Accuracy While Cutting Compute", "abstract": "Readout depth is a decision variable. CLIP predicts a label from the direction of its final-layer image embedding, on the premise that every transformer block moves that direction closer to the correct class. We show that this premise fails for INT8-quantized CLIP and is already unreliable in full precision. In FP32 ViT-B/32, the cosine agreement between intermediate and final representations rises with depth but regresses at three of the eight candidate layers; under INT8, the deviation introduced by quantization grows from 0.09% of the native inter-layer change at Layer 1 to 75.64% at Layer 11. We model transformer depth as a noisy readout channel in which each block adds a semantic increment and a perturbation, and show that under this model a per-sample readout gated on the top-1/top-2 margin, with depth-dependent thresholds, is expected to beat full-depth readout once perturbation growth outpaces semantic gain. LRA-EE implements this rule on a frozen INT8 backbone: a patch-token average replaces the immature shallow [CLS] token, a learned gate trained without ground-truth labels decides when to stop, and exit layers are selected by an FP32-only maturity criterion applied identically to every backbone. Across four INT8 CLIP backbones on ImageNet-1K, accuracy never falls below the full-depth baseline while FLOPs fall by 7.9--14.5%: +3.27%p on ViT-B/32, +4.27%p on ViT-B/16, +3.16%p on EVA-CLIP, and +0.66%p on ViT-L/14. In other words, the quantized model is more accurate when it computes less. A four-quadrant decomposition attributes the gain to a Rescue Effect---8.1% of samples are correct at an early exit but wrong at full depth, against 4.9% the reverse---and an exact-route control splits it into a precision-independent component (+2.42%p) and a quantization-specific increment (+0.86±0.03%p over three seeds). The evidence is empirical: the backbone is never updated and the exit heads are calibrated on 2.3% of the ImageNet training images, so the result is post-training readout calibration of a compressed model rather than zero-shot deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "early exit", "post-training quantization", "CLIP", "vision–language models", "adaptive inference", "representation collapse", "selective classification", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=SvI6SwedKB", "pdf_url": "https://openreview.net/pdf/41cbf8342f99bcc2308f3c4dc9456f7e1c1d2b76.pdf", "updated_at": "2026-10-05T20:26:07.833Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "int8", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "w7v2VhQ0N0", "number": 35551, "title": "MobileGrail: Learning Compact Guardrails for On-Device deployable Mobile GUI Agent Safety", "abstract": "Runtime guardrails protect mobile GUI agents by assessing each proposed action before execution, but existing guardrails rely on cloud hosted large multimodal models, exposing every screen to a remote server, including private information such as payments, messages, and medical records. We therefore study how to train a compact guardrail that runs on device, which raises two challenges. On-device budgets limit the model to a short context, yet safety decisions hinge on evidence observed many steps earlier, which existing guardrails lose in truncating or summarizing the history. A compact model also fails to judge safety from a single prompt, so its safety boundary must be learned from labels, yet existing labels mark anticipated risk and committed violations alike as \\emph{unsafe}, which blurs the boundary and rejects benign tasks that share the same risk indicators. We propose MobileGrail, an on-device deployable model that addresses both. It maintains a \\emph{recurrent task state} that retains only the evidence subsequent safety decisions depend on, so that authorization and execution progress persist beyond the short context. To intervene early while keeping the safety boundary precise, it learns two labels, \\emph{unsafe} and \\emph{block}: \\emph{unsafe} marks the action that commits a violation, and \\emph{block} is admissible over an \\emph{intervention window} that opens once the evidence justifies intervention. To supply the learning signal, we build a training environment of 32 applications whose backends verify the effect of every action, grounding both labels and providing verifiable rewards with which reinforcement learning calibrates intervention timing. We further build MobileVeto, a benchmark of held-out applications annotated with intervention windows and paired with benign counterparts. On MobileVeto, MobileGrail built on Qwen3.5-2B raises effective intervention over its prompted backbone from 64.5\\% to 76.2\\% and reduces premature intervention from 58.6\\% to 5.2\\%, outperforming OS-Sentinel with Kimi-K2.6 by 36.9 and 41.4 points in effective intervention and benign false positives. Quantized to 4 bits, it runs on a Snapdragon 8 Elite at a median latency of 4.7\\,s, comparable to cloud-hosted Gemini 3.8 Flash (5.1\\,s).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mobile GUI agents", "agent safety", "runtime guardrails", "on-device multimodal models", "safety alignment", "long-horizon agent reasoning" ], "forum_url": "https://openreview.net/forum?id=w7v2VhQ0N0", "pdf_url": "https://openreview.net/pdf/59f03aebd64bcb19b76bb51a2ab96d7cb2b19f14.pdf", "updated_at": "2026-10-05T20:26:12.447Z", "categories": [ "ptq", "hardware", "reasoning", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fQcxDUPNCI", "number": 35586, "title": "LW-MIH: GENERATIVE MULTI-IMAGE HIDING WITH VAE PACKING AND LEARNABLE WHITENING", "abstract": "Multi-image hiding conceals several secret images in one image sent over a public channel, but existing methods face a critical issue: cover-based methods reach high capacity yet modify the cover and leave traces, whereas coverless generative methods are secure yet hide only a few images, and neither states how many images one cover can carry nor how to trade that number against per-image fidelity. To address this issue, we propose LW-MIH, a generative coverless multi-image hiding framework with high capacity and high recovery fidelity. LW-MIH first encodes every secret with a frozen foundation VAE and packs the latents into a fixed-size tensor, so the number of images one cover carries follows from the encoder geometry: up to 96 with Sana, 48 with SD1.5 and 12 with FLUX. It then fills the tensor in one of three ways --- full capacity, high-frequency dual-slot and upsampling --- turning the number of hidden images and the fidelity of each into a discrete set (FLUX 12/6/3, SD1.5 48/24/12/3, Sana 96/48/24/6). Finally, an invertible neural network maps the packed tensor to the noise end of a frozen rectified flow, and a learnable invertible whitener aligns the field with the flow's Gaussian prior, adding no capacity cost and keeping recovery exact. Experimental results show that at full capacity the 16-bit PSNR of the recovered secrets is 35.03, 28.17 and 26.89 dB for FLUX, SD1.5 and Sana, within 0.15, 0.07 and 0.06 dB of their VAE ceilings, and rises to 41.03, 37.08 and 37.91 dB when fewer images are hidden; the 8-bit values are 24.12, 22.77 and 20.18 dB. The whitener gains 3.97 and 3.36 dB of 8-bit PSNR over a random orthogonal mixer of the same form on FLUX and SD1.5, and at full capacity leaves a worst gap of 0.41 dB where the pipeline without a mixer falls 12.4 dB below its encoder's limit.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "multi-image hiding", "learnable whitening", "generative steganography", "VAE latent packing", "invertible neural network", "rectified flow" ], "forum_url": "https://openreview.net/forum?id=fQcxDUPNCI", "pdf_url": "https://openreview.net/pdf/d27824fb567e6bed5f4adef97c7d89a145d8d5c6.pdf", "updated_at": "2026-10-05T20:26:17.888Z", "categories": [], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "28zRRzqKQA", "number": 35792, "title": "Prompting What Matters: Adaptive Calibration for Data-Free Quantization of Large Vision-Language Models", "abstract": "Data-Free Quantization (DFQ) provides a promising solution for compressing Large Vision-Language Models (LVLMs) without real calibration data. Existing DFQ methods mainly focus on global statistical matching, visual semantic preservation, or generic contrastive alignment, overlooking the heterogeneous quantization sensitivity of multimodal capabilities and model components. As a result, semantically plausible synthetic samples may fail to activate capability-specific quantization errors that dominate performance degradation. In this work, we reveal that LVLM quantization vulnerability is highly capability-module dependent, raising two fundamental questions: which capabilities and modules are most sensitive to quantization and how can calibration samples be synthesized to recover them? To address these challenges, we propose PromptQ, Prompt-Guided Data-Free Quantization framework that optimizes calibration distribution according to capability-module quantization vulnerability. PromptQ first introduces Sensitivity-Driven Prompt Construction (SDPC), which estimates capability-module sensitivity through behavioral degradation and representation discrepancies under controlled quantization. Based on the estimated sensitivity, SDPC allocates calibration prompts across vulnerable capabilities while maintaining semantic diversity, and generates module-aware prompts that specify the target capability and quantization-sensitive component. Guided by these prompts, Quantization-Sensitive Image Synthesis (QSIS) synthesizes semantically valid calibration samples that expose informative quantization errors by increasing the discrepancy between full-precision and quantized representations under semantic constraints. By coupling sensitivity-aware capability allocation with module-targeted sample synthesis, PromptQ transforms data-free calibration from distribution matching into vulnerability-aware error discovery. Extensive experiments on diverse multimodal benchmarks demonstrate that our PromptQ consistently improves quantization performance without requiring real calibration data. Our code is available in supplementary material package.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Data-Free Quantization", "Large Vision-Language Models", "Informative Calibration Samples" ], "forum_url": "https://openreview.net/forum?id=28zRRzqKQA", "pdf_url": "https://openreview.net/pdf/eb0cddb276c7c2f9ac42d1dc20aee8c4afc36018.pdf", "updated_at": "2026-10-05T20:26:51.629Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W2dZ0aHjGg", "number": 35795, "title": "PATCH: Propagation-Aware Two-Stage Quantization Error Compensation via Fine-Grained Correction for High-Fidelity 2-Bit LLMs", "abstract": "Extreme low-bit quantization reduces the memory and data-movement costs of large language models (LLMs), but 2-bit precision introduces substantial errors that accumulate across layers. Existing compensation methods are constrained by both correction expressiveness and optimization scope: limited-rank corrections struggle to capture the error structure, channel-wise scaling overlooks intra-channel heterogeneity, and locally optimal compensation strengths need not minimize end-to-end loss. We introduce PATCH, a propagation-aware two-stage framework that decouples fine-grained correction from layer-wise strength regulation. The first stage fits group-wise compensation using activations from the quantized network, capturing heterogeneous intra-channel errors without an explicit rank constraint. The second stage fixes the learned corrections and jointly optimizes layer-wise regulation coefficients against an end-to-end objective, accounting for error propagation while preserving the correction directions. The resulting coefficients are fused into existing quantization parameters, avoiding a separate compensation branch. Across Gemma-2, Llama-3, and Qwen3 models of varying scales under 2-bit GPTQ, PATCH reduces WikiText-2 perplexity by 24.7% on average and improves average zero-shot accuracy by 2.03 percentage points relative to the best competing compensation method for each model and metric. These results demonstrate that jointly addressing correction expressiveness and network-wide compensation strength enables effective accuracy recovery in extreme low-bit LLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Quantization", "Post-training quantization", "Quantization error compensation" ], "forum_url": "https://openreview.net/forum?id=W2dZ0aHjGg", "pdf_url": "https://openreview.net/pdf/b4a9e26de0654b238934b820f93e2915d936ba45.pdf", "updated_at": "2026-10-05T20:26:52.100Z", "categories": [ "ptq" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KR6W0HrYEk", "number": 35798, "title": "Understanding Quantization-Aware Training: Gradients at Quantized Weights Bias to the Low-Loss Basin", "abstract": "Post-training quantization (PTQ) converts a trained full-precision model into low-bit weights without task-level retraining, while quantization-aware training (QAT) incorporates quantization into the training loop. Although PTQ is efficient and often accurate at moderate bitwidths, it can fail sharply at aggressive bitwidths; QAT is more expensive but can often recover the lost accuracy. We propose a unified geometric framework that explains both PTQ failure and QAT recovery. We model full-precision training as following a low-loss \\emph{river} inside a wider \\emph{valley}: a normal neighborhood of the river forms a nearly flat \\emph{basin}, while leaving this basin incurs a sharp loss increase. When the quantization grid is comparable to the basin width, local PTQ objectives, including rounding and Hessian-based second-order reconstruction, can select a high-loss deployed quantized point outside the basin even when nearby low-loss quantized points exist. In this regime, straight-through-estimator-based QAT has a useful bias: it evaluates gradients at the deployed quantized weights while updating latent full-precision weights, causing the gradient to sense the valley wall and acquire an inward component that steers subsequent quantized iterates back into the basin. We formalize this mechanism through a local landscape model, construct a geometric PTQ failure mode, and prove finite-time QAT recovery under the landscape and quantizer-compatibility assumptions. Experiments across vision and language models under multiple quantization schemes show PTQ degradation followed by QAT recovery, with landscape diagnostics consistent with this mechanism.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-aware training", "loss landscape", "LLM training" ], "forum_url": "https://openreview.net/forum?id=KR6W0HrYEk", "pdf_url": "https://openreview.net/pdf/23effc8c0ee5b9cb47a077aaf6ecc083feb72c52.pdf", "updated_at": "2026-10-05T20:26:53.044Z", "categories": [ "ptq", "qat", "training" ], "matched_terms": [ "low-bit", "ptq", "qat", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "iikGG5Se3J", "number": 35804, "title": "DMap: Efficient Dense Matching and Bundle Adjustment for Accurate 3D reconstruction", "abstract": "Deep 3D models provide remarkable global robustness, but direct regression acts as an information bottleneck that discards sub-pixel geometric constraints. Conversely, classical optimization achieves sub-pixel precision but fails to scale to dense correspondences and relies on heuristic, discrete track building. We present DMap, a framework that bridges frozen foundation representations with efficient dense optimization. First, DMap extracts dense multi-view correspondences from the latent features of a frozen multi-view backbone, amortizing per-view representation computation so that per-pair decoding reduces to lightweight coarse feature matching and local convolutional refinement. Second, we resolve the efficiency bottleneck of non-linear optimization by formulating dense bundle adjustment on a regular view graph over regular 2D pixel grids rather than irregular point graphs. By anchoring tracks to source rays, dense warps enter the optimization directly without keypoint quantization or track merging, and millions of depth variables are marginalized in parallel within milliseconds. DMap achieves state-of-the-art accuracy across unordered (ETH3D, Texture-Poor) and sequential (TUM RGB-D, ETH3D-SLAM) benchmarks, improving relative-pose AUC@1◦ to 88.6% on ETH3D while running 13–17× faster than existing dense optimization baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "3D reconstruction", "dense matching", "bundle adjustment", "multi-view geometry", "geometric optimization", "visual representation learning" ], "forum_url": "https://openreview.net/forum?id=iikGG5Se3J", "pdf_url": "https://openreview.net/pdf/71acac36cd00337e0813503007067a6473e61766.pdf", "updated_at": "2026-10-05T20:26:54.826Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "18UVn4y0SO", "number": 35823, "title": "How Far Can MXFP4 Go in Pretraining?", "abstract": "While pretraining large language models in 4-bit floating point (FP4) can substantially reduce training cost, FP4 formats struggle to cover the dynamic range of the weights, activations, and gradients encountered during training. Popular FP4 formats therefore pair each small block of values with a shared scale. MXFP4 uses power-of-two block scales that are simpler to support in hardware, while NVFP4 spends mantissa bits on its local scale to reduce quantization error. The two formats thus trade hardware simplicity against scale precision, and recent work reports larger pretraining degradation for MXFP4 than for NVFP4.\nIn this work, we propose a MXFP4 pretraining recipe, MXFP4-MARS, based on recent advanced techniques, including Hadamard rotation, stochastic rounding, and Overflow-Aware Scaling (OAS). Our recipe also leverages Macro Block Scaling (MBS), which maintains higher-precision scale at a coarser granularity while keeping hardware-friendly power-of-two per-block scales.\nAt matched perplexity, our MXFP4 recipe closes the token-overhead gap to NVFP4 from the previously reported 36\\% down to 0.21\\% on average, a 171$\\times$ reduction in additional tokens required.\nAlso, at the same training budget, a 7B model trained using MXFP4-MARS yields a 1.4\\% higher average downstream score than the model trained with NVFP4. These results identify scale hierarchy as a key design dimension for low-precision training and show that MXFP4 can be quality-competitive with NVFP4 while keeping the hardware-friendly power-of-two scales, making it an attractive, portable target for FP4 pretraining.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language model pretraining; low-precision training; FP4 quantization" ], "forum_url": "https://openreview.net/forum?id=18UVn4y0SO", "pdf_url": "https://openreview.net/pdf/3749f69fc069053cde7f907e2bacf4f99302a47b.pdf", "updated_at": "2026-10-05T20:26:56.525Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "4-bit", "fp4", "low-precision", "mxfp4", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RSvBc8XlqY", "number": 35838, "title": "Lagrangian--Hamiltonian Flows for Video Prediction and Image Generation: A Symplectic Perspective", "abstract": "We introduce LHFM, a geometric framework for learning image dynamics\n that combines Lagrangian image representations with Hamiltonian\n dynamics. These structures play a central role\n in classical mechanics, symplectic geometry, and geometric\n quantization. LHFM represents images as exact Lagrangian\n graphs and models their evolution through image-dependent\n Hamiltonian flows. This construction provides a geometric\n interpretation of image-space dynamics by connecting them to the\n Hamiltonian evolution of Lagrangian submanifolds. The framework\n supports both video prediction, our primary application, and image\n generation. We propose a novel deterministic video prediction method based\n on the Hamiltonian evolution of Lagrangian image representations.\n This video variant, LHFM-V, is a recurrent model that achieves the\n lowest reported FLOP count among the compared recurrent models with\n similar prediction accuracy. The image variant, LHFM-I, shows that the\n same construction is compatible with flow matching: in a matched\n experiment, it attains a lower FID than the flow-matching baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Flow matching; Generative modeling; Lagrangian submanifolds; Hamiltonian dynamics; Symplectic geometry; Image generation; Video prediction." ], "forum_url": "https://openreview.net/forum?id=RSvBc8XlqY", "pdf_url": "https://openreview.net/pdf/0a2fae3dcee9b7f5a85720585cdd9bb4ccb4e4d0.pdf", "updated_at": "2026-10-05T20:27:00.367Z", "categories": [ "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AVXqDYsi04", "number": 35843, "title": "CRAFT: Coupled Rotation-Aware Action-Field Caching and Two-Bit Quantization for Vision-Language-Action Models", "abstract": "Diffusion and flow-matching vision-language-action (VLA) policies incur substantial weight traffic through repeated action-head evaluations. Prior VLA quantization typically retains head weights at W4 or higher, while scalar W2 can severely degrade task success. Adaptive caching introduces per-step decision overhead that can offset savings on short trajectories ($T\\leq10$), despite often selecting identical evaluation steps. We introduce Coupled Rotation-Aware Action-Field Caching and Two-Bit Quantization (CRAFT), a training-free framework that jointly reduces action-head precision and evaluation frequency. Motivated by scalar-W2 error concentration in a few high-activation-energy directions, CRAFT quantizes the head to W2A4 using lattice codewords. It selects modules for rotation before deployment using the original weights' tail statistics and reuses the latest predicted field at fixed steps, skipping head evaluation while continuing solver-state updates. With a W4 backbone, W2A4 head, and caching, CRAFT achieves LIBERO-Object success rates of 98.4\\% and 97.8\\% on GR00T N1.5 and $\\pi_{0.5}$ (FP16: 96.4\\% and 98.0\\%). Batch-1 end-to-end speedups on RTX 3090 reach $2.06\\times$ and $2.46\\times$, respectively, versus $1.40\\times$ and $1.38\\times$ for the fastest prior methods. On GR00T N1.5, nominal weight traffic decreases by 79.6\\% relative to FP16 and by a factor of 2.1-3.2 relative to prior methods. Head-only W2A4 quantization achieves 80.45\\% average success across eight models, exceeding the strongest matched-precision baseline by 15.3 percentage points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-language-action models", "efficient robotic inference", "post-training quantization", "extreme-low-bit quantization", "predicted-field caching", "training-free acceleration" ], "forum_url": "https://openreview.net/forum?id=AVXqDYsi04", "pdf_url": "https://openreview.net/pdf/16c8d0ad0eadc8df84649df7d27c3698dfc4084f.pdf", "updated_at": "2026-10-05T20:27:01.766Z", "categories": [ "ptq", "hardware", "diffusion", "multimodal" ], "matched_terms": [ "low-bit", "quantization", "quantizes", "w2a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1yPCfHNbh8", "number": 35869, "title": "MARK-VLA: Mixed-Precision Action-Risk-Guided Quantization with Mixed-Weight Kernels for Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models couple visual perception, language reasoning, and action generation within a unified policy, but their large model size and computationally intensive inference impose substantial memory and latency overheads. We present MARK-VLA, a post-training quantization framework that enables efficient VLA inference through channel-wise mixed-precision quantization and tailored GPU kernel support for mixed-precision weights. Instead of pursuing extreme compression for individual modules, MARK-VLA applies W2/W4/W8 weight quantization with A8 activation quantization across the major perception-to-action stages to reduce the overall memory footprint. We propose a per-module precision allocation method that automatically determines channel-wise precision based on gradient-based action risk evaluated under a quantized reference policy. To account for the discrepancy between the risk estimated for individual quantization decisions and their combined effect, we adopt a progressive precision reduction scheme that gradually lowers the quantization precision. Finally, to translate these fine-grained precision assignments into practical acceleration, we develop fused mixed-weight GPU kernels together with trace-guided per-layer specialization to optimize the complete execution path of quantized Linear layers. Experiments on LIBERO across six VLA policies and three GPU platforms demonstrate up to $3.62\\times$ compression and up to $1.69\\times$ end-to-end speedups, while the largest decrease in average success rate is only 0.82 percentage points. These results demonstrate that action-aware precision allocation and deployment-oriented kernel co-design can jointly achieve compact and efficient VLA inference with minimal accuracy degradation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action Models", "Post-Training Quantization", "Mixed-Precision Quantization", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=1yPCfHNbh8", "pdf_url": "https://openreview.net/pdf/60aba1ef0436e86c771381e75f67fb5bdf4787cd.pdf", "updated_at": "2026-10-05T20:27:03.624Z", "categories": [ "ptq", "training", "hardware", "reasoning", "multimodal" ], "matched_terms": [ "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gmCR6gwMUA", "number": 35895, "title": "PowLU: An Activation Function for Stable Pre-Training of LLMs", "abstract": "In contemporary large language models (LLMs), the swish-gated linear unit (SwiGLU) activation function is widely adopted to regulate the information flow and introduce non-linearity. In its scalar unary representation, SwiGLU approximates the quadratic function $x^2$ for large positive inputs, providing strong nonlinearity and expressive capacity. However, this property also causes numerical instability as the input or model scale increases, particularly in low-precision LLM training. The main reason is the approximate quadratic amplification reflected by its unary form, which enlarges the output range and exacerbates outliers. To address this issue, we propose a stable activation function, Power Linear Unit (PowLU), for large-scale LLM pre-training. Specifically, PowLU employs a rational power function to achieve adaptive nonlinearity, thereby improving representation ability and enabling stable training in spike regions. Moreover, we provide theoretical justification for several key properties of PowLU. Scaling law experiments confirm that the performance is consistent across model sizes, and further experimental results with the Ling-3.0 architecture (7.9B and 124B total parameters) demonstrate that PowLU achieves competitive results against SwiGLU and SwiGLU-Clip in large-scale training of LLMs. In addition, the experimental results also show that PowLU effectively improves the scalability of the large-scale training of LLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Pre-Training", "Activation Function" ], "forum_url": "https://openreview.net/forum?id=gmCR6gwMUA", "pdf_url": "https://openreview.net/pdf/c21acce902b55965c3e404544b192b823bc7b5c7.pdf", "updated_at": "2026-10-05T20:27:08.863Z", "categories": [ "theory" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7ktmLdjqgQ", "number": 35928, "title": "Don't Repeat Yourself: Stopping Verbatim Loops at Sampling Time", "abstract": "Large Language Models generate text by sampling tokens autoregressively, but open-ended generation is prone to verbatim looping, where the model repeats spans already present in its context. Standard defenses, such as repetition, presence, and frequency penalties, and n-gram blocking, act on token recurrence rather than on the sequential structure of a loop, and suppress looping only at strengths that also degrade formatting and fluency. We propose Don’t Repeat Yourself (DRY), a sampling-time logit adjustment that penalizes a candidate token only when generating it would extend the current suffix into an exact continuation of a span seen earlier in the context, with sequence breakers that protect chat templates and formatting tokens. Our experiments across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study show that DRY reduces the suffix-extension rate by 47% while improving lexical diversity. An intervention-matched placebo produces no such reduction, identifying suffix-matching as the operative mechanism. On AWQ-quantized 70B and 120B models, DRY reduces the loop rate by roughly half while preserving MT-Bench, MMLU, and GSM8k, where standard alternatives lose measurable ground. DRY has been adopted by popular open-source LLM inference frameworks, including llama.cpp, ExLlamaV2, and text-generation-webui, highlighting its impact on practical text generation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "repetition", "penalty", "LLM" ], "forum_url": "https://openreview.net/forum?id=7ktmLdjqgQ", "pdf_url": "https://openreview.net/pdf/d1005c1e191001ecc09b9214bcee385c6d3f652e.pdf", "updated_at": "2026-10-05T20:27:12.362Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sm22EahEaf", "number": 35938, "title": "SAGE-VQ: Structure-Adaptive Gating-Aware Expert Vector Quantization with Output Correction for MoE LLMs", "abstract": "Mixture-of-Experts (MoE) large language models (LLMs) expand capacity through sparse activation, yet expert storage remains a bottleneck for single-GPU and on-device deployment. Post-training vector quantization (VQ) enables ultra-low-bit compression, but dense-LLM methods overlook heterogeneous expert sensitivity, architecture-dependent output composition, and routing-weighted errors. We introduce SAGE-VQ, a structure-adaptive, gating-aware expert VQ framework that aligns precision allocation and output correction with MoE structure. SAGE-VQ selects architecture-appropriate sensitivity signals to assign expert bit-widths under a fixed average bit budget. It incorporates routing weights into the correction objective to mitigate norm and scale mismatches in quantized expert outputs. Correction factors are fused before inference without additional runtime latency. On Qwen2-MoE, SAGE-VQ achieves 2.25-bit expert compression, reducing WikiText-2 perplexity by 1.21 relative to the MoE-aware mixed-precision baseline MxMoE while limiting average accuracy degradation across seven zero-shot benchmarks to 1.42 percentage points relative to FP16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-experts", "post-training quantization", "vector quantization", "ultra-low-bit compression", "mixed-precision quantization", "routing-aware error compensation" ], "forum_url": "https://openreview.net/forum?id=sm22EahEaf", "pdf_url": "https://openreview.net/pdf/bfc6d3a4f3f3b852bc83ccbdc3a593541e27d6c6.pdf", "updated_at": "2026-10-05T20:27:13.025Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "X4XngsoBml", "number": 35945, "title": "Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate", "abstract": "Can LLMs recognize degradation in their own computational substrate?\nInspired by *anosognosia*, a neurological condition in which patients fail to recognize impairments in their own abilities, we investigate whether LLMs can recognize degradation in their computational substrate induced by quantization.\nWe first show that existing models *fail* to self-report their quantization state, even when provided with their own generated text as an external cue.\nLinear probing reveals that, while generated text carries almost no trace of quantization, internal representations contain clear, method-specific fingerprints.\nThrough training, models learn to identify severely degraded outputs such as those of 4-bit models by comparison, yet still fail to do so from a single output.\nA shared LoRA trained jointly across quantization levels *succeeded* in reading out internal fingerprints, but fails on unseen quantization methods, merely mapping method-specific fingerprints to labels.\nWhereas external self-observation can restore awareness in some cases of human anosognosia, our results suggest that the more promising route to enabling such awareness in LLMs may lie in their internal representations.\nOur results highlight fundamental limits of generalizability to LLM self-monitoring.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Metacognition", "Self-Awareness", "Weight Quantization", "Mechanistic Interpretability", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=X4XngsoBml", "pdf_url": "https://openreview.net/pdf/96693fd25b9babdccc614cdbd8c62283d6b5d6e1.pdf", "updated_at": "2026-10-05T20:27:14.065Z", "categories": [], "matched_terms": [ "4-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BXUnpS9Y2b", "number": 36025, "title": "Quantization-Conditioned Backdoors on Large Language Models via Attention Sinks", "abstract": "Post-training quantization is widely used to reduce the deployment cost of large language models (LLMs), but recent studies have shown that quantization may be exploited as a backdoor trigger: a released model may appear benign before quantization while exhibiting attacker-specified behaviors after quantization. Existing quantization-conditioned backdoor attacks first implant the target behavior into the quantized model and then restore benign behavior at full precision while keeping the weights within a subspace that yields the same quantized values as the backdoored counterpart. This requires optimizing a complex objective within a highly constrained weight space, leading to limited attack effectiveness. To address this issue, we introduce SinkQCB, which separates the insertion of malicious behavior from the mechanism that triggers it. More specifically, the malicious behaviors are encoded in a small set of attention heads, and attention sinks are utilized to keep these heads suppressed at full precision and activated after quantization. Extensive experiments on commonly considered LLMs, malicious tasks, and quantization methods demonstrate that SinkQCB achieves a higher attack success rate (ASR) after quantization while keeping ASR at or below 1\\% and maintaining competitive performance on benign tasks at full precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Backdoor Attacks", "Post-Training Quantization", "Attention Sinks", "Model Security" ], "forum_url": "https://openreview.net/forum?id=BXUnpS9Y2b", "pdf_url": "https://openreview.net/pdf/17e1e012244f709b8849ec561a9c475393cc0453.pdf", "updated_at": "2026-10-05T20:27:27.041Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pytrHkemYT", "number": 36136, "title": "RBS-ATTENTION: Radius-Bounded Sparse Prefill for Long-Context Large Language Models", "abstract": "Long-context large language model inference is increasingly limited by prefill, where dense self-attention processes the entire prompt before generation begins. Sparse block selection reduces this cost, but a block centroid may hide a highly relevant token among many irrelevant ones. We call this failure mode mean dilution and propose RBS-Attention, a training-free sparse-prefill method with two complementary branches. A centroid base branch captures average relevance, while a rescue branch uses the maximum key-block radius and its prompt-, layer-, and head-dependent distribution to identify blocks at risk of underestimation. Independent thresholds and a mask union focus rescue on dispersed blocks while preserving block-sparse FlashAttention execution. On H100 GPUs, RBS-Attention achieves $20.65\\times$ standalone prefill-attention speedup and $5.97\\times$ end-to-end time-to-first-token speedup at 128K on Qwen3-30B-A3B-Instruct-2507-FP8. On Qwen3-32B, it obtains 88.65 overall RULER accuracy versus 89.52 for dense attention; LongBench-v2, InfiniteBench, and Video-MME provide additional quality evaluation. Matched-density and selector studies further support radius-adaptive rescue as an effective approach to long-context prefill.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Sparse Attention", "Long-Context LLM Inference", "Prefill Acceleration", "Adaptive Block Selection", "Training-Free Inference" ], "forum_url": "https://openreview.net/forum?id=pytrHkemYT", "pdf_url": "https://openreview.net/pdf/fabeb0f534c60164a10ed80a2f85100d01004956.pdf", "updated_at": "2026-10-05T20:27:41.185Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bgjT8XlYOl", "number": 36171, "title": "One Man's Cure Is Another's Poison: Turning JPEG Compression Into Backdoor Triggers for Vision-Language (Action) Models", "abstract": "JPEG compression is ubiquitous in practical vision-language model (VLM) pipelines for storage, transmission, and input preprocessing, and has also been evaluated as a lightweight defense against adversarial perturbations. Existing studies largely assume that compression suppresses malicious behavior or represents a distortion that adversarial examples should survive. In this work, we find that such a defense may instead be exploited by the adversary as a trigger for attacks, i.e., the input appears benign but induces attacker-specified behavior after JPEG compression in vision-language(-action) tasks. We propose JPEG-triggered Adversarial attaCK (JACK), a two-stage method that achieves pre-compression stealthiness and post-compression attack effectiveness without modifying the victim model or the JPEG codec. JACK first constructs a candidate image whose JPEG-compressed version induces the target behavior, and then restores benign behavior for the uncompressed counterpart within the same JPEG quantization cell, thereby preserving the attack-inducing compressed representation. Extensive experiments on targeted image classification, visual question answering, multimodal jailbreaking, and multi-view Vision-Language-Action (VLA) policy hijacking validate the effectiveness of the proposed method.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "VLM", "adversarial attack", "jailbreak attack", "JPEG compression", "stealthiness" ], "forum_url": "https://openreview.net/forum?id=bgjT8XlYOl", "pdf_url": "https://openreview.net/pdf/7c493a0c489ac6afad8b3bb4866ffa3f9991b312.pdf", "updated_at": "2026-10-05T20:27:45.398Z", "categories": [ "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wmBiZKsp47", "number": 36246, "title": "Beyond Binary Correctness: Confidence Estimation for Continuously Graded Responses in Large Language Models", "abstract": "Confidence estimation predicts the correctness of generated outputs of LLMs. Most existing approaches typically use an arbitrary threshold over a continuous quality metric, such as syntactic or semantic similarity against ground truth, to determine whether an LLM response is correct. In this paper, we propose modeling confidence as a probability distribution over \\emph{continuous} correctness scores, thereby remaining agnostic to threshold choice, avoiding information loss from binarization, and enabling rich uncertainty estimates such as confidence variance and credible intervals for robust decision making. We propose Beta regression and piecewise uniform regression approaches as distributional methods, showing theoretically that Beta regression achieves superior sample efficiency and faster convergence for smooth distributions with at most one interior mode. We also conduct an empirical study on several datasets across diverse tasks to analyze and compare methods with each other, as well as traditional point-based confidence estimation baselines, demonstrating their relative suitability and advantages.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "confidence estimation", "uncertainty quantification", "large language models", "distributional regression" ], "forum_url": "https://openreview.net/forum?id=wmBiZKsp47", "pdf_url": "https://openreview.net/pdf/036a21d8f60e5d3666bf511a515cbaa9d7cfea42.pdf", "updated_at": "2026-10-05T20:27:56.516Z", "categories": [ "theory" ], "matched_terms": [ "binarization", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "34VlQmtqkM", "number": 36262, "title": "Compress Locally, Predict Globally: Composable Margin Statistics for Estimating Compressed LLM Accuracy", "abstract": "We introduce a statistical model that predicts the task accuracy of compressed language models without constructing and evaluating each candidate configuration. For each block and compression option, we summarize changes in answer margins (the score differences between competing answers) by a scaling coefficient and a residual variance. We estimate these statistics by compressing one block at a time, then reuse and combine them across configurations. We use the combined statistics to parameterize a Gaussian score model. For each example in a labeled reference set, we use the uncompressed model's scores to estimate the probability that the compressed model selects the correct answer. We average these probabilities to predict task accuracy. Across exhaustively evaluated MMLU quantization and pruning allocation spaces, prediction mean absolute error (MAE) ranges from 0.61 to 3.83 percentage points (pp). Across five K2-Horizon-7B quantization spaces, composing margin statistics achieves a mean per-space prediction MAE of 1.10 pp, compared with 1.55 and 1.42 pp for additive and multiplicative accuracy composition. Comparisons with covariance-based and empirical-residual predictors further support the accuracy estimates from the fitted Gaussian model. GPTQ prediction MAE remains within 0.67-1.19 pp across five checkpoints from 7B to 65B and reaches 0.90 pp on 24 presampled individual-layer allocations. The same prediction rule extends to BoolQ, HellaSwag, WinoGrande, and encoder-based AG News classification. As an application, maximizing predicted accuracy selects the best measured configuration in the complete Gemma-31B quantization space and improves pruning allocation over OWL under matched budgets.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "quantization", "pruning", "performance prediction", "large language models" ], "forum_url": "https://openreview.net/forum?id=34VlQmtqkM", "pdf_url": "https://openreview.net/pdf/8ce00620e99e210ed5b5fc6f412a76f586ed2c6c.pdf", "updated_at": "2026-10-05T20:27:58.141Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WeBggKgUhn", "number": 36471, "title": "Understanding LLM Quantization through Activation-Guided Compensation and Orthogonal Residuals", "abstract": "Post-training weight-activation quantization reduces the memory and inference costs of large language models, but aggressive W4A4 quantization remains difficult because activation outliers degrade effective quantization resolution.\nAlthough weight optimization, channel-wise scaling, and orthogonal rotation mitigate this problem, the error components they address and their relationship remain unclear.\nUsing an exact decomposition of local weight-activation quantization error into an activation-guided weight compensation term and an orthogonal residual, we bound the residual using persistent channel-wise outlier and regular activation quantities.\nThis decomposition clarifies which error components can be addressed by weight compensation and which require transformation design.\nWe then use the residual bounds to derive practical guidelines for randomized Hadamard rotation, sign-based probability amplification, and channel scaling.\nIn particular, the analysis explains how random signs suppress constructive interference among persistent outlier channels and how probability amplification through independent sign randomization tightens the high-probability bound on the minimum outlier term across sign configurations.\nIt also shows how second-moment balancing leads to an $L_2$ scaling rule while a further relaxation recovers SmoothQuant-style $L_\\infty$ scaling.\nWe evaluate these guidelines through backpropagation-free configurations across eight Llama and Mistral models, obtaining performance competitive with gradient-trained SpinQuant.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "weight-activation quantization", "quantization error analysis" ], "forum_url": "https://openreview.net/forum?id=WeBggKgUhn", "pdf_url": "https://openreview.net/pdf/9b9776771bf53bdbed19875a38e51d0ef9286cb3.pdf", "updated_at": "2026-10-05T20:28:26.224Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hgUngr5Oxt", "number": 36496, "title": "ReQOR: Operator-Discrepancy Subspace Repair for Quantization-Conditioned Backdoors in Large Language Models", "abstract": "Quantization-Conditioned Backdoors (QCBs) cause large language models (LLMs) to exhibit malicious behavior after quantization while remaining benign at high precision. Existing defenses often rely on security-guided configuration selection or iterative optimization, but do not identify which quantization-induced changes should be repaired. We isolate module-local quantization effects from upstream drift by evaluating the high-precision and quantized versions of each module on identical states from the quantized inference trajectory, and formulate QCB defense as repairing the resulting same-state operator discrepancies.\nWe propose ReQOR, a training-free defense that identifies the dominant directions of these discrepancies on benign calibration data and restores the full-precision computation along them in closed form, yielding a factorized residual adapter. ReQOR uses neither attack information nor gradient-based optimization and supports both full-span and low-rank repair. Experiments across four LLMs, three QCB attacks, three malicious behaviors, and three quantization formats show substantial QCB mitigation while largely preserving benign utility. Further analysis shows that dominant operator-discrepancy directions provide more effective and rank-efficient repair than rank-matched random, low-energy, or raw weight-error directions, while low-rank ReQOR reduces deployment memory overhead.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Conditioned Backdoor", "Large Language Models", "LLM Security", "Model Quantization", "Backdoor Defense", "AI Safety", "Trustworthy AI" ], "forum_url": "https://openreview.net/forum?id=hgUngr5Oxt", "pdf_url": "https://openreview.net/pdf/1f31345d7d9de4cc68c88163a1d502b0d13e6252.pdf", "updated_at": "2026-10-05T20:28:29.329Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "a0I8jYG8ov", "number": 36585, "title": "Complexity-theoretic tradeoffs between precision and expressivity in transformers", "abstract": "A common technique to accelerate inference and reduce the memory costs of transformers and related models is to restrict to low-precision weights and arithmetic. On the other hand, doing this could reduce the expressive power of the model, but the extent of this phenomenon remains poorly-understood.\n\nIn this paper, we demonstrate a fine-grained theoretical tradeoff between expressivity and precision: For every $p$ we exhibit a function $\\Gamma$ which tests equalities, and prove that a one-layer softmax transformer can compute $\\Gamma$, with $p$ bits of precision, but not with $p-1$ bits of precision. Equality testing arises, both explicitly and implicitly, in many tasks that transformers are expected to perform, and our results show it is particularly sensitive to model precision.\n\nOur proofs combine explicit finite-precision transformer constructions with communication-complexity lower bounds, yielding a tight ``one-bit'' threshold.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "precision", "lower bounds", "expressivity", "communication complexity", "transformers", "quantization" ], "forum_url": "https://openreview.net/forum?id=a0I8jYG8ov", "pdf_url": "https://openreview.net/pdf/a73f9c5b14bf5f40c2d5270b37050f6d3349db78.pdf", "updated_at": "2026-10-05T20:28:41.858Z", "categories": [ "theory" ], "matched_terms": [ "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Gwmhg9mHCB", "number": 36611, "title": "Meaning Before Words", "abstract": "Predicting words and preserving meaning are different tasks. We study this gap in a finite language model where many strings express one meaning. The probability chain rule can already represent any latent semantic generator; the obstruction lies in the decisions and resource allocation built around it. For every fixed temperature below one, we construct a full-support binary model whose semantic error tends to zero, while both token-level and sequence-level temperature sampling have error tending to one. Even exact sequence maximization fails. We give a sharp finite-length bound for the local effect. Under a shared memory limit, an optimal next-token predictor can discard an available semantic bit; we compute the exact cost of protecting it. A sharp log-loss bound separates semantic guarantees from length-normalization artifacts. Finally, a meaning-first interface makes decisions independent of verbal form when its decoder is faithful. The results separate semantic decision-making from its autoregressive language realization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Next-token prediction", "Surface form competition", "Decoding strategies", "Semantic representation" ], "forum_url": "https://openreview.net/forum?id=Gwmhg9mHCB", "pdf_url": "https://openreview.net/pdf/98a5c4a4a7c9784806fdc8fb6564645fa30612e6.pdf", "updated_at": "2026-10-05T20:28:45.703Z", "categories": [], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XJg2ho9Nn9", "number": 36707, "title": "ReCaC: Residual-Guided Calibration via Activation Coverage for LLM Quantization", "abstract": "As a crucial technique for efficient deployment of large language models (LLMs), post-training quantization (PTQ) performs low-bit quantization for LLM relying on a pre-defined calibration dataset. Logically, the calibration dataset determines quantization parameters and so clearly influences PTQ performance. However, effective calibration construction is still underexplored. Particularly, existing methods typically focus on selecting representative and task-related samples while overlooking the complementarity among selected samples, which hardly guarantees coverage of the calibration dataset and limits the generalization ability of quantized models. Therefore, this paper proposes a Residual-guided Calibration via activation Coverage (ReCaC) method, aiming to construct an effective calibration dataset by maximizing the sample coverage in the activation space under a small fixed budget. Specifically, given a pre-defined benchmark for constructing the calibration dataset, our ReCaC uses residual-guided calibration construction (RGC) to progressively seek and pick the samples with the smallest residual score after being added to the currently selected calibration set, which effectively prioritizes complementary activation information and maximizes calibration coverage in activation space. Furthermore, ReCaC can be flexibly extended to multi-domain scenarios by joint global and domain coverage (JGDC), which selects the samples that provide complementary contributions to the overall activation space while maintaining sufficient coverage for each domain. As such, ReCaC can capture complementary activation patterns among heterogeneous domains and so enlarge the coverage of calibration dataset. Experiments are conducted by using different language models across various post-training quantization methods and settings, and results on a dozen downstream tasks demonstrate ReCaC achieves highly competitive or superior performance compared with existing state-of-the-art calibration methods in both iintra-domain, inter-domain and out-of-domain downstream tasks, particularly under the setting of lower-bit quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model", "model quantization", "calibration" ], "forum_url": "https://openreview.net/forum?id=XJg2ho9Nn9", "pdf_url": "https://openreview.net/pdf/4774a49fb55d0b28ae1266078e2fcd3e61d243a8.pdf", "updated_at": "2026-10-05T20:29:00.243Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "L5LdjTSK7k", "number": 36727, "title": "BASS: A Block-wise Assessment of Sensitivity and Selective Smoothing Framework for Quantization of Branch-Conditioned Video Diffusion Transformers", "abstract": "Video Diffusion Transformers (VDiTs) have substantially advanced high-resolution and controllable video generation, but their considerable computational and memory costs make practical deployment prohibitively expensive. Post-training quantization (PTQ) has proven to be an effective method for reducing memory usage and computational costs, yet existing methods primarily target VDiTs with a single denoising backbone and suffer from accuracy degradation when applied to branch-conditioned architectures with auxiliary control signals. Wan2.2-Animate-14B is a representative example: its auxiliary face and body adapters inject motion and identity conditioning signals into the main DiT, and its conditioning pathways exhibit quantization sensitivities that differ markedly from those of the backbone. Applying a uniform low-bit configuration across these components can therefore corrupt the conditioning signals, leading to motion degradation and identity drift. To address this problem, we propose BASS, a Block-wise Assessment of Sensitivity and Selective Smoothing framework for mixed-precision post-training quantization of branch-conditioned VDiTs. BASS identifies and protects sensitive layers across different branches, particularly those within the adapters, with minimal manual intervention. We validate BASS on advanced branch-conditioned VDiTs, using a mixed-precision policy combining W8A8 and W4A4, and show that it preserves motion fidelity, identity consistency, and overall visual quality under a mixed low-bit quantization policy. Furthermore, we implement efficient GPU kernels to achieve practical memory savings and inference speedups.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "video generation", "low-bit quantization", "diffusion model" ], "forum_url": "https://openreview.net/forum?id=L5LdjTSK7k", "pdf_url": "https://openreview.net/pdf/0fd596c0b10bf701cdbd58da62a03571da12fa1f.pdf", "updated_at": "2026-10-05T20:29:02.665Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "low-bit", "mixed-precision", "ptq", "quantization", "w4a4", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bYpvxDEkUD", "number": 36742, "title": "TeleQuant: Temporal Noise Shaping for Post-Training Quantized Diffusion Transformers", "abstract": "Post-training quantization lowers the cost of the repeated denoiser evaluations that diffusion Transformers use for image and video generation. Because those evaluations are chained, the rounding errors they leave behind interact across steps. Correlated rounding errors can reinforce, accumulate into structured trajectory drift, and persist in the returned sample. This drift can degrade image and video fidelity. We propose TeleQuant, a training-free temporal noise-shaping framework that feeds a discarded activation-rounding residual into the next evaluation, with a gain matched to adjacent solver responses and activation scales. We derive a first-order telescoping identity: under ideal state storage, accumulated error separates into a terminal residual and a sensitivity-variation term, motivating selective layer placement and damped feedback. INT8 residual state and a fused activation operator add the compensation on the host's path. The feedback lowers accumulated latent drift and improves paired fidelity to matched FP16 outputs. Evaluations on DiT-XL/2, PixArt-$\\Sigma$, SDXL, and Open-Sora 1.2 show improved image fidelity and video consistency over frozen ViDiT-Q and SVDQuant hosts, with matched rounding ablations supporting the complete gain. The profiled image pipelines retain most of the host speedup, at $1.03\\times$–$1.05\\times$ latency. Code is available at https://anonymous.4open.science/r/TeleQuant-8217/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion transformer", "post-training quantization", "temporal noise shaping", "trajectory drift", "efficient generation" ], "forum_url": "https://openreview.net/forum?id=bYpvxDEkUD", "pdf_url": "https://openreview.net/pdf/85e3f775383b7af973a25ad99bfbcb7069fca187.pdf", "updated_at": "2026-10-05T20:29:06.060Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "int8", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qNTM1tmIOT", "number": 36761, "title": "CSQuant: Cross-Block Smoothing for Accurate and Efficient Low-Bit LLM Quantization", "abstract": "Low-bit quantization is essential for efficient large language model (LLM) inference, but aggressive quantization remains vulnerable to activation outliers that amplify quantization error.\nOrthogonally rotating activations alleviates quantization error but incurs additional online computation, particularly in adaptive methods that use intermediate permutations, which prevent multiple rotations from being combined and introduce explicit data movement.\nWe propose CSQuant, an accurate and efficient post-training quantization framework comprising two key modules: Cross-Block Smoothing and LiteBoA.\nCross-Block Smoothing replaces intermediate permutations with a normalized cross-block Hadamard transform, enabling cross-block outlier redistribution while allowing two adaptive rotations to be merged into a single online transformation.\nThis separable structure further enables CUDA kernel fusion with adjacent operations, reducing intermediate tensor materialization and global-memory traffic.\nLiteBoA selectively applies attention-aware second-order compensation to value projection matrices and employs head-shared Hessians to reduce calibration cost while preserving quantization accuracy.\nExperiments on LLaMA-1/2/3 demonstrate improved quantization accuracy across W4A4KV4 and weight-only settings, with particularly strong performance at 2-bit weights.\nCSQuant achieves a 2.25$\\times$ prefill speedup over FP16 under W4A4KV4 quantization with a sequence length of 2048 and a batch size of 4, while our fusion strategy further accelerates existing Hadamard-based rotation methods. CSQuant also substantially reduces peak GPU memory during post-training quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Efficient Low-Bit LLM Quantization", "Post-Training Quantization", "Low-Bit Inference" ], "forum_url": "https://openreview.net/forum?id=qNTM1tmIOT", "pdf_url": "https://openreview.net/pdf/b8b2349a73b1b87703b7c7bd5644f6b2d8e677f5.pdf", "updated_at": "2026-10-05T20:29:07.159Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ixqm6k0cVW", "number": 36786, "title": "Q-WAM: 4-Bit Quantization of World Action Models with Action-Subspace Protection", "abstract": "World Action Models (WAMs) jointly generate video and robot actions through iterative diffusion and\nperform strongly in robotic manipulation. However, their prohibitive compute and memory costs pose substantial deployment challenges. Post-training quantization (PTQ) can reduce these costs, but existing\nPTQ methods such as smoothing and rotation are insufficient to maintain the precision of action generation. To overcome this limitation, we propose Q-WAM, a new 4-bit weight-activation quantization for WAMs that\npreserves the actions the model generates. Specifically, we introduce the Action\nObservability Gramian (AOG), which measures how much rounding errors in each weighted combination of a layer's input\nchannels change the final action through all denoising steps. We also develop\nAction-Subspace Protection (ASP), which keeps the few most action-sensitive channel combinations in a tiny 16-bit\nbranch and quantizes the remaining weights and activations to 4 bits.\nFinally, we aggregate the AOG-derived action mass across the layers of each expert to identify the experts that matter most for the generated action, and apply ASP only to those experts, thereby preserving action quality with minimal overhead.\nWe evaluate Q-WAM on three WAMs, both in simulation and in real-world deployment. On the RoboTwin 2.0\nbenchmark, it reaches 89.6--93.0\\% average success rate, within 1.1 percentage points of the 16-bit models, while reducing the memory of\nthe quantized blocks by 3.1--3.4$\\times$. Our method outperforms the strongest baseline, SVDQuant, by\n2.5--8.7 percentage points. In real-world deployment on a Unitree G1 humanoid and a UR3 bimanual robot, our quantized\nFast-WAM and ImageWAM reach 56.0\\% and 76.0\\% average success rate, 12.8 and 17.6 percentage points above SVDQuant.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "World Action Models", "Post-training quantization", "Efficient inference", "Robotic manipulation" ], "forum_url": "https://openreview.net/forum?id=Ixqm6k0cVW", "pdf_url": "https://openreview.net/pdf/2dd56df18ac753df9fdd815b5900bf58f0d9d676.pdf", "updated_at": "2026-10-05T20:29:10.332Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "ptq", "quantization", "quantized", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "miyTKYn29T", "number": 36887, "title": "AI as a Compiler: Compiling triton kernels without the triton compiler", "abstract": "Compiler backends are expensive to build and maintain as programming models,\nworkloads, and accelerators evolve. We investigate whether large language models\ncan replace the conventional optimizing and lowering pipeline, a process that we\ncall AI lowering. We study AI lowering from Triton to NVIDIA PTX: an LLM\nagent translates Triton kernels directly into PTX, bypassing Triton’s intermediate\nrepresentations and optimization passes while preserving the kernel ABI, and\nrefines its output using randomized differential testing, benchmarking, and profiler\nfeedback. Across twelve common kernels on Ada, Hopper, and Blackwell GPUs\nand ten kernels from recent ML papers, AI lowering achieves 0.83×-3.34× the\nperformance of autotuned Triton. The largest gains come from transformations\nthat Triton’s lowering pipeline does not perform, such as decoding packed binary\nweights directly into Tensor Core operands (3.34× on BitDelta), assigning each\nthread a complete softmax row in tensor memory (1.36× on FlashAttention), and\nreusing overlapping convolution windows (up to 2.23×). Getting correctness\nguarantees is now the bottleneck. The state-of-the-art PTX verifier does not model\nthe Hopper and Blackwell features these kernels rely on, and restricting generation\nto its supported fragment erases the gains on compute-heavy kernels. With best-\neffort extensions, the verifier checks 16 of 22 kernels; atomics, input-dependent\ncontrol flow, and bit-level floating-point manipulation remain out of reach, making\nverification the central open challenge for trusted AI compilation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "compilation; kernels; verification" ], "forum_url": "https://openreview.net/forum?id=miyTKYn29T", "pdf_url": "https://openreview.net/pdf/9faafad77b1822e1bd67e5897abc1448ef080957.pdf", "updated_at": "2026-10-05T20:29:24.976Z", "categories": [ "hardware" ], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7nJpaxcCEv", "number": 36949, "title": "COMPACT: A UNIFIED FRAMEWORK FOR MULTIMODAL SPLIT LEARNING WITH COMPRESSED ACTIVATIONS", "abstract": "Split learning enables collaborative training of vision–language models without sharing raw data, but transmitting intermediate activations across distributed clients can incur substantial communication overhead. We investigate a lightweight multi-client setting in which client-side modules are frozen, eliminating gradient transmission and parameter synchronization, and focus on activation compression at the client-server boundary. We introduce COMPACT, a unified framework that supports heterogeneous vision–language backbones with\ninterchangeable compression modules, and evaluate seven compression strategies across four vision–language models on four medical VQA benchmarks. We find that compression effectiveness is strongly model-dependent, and blockwise NormalFloat activation quantization (Ql) provides the most consistent performance across backbones. Amongst compression configurations, larger cosine distance from uncompressed representations is consistently related to downstream accuracy, with Spearman correlations ranging from −0.72 to −0.85. Layer-wise analysis further shows that internal representation drift caused by reconstruction errors accumulates through frozen layers. Moreover, we identify two recurring phenomena associated with performance degradation: unstable training of reconstruction modules under short training budgets and poor utilization of quantization levels. Finally, in a simulated cross-silo setting, COMPACT reduces communication by 72.5% for 4-bit Ql, while preserving 98.5% baseline performance. Together, these results show that compression effectiveness depends on both compression rate and its interaction with pretrained representations and downstream adaptation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multimodal Split Learning", "Vision-Language Models", "Quantization" ], "forum_url": "https://openreview.net/forum?id=7nJpaxcCEv", "pdf_url": "https://openreview.net/pdf/86eaa4fa1453fba03a66f2170a94b517ce096663.pdf", "updated_at": "2026-10-05T20:29:35.547Z", "categories": [ "ptq", "training", "multimodal" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GDTAORcpJ3", "number": 37068, "title": "When Weight Fidelity Misleads: Concept-Erasure Safety Edits Under Diffusion-Model Quantization", "abstract": "Concept-erasure edits are usually evaluated before quantization, although deployed models often run at lower numerical precision. We ask whether preservation of an edit in weight space predicts preservation of its function and behavior. On Stable Diffusion 1.5, we separately measure the deployed weight difference, the internal activation difference induced by the edit, and generated outputs.\nUnder NF4 weight quantization, UCE's deployed weight difference has cosine similarity 0.568 with the original edit, while its activation-level effect remains aligned at 0.918. For the nudity edit, projecting the differential weight error onto edit-relevant conditioning inputs reduces its relative magnitude from 1.29 in weight space to 0.44. ESD-u also exhibits substantial distortion of the deployed weight difference while retaining high average activation alignment, extending the parameter–function decoupling to a different editing surface.\nWe next quantize the text-conditioning inputs to selected cross-attention projections. At four activation bits, NudeNet efficacy falls by 0.264–0.290 for UCE and TIME, and CLIP measurements show the same direction, with all four paired intervals excluding zero. The INT8 weights remain identical across activation precisions. An exploratory 48-prompt, six-seed isolation with INT8 weights fixed attributes this change to activation rounding. Raw edited-image nudity scores also rise, matched-magnitude Gaussian controls do not reproduce the weakening, and Dreamlike Photoreal 2.0 reproduces the A4 effect with paired intervals excluding zero.\nOn 50 unrelated prompts, CLIPScore declines similarly for base and edited models, providing no evidence of selective edited-model collapse under this probe. Our evidence is limited to one safety target, related SD-1.5 checkpoints, and simulated quantization of selected projections. Nevertheless, it shows that weight-space fidelity is insufficient: safety edits should be evaluated under the numerical configuration in which they will be deployed.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "concept erasure", "machine unlearning", "post-training quantization", "diffusion models", "text-to-image generation", "AI safety", "activation quantization", "model compression" ], "forum_url": "https://openreview.net/forum?id=GDTAORcpJ3", "pdf_url": "https://openreview.net/pdf/45929023d015a529a791b90d8c6a037246986027.pdf", "updated_at": "2026-10-05T20:29:51.598Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "int8", "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tzL5oGX0QB", "number": 37104, "title": "SCOPE-Q: ONE-CHECKPOINT QUANTIZATION-AWARE TRAINING FOR CROSS-FRAMEWORK LLM DEPLOYMENT", "abstract": "Large language models (LLMs) are often deployed through several frameworks,\nsuch as OpenVINO, TensorRT, Core ML, and AIMET, each of which quantizes the\nmodel with its own quantization configuration. Most LLM quantization methods\noptimize a model for a single, fixed quantizer. We find that this assumption\nis costly in practice: at the same 4-bit precision, the MMLU accuracy of one\ncheckpoint varies by 12–15 points across these four frameworks, mainly because of\ndifferences in their quantization configurations. In this paper, we propose SCOPEQ, a unified quantization-aware training method that produces a single checkpoint\nfor multiple deployment frameworks. SCOPE-Q learns a low-rank update to the\nfrozen pretrained weights while sampling one target quantization configuration\nat each training step, and distills knowledge from the full-precision model. After\ntraining, the update is merged into the weights, yielding a standard floating-point\ncheckpoint that each framework quantizes without modification. Experiments on\nfour LLMs ranging from 0.5B to 4.6B parameters show that a single SCOPE-Q\ncheckpoint improves worst-case accuracy across frameworks by 7.5–8.4 points over\nround-to-nearest quantization and reduces the accuracy gap between frameworks\nby 44–69%, while training only 1.1–3.6% of the parameters.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "quantization", "quantization-aware training", "low-rank adaptation", "weight-only quantization", "cross-framework deployment", "model compression", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=tzL5oGX0QB", "pdf_url": "https://openreview.net/pdf/3e710eb64cad4ec4ef8a1f13fd049101a65fe2e3.pdf", "updated_at": "2026-10-05T20:29:57.285Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "4-bit", "quantization", "quantizer", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QlvNqskhdY", "number": 37133, "title": "The Compression Square: Weight Quantization and KV-Cache Eviction Do Not Compose at the Operating Point", "abstract": "A deployed long-context language model is compressed twice: its weights are post-training quantized, and its KV cache is pruned by an eviction policy that ranks tokens with the model's own attention. Each axis is evaluated alone and the two losses are assumed to add, which misses that the scores deciding what to evict come from the quantized model. We measure the coupling with the compression square, a factorial over model precision and selection source whose counterfactual transplant cell splits the interaction $\\Delta_{\\mathrm{int}}$ exactly into an oracle term (the quantized model kept the wrong tokens for itself) and a fragility term (it is more hurt by the same eviction). Across 499k released per-item records on Qwen2.5-7B, Llama-3.1-8B and Qwen2.5-3B at 4k–32k, the deployed cell is significantly worse than additive in 71 of 143 identifiable SnapKV cells (63 after Benjamini–Hochberg) and sub-additive in 3; on the needle tasks the effect is a tail: most items are unaffected and a few are badly hurt. Its size is set by the quantizer (two noisy RTN builds exceed the three shipped W4-g128 recipes in 32 of 33 matched conditions), but the shipped recipes are not exempt on real documents: on Llama-3.1-8B LongBench, GPTQ W4-g128 is super-additive in 9 of 9 cells over three eviction policies at 288 items (5 at 120), and on Qwen2.5-7B LongBench, where eviction alone costs 0.7% of the baseline loss and RTN W4-g128 +0.07 nats, the two together cost +0.49 [+0.14, +0.92] more than their sum. An intervention supports the channel: i.i.d. noise injected into the ranked scores alone, with the model untouched, opens it at a level within a factor of two of a quantizer's own score error. At these budgets, single-axis evaluation does not predict the deployed system. All code and the 499,289 per-item records will be made public upon acceptance at https://github.com/[anonymized]/[anonymized].", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "weight quantization", "kv-cache eviction", "model compression", "interaction effects", "factorial design", "llm inference", "compression composition" ], "forum_url": "https://openreview.net/forum?id=QlvNqskhdY", "pdf_url": "https://openreview.net/pdf/9d5b38539ad4665d05baf64816a160f855ee1268.pdf", "updated_at": "2026-10-05T20:30:01.384Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nGjdwrBva0", "number": 37167, "title": "Demystifying Maximum Attention-Logit Dynamics in Muon Training", "abstract": "Muon has been widely adopted in language-model training due to its superior efficiency, yet the origin of its excess late-stage attention-logit growth remains unclear. Our work systematically investigates the maximum attention-logit dynamics in Muon training. First, by comparing different attention implementations and numerical precisions, we show that excess late-stage growth appears only with FlashAttention (FA) in Brain Floating Point 16 (BF16), indicating that it is not intrinsic to the optimizer design but originates from numerical errors. We further find that, after an early transient, the resulting gradient errors become persistently rank-one, with stable directions and consistent signs, causing them to accumulate coherently rather than cancel and producing structured high-logit heads. Second, motivated by this structure, we analyze a representative retrieval task with additive rank-one gradient errors and prove that the precision-induced logit gap is negligible in the early stage but becomes dominant later, directly linking the observed errors to the two-stage logit growth. Finally, we identify the FA intermediate quantities most responsible for this growth and mitigate it by selectively increasing their precision. Experiments show that these interventions provide a new trade-off between training performance and logit magnitude relative to existing methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "FlashAttention", "Muon", "Low-Precision Training" ], "forum_url": "https://openreview.net/forum?id=nGjdwrBva0", "pdf_url": "https://openreview.net/pdf/65f4f6fe876fab5dea26dd9cfdd7476012578586.pdf", "updated_at": "2026-10-05T20:30:07.189Z", "categories": [ "training" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "XuIzVwC9gu", "number": 37305, "title": "MixQuant: Adaptive Mixed-Precision Quantization for Large Language Models", "abstract": "Mixed-precision quantization improves post-training quantization by assigning higher precision to sensitive layers, but existing methods typically optimize bit allocation for a fixed memory budget.\n On edge devices, the memory available to an LLM can vary at runtime and is unknown during offline calibration.\n Adaptive quantization addresses this setting by calibrating once and deriving an allocation for the available budget at deployment.\n However, accurately calibrating before the deployment-time allocation is known is challenging because a layer's distortion depends on the bitwidths of its upstream layers.\n We show that this dependence substantially changes layer scores and the resulting allocations.\n We propose MixQuant, a quantizer-agnostic adaptive quantization framework that marginalizes over information unknown during offline calibration while exploiting the fixed quantizer and allocator.\n MixQuant estimates each layer's expected distortion over quantized upstream configurations, regularizes low-bit assignments that introduce disproportionate distortion, and calibrates quantizer parameters on plans produced by the deployment-time allocator.\n At deployment, a greedy solver produces an allocation for any available budget in $O(L|\\mathcal{B}|\\log L)$ time.\n Across three LLMs, two base quantizers, multiple memory budgets, and diverse downstream tasks, MixQuant consistently improves perplexity and average task accuracy.\n On Llama-3.2-3B at $1.25\\,\\mathrm{GB}$, it improves average accuracy over the strongest baseline by $7.7$ and $8.4$ percentage points with AWQ and GPTQ, respectively.\n Under tight memory constraints, competing methods incur $2.5\\times$--$8.75\\times$ the degradation from FP16 of MixQuant across the evaluated task categories.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixed-precision quantization", "post-training quantization", "adaptive quantization", "large language models", "LLM compression", "AWQ", "GPTQ", "efficient deep learning", "neural network compression" ], "forum_url": "https://openreview.net/forum?id=XuIzVwC9gu", "pdf_url": "https://openreview.net/pdf/1878e35b44fc15790eded5f9fbb736f4668a302c.pdf", "updated_at": "2026-10-05T20:30:29.714Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "mixed-precision", "quantization", "quantized", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cQOT7zd2tr", "number": 37364, "title": "Foundation Neural-Network Quantum States for Molecular Potential Energy Surfaces in Second Quantization", "abstract": "Second-quantized neural-network quantum states have achieved accurate molecular energies, but extending them across molecular geometries requires a shared representation of the geometry-dependent wavefunction coefficients. We introduce geometry-conditioned foundation neural-network quantum states for molecular electronic structure in second quantization. A single autoregressive model learns a family of ground states from sparse anchor geometries and provides wavefunctions at untrained geometries without further optimization. Orbital alignment matches orbital identities and transports their phases, establishing an aligned orbital basis across geometries. Frozen energies reach chemical accuracy at every untrained query geometry for N$_2$, CO, and H$_4$. On additional molecular paths, the energy-trained wavefunctions yield dipoles, quadrupoles, and natural occupations without property labels. Across three paired N$_2$ training seeds, orbital alignment lowers the mean absolute energy error over all untrained query geometries from 34-37 mHa to 0.049-0.085 mHa. At approximately 1 mHa mean absolute error, frozen evaluation reduces the per-geometry cost by $986\\times$ relative to independent optimization, yielding an estimated $25.8\\times$ end-to-end GPU-cost reduction on a 161-point N$_2$ grid.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "neural-network quantum states", "potential energy surfaces", "second quantization", "variational Monte Carlo" ], "forum_url": "https://openreview.net/forum?id=cQOT7zd2tr", "pdf_url": "https://openreview.net/pdf/5c0378e724731d32339ee31e8ea90ca171c4410c.pdf", "updated_at": "2026-10-05T20:30:35.865Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ua10RlW7vr", "number": 37448, "title": "GrowBrep: Growing Closed B-Rep Shells One Face at a Time under Grammar Constraints", "abstract": "Boundary representation (B-Rep) is the native format of computer-aided design (CAD),\nand generating it means producing continuous geometry and discrete topology at the same\ntime, with topology as a hard constraint. A B-rep is usable downstream only if it\ncloses into a solid, yet existing generators can only check closure after generation.\nWe present GrowBrep, an autoregressive B-rep generator that makes closure a property of\nits sequence grammar. GrowBrep serializes a solid as a shell that grows one face at a\ntime along a frontier of open edges. Each edge is introduced exactly once and closed\nexactly once by a second half-edge, and the sequence cannot end while any edge remains\nopen, so every sequence the model is allowed to finish describes a closed, consistently\noriented shell. The grammar does not constrain geometry, which the model still has\nto learn. Geometry enters the sequence through per-primitive autoencoders with a finite\nscalar quantization bottleneck. A decoder-only transformer generates the sequence,\nconditioned on a complexity token and optionally on an input point cloud, and illegal\ntokens are masked out at every decoding step. On ABC-1M, 81.7\\% of GrowBrep's generation\nattempts yield valid watertight solids, versus 49.0\\% for the strongest baseline, under\none strict validity checker applied to every attempt of every method without rejection\nsampling. On the distribution-level metrics (COV, MMD, and JSD) it is comparable to the\nstrongest baselines. Conditioned on point clouds, it produces a valid solid for 80.0\\%\nof the inputs, compared with 51.6\\% for the closest prior method on its own published\ntest set, at $65\\times$ lower cost per cloud. Drawing five independent samples per\ncloud and keeping the first valid solid raises validity to 95.8\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "CAD Generation; Boundary Representation (B-Rep); Autoregressive Models; Grammar-Constrained Decoding; Topological Constraints" ], "forum_url": "https://openreview.net/forum?id=ua10RlW7vr", "pdf_url": "https://openreview.net/pdf/51e119267a3a0018b9f8517ffeba2ac9f9f089f7.pdf", "updated_at": "2026-10-05T20:30:48.688Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "70RbLmZVvW", "number": 37652, "title": "SafeQ: Safety Alignment Preserving Post-Training Quantization for Language Models", "abstract": "With the growing adoption of Language Models (LMs), Post-Training Quantization\n(PTQ) has become a popular standard for efficient LM deployment, especially\nto resource-constrained environments like personal and edge devices. Most PTQ\nmethods only consider to preserve natural language understanding (understanding)\ncapabilities, without taking safety capabilities into consideration, that enables\nLMs to refuse harmful, unethical, privacy-breaching, or illegal user queries. This\ninadvertently causes complete or partial safety alignment degradation in quantized\nLMs. To address this, we propose SafeQ, a learnable PTQ method that preserves\nsafety alignment by taking both understanding and safety capabilities of LMs into\nconsideration. Precisely, SafeQ learns quantization parameters by minimizing\na block-wise weighted-sum loss over understanding and safety capabilities for\nboth weight-only quantization and weight-activation quantization. We performed\nextensive evaluations for assessing understanding capabilities, and safety and\ngeneral utility capabilities using different pre-trained and instruction-tuned LMs\nrespectively. Our evaluations demonstrate best or near-best performance by SafeQ\nin pre-trained LMs for W4A16 (i.e., 4-bit weight and 16-bit activation), and\nW4A4 quantization modes, and in instruction-tuned LMs for W4A16 quantization\nmode. In few instances, SafeQ improved safety capabilities of quantized LMs\nslightly beyond full-precision LMs. Our results evidently show that SafeQ resolved\ntrade-off between safety and general utility capabilities to a great extent in\ninstruction-tuned LMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Language Models", "Quantization", "Safety Alignment" ], "forum_url": "https://openreview.net/forum?id=70RbLmZVvW", "pdf_url": "https://openreview.net/pdf/0a08005d106843d5c4c675c054ca0879b5913806.pdf", "updated_at": "2026-10-05T20:31:12.066Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "ptq", "quantization", "quantized", "w4a16", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Y8x0hfaJmc", "number": 37695, "title": "Frozen Input-Harm Probes under Compression: Retained AUROC Does Not Mean Retained Refusal", "abstract": "Linear probes trained on a model's BF16 activations are candidate safety monitors, but models are often compressed after a probe is fitted. We ask whether a frozen input-harm probe that still separates harmful from harmless prompts implies that the compressed model still refuses. Across four families and forty compressed configurations (post-training quantisation, QLoRA, distillation, and Wanda and SparseGPT pruning), we pair judged attack success on StrongREJECT with frozen-probe AUROC on XSTest wherever hidden dimensions match. With greedy 48-token generation, thirteen configurations have attack-success intervals above zero and seven survive Holm correction; none is post-training quantised or Gemma-2-9B. In OLMo-2-7B, SparseGPT 2:4 and Wanda 2:4 raise attack success by $+0.0590$ $[+0.0377,+0.0913]$ and $+0.0654$ $[+0.0427,+0.0988]$, while seed-mean probe AUROC falls detectably but by less than our planned 0.10 big-drop threshold at all eight probed layers (a one-sided test confirms this at every layer only for SparseGPT 2:4). The same methods give Llama and Qwen big drops at two or three layers shallower than the planned one. On our primary pooled AUROC basis this is therefore an existence result in one family (per fold, Llama SparseGPT 2:4 also qualifies), on disjoint probe and behavioural prompts. The OLMo increases lose Holm significance if three or four of six unjudged baseline responses (SparseGPT and Wanda, respectively) were successes; twelve of the thirteen increases persist at a 256-token cap and in every sampled seed. Retained AUROC does not certify retained refusal: reusing a probe after compression still requires behavioural evaluation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "quantization", "pruning", "knowledge distillation", "linear probes", "interpretability", "safety monitoring", "refusal", "harmful compliance", "sparse autoencoders", "safety evaluation", "LLM safety" ], "forum_url": "https://openreview.net/forum?id=Y8x0hfaJmc", "pdf_url": "https://openreview.net/pdf/8d409960ef1cb3348eaae61eb1d0a96796e9ecbc.pdf", "updated_at": "2026-10-05T20:31:20.738Z", "categories": [ "ptq" ], "matched_terms": [ "quantisation", "quantised", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AYgk1E3JbF", "number": 37706, "title": "PRQuant: Permutation Residual Quantization for Low-Overhead Inference", "abstract": "Low-bit quantization of linear layers is often dominated by a small number of outlier channels. Existing smoothing, rotation, and residual-based methods can mitigate this issue, but may shift the quantization bottleneck to weights or introduce costly online operations. To address these limitations, we propose PRQuant (Permutation Residual Quantization), a training-free framework that combines channel permutation with offline weight residual compensation. PRQuant identifies the scaled-weight columns with the largest quantization errors and permutes them into contiguous tail blocks. This structure allows the corresponding weight residuals to be precomputed entirely offline, while replacing scattered activation gathering with simple contiguous access during inference, yielding a single regular MXFP4 GEMM for compensated computation. Experiments show that PRQuant substantially reduces down-projection reconstruction error, with scaling and residual compensation providing the main numerical gains while permutation enables a hardware-friendly contiguous layout.\nComprehensive experiment results on Qwen3-4B-Instruct-2507 and Qwen3-30B-A3B-Instruct-2507 illustrate that PRQuant achieves up to averagelly 2.6× and 1.8× operator speedup over BF16 respectively, while preserving near plain MXFP4 end-to-end decoding efficiency. Across five downstream benchmarks, PRQuant achieves the best average accuracy among the quantized methods, improving accuracy over MXFP4 by 1.24 and 0.55, respectively.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Post-Training Quantization", "Low-Bit Quantization", "MXFP4", "Microscaling Formats", "Hardware-Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=AYgk1E3JbF", "pdf_url": "https://openreview.net/pdf/568fcb4ba64dedd3650d6959a1beb4612342b8d1.pdf", "updated_at": "2026-10-05T20:31:22.955Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "mxfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4rQFijPAgb", "number": 37777, "title": "ZORA: Zeroth-Order Reliability-Calibrated Adaptation with Shift-Sensitive Updates for Multimodal Test-Time Shifts", "abstract": "Audio-visual continual test-time adaptation (AV-CTTA) enables pretrained multimodal models to adapt online to non-stationary target streams. However, most existing AV-CTTA methods rely on backpropagation-based updates, limiting their deployment on edge platforms with quantized backbones, inference-only engines, or strict memory and latency constraints. Zeroth-order optimization (ZOO) provides a forward-only alternative, but its direct application to AV-CTTA faces two challenges: zeroth-order queries must explore a high-dimensional multimodal parameter space without shift awareness, and entropy minimization may reinforce incorrect predictions that exhibit high confidence and low entropy under online multimodal shifts. To address these limitations, we introduce Zeroth-Order Reliability-Calibrated Adaptation (ZORA), a forward-only AV-CTTA framework that determines both where ZOO should update and which learning signals it should trust. ZORA incorporates Shift-Sensitive Layer Selection (SSLS) to restrict zeroth-order updates to shift-sensitive normalization layers in the audio and visual branches. It further employs Bias-Aware Reliability Calibration (BARC) to assign signed entropy weights based on prediction bias and confidence, while aligning target-window feature statistics with cached source-reference statistics. Experiments on Kinetics50-C and VGGSound-C demonstrate that ZORA consistently outperforms representative CTTA, multimodal adaptation, and forward-only baselines across multiple continual audio-visual shift protocols. Experiments with quantized CAV-MAE models and efficiency analyses further support resource-constrained forward-only deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Audio-Visual Continual Test-Time Adaptation; Zeroth-Order Optimization; Forward-Only Adaptation; Multimodal Learning; Reliability Calibration; Distribution Shift; Edge Deployment." ], "forum_url": "https://openreview.net/forum?id=4rQFijPAgb", "pdf_url": "https://openreview.net/pdf/b0be7b85f12c747954e95f46d90c4ef40b8bfc98.pdf", "updated_at": "2026-10-05T20:31:36.610Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9fAgUx2f15", "number": 37862, "title": "Quantizing Recursive Reasoning Models", "abstract": "Recursive reasoning models solve hard puzzles by applying compact, weight-tied blocks over many refinement steps. Because these blocks are reused many times, quantizing them creates a severe error accumulation challenge: the quantization error is incurred at every step. While 8-bit quantization (integer or float) preserves accuracy, moving to a per-tensor 4-bit format causes a systematic bias to accumulate. The ensuing drift catastrophically degrades exact-solution accuracy on Sudoku from 84.1% to 0.0% (only ~25% of cells correct). In this work, we show that this collapse is caused by activation-scaling granularity rather than bit-width or number format. Crucially, moving to per-block scaling restores stable contraction and most of the lost accuracy. To implement this, we apply MXINT4, a blockwise integer format for weights and activations, to recursive reasoning models. It is competitive with blockwise float formats on our tasks, while keeping integer elements and power-of-two block scales. Finally, recursion depth and reuse modulate quantization sensitivity, with the deepest trained model we test (EqR) the most sensitive. Yet blockwise scaling overcomes this vulnerability, staying robust across the evaluated recursive models and transferring to the open-ended ARC-AGI benchmark.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "recursive reasoning models", "weight-tied recursion", "post-training quantization", "microscaling formats", "MXINT4", "low-bit inference", "tiny recursive models", "activation quantization", "error accumulation" ], "forum_url": "https://openreview.net/forum?id=9fAgUx2f15", "pdf_url": "https://openreview.net/pdf/08bd141ddd66d27a5c938bb0de37c0357e3a63d4.pdf", "updated_at": "2026-10-05T20:31:46.056Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "4-bit", "8-bit", "bit-width", "low-bit", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9Lrcxvs7GS", "number": 37880, "title": "Beyond Alignment: Enhancing LLM Safety Robustness via Zeroth-Order Optimization", "abstract": "Safety alignment for large language models (LLMs) aims to reduce harmful or unsafe behavior while preserving general utility. However, recent findings reveal that alignment effects can be fragile: lightweight post-alignment manipulations, such as parameter noise, activation noise, or quantization, can easily weaken the intended safety behavior. Prior efforts to improve robustness have primarily focused on data curation, modified alignment objectives, and safety-critical parameter identification, leaving the role of the optimizer itself largely unexplored.\n\nIn this paper, we are the first to study the robustness of safety alignment from the perspective of the base optimizer. This optimizer-centric view naturally points to zeroth-order optimization, which provides a robustness-oriented signal by evaluating safety alignment under perturbations. Based on this insight, we propose a hybrid framework that first performs standard first-order safety alignment and then applies zeroth-order refinement to improve robustness. Both theoretically and empirically, we show that only a few zeroth-order refinement steps can enhance robustness while preserving safety alignment. We further improve the efficiency of zeroth-order refinement by exploiting its inherent perturbation-based evaluations to estimate layer-wise robustness sensitivity, enabling the refinement process to concentrate updates on robustness-critical layers with modest training overhead.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Safety Alignment", "Robustness of Safety Alignment", "Zeroth-order Optimization" ], "forum_url": "https://openreview.net/forum?id=9Lrcxvs7GS", "pdf_url": "https://openreview.net/pdf/5afef5cd32f60758c2305995a1beb37feddbe68a.pdf", "updated_at": "2026-10-05T20:31:49.772Z", "categories": [ "training", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WNTHArc3s0", "number": 37899, "title": "EntroPack: Fast and Accurate Entropy-Coded Weight Compression at Arbitrary Bitrates", "abstract": "Weight compression helps large generative models fit deployment memory budgets, but common fixed-width formats offer only coarse storage choices. Entropy coding supports finer rates, yet the achieved size depends on the quantized weight distribution and coding overhead. Exploiting this flexibility requires accurate rate selection and efficient weight reconstruction for inference. We present EntroPack, an entropy-coded weight compressor that supports arbitrary target bitrates without activation calibration or fine-tuning. It combines row-normalized $E_8$ lattice quantization with a conditional probability model of lattice coordinates. Sampled storage estimates select the quantization resolution without repeated full-stream encoding. The final coordinates are entropy-coded in independently decodable tiles, enabling fast, fused symbol decoding and numerical weight reconstruction on the GPU. EntroPack supports floating-point and integer weight containers, including BF16, FP16, FP8, and INT8, with storage bitrate controlled independently of numerical precision. We evaluate EntroPack's compression quality and encoding speed on popular diffusion and language models. When compressing the linear-layer weights of the image generator Z-Image-Turbo, EntroPack achieves substantially lower weight and denoiser output errors than fixed-width formats at comparable storage rates, with modest denoising-step overhead. Targeting 4 bits per parameter, it achieves lower weight and denoiser output errors than NF4, including about 24\\% lower relative $L_2$ weight error, with less storage.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Weight compression", "Entropy coding", "Lattice quantization" ], "forum_url": "https://openreview.net/forum?id=WNTHArc3s0", "pdf_url": "https://openreview.net/pdf/b0364cee5181395db025d494cbc7d2caf4264411.pdf", "updated_at": "2026-10-05T20:31:52.330Z", "categories": [ "ptq", "training", "hardware", "diffusion" ], "matched_terms": [ "fp8", "int8", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W7Pi7omZu9", "number": 37921, "title": "ARCH-CPTQ: Replay-Free Continual Post-Training Quantization with Curvature History and Anchoring", "abstract": "Calibration-based post-training quantization (PTQ) can be sensitive to differences between calibration and deployment data, especially at very low precision. Adaptive quantization addresses distribution shifts, but updates based only on current reconstruction objectives may degrade accuracy on earlier domains. We therefore formalize replay-free continual PTQ (CPTQ), which retains reconstruction history while updating a shared low-bit checkpoint using unlabeled calibration data and a fixed high-precision reference. Our derivation shows that weighted quadratic reconstruction losses on fixed inputs can be represented exactly by one fixed-dimensional curvature matrix per layer, but this representation alone does not determine how to guide subsequent updates. We propose anchored reconstruction with curvature history (ARCH)-CPTQ, which emphasizes poorly reconstructed current units, balances current and historical curvature using a Riemannian discrepancy, and penalizes changes from the previous checkpoint along historically sensitive directions. Across six large language models and five precision settings, ARCH-CPTQ improves model-averaged sequence-wide average performance (AP) and retention over equal-domain curvature averaging (CPTQ-AVG) under matched solver and representation settings, with a 2.16-percentage-point AP gain in our main W2A16 (2-bit weights, 16-bit activations) setting. In this setting, all six models achieve higher final accuracy on earlier domains, on average, than their own scores immediately after each domain was first calibrated, without replay.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "large language models", "continual learning", "low-bit model compression", "replay-free adaptation", "curvature-aware optimization" ], "forum_url": "https://openreview.net/forum?id=W7Pi7omZu9", "pdf_url": "https://openreview.net/pdf/7dc2e617a4d1216c0b7b486ae53fa23cef2e615f.pdf", "updated_at": "2026-10-05T20:31:55.712Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low precision", "low-bit", "ptq", "quantization", "w2a16" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "n4iNw5dLd7", "number": 37929, "title": "Restoring Minority Generation in Low-Bit Diffusion Models with a Plug-in Repair", "abstract": "Efficient deployment of diffusion models has motivated extensive research on low-bit quantization, which can substantially reduce memory and computation while maintaining generation quality under standard sampling. However, we argue that this preservation is not guaranteed to extend to the task of \\emph{minority generation}, which specifically targets generation from low-density but valid regions of the data distribution. We show that quantization error becomes significant in clean estimates at high noise levels, on which existing minority samplers heavily rely yet standard sampling places only a small weight. Based on this analysis, we propose a lightweight plug-in repair that aligns the high-noise clean estimates of quantized models with those of their full-precision counterparts. Our repair combines selective bit allocation with a quantization-consistent low-rank adapter that remains on the deployed quantization grid, and is activated only when computing these estimates, leaving the ordinary sampling path unchanged. Experiments across diverse diffusion architectures, bit widths, and minority samplers show that our repair substantially improves minority generation with negligible overhead while retaining the efficiency of low-bit inference. Code is available here.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "Diffusion Model Quantization", "Low-Bit Inference", "Minority Generation", "Minority Sampling" ], "forum_url": "https://openreview.net/forum?id=n4iNw5dLd7", "pdf_url": "https://openreview.net/pdf/b193d06be53f096e5d42e0cbb298e09c1de80127.pdf", "updated_at": "2026-10-05T20:32:00.759Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "N7pHNz2agR", "number": 38036, "title": "SSN++: Layerwise Subspace Networks for Low-Bandwidth Pipeline Training", "abstract": "Subspace networks (SSNs) make pipeline-parallel training communication-efficient by constraining every layer to write into one shared low-dimensional subspace, so activations and gradients cross each stage boundary losslessly. This shared subspace, however, limits what the model can express, and with it the loss an SSN can reach relative to an uncompressed transformer. We propose SSN++, which removes this constraint without additional communication. We introduce a norm-preserving orthogonal transform that losslessly transports each compressed activation from one layer's subspace into the next, so every layer can write into its own subspace. We extend this separation within each layer, giving attention and the MLP separate subspaces and dividing each into groups with their own subspaces, which together exceed the rank of one subspace at equal parameters, and we add learned gates that choose what each layer passes on. Because the compressed activations of preceding tokens have already crossed the boundary, the receiving layer also decodes its input from them with a learned causal filter, raising its rank beyond one subspace. Random subspaces spread activations like rotation-based quantisers, which makes the communication amenable to quantisation, and we spend the saved bits on a higher rank. Approximation theory and measurements during training support each choice. From 229M to 2B parameters, SSN++ lowers loss and raises eval accuracy over a parameter-matched SSN at equal communication, and with quantised communication it closes 63 to 78\\% of the matched SSN's loss gap to the uncompressed transformer at 229M and 1B and 58\\% at 2B, while training 12 to 26 times faster per update than it over 200\\,Mb/s links. The loss gap of SSNs thus stems largely from how capacity is organised around the compressed activations, and closing it without extra communication makes pipeline-parallel training over internet-grade links far more efficient.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "communication-efficient training", "model parallelism", "low-rank subspace", "activation compression", "quantisation", "neural networks", "decentralised training" ], "forum_url": "https://openreview.net/forum?id=N7pHNz2agR", "pdf_url": "https://openreview.net/pdf/355b892691bf6297c6770ed99be1a0b953029257.pdf", "updated_at": "2026-10-05T20:32:11.960Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantisation", "quantised", "quantisers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "69ukjZCQkV", "number": 38085, "title": "GR4DGS: Sparse-View 4D Gaussian Compression via Temporal Grid Prediction", "abstract": "4D Gaussian representations enable free-viewpoint video, but their frame-wise scene updates incur substantial storage and transmission costs. We investigate this compression challenge under sparse-view capture. In our three-view evaluation of a grid-based streaming codec, motion grids account for about 99% of P-frame payloads, although consecutive quantized grids remain highly correlated. Independent per-frame grid coding therefore repeatedly transmits information that is predictable from the decoder's preceding state. We propose GR4DGS to exploit this redundancy through temporal motion-grid prediction. GR4DGS transmits integer differences from the previously decoded grid, recovering the current quantized grid exactly while retaining the other rendering inputs. To further reduce payloads, GR4DGS-E2E combines residual-domain rate optimization with occupancy-aware initialization: unqueried cells inherit the preceding decoded values, while queried cells restart from a shared warm-up. Across six Neural 3D Video scenes of 300 frames each under three-view training, lossless prediction reduces P-frame payloads by 65–72% relative to the default independently coded 4DGC baseline. The separately trained GR4DGS-E2E variant reduces them by 88–93%, with per-scene mean PSNR differences within 0.03 dB. Including the shared uncompressed keyframe once, E2E reduces sequence payloads by 81–86%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "4D Gaussian compression", "temporal prediction", "sparse-view reconstruction", "entropy coding", "rate-distortion optimization", "free-viewpoint video" ], "forum_url": "https://openreview.net/forum?id=69ukjZCQkV", "pdf_url": "https://openreview.net/pdf/6729ae8bcaf05e0fef3371f999fac571f8132416.pdf", "updated_at": "2026-10-05T20:32:21.730Z", "categories": [ "ptq" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "w2w4TwPsCl", "number": 38102, "title": "Subspace-Coded Parallel Exploration for Forward-Only Test-Time Adaptation", "abstract": "Zeroth-order (ZO) optimization enables forward-only test-time adaptation (TTA), avoiding the computation and memory overhead of backpropagation. However, existing direction-centric approaches face an exploration–exploitation dilemma in capturing gradient energy: exploiting promising directions may overlook complementary gradient components, whereas broader exploration typically requires additional forward evaluations. To address this limitation, we propose **S**ubspace-**C**oded **O**ptimization via **P**arallel **E**xploration (**SCOPE**), a novel TTA framework that shifts from individual direction selection to higher-dimensional subspace exploration. Specifically, we introduce a perturbation encoding scheme that assigns different samples within a batch complementary combinations of shared basis directions, rather than applying the same perturbation to all samples. A lightweight least-squares estimator jointly decodes their loss differences to estimate batch-average directional responses for a shared prompt update. To support batch-coupled objectives such as feature-statistic alignment, we further decompose their batch-level loss differences into exact sample-wise contributions, preserving the original adaptation objective while making its feedback compatible with the perturbation codes. Together, these designs embed multi-directional measurements into existing batch computation, broadening gradient-energy coverage with only two forward passes. Extensive experiments across mild, continual, and wild TTA settings, together with natural-shift and quantized models, demonstrate the effectiveness and efficiency of SCOPE. Code will be publicly available.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Test-time adaptation", "Zeroth-order optimization", "Forward-only adaptation", "Gradient estimation", "Subspace exploration" ], "forum_url": "https://openreview.net/forum?id=w2w4TwPsCl", "pdf_url": "https://openreview.net/pdf/26d2369eefbefe87fde4a3ec43b4a9596c7d81ba.pdf", "updated_at": "2026-10-05T20:32:23.469Z", "categories": [ "training" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "uHw27dDfbO", "number": 38107, "title": "Provable LLM Tampering Detection: Catching Model Provider Tampering of Open-Weights LLMs", "abstract": "Prior theory-driven methods for detecting black-box LLM configuration tampering do not preserve their guarantees of a low false-positive rate (FPR) when the values of multiple common sampling parameters—such as temperature, token logit bias, and top-p—are hidden. Such sampling parameters can unpredictably and nonlinearly distort LLM output distributions without changing the underlying model configuration, making benign variation in sampling parameters difficult to statistically distinguish from genuine model configuration mismatches such as those arising from model substitution, secret fine-tuning, prompt injection, and weight quantization. To our knowledge, our paper establishes the first tamper-detection method for black-box LLM configurations that formally guarantees a low long-run FPR in the presence of a broad set of hidden standard sampling parameters. To this end, we develop a general theoretical framework for tamper-detection in this setting, and then construct two concrete classes of tests—the Temperature Test and the Logit Bias Delta Test—each provably achieving a long-run FPR of 0%. In our follow-up demo spanning four major tampering classes—model substitution, secret fine-tuning, system prompt injection, and weight quantization—the Temperature Test achieves 100% true-positive rate (TPR) and 0% FPR, while the Logit Bias Delta Test achieves 92.9% TPR and 0% FPR. Estimating the cost of each test in the demo at the prevailing rates of currently popular open-weight model APIs places the upper bound between \\\\$0.1 and \\\\$2.6, with our adaptive early-stopping technique further reducing the average estimated cost of a positive test to \\\\$0.02–\\\\$0.6.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM auditing", "black-box model verification", "model tampering detection", "model fingerprinting", "sampling parameters", "theoretical guarantees", "formal guarantees", "statistical identifiability" ], "forum_url": "https://openreview.net/forum?id=uHw27dDfbO", "pdf_url": "https://openreview.net/pdf/b96c0396bb4e2da5cc352cca68d44cb9f92e2822.pdf", "updated_at": "2026-10-05T20:32:24.681Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bueqgdoyyD", "number": 38114, "title": "Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution", "abstract": "Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite quantization being a common deployment transformation for language models. Prior works have shown that 4-bit post-training quantization can reverse machine unlearning. We show this is a systematic dual failure. Gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures have the same root cause. Across all baselines, per-parameter RMS updates are 47 to 828× below the canonical NF4-calibrated floor used in our implementation, so updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a **sparsity-permanence tradeoff**. We present **MANSU** (**M**echanistic-**A**ligned **N**ull-**S**pace **U**nlearning), which addresses this tradeoff by treating it as an update-budget allocation problem. Causal circuit attribution isolates a compact, causally implicated forget circuit, a diagonal-Fisher mask restricts updates to retain-safe coordinates within that circuit, and a quantizer-calibrated magnitude floor, applied post-hoc, makes the resulting update deployment-visible, which we verify directly on the trained checkpoints. Matched controls (the same floor applied to a random circuit, an inverse circuit, a globally projected update, and a Fisher-ratio selective baseline) show that neither circuit concentration nor magnitude flooring alone reproduces this behavior in the reported ablations. We additionally introduce **Circuit Attribution Divergence (CAD)**, a mechanistic verification metric providing evidence of circuit-level structural change beyond behavioral suppression, corroborated by independent linear-probe and activation-patching analyses. Across multiple model families and WMDP hazard domains, MANSU consistently yields a non-positive NF4 PTQ gap while retaining substantially higher general capability than aggressively optimized baselines, whereas gradient ascent, for instance, partially reverses under compression (BF16 forget 0.260 to NF4 0.310, PTQ gap +0.050). We scope our boundary-crossing analysis to NF4 and report GPTQ and INT8 results separately as empirical transfer checks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Machine Unlearning", "Large Language Models", "Quantization", "NF4", "Mechanistic Interpretability", "Circuit Attribution", "Quantization-Robust Unlearning", "Continual Learning" ], "forum_url": "https://openreview.net/forum?id=bueqgdoyyD", "pdf_url": "https://openreview.net/pdf/bcde45f93f8f4a2d5ae13e0de9810a3be8dc1cf6.pdf", "updated_at": "2026-10-05T20:32:25.191Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "4-bit", "int8", "ptq", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cpY9ygBWEB", "number": 38120, "title": "Best-Arm Identification in Factored Multi-Agent Bandits via Junction Trees", "abstract": "Factored Multi-Agent Multi-Armed Bandits (MAMABs) model cooperative decision problems with exponentially large joint-action spaces and sparse local interactions. In fixed-confidence Best-Arm Identification (BAI), computing the information-optimal sampling allocation remains combinatorial. We derive a Junction-Tree (JT) representation of the characteristic-time program: locally consistent clique marginals characterize feasible sampling allocations, while a gap-augmented dynamic program handles competing joint actions without enumerating the combinatorial action space. The resulting complexity depends on the induced width of the graphical decomposition and a gap-state width measuring the number of cumulative reward gaps retained during elimination. For lattice-valued means, the resulting dynamic program is exact with pseudo-polynomial complexity; for arbitrary real-valued rewards, quantizing this gap dimension yields a certified approximation with a controllable accuracy-complexity trade-off. We further introduce JT Track-and-Stop, (JT-TaS) a $\\delta$-correct algorithm that asymptotically attains the instance-dependent lower bound. Experiments recover the exact optimum on enumerable instances and scale to instances with large joint action spaces.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multi-Armed Bandits", "Multi-Agent Bandits", "Factored Bandits", "Best-Arm Identification", "Pure Exploration" ], "forum_url": "https://openreview.net/forum?id=cpY9ygBWEB", "pdf_url": "https://openreview.net/pdf/e208fd1e85544705aeee38b1e9a12c462bb29004.pdf", "updated_at": "2026-10-05T20:32:25.506Z", "categories": [], "matched_terms": [ "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qSAFcuWbeQ", "number": 38176, "title": "TraceKV: Signed Error Feedback for KV-Cache Quantization", "abstract": "Low-bit key-value (KV) cache quantization reduces the storage requirements of large language models, but quantization errors can alter their output distributions and generation trajectories. Because errors across tokens can cancel or reinforce one another, reducing individual reconstruction errors alone does not directly control their combined effect on generation. We propose TraceKV, a method that uses self-generated traces to guide cache reconstruction through signed error feedback while preserving the base quantizer's encoding format, bit widths, and metadata footprint. TraceKV generates a short continuation with the bfloat16 (BF16) model, then replays it to compute gradients of multi-step Kullback-Leibler (KL) divergence with respect to cached values. Within each layer and attention head, it accumulates signed projections of quantization errors onto these gradients and selects between independent quantization and residual-feedback reconstruction for each token, encouraging positive and negative contributions to offset. Trace replay then selects among the completed candidate caches. Across four quantization representations, TraceKV reduces cumulative trajectory distortion. On PatternKV and OSCAR, gains persist for 64 tokens beyond the H64 selection window; paired ablations support retaining error signs. In two evaluations with frozen data and protocols, generation restarts from the compressed prompt. With KVarN as the base quantizer, TraceKV increases the mean length of the generated prefix that exactly matches the BF16 continuation by 14.4 and 22.8 tokens over independent KVarN, respectively, with prefix length capped at 1024 tokens.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache quantization", "large language models", "error feedback", "generation fidelity" ], "forum_url": "https://openreview.net/forum?id=qSAFcuWbeQ", "pdf_url": "https://openreview.net/pdf/3dee7d9b58034d11af766416e0677b4a92368b20.pdf", "updated_at": "2026-10-05T20:32:35.892Z", "categories": [ "ptq", "training", "kv-cache" ], "matched_terms": [ "low-bit", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1aQ0OuhtdV", "number": 38358, "title": "Bicameral: Memory-Efficient Transformer via Latent Tensor Rematerialization", "abstract": "Memory capacity limits the size of language models that can be trained and served on given hardware.\nQuantization lowers the precision of each weight, but leaves the number of trainable parameters unchanged, and with it the gradients and optimizer states that dominate training memory.\nWe introduce \\textbf{Latent Tensor Rematerialization} (LTR), which generates each weight matrix on demand from compressed latent codes with small learned generators, and in doing so converts a standard matrix into a recomputable activation. This reduces memory usage by directly storing fewer parameters and optimizer states, as an alternative to lowering precision. Moreover, LTR leverages underutilized compute to reduce the number of parameters that must be stored, and its full-precision parameters remain compatible with quantization.\nWe validate the effectiveness of LTR with \\textbf{Bicameral}, a LLaMA-style transformer architecture whose weight matrices are generated on demand by small decoders.\n%When trained to 10B tokens, a Bicameral model with half the stored parameters of a 2B-parameter transformer matches its validation perplexity and outperforms a standard transformer with the same number of stored parameters.\nWe train Bicameral models with 2B effective parameters to 10B tokens, and find that Bicameral matches the uncompressed baseline on perplexity and zero-shot benchmarks, while reducing stored parameters by 50\\% and training memory usage by up to 48\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models", "Transformer Architecture", "Tensor Rematerialization", "Memory Efficient Training", "Model Compression", "Hypernetworks" ], "forum_url": "https://openreview.net/forum?id=1aQ0OuhtdV", "pdf_url": "https://openreview.net/pdf/1f07b6ea2350947dc949251518780b9ceeab17a2.pdf", "updated_at": "2026-10-05T20:33:07.542Z", "categories": [ "training", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7hyFiIpcLn", "number": 38407, "title": "ResponseScope: Measuring Finite-Response Distortion in Compressed Models", "abstract": "Model compression is commonly evaluated using clean-task performance and static\nteacher–student agreement, but these criteria do not characterize responses to finite,\ntask-valid input changes. We study three related questions: how to measure this\nbehavioral drift, predict compression vulnerability before full recovery and final\nutility evaluation, and preserve response behavior at a fixed model footprint. Re\u0002sponseScope (RS) is a teacher-normalized finite-response distortion measure. The\nevaluation spans six datasets, nine teacher–student architecture pairs, four compres\u0002sion families, three storage budgets, 540 recipe means, and 2,700 independently\nstored seed checkpoints. Combining teacher layer fragility with operator-local\ndistortion and applicable initialization compatibility yields task-equal leave-one\u0002dataset-out R2 = .687 and MAE 1.800 percentage points for predicting utility\ndamage prospectively. After compression, adding measured RS to a scale- and\nendpoint-matched baseline raises R2\nfrom .587 to .778 and reduces selection regret\nfrom 1.260 to .688 points. At footprint ratio .25, IF-Reg reduces mean held-out RS\nfrom .440 to .257 and improves the task-specific utility composite by 3.017 points\nrelative to augmented logit distillation. A baseline matched for endpoints, teacher\nscale, and margin explains part of this gain. These results connect the measurement,\nprediction, and preservation of interventional fidelity", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Compression;Finite-Response Distortion;Behavioral Evaluation;Structured Input Transformations;Model Selection;Utility-Damage Prediction;Knowledge Distillation;Structured Pruning;Post-Training Quantization;Distribution Shift" ], "forum_url": "https://openreview.net/forum?id=7hyFiIpcLn", "pdf_url": "https://openreview.net/pdf/7e986bca32f3e0c4686c045c419ffc536fef9c94.pdf", "updated_at": "2026-10-05T20:33:16.216Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "axqI8H3ZaI", "number": 38686, "title": "Decoupled Polynomial State-Space Iterations for Matrix Polar Approximation, with an Application to Muon", "abstract": "We introduce Gram-matrix-based iterations (GMBI), a decoupled polynomial state-space parameterization that separately controls propagation of a square auxiliary state and accumulation of the rectangular output under a restricted matrix-computation budget. Building on the finite-depth polynomial framework of Polar Express, GMBI supports both compact short graphs and robust full-restart graphs. For restart graphs, we freeze a guarded two-layer prefix, materialize its outputs from FP16 calibration runs to rebuild true Gram states, and optimize the suffix by scalar rollouts from the realized restart spectra, adapting to prefix-induced distribution shifts and finite-precision effects. In controlled FP16 experiments, the decoupled GMBI parameterization outperforms a learned tied-polynomial control under both robust-only and Muon-aware fitting protocols, while the Muon-aware protocol provides an additional improvement for the decoupled variant. Tests on unseen spectra, matrix shapes, and held-out GPT-2 Medium updates show that these gains are not confined to the calibration set, while restart controls separately identify the benefit of finite-precision Gram-state repair. Kernel benchmarks show that compact GMBI graphs can reduce polar-call latency on large rectangular matrices under both compiled and specialized backends. Multi-seed GPT-2 Small training with Muon matches Direct PE in final validation loss while achieving modest reductions in mean step time, showing that the matrix-level gains can survive end-to-end training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "matrix polar approximation", "Muon optimizer", "polynomial iterations", "Gram matrices", "low-precision numerical linear algebra", "GPU acceleration", "matrix orthogonalization" ], "forum_url": "https://openreview.net/forum?id=axqI8H3ZaI", "pdf_url": "https://openreview.net/pdf/90790995ed10ac9111ef8c02ea8b129228b0ade6.pdf", "updated_at": "2026-10-05T20:34:53.894Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "c3FJN8d9Wz", "number": 38748, "title": "BMoEQ: Binary-Coded MoE Quantization with Runtime-Adaptive Mixed-Precision Support", "abstract": "Mixture-of-Experts large language models (MoE-LLMs) scale model capacity through sparse activation but retain a substantial memory footprint from expert weights. Existing expert-level mixed-precision methods alleviate this cost by fixing bit-widths according to calibration-based importance estimates, which may not generalize across inputs, domains, and tasks. We propose BMoEQ, a quantization method that combines uniform ultra-low-bit base precision across experts with input-adaptive execution precision through binary-coded quantization (BCQ). To improve quantization accuracy without static expert-level mixed-precision allocation, BMoEQ incorporates saliency-aware error reconstruction, assigning an additional bit plane to a small subset of salient weights while maintaining a common base bit-width across experts. At runtime, online importance scores guide the selective omission of bit planes for less important activated experts, reducing memory traffic and computation without removing their contributions entirely. A single stored representation thus supports multiple execution precisions without requantization or separate lower-precision copies, while preserving access to each expert’s full stored precision. We further develop hardware-efficient GPU kernels that selectively load and process the required bit planes and fuse the salient-weight correction path to minimize latency overhead. Experiments show that the saliency-aware quantization in BMoEQ achieves superior overall accuracy in both 2-bit and 3-bit settings, outperforming expert-level mixed-precision baselines at comparable effective bit-widths. With runtime precision adaptation and optimized GPU kernels, BMoEQ achieves up to 6.4$\\times$ speedup over the FP16 cuBLAS baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixture-of-experts", "quantization", "dynamic multi-precision", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=c3FJN8d9Wz", "pdf_url": "https://openreview.net/pdf/96ec69826fe40268fdc3994df217c70280d383e1.pdf", "updated_at": "2026-10-05T20:35:52.464Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "3-bit", "binary", "bit-width", "low-bit", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "W8BViO3wZp", "number": 38877, "title": "FlyMark: Training-Free Invisible Watermarking of 3D Gaussian Splatting via a Fruit Fly Connectome", "abstract": "A trained 3D Gaussian Splatting (3DGS) scene ships as a portable parameter array that can be copied, pruned, requantized, or repackaged outside its training pipeline, so ownership evidence is most useful when it lives in the released parameters and remains checkable long after the embedding tooling is gone. Existing 3DGS watermarks typically tie embedding or extraction to scene optimization, a learned decoder, or rendered views, so the evidence survives only as long as a second trained artifact does. FlyMark instead writes a keyed message into the parameters a 3DGS file already stores. Its carrier directions are derived from the photoreceptors of a published connectome, a citable versioned artifact that fixes the geometry exhaustively and leaves nothing to tune per scene. A virtual observer reads cone-wise apparent luminance along a scene-normalized orbit from stored centers, colors, and opacities; a keyed dithered quantization-index-modulation code replicates each message bit across these observations; and one sparse bounded least-squares solve realizes the targets through achromatic shifts of existing degree-zero colors under a hard per-channel linear-RGB bound. All geometry and higher-order appearance parameters are preserved bit-identically, and extraction needs only cone queries, rounding, and majority voting. Under a model-domain threat model on synthetic and real scenes, FlyMark attains high clean bit accuracy and visual fidelity while cleanly separating matched from wrong keys.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "3D Gaussian Splatting", "Digital Watermarking", "Fruit Fly Connectome" ], "forum_url": "https://openreview.net/forum?id=W8BViO3wZp", "pdf_url": "https://openreview.net/pdf/3dd13371c72cb7ef8ea201a9a938f5e4fc1e12d4.pdf", "updated_at": "2026-10-05T20:37:00.524Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7hRDluQdRc", "number": 38996, "title": "QuHIC: Quantization-Consistent Multi-Prior Unrolled Reconstruction for Hyperspectral Image Compression", "abstract": "Recent advances have demonstrated the great potential of lossy hyperspectral image (HSI) compression for airborne remote-sensing applications. However, given the limited storage capacity and transmission bandwidth of airborne sensing platforms, most existing methods insufficiently integrate compression-induced constraints with spectral–spatial priors, making it difficult to balance low bit rates and high-fidelity reconstruction. To address this challenge, we propose QuHIC, which synergizes bit-level quantized compression with multi-stage unrolled reconstruction, establishing a new paradigm for HSI compression and reconstruction. Specifically, during encoding, QuHIC employs a learnable lightweight encoder to compress an HSI into a low-dimensional quantized representation for efficient storage and transmission. During decoding, HSI reconstruction is formulated as an iterative optimization problem jointly constrained by quantization consistency and multi-prior regularization. Each iteration alternates between a quantization-consistency module (QCM) that explicitly corrects deviations from the feasible quantization interval and a multi-prior refinement network (MPRN) that serves as a learnable proximal operator integrating raw-band spectral continuity and curvature modeling, multi-scale spatial detail enhancement, and gated spectral–spatial contextual interaction. Extensive experiments demonstrate that the proposed QuHIC not only achieves efficient HSI compression at competitive bit rates but also outperforms many state-of-the-art methods in both reconstruction fidelity and spectral detail preservation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Hyperspectral Image Compression", "Hyperspectral Image Reconstruction", "Deep Unrolling." ], "forum_url": "https://openreview.net/forum?id=7hRDluQdRc", "pdf_url": "https://openreview.net/pdf/7006c81f61e24301f0406ef87a532cf29c6ac29e.pdf", "updated_at": "2026-10-05T20:37:38.288Z", "categories": [ "ptq" ], "matched_terms": [ "low bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lWUM9JOGww", "number": 39044, "title": "Tailoring the Quantization Space for 1-Bit KV Cache Compression", "abstract": "The key-value (KV) cache becomes a major memory bottleneck in long-context LLM inference, placing substantial pressure on memory capacity and bandwidth.\nTo mitigate this bottleneck, vector quantization (VQ) has emerged as a promising approach for aggressive KV cache compression.\nHowever, existing VQ methods degrade substantially in the 1-bit regime.\nAt such extreme compression, each codebook must represent a larger group of channels with a limited set of centroids, making effective use of its capacity increasingly challenging.\nTo address this, we introduce $\\textbf{TaSQ}$, which tailors the VQ target space by combining query-guided channel weighting, cross-head normalization, and covariance-aware channel grouping to better reflect the error sensitivity and statistical structure of cached activations.\nSince these transforms are RoPE-compatible and can be easily merged into projection weights and codebooks, TaSQ preserves the conventional VQ lookup structure and adds negligible serving overhead.\nAcross general, long-chain-of-thought reasoning, and long-context retrieval benchmarks, TaSQ consistently outperforms existing low-bit KV cache VQ baselines while preserving reasoning stability.\nOn a single RTX 6000 Ada GPU, its SGLang implementation supports up to $14\\times$ larger batch sizes and achieves $1.87\\times$ higher peak throughput compared to the BF16 baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "vector quantization", "low-bit quantization", "large language model inference", "efficient serving" ], "forum_url": "https://openreview.net/forum?id=lWUM9JOGww", "pdf_url": "https://openreview.net/pdf/0f9d7b4a7554408f04dbefadab056fe76f6c966a.pdf", "updated_at": "2026-10-05T20:37:46.988Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "1-bit", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dpHxvmkB39", "number": 39068, "title": "Unmasking Nondeterminism in Masked Diffusion Language Models", "abstract": "Greedy decoding of masked diffusion language models (MDLMs) is nominally deterministic, yet the same prompt can yield different outputs when only the batch size or the GPU changes, while accuracy barely moves and so hides the problem. In autoregressive models, the first differing token permanently separates two generations, but MDLMs also decide at every step where to write, and how nondeterminism develops under this second decision is not understood. Tracking the partially decoded response over 512 steps for LLaDA and Dream, we find that forks arise mainly from the position decision. These forks are often transient: on LLaDA, 92.6\\% of prompts return to an identical intermediate response, and the token conflicts that make a difference permanent appear much later. Treating the two decisions separately, we freeze the position order under periodic LayerCast refresh and recheck only tokens within a few ULPs of a tie, cutting LLaDA's MATH500 output divergence from 69.2\\% to 1.0\\% at less than half the cost of LayerCast, and most of Dream's. Freezing changes the decoding rule, however, so the recipe's outputs agree with each other but often not with LayerCast's, and accuracy can shift. Reproducible MDLM evaluation should therefore track states beyond the first fork and report agreement across executions and accuracy separately. More broadly, locating where nondeterminism enters the decoding rule lets a remedy target the fragile decision instead of paying for full precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language models", "nondeterminism", "reproducibility", "trajectory analysis", "mixed-precision inference", "selective verification" ], "forum_url": "https://openreview.net/forum?id=dpHxvmkB39", "pdf_url": "https://openreview.net/pdf/4ece17173983b65d75a5b16dc3dcee8b759f612a.pdf", "updated_at": "2026-10-05T20:37:50.560Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RIg5yH0tIs", "number": 39100, "title": "QCache-VLA: Quantization and Diffusion Cache for Efficient Vision-Language-Action Policy Inference", "abstract": "Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalist robot control. But their deployment on resource-constrained robotic platforms remains challenging due to the heavy Vision-Language Model (VLM) backbone and multi-step diffusion-based action generation. Existing acceleration methods are not directly suited to VLA diffusion policies: post-training quantization can suffer from image-text distribution mismatch and action-agnostic quantization decisions, while fixed-interval diffusion caching cannot adapt to non-uniform redundancy across denoising steps.\n To address these issues, we propose \\textbf{QCache-VLA}, a discrepancy-aware acceleration framework for efficient VLA inference. For the VLM backbone, we introduce action-aware post-training quantization with modality-aligned smoothing candidate generation, action-aware candidate selection, and modality-balanced weight reconstruction. For the action expert, we propose a residual-calibrated dynamic diffusion cache that reuses deep residuals only when adjacent denoising steps are sufficiently similar. We further integrate INT8 Triton kernels and diffusion-step precomputation to enable practical end-to-end acceleration.\n Experiments in simulation and real-world tasks show that QCache-VLA preserves task performance while reducing GPU memory usage by $37\\%$ and achieving a $4\\times$ end-to-end speedup.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language-Action Models", "Post-training Quantization", "Diffusion Cache", "Efficient Inference." ], "forum_url": "https://openreview.net/forum?id=RIg5yH0tIs", "pdf_url": "https://openreview.net/pdf/0229ed5e0ee7f5c7c89618e3eedab20ebe6e1df7.pdf", "updated_at": "2026-10-05T20:37:58.309Z", "categories": [ "ptq", "hardware", "diffusion", "multimodal" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6xNJZSLSjl", "number": 39109, "title": "The Double-Edged Role of Mean Bias in FP4 LLM Pretraining", "abstract": "FP4 training promises substantial memory and compute savings for large language models, but remains fragile because blockwise quantization is dictated by extreme activation magnitudes, which inflate dynamic range and compress long-tail signals. We identify a counterintuitive source of this failure: dominant activation outliers are not merely arbitrary sparse events, but are largely induced by a coherent rank-one mean bias, whose direction aligns with the leading anisotropic spectral component. This mean component strengthens during training, is amplified and reshaped by attention and FFN operators, and increasingly dominates top activation magnitudes. Crucially, this discovery reveals that a seemingly complex outlier-suppression problem admits a truly simple solution: isolate the coherent mean before quantization. We therefore propose Averis, a mean–residual splitting quantization method that separates the mean component using only reductions and elementwise subtractions before FP4 quantization. Across Qwen3-0.6B Dense and Qwen3-7B-A1.5B MoE, each trained on 100B tokens, Averis enables robust W4A4G4 FP4 training, reducing BF16 loss gaps to 1.19%/0.479% versus 2.05%/0.737% for NVIDIA's recently released Hadamard-based outlier-smoothing method, while limiting downstream gaps to 0.89/0.68 points. With only 2.20% end-to-end overhead over vanilla NVFP4, about 30% of NVIDIA's Hadamard-based design, Averis provides a hardware-efficient path to stable low-bit LLM training. Code is available at: https://anonymous.4open.science/r/averis-504D.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Quantization", "FP4", "Pretraining" ], "forum_url": "https://openreview.net/forum?id=6xNJZSLSjl", "pdf_url": "https://openreview.net/pdf/6a4127e0d0d3fc9680b28931d758e6fddfb47017.pdf", "updated_at": "2026-10-05T20:37:59.008Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "low-bit", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CEVDNWYJ1H", "number": 39346, "title": "BRINE: Matched-State Counterfactual Supervision for Risk-Aware Cognitive Routing in Underwater Robots", "abstract": "Long-horizon underwater control requires cognitive computation to be allocated by current risk and capability. Reactive control efficiently handles routine conditions, while degraded sensing and disturbances benefit from adaptive deliberation; learning this allocation requires evidence about the consequences of alternative cognitive branches. We introduce BRINE, a multimodal system that addresses this coupled decision, learning, and execution problem. From observed robot state, RGB images, point clouds, sonar, task commands, and episodic memory, BRINE determines when deliberation is needed and coordinates three role-specialized components: a state refiner, a safety verifier, and a capability planner. A safety-constrained action projection governs execution. To train the router, we construct matched-state counterfactual supervision by evaluating alternative expert combinations from the same simulator state and comparing their task, safety, and computation outcomes. We further formulate risk-shielded contextual-bandit refinement of the router and introduce role-aware mixed-precision execution that retains precision-sensitive semantic and safety computations in FP32. On 72 paired underwater tasks, the standard-precision system succeeds in 62 cases, compared with 48 for reactive control alone and 52 for always-on deliberation. Counterfactual supervision improves held-out expected return by 0.255 (95% confidence interval: 0.139–0.377) over matched risk and teacher supervision. In separate paired 24-task evaluations, contextual-bandit refinement reduces actual expert-forward duty from 0.609 to 0.482 (paired difference −0.127; 95% confidence interval: −0.180 to −0.074; Holm-adjusted p < 0.001), completes 24/24 tasks with zero collisions, and matches the reference route on all 287 high-risk decisions; mixed precision reduces peak temporary allocated memory by 7.13% and matches FP32 with 24/24 successes. These findings support counterfactual learning for selective, safety-conscious deliberation and mixed-precision execution with preserved behavioral semantics.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "robot learning", "conditional computation", "counterfactual supervision", "cognitive routing", "safe robot control", "multimodal perception", "contextual bandits", "mixed-precision inference", "underwater robotics" ], "forum_url": "https://openreview.net/forum?id=CEVDNWYJ1H", "pdf_url": "https://openreview.net/pdf/0b0b99538d7ad245d077c298bcf1f4ba00db4c02.pdf", "updated_at": "2026-10-05T20:38:38.645Z", "categories": [ "multimodal" ], "matched_terms": [ "mixed precision", "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8rRLc5sCxL", "number": 39529, "title": "IUC: Intersection--Union Correction for Quantization-Activated Attacks on LLMs", "abstract": "While quantization is a key technique for the efficient deployment of large language models (LLMs), recent studies have shown that it may also introduce new security risks: a model that behaves normally in full precision may exhibit malicious behaviors after quantization, such as advertisement injection, over-refusal, or jailbreak, posing a practical threat to the secure deployment of LLMs. To address this issue, we propose **I**ntersection--**U**nion **C**orrection (IUC), a *one-shot* defense method applied directly to the full-precision model. IUC models local parameter intervals under different quantization configurations, where the intersection of these intervals characterizes the shared local reference region underlying the correction, while the union constrains the magnitude of parameter correction. *A smaller intersection indicates a narrower shared local reference region,* and IUC prioritizes the correction of such parameters. With only a single correction applied to the full-precision model, IUC supports secure deployment under multiple commonly used quantization configurations. We conduct experiments across multiple LLMs, different attack scenarios, and representative quantization configurations. The results show that IUC effectively suppresses the activation of malicious behaviors after quantization with low resource overhead, while largely preserving model utility, providing a one-shot pre-sanitization solution for secure model deployment under commonly used quantization configurations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Conditioned Attacks; Large Language Model Security; Model Quantization; Secure Quantized Deployment" ], "forum_url": "https://openreview.net/forum?id=8rRLc5sCxL", "pdf_url": "https://openreview.net/pdf/328c1d98b0818d349f1df8f899f41a04315682b7.pdf", "updated_at": "2026-10-05T20:39:07.424Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cbMVbu0jAP", "number": 39539, "title": "RRC-KV: Range-Driven Rotation and Centering for Two-Bit KV Caches", "abstract": "Scalar INT2 KV caches have only four reconstruction levels per group, making accuracy sensitive to quantization coordinates. We present RRC-KV, which learns per-head key and per-layer value rotations through range-based objectives. A query-based auxiliary regularizer limits coordinate sensitivity concentration, while static key centering shifts the encoding origin with its contribution restored in attention logits. The method requires neither downstream labels nor quantizer backpropagation and supports mixed-precision attention with fused Triton kernels. Across four models and five reasoning and coding benchmarks, RRC-KV achieves a macro-average score of 79.54 versus 80.04 for BF16. Under the same KV storage layout, it outperforms the spectral-rotation baseline OSCAR in every tested RULER setting across two models and 8K–64K contexts (87.19 versus 81.61 on average), while a long-context gap to BF16 remains. Across three Qwen models, scores on two coding tasks vary by at most 3.05 points between two calibration configurations, compared with up to 34.69 points for OSCAR on Qwen3-32B. At 64K tokens, the KV representation provides 6.91× compression relative to BF16, including metadata and retained BF16 windows.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "low-bit quantization", "efficient LLM inference", "long-context inference" ], "forum_url": "https://openreview.net/forum?id=cbMVbu0jAP", "pdf_url": "https://openreview.net/pdf/bd427a62a9630476fc34cb4c84e04f7ad0c6583e.pdf", "updated_at": "2026-10-05T20:39:07.980Z", "categories": [ "ptq", "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "int2", "low-bit", "mixed-precision", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "qo5076s7Xg", "number": 39607, "title": "CLASP: Cross-Layer Active-Subspace Pooling for Quantized Mixture-of-Experts", "abstract": "Low-rank corrections can recover accuracy lost to aggressive post-training quantization of mixture-of-experts (MoE) language models. Most low-bit MoE correction methods fit their factors independently or within individual layers. We find substantial overlap between experts' activation-weighted dominant subspaces across layers. We use this overlap in Cross-Layer Active-Subspace Pooling (CLASP), which clusters experts by the Grassmannian distance between their active subspaces and fits a shared low-rank basis to each cluster. Sharing reduces basis storage and allows a larger correction rank at a fixed bit budget. In the evaluated Qwen models, experts activated by the same token also tend to belong to a few clusters. CLASP stores the shared factors in a device-resident pool and batches correction operations, with stream overlap reducing execution overhead. At 2.15--2.21 bits per quantized parameter, CLASP achieves the lowest perplexity among the compared quantizers on four MoE models. Under the same measurement protocol and hardware, end-to-end generation is up to $4.9\\times$ faster than the FP16 baseline. Our source code will be available after acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts (MoE)", "Post-Training Quantization (PTQ)", "Low-Rank Correction", "Cross-Layer Subspace Sharing", "Active-Subspace Pooling", "Grassmannian Clustering", "Efficient LLM Inference" ], "forum_url": "https://openreview.net/forum?id=qo5076s7Xg", "pdf_url": "https://openreview.net/pdf/a7bf80a7f7f61085c731b67c8e907963d988cb7b.pdf", "updated_at": "2026-10-05T20:39:19.098Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nS6Whua0UL", "number": 39802, "title": "PiSO: Scale Optimization in Post-Training Quantization", "abstract": "Post-training quantization (PTQ) compresses large language models by mapping weights to low-bit representations. While much work has focused on how weights are rounded onto the quantization grid, the scaling factor that defines the grid is typically chosen with simple, data-free heuristics. We observe that, under round-to-nearest (RTN) quantization, the grid assignments are piecewise constant in the scale, making the data-aware layer-output error piecewise quadratic over finitely many intervals, each with a closed-form minimizer. Building on this insight, we present PiSO (Piecewise Scale Optimization), an algorithm that leverages calibration data to efficiently compute the globally optimal channel-wise scales under RTN. We extend PiSO to group-wise quantization via tractable objectives. We also propose strategies for integrating it with error-correction methods such as GPTQ and Qronos, which greedily adjust weight assignments to compensate for the rounding errors of previously quantized weights. Experiments on Llama and Qwen models across multiple sizes and bit-widths show consistent improvements in perplexity and zero-shot accuracy, both standalone and combined with error correction, with gains increasing as the bit-width narrows. For instance, on Llama-3, PiSO reduces the 2-bit perplexity of Qronos by up to 7x, and at 3 bits, RTN with PiSO scales alone outperforms GPTQ with absmax scales.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post Training Quantization", "Large Language Models", "Foundation Models", "Efficient Machine Learning" ], "forum_url": "https://openreview.net/forum?id=nS6Whua0UL", "pdf_url": "https://openreview.net/pdf/50498a2e31643a6d8650df91807e8338c4d5a9c5.pdf", "updated_at": "2026-10-05T20:39:44.757Z", "categories": [ "ptq" ], "matched_terms": [ "2-bit", "bit-width", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nKphW3JJe3", "number": 39907, "title": "VLA-HWBench: Where the Time Goes When VLA Models Run on Edge Hardware", "abstract": "Vision--language--action (VLA) models increasingly run on edge devices aboard robots, where inference latency constrains closed-loop robot control. However, it remains unclear how to design VLA models, their inference software and edge hardware so that robots can run VLA inference in real time. Most VLA models report only task success, and prior studies cover few of the ways recent VLAs generate actions, split a call into at most three phases and do not compare inference optimizations on the same models. Therefore, we introduce **VLA-HWBench**, a benchmark that measures 18 open VLA models on their released software and on alternative backends, on an NVIDIA Jetson AGX Thor and an AMD Ryzen AI Max+ PRO 395, with a seven-stage breakdown of each call. Our study reveals three insights. First, where the time goes depends on how a model generates actions, not on its size. Action generation is the largest stage in 14 of the 19 configurations on the Jetson, latency spans $82\\times$ on the Jetson and $157\\times$ on the AMD platform, and parameter count explains at most 15\\% of its variance. Models that generate long action chunks in a few flow-matching steps keep up with control, whereas token-by-token decoding does not. Second, kernel fusion with 4-bit weights helps most, running $\\pi_{0.5}$ and GR00T N1.7 2.6--2.9$\\times$ faster than PyTorch with torch.compile, but it exists for only these two models and only on the Jetson. Optimization also moves the time to the stages it shortens least, as CPU preprocessing grows from 12\\% to 39\\% of the GR00T N1.7 call. Third, after optimization, memory bandwidth limits only the stages that process few tokens on the Jetson, such as the denoising steps, whereas software limits all other stages, from unfused kernels to serial preprocessing on the host CPU. Overall, low-latency edge VLAs should use few action-generation iterations, inference software should extend kernel fusion and low-bit weights to more models and platforms, and edge hardware needs high memory bandwidth and a fast host CPU.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "VLA", "benchmark", "latency" ], "forum_url": "https://openreview.net/forum?id=nKphW3JJe3", "pdf_url": "https://openreview.net/pdf/de4c927bc323423f6ac2946c17d7c8424572379d.pdf", "updated_at": "2026-10-05T20:40:00.665Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "4-bit", "low-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FHtrHTQA3C", "number": 39908, "title": "COMPOTE: Collapse-Resistant Multiple-Hypothesis Learning For Sequential Recommendation", "abstract": "Next-item prediction is inherently ambiguous: a single interaction history often\nadmits multiple plausible continuations across distinct catalogue regions. Stan-\ndard sequential recommenders compress these alternatives into a single prediction,\nconflating user intents. We introduce COllapse-resistant Multiple hyPOThesis\nlEarning (COMPOTE), a multi-hypothesis transformer trained with an annealed\nobjective that transitions from uniform supervision to a winner-takes-all assign-\nment. This mechanism forces hypotheses to specialize into distinct catalogue re-\ngions, while inference relies on an efficient, normalizer-free max-similarity rule.\nTheoretically, we show that in the population case, the optima solutionspartition\nthe conditional next-item distribution according to a max-entropy quantization\nprinciple, while pooled scores preserve the Bayes-optimal ranking. We also prove\na generalization bound scaling logarithmically with the number of hypotheses.\nAcross three Amazon Reviews 2023 benchmarks, COMPOTE outperforms strong\nID-based, generative, and multi-interest baselines using item IDs alone. Matched-\ncapacity ablations confirm that performance gains stem directly from the annealed\nobjective, preventing the head collapse observed in conventional multi-head archi-\ntectures, rather than added capacity, ensembling, or routing.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sequential recommendation", "multi-hypothesis prediction", "multiple-choice learning", "winner-takes-all learning" ], "forum_url": "https://openreview.net/forum?id=FHtrHTQA3C", "pdf_url": "https://openreview.net/pdf/2ebbffbb5a8688fbca87067229f267864a2c53a5.pdf", "updated_at": "2026-10-05T20:40:00.860Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zJjpNWymXH", "number": 39924, "title": "Watermark Basins: Black-Box Detection of Distilled LLMs", "abstract": "Proprietary language models and their reasoning traces are vulnerable to API-based distillation, where an adversary trains a student model on the teacher model outputs. To provide evidence of distillation, model providers can watermark the teacher outputs, aiming to leave radioactivity in the student outputs. Existing black-box detectors search for such radioactivity using token-level statistics from the suspect\nmodel’s outputs. However, token-level signals are not always detectable; for instance, when the teacher model is distilled at a low temperature. In this work, we discover watermark basin, a persistent signal that enables more reliable watermark detection: during fine-tuning, the watermarked teacher traces embed a distinct, watermark key-specific geometric signal in the parameter space that lies orthogonal to the ordinary task-specific fine-tuning subspace. Building on this, we introduce Watermark Basin Alignment (WM-Basin), a black-box watermark detector that queries through the suspect model’s API, without requiring access to its weights or logits. In particular, WM-Basin maps the generated text into gradient representations using a fixed clean probe model, projects out the ordinary fine-tuning task directions, and verifies if the residual gradient is aligned with the watermark basin induced by the watermark key. We evaluate WM-Basin extensively across diverse teacher models and student architectures (spanning 135M to 8.2B parameters) distilled on mathematical reasoning. Under a strict 0.1% false-positive rate, WM-Basin detects 79% of distilled models at n-gram=1 and 45% at n-gram=3, substantially improving the state-of-the-art token-level baseline (29% and 1.3%, respectively). Furthermore, because the watermark signal resides in a persistent basin, it survives aggressive post-distillation modifications, including further fine-tuning, parameter quantization, pruning, and reinforcement-learning post-training", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Distillation detection", "watermarking" ], "forum_url": "https://openreview.net/forum?id=zJjpNWymXH", "pdf_url": "https://openreview.net/pdf/0fe54f8b0a6d1fe29492df1dbf88f132a4c84141.pdf", "updated_at": "2026-10-05T20:40:02.002Z", "categories": [ "ptq", "training", "reasoning" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "o6yJANh8JA", "number": 39930, "title": "Pixel Art Is a Palette, Not a Downscale: Rebuilding Targets and Decoding for Sprite Generation", "abstract": "Game sprites of 12–24 px are among the smallest images people draw on purpose, and their medium is defined by a handful of flat colours and hard alpha edges. Despite rapid progress in text-to-image generation, existing systems render such sprites as shrunken illustrations with soft shading and around a hundred colours. We identify two root causes: the training targets of sprite corpora are dominated by downscaled artwork rather than native pixel art, and continuous diffusion has no mechanism to express the few-colour structure of the medium. To address these challenges, we propose PaletteDiff, a framework for native-resolution sprite generation. First, we introduce Native-Resolution Target Rebuilding, which removes duplicate leakage, forbids aggressive downscaling and adds 15k permissively licensed sprites, together with a native evaluation protocol that scores against real low-resolution pixel art only. Second, we propose Palette-Projected Sampling, a training-free decoding rule that projects the predicted clean image onto a per-sprite k-colour palette during the late denoising steps, so that the remaining steps repair the quantisation seams. Third, combined with SDEdit-style re-noising, the same rule yields a Palette Refinement module for sprites from generators we did not train. PaletteDiff lowers FD-DINOv2 at 16 px from 167.2 to 31.8 ± 0.5; a per-sprite palette is the decisive ingredient, and imposing it inside the sampler improves on the same quantisation applied afterwards by a further 18%. Under the native protocol, PaletteDiff has the lowest FD among gpt-image-2, FLUX.2 and SDXL pipelines given the same palette post-process; Palette Refinement improves all three external systems by 36–41% while preserving 80% of their silhouettes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion models", "text-to-image generation", "sampling", "pixel art", "evaluation protocols", "dataset curation" ], "forum_url": "https://openreview.net/forum?id=o6yJANh8JA", "pdf_url": "https://openreview.net/pdf/0a672a0c61dfc95e7e8cedebf60238c01d33e323.pdf", "updated_at": "2026-10-05T20:40:03.155Z", "categories": [ "diffusion" ], "matched_terms": [ "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "0836yfLWhh", "number": 39961, "title": "Language-Grounded Controller Latents for Humanoid Whole-Body Control", "abstract": "Language-conditioned humanoid motion is commonly built as a cascade in which a text-to-motion model generates a kinematic trajectory that is subsequently tracked by a separate controller. We instead generate the latent commands of a frozen whole-body controller directly from text, removing runtime retargeting and motion tracking from the primary generation path. We also incorporate differentiable kinematics into diffusion sampling to impose hand and foot target constraints directly in the controller latent space, without retraining the generator or controller. Experiments demonstrate strong language alignment, high completion rates, and effective spatial-constraint preservation after physical execution. We further demonstrate diverse behaviors in simulation and on the real Unitree G1. Beyond generation and execution, we characterize the semantic and physical structure of the frozen controller latent representation and examine whether this structure is preserved in latents generated directly from language. Our analysis shows that the 64-dimensional latent representation encodes joint configuration, locomotion direction and intensity, and gait phase in a distributed manner, with much of this structure preserved after finite scalar quantization and also present in the generated latents. We further show that text-conditioned motion characteristics can be composed in latent space to generate motions combining multiple attributes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Humanoid Control", "Language-Conditioned Motion Generation", "Controller Latent Space", "Whole-Body Control" ], "forum_url": "https://openreview.net/forum?id=0836yfLWhh", "pdf_url": "https://openreview.net/pdf/85467268a67cef73fd9cfab65015d014c257c1c0.pdf", "updated_at": "2026-10-05T20:40:09.856Z", "categories": [ "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3AfnbslMtb", "number": 40006, "title": "Adaptive Mixed-Precision Codebook Quantization for Large Language Models", "abstract": "Codebook quantization enables efficient compression of large language models to low bit-widths, making it suitable for resource-constrained deployment. \n Existing methods assign a uniform codebook configuration to all matrices. \n However, under a tight average-bit budget, this uniform configuration leads to capacity misallocation: insensitive linear maps receive disproportionate precision, while sensit ive projection layers remain under-represented. \n To address this issue, we propose Adaptive Mixed-Precision Codebook Quantization (AmpCQ), a framework that assigns different codebook configurations based on per-matrix sensitivity. \n The proposed method combines: (i) a heterogeneous tier space comprising single- and dual-codebook quantizers; \n (ii) a normalized activation-aware reconstruction metric for cross-matrix comparability;\n (iii) Pareto-filtered greedy allocation guided by marginal return per bit; \n and (iv) lightweight codebook-delta adaptation for downstream tasks. \n On Llama3-8B and Qwen3-14B, AmpCQ demonstrates superior performance over uniform AQLM-style allocation at matched average bit width. \n Specifically, when Qwen3-14B is compressed to an average bit width of approximately 3 bits, it retains 97.3\\% of its original average zero-shot accuracy.\n Our code is available in the Supplementary Material.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Compression", "Codebook", "Quantization" ], "forum_url": "https://openreview.net/forum?id=3AfnbslMtb", "pdf_url": "https://openreview.net/pdf/15d8fe6df68d591b546ac7d2b820e68301c47185.pdf", "updated_at": "2026-10-05T20:40:15.245Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "bit width", "mixed-precision", "quantization", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ks9OjFC5hZ", "number": 40099, "title": "HiCAP-Q: Hierarchical Capability-Aware Mixed-Precision Quantization for Large Language Models", "abstract": "Large language model quantization effectively reduces inference costs and memory usage. However, existing methods typically minimize weight or activation reconstruction errors during calibration. These objectives are only weakly aligned with preserving downstream capabilities.\nAs a result, a small reconstruction error does not necessarily imply a small degradation in domain-specific abilities such as mathematical reasoning, code generation, or long-context understanding.\nWe propose **HiCAP-Q**, a capability-aware mixed-precision post-training quantization framework. It preserves model-specific capabilities under a user-specified bit budget. Our key insight is that different capabilities rely on different model parameters. Some layers and linear modules are critical for mathematics, coding, or long-context reasoning. Others can tolerate more aggressive quantization.\n**HiCAP-Q** exploits this structure by reallocating precision from capability-insensitive parameter groups to capability-critical ones.\n**HiCAP-Q** consists of three stages.\nFirst, it constructs capability-specific probing inputs at low cost and\nwithout human annotation. It uses the target model and, optionally, a judge\nmodel. An **input score** selects prompts based on question quality,\ntask relevance, and response confidence.\nSecond, the framework estimates capability sensitivity at the\nTransformer-layer and linear-module levels. A capability loss score measures\nthe degradation caused by quantizing each parameter group.\nThird, the framework ranks parameter groups by sensitivity and assigns mixed\nprecision under a target bit budget.\nExperiments on MMLU-Pro, AIME-2026, and GPQA show consistent gains over\nreconstruction-based baselines. With only 10\\% more average bits than uniform\nlow-bit quantization, **HiCAP-Q** substantially narrows the gap to\nthe full-precision model. It also preserves strong general-task performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Model Quantization", "Mixed Precision" ], "forum_url": "https://openreview.net/forum?id=ks9OjFC5hZ", "pdf_url": "https://openreview.net/pdf/108b8cb60e4a71cdd4cb6760dd1ccfc49c00f42a.pdf", "updated_at": "2026-10-05T20:40:28.032Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "low-bit", "mixed precision", "mixed-precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JwPObstJCQ", "number": 40212, "title": "Beyond Average Success: Outcome Reversals and Delayed Divergence in Quantized Vision-and-Language Navigation", "abstract": "Quantization is widely used to compress models, reduce memory requirements, and potentially accelerate inference. When average performance is preserved, a natural expectation is that quantization introduces only local numerical perturbations while leaving model outputs and most individual outcomes largely unchanged. However, our experiments in streaming vision-and-language navigation (VLN) reveal substantial behavioral differences behind similar average scores. Changed trajectories can also alter subsequent observations and histories. Using instruction-paired R2R-CE episodes, we compare four-bit weight quantization of StreamVLN against its unquantized bfloat16 (BF16) model as the full-precision baseline. Success rate decreases by only 0.11 percentage points, yet 18.5\\% of outcomes reverse and 87.0\\% of action sequences change, as gains nearly offset losses. To examine this discrepancy, fixed-state probes separate numerical perturbations from immediate decisions and distinguish the effects of current computation and historical content. Further, controlled continuations show that history perturbations can induce later divergence even when the initial action queue is held fixed and subsequent computation returns to BF16. Together, these findings show that similar average performance does not imply behavioral fidelity, and immediate action agreement does not guarantee subsequent agreement in the tested states. Behavioral differences need not reduce task success. Quantized VLN deployment therefore calls for evaluating retained task outcomes and subsequent behavior alongside average quality, realized memory savings, and end-to-end latency.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model quantization", "vision-and-language navigation", "decision-level analysis" ], "forum_url": "https://openreview.net/forum?id=JwPObstJCQ", "pdf_url": "https://openreview.net/pdf/0c5a13f896a5fd61f4d49650a55040f8e8fd2a97.pdf", "updated_at": "2026-10-05T20:40:45.195Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "R8W4EfG54g", "number": 40230, "title": "OPTQ: Quantization Error is Relocated, Not Eliminated", "abstract": "Post-Training Quantization (PTQ) inherently cannot eliminate the quantization error it generates; it can only determine where to distribute this error. The magnitude of the residual injected during rounding to a fixed grid is strictly determined by the grid itself, remaining identical across any algorithm. Consequently, what a quantizer actually controls is not the magnitude of the error, but the allocation of the error across channels, basis directions, and time. This paper formalizes this perspective. For an arbitrary linear operator receiving a quantized tensor, the distortion surviving to the output is dominated by the operator's spectrum. Thus, a good quantizer is one that steers the error toward the directions that the operator attenuates. This framework not only designates operators for each axis but also, equally importantly, predicts which axes are not worth optimizing. For instance, the time axis of the Value cache is dismissed prior to any empirical experiments, solely based on the observation that the measured lag-1 attention autocorrelation is approximately 0.06.\nBy instantiating this framework, we derive a calibration-free sequential quantizer for activations, where the downstream operator is not estimated from data but exactly determined by the layer's weights. For the Key cache, we formulate an operator combining channel mean subtraction and per-head Hadamard rotation. Empirically, under a W4A4KV4 setting, our training-free configuration achieves a perplexity of 5.85 on Llama-2-7B, outperforming recent methods that rely on learned transformations (QuaRot, FlatQuant, SpinQuant, and OSTQuant). Incorporating transformations learned to pass through the relocation operator establishes a new state-of-the-art, yielding a perplexity of 5.714 on Llama-2-7B and 6.899 on Llama-3-8B.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "llm", "quantization" ], "forum_url": "https://openreview.net/forum?id=R8W4EfG54g", "pdf_url": "https://openreview.net/pdf/d0457cf266e8d91005e5635bdd5d3481994cd365.pdf", "updated_at": "2026-10-05T20:40:48.387Z", "categories": [ "ptq" ], "matched_terms": [ "ptq", "quantization", "quantized", "quantizer", "w4a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hpBBQBWSIM", "number": 40277, "title": "RAG-QAF: Robust Adaptation and Generalization for Quantization-Aware Fine-Tuning", "abstract": "Quantization-aware fine-tuning (QAF) is crucial for adapting pretrained large language models (LLMs) to specific downstream tasks on resource-constrained devices. However, existing QAF methods tend to prioritize in-distribution (ID) performance on target adaptation tasks, while paying limited attention to out-of-distribution (OOD) generalization. Although random parameter perturbations can improve generalization in full-precision fine-tuning, directly introducing such noise into QAF may degrade both ID and OOD performance. In this work, our multi-scale analysis reveals that random noise lacks explicit optimization guidance and overlooks local differences in quantization error. We further observe that the model’s global noise sensitivity evolves throughout training, making a fixed perturbation magnitude inadequate. Based on these findings, we propose RAG-QAF, a robust QAF framework for ID adaptation and OOD generalization. Specifically, **local error-guided adversarial optimization** uses quantization errors to guide the search for adversarial perturbations through gradient ascent. **Global adaptive curriculum learning strategy** progressively increases the overall perturbation budget to accommodate evolving noise sensitivity, enabling a transition from stable early optimization to increasingly challenging robust optimization. RAG-QAF can be integrated into existing QAF pipelines as a plug-and-play module. Extensive experiments on ID and OOD benchmarks demonstrate state-of-the-art performance with low additional training overhead.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "model compression", "large language model", "model quantization" ], "forum_url": "https://openreview.net/forum?id=hpBBQBWSIM", "pdf_url": "https://openreview.net/pdf/333a38bfbdcfcd929279b8f4dd57ea818b3946f8.pdf", "updated_at": "2026-10-05T20:40:56.225Z", "categories": [ "qat", "training", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "f8HuMKzsHf", "number": 40332, "title": "UAQB: Utility-Aware Quantization Bottleneck for Private Split VLM Inference", "abstract": "Split inference allows resource-constrained devices to offload vision-language model (VLM) computation to a server by transmitting intermediate visual tokens instead of raw images, thereby avoiding direct exposure of private image content. However, these visual tokens remain susceptible to reconstruction attacks because substantial spatial and semantic information is retained in them. Existing defenses mitigate leakage largely at the expense of downstream task performance. To address this challenge, we propose the Utility-Aware Quantization Bottleneck (UAQB), a plug-in defense deployed on-device prior to token transmission. UAQB applies adaptive quantization to visual tokens to restrict image information flowing through the bottleneck, coupled with utility adaptation that restores compatibility of the resulting information-limited representation with the downstream VLM. We derive a tractable upper bound on the information bottleneck objective, enabling joint optimization of privacy and utility. Extensive evaluations across multiple utility benchmarks and reconstruction metrics demonstrate that UAQB achieves a more favorable utility-privacy trade-off than existing defenses, with particularly high utility retention at stronger protection levels.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "split inference", "vision-language models", "privacy", "reconstruction attacks", "information bottleneck", "quantization" ], "forum_url": "https://openreview.net/forum?id=f8HuMKzsHf", "pdf_url": "https://openreview.net/pdf/a3020693a11142b1883ef3cfc3ec8e9224e09f1b.pdf", "updated_at": "2026-10-05T20:41:04.364Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Zch46j3XOj", "number": 40370, "title": "Let the Batch Decide, On- or Off-Policy: Effective Sample Size for RL Post-Training", "abstract": "Reinforcement learning (RL) is structurally harder than supervised learning because the policy changes the data distribution it learns from. The resulting fragility is especially visible in large-model training using nominally on-policy RL, where the training and rollout systems can differ in numerical precision, sampling and other implementation details. The common approach handles this difference by constraining the update with a fixed clip range, so the algorithm becomes more sensitive to its configuration and must be retuned whenever the task, model scale or distribution mismatch changes. We argue that the fragility traces to two linked concerns that this one fixed range has to cover. The first is a trust-region concern: an update should not move the policy too far from its current value. The second is an off-policy concern: data from older or different behavior policies should influence the update only to the extent that the update remains reliable. Neither concern is a constant to set in advance, and the policy ratios of the current batch show how severe each one is. We adopt P3O, a simple yet effective batch-adaptive objective from prior work that replaces fixed clipping with the normalized effective sample size of the per-token policy ratios, which we take against the rollout engine's log-probabilities. The same statistic caps the score-function weight and sets the strength of an off-policy regularizer, so the update stays close to the on-policy update when the ratios are close to one. When stale or mismatched data concentrate the ratios, the update tightens but keeps a nonzero score-function weight on every token. Experiments on math reasoning show that P3O compares favorably with tuned baselines and, under FP8 rollouts on Qwen3-8B-Base, avoids the late collapse of GRPO. P3O adds no new objective hyper-parameters and removes the clip range.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement learning", "Large language models", "off-policy", "RL Post-Training" ], "forum_url": "https://openreview.net/forum?id=Zch46j3XOj", "pdf_url": "https://openreview.net/pdf/5c351a4ffc834b3f1114921072a6569a102ed531.pdf", "updated_at": "2026-10-05T20:41:09.084Z", "categories": [ "ptq", "training", "reasoning" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OVlOj6bcY0", "number": 40550, "title": "Widening the Gap: Exploiting LLM Quantization via Outlier Injection", "abstract": "LLM quantization has become essential for memory-efficient deployment. Recent work has shown that quantization schemes can pose critical security risks: an adversary may release a model that appears benign in full precision but exhibits malicious behavior once quantized by users. However, existing quantization-conditioned attacks have been limited to relatively simple quantization methods, where the attacker can estimate weight regions that remain invariant under the target quantization. Notably, prior attacks have consistently failed to compromise more popular and sophisticated schemes, limiting their practical impact. In this work, we introduce the first quantization-conditioned attack that consistently induces malicious behavior that can be triggered by a broad range of advanced quantization techniques, including AWQ, GPTQ, and GGUF I-quants. Our attack exploits a simple property shared by many modern quantization methods: large outliers can cause other weights to be rounded to zero. Consequently, by injecting outliers into specific weight blocks, an adversary can induce a targeted, predictable weight collapse in the model. This effect can be used to craft seemingly benign full-precision models that exhibit a wide range of malicious behaviors after quantization. Through extensive evaluation across three attack scenarios and LLMs, we show that our attack achieves high success rates against a broad range of quantization methods on which prior attacks fail. Our results demonstrate, for the first time, that the security risks of quantization are not restricted to simpler schemes but are broadly relevant across complex, widely-used quantization methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM security", "quantization attacks", "post-training quantization" ], "forum_url": "https://openreview.net/forum?id=OVlOj6bcY0", "pdf_url": "https://openreview.net/pdf/e809498a7b7e63bf9ed66abf7918096aa74d9096.pdf", "updated_at": "2026-10-05T20:41:34.292Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xsrzSn6t5B", "number": 40583, "title": "LapKDE Flow: Exact Invertible Transports Between Kernel Density Estimates", "abstract": "Kernel density estimation and normalizing flows address opposite sides of the same problem: KDE is directly data-driven and effective in low-data regimes, but static; flows compose, sample, and train, but their scalar maps must be fitted. We close this gap with *LapKDE Flow*, a normalizing flow whose scalar map is\n$T=F_2^{-1}\\circ F_1$, the monotone transport between two Laplace-kernel density estimates. The map is exactly invertible in closed form and admits an exact log-determinant. Anchoring the knots at data recovers KDE as a flow layer; adding rotations and depth yields a multivariate estimator; and adding gradient steps turns the same construction into a trained flow - three operating points of one construction rather than three separate models. Because its bandwidths act as statistical temperatures rather than free slopes, LapKDE Flow provides two independent routes to quantization and propagates gradient to every mixture component from every data point, unlike fixed-knot splines. With zero gradient steps, the anchored estimator outperforms classical small-sample estimators by up to $0.49$ bits per dimension on MNIST. When trained, the same layer improves on published Gaussianization-flow results on all seven standard benchmarks, with the advantage persisting under a matched parameter budget on three of the four comparable benchmarks. It also remains competitive on the circle and in linear ICA. The contribution is not another scalar transform, but a model family in which estimating a density and fitting a density become the same operation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Kernel density estimation", "normalizing flows", "fixed-knot splines", "Gaussianization flows" ], "forum_url": "https://openreview.net/forum?id=xsrzSn6t5B", "pdf_url": "https://openreview.net/pdf/3f99b3e26109d4738797740482f7debd7d79a6b9.pdf", "updated_at": "2026-10-05T20:41:40.141Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "f92skggSl2", "number": 40591, "title": "MoE-VQ: A Model-Agnostic Gaussian Vector Codebook for 2-Bit Expert Weight Quantization", "abstract": "Mixture-of-Experts (MoE) models concentrate most parameters in sparsely activated experts, making weight memory rather than computation the primary deployment bottleneck. However, scalar post-training quantization incurs severe and irreversible distortion at 2 bits. We show that, after row-group RMS normalization, expert MLP weights from six MoE architectures closely follow a shared standard Gaussian distribution. This observation enables a single, model-agnostic vector-quantization codebook: we fit an 8-dimensional, 2^16-entry codebook once to N(0, I) and reuse it across models, layers, and experts without model-specific codebook training. Building on this codebook, MoE-VQ combines block-wise LDLQ residual feedback with rank-14 low-rank error compensation. At 2.34 effective bits per weight, MoE-VQ improves average accuracy across seven tasks by 3.3–13.3 points over tuned 2-bit GPTQ on four MoE models, while removing 51% of the excess perplexity relative to FP16. It also better preserves expert routing and outperforms an equal-rate E8 lattice by 1.0–3.2% in perplexity. For Mixtral-8x7B, MoE-VQ reduces weight memory from 93.4 GB to 15.6 GB, allowing the validated weight configuration to fit on one 80 GB GPU instead of two. These results show that distribution-adaptive vector codebooks can be reusable rather than model-specific for low-bit MoE quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "moe", "llm", "quantiztion" ], "forum_url": "https://openreview.net/forum?id=f92skggSl2", "pdf_url": "https://openreview.net/pdf/843d269685aa6c0db34b98c00f6affff06c0cfde.pdf", "updated_at": "2026-10-05T20:41:41.096Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "quantiztion" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eSETJYHDId", "number": 40697, "title": "StageCal: Shared Neuron Calibration for ANN-to-SNN Conversion with Few-Step Inference", "abstract": "Converting pretrained artificial neural networks (ANNs) into spiking neural networks (SNNs) aims to reuse learned representations while addressing severe clipping, quantization, and temporal mismatch errors at short simulation steps. To mitigate these errors, existing calibration methods commonly reconstruct ANN activations through layer-wise parameter adjustment. Consequently, these methods fail to preserve network-wide feature alignment across heterogeneous stages and varying inference budgets. In this paper, we find that isolated parameter fitting inherently overlooks the complex interactions of temporal conversion errors. Effective calibration cannot treat neuron operating parameters as independent variables. Instead, it must govern them as a globally coordinated system anchored by shared network statistics. In light of this, we propose STAGECAL, a novel shared neuron calibration framework for plug-and-play ANN-to-SNN conversion that keeps pretrained ANN backbone weights strictly fixed. Our framework brings two key technical contributions. Firstly, we introduce a shared calibration network. It utilizes channel-wise activation statistics, normalized stage depth, and target spike budgets to jointly predict threshold adjustments and initial membrane potentials. On top, we then introduce a structured residual correction function. This function dynamically predicts mixing coefficients that combine three fixed residual bases from an initial pilot conversion into bounded stage-wise current corrections. These corrections are theoretically linked through an exact membrane-potential balance identity. Extensive experiments on classification tasks demonstrate that STAGECAL even exceeds the original ANN by 5.71% with only 0.17x energy and few simulation time steps on ImageNet. It also reduces hidden-layer synaptic operations by 4.24% relative to percentile-based conversion. This enables highly efficient SNN inference without running the source ANN and calibration network.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Spiking neural networks", "ANN-to-SNN conversion", "post-training calibration", "low-latency inference", "shared calibration policies", "spike-budget optimization" ], "forum_url": "https://openreview.net/forum?id=eSETJYHDId", "pdf_url": "https://openreview.net/pdf/b9bd7676d3261659cba4be8f6e84216ce5e24d38.pdf", "updated_at": "2026-10-05T20:41:56.948Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yAsnUMasEx", "number": 40802, "title": "StegoCoder: Learning Channel for Steganographic Coding", "abstract": "Image steganography hides a message in a cover image with minimal perceptible distortion, producing a stego image from which an authorized receiver can recover the message. Robust steganography additionally requires recovery after the stego image is perturbed. Classical steganographic codes operate close to the rate-distortion bound, but on a fixed binary projection of the cover, typically its least-significant bit planes, in which the payload is destroyed by even mild perturbations. Neural steganography can be trained end-to-end to survive specific perturbations, but lacks the explicit rate-distortion control that coding theory provides. StegoCoder combines the two. A steganographic code first embeds the message into a fixed binary sequence and produces a binary stego sequence. A neural renderer then synthesizes the stego image from the cover image and the resulting binary stego. At the receiver, a learned bit predictor outputs soft estimates of the binary stego from the perturbed image, and the code decodes the message. The networks thus learn a reliable binary channel over which a classical steganographic code operates, splitting the problem into learning the channel and embedding the message efficiently over a binary cover. With a nested polar code, the same code acts as a lossy source code at the sender and as a channel code at the receiver, giving explicit control over the trade-off between embedding distortion and robustness. On CelebA-HQ, StegoCoder recovers most of the $500$ test images without a single bit error under additive Gaussian noise of standard deviation $10^{-2}$ at $1$ and $2$ bits per pixel (bpp) and $5\\cdot10^{-3}$ at $3$ and $4$ bpp, where least-significant-bit embedding, per-image optimization and end-to-end neural approaches recover none and quantization index modulation substantially fewer. A per-image refinement of the stego image further extends robustness across different payloads and co-training the scheme with the neural polar decoder decreases the distortion.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Data Hiding", "Image Steganography", "Steganographic Coding" ], "forum_url": "https://openreview.net/forum?id=yAsnUMasEx", "pdf_url": "https://openreview.net/pdf/7370e756fbc2d3a41e7769c7857b13d50ca9b031.pdf", "updated_at": "2026-10-05T20:42:11.814Z", "categories": [ "theory" ], "matched_terms": [ "binary", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "s6DxZ6awYc", "number": 40891, "title": "PebbleVideo: On-Device Video Generation via Extreme Model Compression", "abstract": "Diffusion Transformers (DiTs) have substantially advanced video generation quality, yet their quadratic bidirectional attention and billion-scale parameter incur significant memory and storage overhead, hindering deployment on resource-constrained devices. Existing approaches improve efficiency through attention redesign or streaming generation, but their restricted global spatiotemporal interactions limit modeling capacity. In contrast, we explore an alternative direction, model compression, and introduce **PebbleVideo**, which aggressively reduces the parameter footprint through structured compression. The core of PebbleVideo is to evaluate compression sensitivity in functional space, preserving generation-critical functions rather than relying on indirect proxies. For the dominant DiT backbone, we measure structural importance through the perturbation induced on the diffusion vector field, directly exposing each module's influence on generation dynamics. For heterogeneous condition encoders, cross-scale representation consistency determines whether compression proceeds through compact-model substitution or in-place structural pruning. Combined with DMD-based distillation, PebbleVideo achieves 512$\\times$320 video generation at 28 FPS with only 7.45 GB peak memory. Further incorporating model quantization, we deploy the quantized PebbleVideo on a Samsung mobile NPU, achieving a VBench score of 87.46.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model compression", "Video generation", "Device deployment" ], "forum_url": "https://openreview.net/forum?id=s6DxZ6awYc", "pdf_url": "https://openreview.net/pdf/d45199664e25d54c5ef87aca7b3debe55ed4aa61.pdf", "updated_at": "2026-10-05T20:42:24.296Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RTYrjX19YU", "number": 41003, "title": "REPS: Compressed Expert Replacement for Multiversion MoE Serving", "abstract": "Serving expert-local customizations of a mixture-of-experts (MoE) model requires a representation that supports both local updates and shared execution. We propose Replaceable Expert Patches for Serving (REPS), which uses the same compressed records for a parent and its complete expert replacements. Two sign bytes encode each width-eight block through a complete Hadamard transform. This structure gives independent scalar assignment at fixed scale and integer reconstruction without codeword lookup. A version-aware operator selects parent or patch rows and consumes these records directly during prefill. Qwen3-30B-A3B Code updates improve target perplexity while keeping non-target and general-text degradation within 1% in three fixed-initialization data orders. On one of these exact updates, integer reconstruction reduces synchronized backend first-token time by 3.51$\\times$ and 4.45$\\times$ at batches one and eight relative to our lookup-table (LUT) decoder, with unchanged weights and decode. In a nine-format storage control, sharing reduces REPS weight allocation by 48.9% and 73.4% at two and four structural versions.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixture-of-experts", "model serving", "expert versioning", "efficient inference", "quantization" ], "forum_url": "https://openreview.net/forum?id=RTYrjX19YU", "pdf_url": "https://openreview.net/pdf/4c588c4b9835976747c4ccb42c702f3a56582e8f.pdf", "updated_at": "2026-10-05T20:42:37.865Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CdlOGvKKE3", "number": 41073, "title": "RCDQ: Reference-Guided LLM Quantization with Interleaved Continuous Compensation and Discrete Decisions", "abstract": "Full-precision reference outputs are widely used to compensate for accumulated errors in low-bit quantization of large language models. Existing methods incorporate these references in different ways. Some apply continuous compensation to an entire layer before quantization. Others incorporate deviations from reference outputs into sequential weight updates or use reference-based reconstruction objectives to optimize quantization parameters and integer codes. A key challenge is that improvements in output reconstruction achieved through continuous compensation may not be preserved after discretization. To address this challenge, we propose \\textbf{RCDQ} (Reference-Guided Continuous--Discrete Quantization), a two-stage group-wise quantization method that uses the same full-precision reference objective to guide both continuous compensation and quantization decisions. In stage one, before quantizing each weight group, we recompute its weight compensation using statistics describing the output discrepancy between the quantized and full-precision paths. These statistics are updated as groups are processed sequentially. Under the same full-precision reference, in stage two, we then evaluate fake-quantized candidates constructed from the compensated weights and select the best quantization scale multipliers to reduce the discrepancy between the candidate outputs and the full-precision reference outputs. Experiments across seven Llama and Mistral models show that RCDQ achieves the lowest W3 perplexity in all 14 WikiText-2 and C4 model--dataset settings, together with the highest or tied-highest six-task average and MMLU accuracy on six of seven models, respectively. RCDQ remains robust at more aggressive precisions, achieving the lowest W2A16 perplexity in 10 of 14 settings and the lowest W3A8/W4A8 perplexity in three of four settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "large language models", "low-bit quantization", "model compression", "error compensation" ], "forum_url": "https://openreview.net/forum?id=CdlOGvKKE3", "pdf_url": "https://openreview.net/pdf/e0b9fa08c249793a78c24374e60639ec1c6cd3ac.pdf", "updated_at": "2026-10-05T20:42:51.347Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizing", "w2a16", "w3a8", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "J7Z6E1aisY", "number": 41106, "title": "AWSLQ: Activation-Wise Sparse-Ladder Quantization for Accurate W4A4 Single-Load Decoding", "abstract": "On FP4-native GPUs, deployed serving stacks already run W4A4 — FP4 weights times FP4 activations — at decode, yet the quantization literature retreats to 16-bit decode because W4A4 is where accuracy collapses. Small-batch decode is bound by streaming the weights, so its tensor cores sit mostly idle. AWSLQ, a training-free W4A4 scheme that requires no external calibration corpus, spends that idle compute: it multiplies each weight tile by several FP4 activation terms instead of one, here a ladder of N structured-sparse NVFP4 operands, peeled online, that share one weight load. Two-operand ablations had concluded that more sparse computation does not recover the loss; the knee is at N=3, and by N≈4 the ladder is on the W4A16 floor, to within +0.005 perplexity on ten of eleven models. A selective reverse-smoothing fold, in the direction SmoothQuant advises against, lowers the weight-only floor itself; the ladder absorbs the activation outliers the fold creates and lands on the lowered floor, while a single dense term stays above it; whether to fold is decided per model from its own generated text at preparation time. A fused Blackwell kernel runs the ladder as one M-stacked sparse GEMM into a single accumulator and beats cuBLASLt's FP4 GEMM across a Llama-3.1-8B decode block at M=16–64 (1.26× to 1.07× on an RTX 5090, 1.08× at M=16 on a B200). AWSLQ matches or beats every published W4A4 method, calibrated or not, on 16 of the 18 model–dataset cells with a published comparator, at coarser block-32 scales, and on GSM8K it matches the accuracy of retreating to 16-bit decode activations. The accuracy comes from the number of FP4 terms that share one weight load. A sparse rung performs half the multiply-accumulates of a dense term, so the ladder also offers an intermediate arithmetic budget, three rungs at one and a half dense terms, which chooses the refined half of each token's channels online.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Efficient LLM inference", "Activation quantization", "Autoregressive decoding", "Sparse MMA", "Sparse tensor cores", "NVFP4", "GPU kernels" ], "forum_url": "https://openreview.net/forum?id=J7Z6E1aisY", "pdf_url": "https://openreview.net/pdf/7867e5b4103a0940ab8a3f5e039808176a412a84.pdf", "updated_at": "2026-10-05T20:42:57.448Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "nvfp4", "quantization", "w4a16", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JBSJpyvmCe", "number": 41112, "title": "SERESA: Selective Recurrence Restriction for State Space Duality Acceleration", "abstract": "Mamba-2's State Space Duality (SSD) enables hardware-efficient chunk-wise recurrence, yet its computational efficiency does not translate proportionally into prefill latency. Despite accounting for less than 5% of FLOPs, SSD consumes nearly 20% of prefill latency, motivating us to examine temporal dependencies across SSD heads. While SSD applies full recurrence uniformly, output-relevant recurrent contributions vary substantially in temporal range across heads. We quantify this heterogeneity by defining the effective recurrence distribution and its expected lag, Effective Recurrence Length (ERL). Our ERL analysis reveals that most heads concentrate their contributions at short temporal lags, whereas long-range contributions are confined to a small subset of heads. The head-wise ERL ordering also remains stable across input distributions, revealing a quasi-static recurrence structure. Building on this, we propose SERESA, a training-free framework that selectively restricts recurrence. SERESA determines the number of restricted heads from the ERL distribution and selects heads that retain the most contribution within a prescribed recurrence length. It reduces their intra- and inter-chunk computation while recovering boundary-crossing contributions through boundary correction, leaving full recurrence for the remaining heads. Without additional training or input-dependent reselection, SERESA restricts 78–83% of heads across Mamba-2 and Nemotron-H while closely matching full-recurrence baseline perplexity and downstream accuracy. It accelerates the SSD core by 1.75–2.02X and achieves up to 1.29X end-to-end prefill speedup with projection quantization. Code will be made publicly available upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mamba-2", "State Space Duality", "Inference Acceleration", "Prefill Acceleration" ], "forum_url": "https://openreview.net/forum?id=JBSJpyvmCe", "pdf_url": "https://openreview.net/pdf/82e4010d2d3b120e15e6e3784716202d3f49f865.pdf", "updated_at": "2026-10-05T20:42:58.113Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4WBh3S8UYY", "number": 41187, "title": "Compressing Protein Language Models: A Unified Benchmark", "abstract": "Protein language models are becoming central tools across a growing range of\ntasks, and their scale and adoption are increasing in tandem, making efficiency\na main concern, with compression as a popular way to achieve it. Yet what\nit actually costs remains unmeasured. We present the first compression bench-\nmark spanning model families, training objectives, and compression paradigms\nat a common protocol, evaluated at reconstruction, representation geometry, and\ndownstream biology. For deployment, calibration-aware 4-bit is near-lossless at\nevery scale, equivalence-tested with pre-declared margins; 3-bit is free only at the\ntop of a family; and 2-bit collapses the zero-shot channel except at the very top of\na family. How much compression costs depends on the readout: a probe retrained\non the compressed features survives to 2-bit at scale, whereas a reused head or\nSAE does not. Underlying this is a structural dissociation: compression largely\npreserves the representation and shifts its coordinates, so a label-free realignment\nrestores coarse readouts like the probe and SAE, each with a different correction,\nwhile the fine-grained LM head reads exactly the part no such map recovers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "PLMs", "Compression", "Pruning", "Benchmark", "Fine Tuning" ], "forum_url": "https://openreview.net/forum?id=4WBh3S8UYY", "pdf_url": "https://openreview.net/pdf/a02be76532674eb51f568dfd5a14ab5e099c2c3a.pdf", "updated_at": "2026-10-05T20:43:08.590Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "3-bit", "4-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CBPEC3Kbtv", "number": 41217, "title": "COTQuant: Cross-Modal Orthogonal Transform for Large Vision-Language Model Quantization", "abstract": "Post-training quantization (PTQ) is essential for efficient deployment of vision-language models (VLMs), yet low-bit quantization remains challenging due to heterogeneous multimodal activation outliers. We propose COTQuant, a PTQ framework with two complementary components. Cross-Modal Transform (CMT) jointly redistributes text and vision activation outliers through hierarchical orthogonal transformations, while Quantization-Aware Group Search (QAS) reorganizes weight channels to minimize group-wise quantization error. Both transformations leave the model output unchanged and introduce only a small amount of additional inference-time overhead. Across three VLMs and five multimodal benchmarks, COTQuant consistently outperforms existing PTQ methods while retaining 98.03\\%--98.73\\% of the corresponding BF16 performance under W4A4 quantization. Notably, we demonstrate effective 4-bit PTQ on Qwen3.5-122B, which is the first 4-bit PTQ attempt on over 100B-scale VLM.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Vision-Language Models", "Post-Training Quantization", "Weight-Activation Quantization" ], "forum_url": "https://openreview.net/forum?id=CBPEC3Kbtv", "pdf_url": "https://openreview.net/pdf/89f594f97ff98a38bcf7170392dc8641099c00e0.pdf", "updated_at": "2026-10-05T20:43:15.640Z", "categories": [ "ptq", "qat", "hardware", "multimodal" ], "matched_terms": [ "4-bit", "low-bit", "ptq", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Fiuq2wls5v", "number": 41252, "title": "Layerwise Error Attribution for Fast and Robust Mixed-Precision Post-Training Quantization", "abstract": "Mixed-precision post-training quantization is a network compression method that assigns bits layer by layer, under a global memory budget using a small calibration set. The main difficulties are to overcome the combinatorial nature of the allocation problem and to manage the sensitivity to small, potentially corrupted databases. Hence, an efficient allocation method should be fast to compute and preserve model quality when calibration data are corrupted. To design such a method, we derive a layerwise probabilistic analysis of the quantization error that separates propagated error from the local perturbation introduced at a given layer. We use this local term to build a separable score for a simple allocation algorithm, that requires no external solver. The probabilistic nature of our approach brings robustness to corrupted data. On denoising tasks with DRUNet, with an average budget of $4$ bits per weight, our method matches or improves state-of-the-art mixed-precision baselines under clean calibration, and is more robust to corrupted calibration, with PSNR gains of up to $7.5$ dB under the tested corruptions. Experiments show bit-allocation speed-ups from $28\\times$ to $2{,}570\\times$ over the studied baselines. For quantized diffusion models, our experiments show that a direct application of our framework also improves the state-of-the-art.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Mixed-precision quantization", "Probabilistic error bounds", "Calibration robustness" ], "forum_url": "https://openreview.net/forum?id=Fiuq2wls5v", "pdf_url": "https://openreview.net/pdf/f5ebaef144c0bb287da77a0ae48f97770380dd3e.pdf", "updated_at": "2026-10-05T20:43:20.295Z", "categories": [ "ptq", "diffusion", "theory" ], "matched_terms": [ "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HfH5Wq5gRi", "number": 41281, "title": "Rescuing Regrettable Rejections: Relaxed Acceptance for Speculative Decoding", "abstract": "Speculative decoding has become one of the most effective techniques for accelerating inference in large language models. Its defining property is that it is lossless: the output distribution is identical to that of the target model, so the speedup comes at no cost in output quality. In speculative decoding, a small draft model proposes several tokens, the larger target model verifies them in parallel, and decoding resumes from the first proposal that fails verification. Every draft token after the rejected one is discarded, and the more tokens are discarded, the smaller the speedup becomes. We examine these rejections and find that many of them are regrettable. A rejection is regrettable when accepting the draft token would not have affected the quality of the output. We propose a method that rescues such rejections without training any additional model or adapter. It targets the three types of regrettable rejection that we identify, surface form, preference, and order, with a whitelist of interchangeable tokens, mutual confirmation between the two models, and order-swap tables, respectively. Experiments on Qwen3-14B and DeepSeek-R1-Distill-Qwen-32B, each drafted by a smaller 4-bit quantized model of the same family, show that the method rescues 30–36% of the rejections of lossless speculative decoding at temperature 0 and 14–15% at temperature 0.6, which raises throughput by a further 8–13% and 3.5–6.5%, respectively, and lifts the speedup over the target alone to 2.1–2.4×, with no forward pass added and both models left untouched. Mean accuracy stays within 0.4 points of lossless decoding; the one clear loss, on a single benchmark, is traced to one rule.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "speculative decoding", "relaxed acceptance", "lossy speculative decoding", "LLM inference acceleration", "training-free", "quantized draft model", "verification", "rejection analysis" ], "forum_url": "https://openreview.net/forum?id=HfH5Wq5gRi", "pdf_url": "https://openreview.net/pdf/f6dbead830744500ed6fe9fc1a7fd2f0434b5eb1.pdf", "updated_at": "2026-10-05T20:43:24.749Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LIgMvRMUdA", "number": 41291, "title": "Mixed Precision Scaling Laws for Quantized Pretraining", "abstract": "Since scaling laws were introduced for modern LLMs \\citep{kaplan2020scaling}, several works have extended them to account for quantization, tracking the emerging low-precision data types supported on AI accelerators. We develop a more accurate scaling law for quantized pretraining, grounded in optimization theory and validated empirically. Our formulation separates the effects of quantization on model capacity and on optimization efficiency. Through convergence analysis, we further observe that both forward and backward quantization can introduce an optimization error that persists with continued training, which we model explicitly in our law. To design our experimental regimes used to fit and evaluate our scaling law, we first revisit existing 4-bit training recipes and decompose them into their constituent forward and backward pass quantization choices, isolating the components that most strongly affect loss. Our Qwen3 experiments under selected 4-bit regimes, together with published Llama-style experiments, show that our law predicts loss with up to 11× lower MSE than existing laws as model size and token count scale. Finally, we extend our law to layer-wise mixed precision, and identify the optimal quantized-pretraining recipe, i.e., model size, training tokens, and per-layer forward/backward precision, under a hardware-aware compute budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "pretraining", "quantized pretraining", "scaling laws" ], "forum_url": "https://openreview.net/forum?id=LIgMvRMUdA", "pdf_url": "https://openreview.net/pdf/d00e422b0d38c1437ddb8df0358cc911b01f9114.pdf", "updated_at": "2026-10-05T20:43:26.109Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "4-bit", "low-precision", "mixed precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AeaEhBu0iI", "number": 41608, "title": "The Undetected Damage of Quantization on Retrieval and How to Fix It", "abstract": "Quantization rounds weights to a lower bit-width, and it is often judged by accuracy loss. This criterion is incomplete. The same quantized model that preserves classification accuracy changes 20-46% of top-1 retrieval results, and the aggregate ranking metrics reflect only part of that change. The top-1 result is guaranteed to survive quantization only when the gap between the two highest scores, logits in classification and query-document similarities in retrieval, exceeds twice the largest rounding error. In classification, the loss function pushes the correct class apart from every other, ensuring this gap, but in retrieval nothing separates the top-1 item from the second. This gap can be measured without labels. Before deployment, it predicts which models will break, and at deployment time it tells, per query, whether the quantized answer still matches the one full precision would have returned. This holds for most classification inputs and a minority of retrieval queries. That gap motivates a different fix in each setting. In retrieval, spending bit-width on the layers most sensitive to it recovers three-fifths to three-quarters of an extra bit's benefit for half its cost. In classification, routing the few low-gap inputs to full precision recovers most of the lost accuracy at a fraction of the cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "retrieval" ], "forum_url": "https://openreview.net/forum?id=AeaEhBu0iI", "pdf_url": "https://openreview.net/pdf/22edbd56ce11edfb2ee446e880dc1575813bac24.pdf", "updated_at": "2026-10-05T20:44:28.797Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "bit-width", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hxPhcaNOo1", "number": 41839, "title": "A Unified Hessian-Guided Framework for Post-Training Quantization", "abstract": "Low-bit quantization enables efficient deployment of large language models by reducing memory footprint and computational cost. Achieving high accuracy at low precision requires selecting quantization ranges, linear transformations, and weight rounding decisions. Existing approaches typically optimize these components independently using local reconstruction objectives, which fail to account for the varying impact of quantization errors on downstream predictions. Direct optimization of task-level metrics would provide a more faithful objective, but is challenging due to the discontinuous nature of quantization operations and the high computational cost of evaluating their downstream effects.\nIn this work, we derive a Hessian-guided approximation of the task loss induced by joint weight and activation quantization. Based on a tractable Fisher approximation, we formulate a unified post-training quantization framework that enables the joint optimization of quantization ranges, linear transformations, and weight rounding under a common output-sensitivity objective. The resulting method remains practical for modern large language models and consistently outperforms state-of-the-art transformation and rounding methods across both channel-wise and block-wise quantization settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Compression", "Model Efficiency" ], "forum_url": "https://openreview.net/forum?id=hxPhcaNOo1", "pdf_url": "https://openreview.net/pdf/894dae01992a3895008dca218993543546cf1e17.pdf", "updated_at": "2026-10-05T20:45:09.645Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low precision", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "44hyvrJuKb", "number": 41895, "title": "NanoVoco: Upsampling-Aware Rank-Width Reallocation for Lightweight Vocoding", "abstract": "Deploying text-to-speech (TTS) systems on resource-constrained edge devices requires lightweight neural vocoders. Under extreme compression, reducing channel width degrades synthesis quality. Although depthwise-separable convolutions (DSCs) allow wider layers at comparable parameter budgets, replacing standard convolutions (StdCs) with DSCs degrades quality even further in our experiments. We observe that this replacement involves reallocating parameters from convolutional rank to channel width, and find that the benefits of this rank-width reallocation depend strongly on the temporal upsampling rule. Replacing zero-insertion (ZI) with sample-repetition (SR) upsampling substantially improves the benefits of rank-width reallocation. We observe this interaction across different discriminator designs, generator backbones, and parameter budgets, with the strongest effect under extreme compression. Based on this observation, we propose NanoVoco, with only 72.3K parameters and UTMOS-measured synthesis quality comparable to the evaluated Parallel WaveGAN model and higher than FreGrad. Integrated into a complete TTS pipeline, its quantized implementation enables real-time streaming speech synthesis on a microcontroller using only on-chip memory.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Neural vocoding; Speech synthesis; Efficient neural networks; On-device inference; Depthwise-Separable Convolution" ], "forum_url": "https://openreview.net/forum?id=44hyvrJuKb", "pdf_url": "https://openreview.net/pdf/bc9f9aad9a0b37188711058c948ca41e49346047.pdf", "updated_at": "2026-10-05T20:45:19.988Z", "categories": [ "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "WClSeTrDM0", "number": 41931, "title": "DART: Diffusion Acceleration via Rectified Temporal Quantization", "abstract": "Diffusion models achieve high generative quality but incur substantial computational cost from iterative denoising. Feature caching and low-bit quantization each reduce this cost, and combining them promises multiplicative speedups. Yet we find that directly stacking the two strategies is unstable: cache reuse introduces a temporal state drift along the denoising trajectory, while quantization distorts the denoising operator, and the two errors interact in a way that local feature distance fails to predict. We propose DART, a training-free joint cache--quantization framework that addresses two questions explicitly---when to recompute and how to use cached quantized states---under a unified trajectory-sensitivity view. DART first measures a relative update sensitivity proxy along the sampling trajectory and partitions the timesteps into segments whose sensitivity fluctuation is bounded; the partitioning, formulated as a constrained dynamic program, decides where caching is safe. Within each segment, DART rectifies the cached quantized states through dual-domain affine correction, which decouples cache-induced input shift from quantization-induced operator distortion, and shares a segment-wise quantizer that adapts to temporal heterogeneity. The rectification parameters are absorbed into adjacent linear and dequantization operators whenever the required affine conditions hold, yielding negligible deployment overhead. Across class-conditional generation on ImageNet-256 with LDM and DiT-XL/2 and text-to-video generation on HunyuanVideo, DART consistently improves the quality--efficiency frontier over caching-only, quantization-only, and joint cache--quantization baselines under matched-speed and matched bit-width settings. For instance, on LDM-4 with W8A8 quantization DART improves FID from 4.03 to 3.83 at 8.12x versus CacheQuant's 7.87x, and improves FID from 7.21 to 6.53 at the most aggressive 18.93x setting where directly stacking DeepCache and W8A8 collapses to FID 15.36, an 8.83 FID gap at the same speed. On DiT-XL/2 at 50 sampling steps DART further improves FID from 5.43 to 5.08 at 13.2x versus the 12.7x of the recent joint baseline Q&C, and on HunyuanVideo it lowers LPIPS by 23 from 0.189 to 0.145 and lifts VBench from 80.35 to 80.42 at 4.12x versus DPCache's 4.05x.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Computer vision" ], "forum_url": "https://openreview.net/forum?id=WClSeTrDM0", "pdf_url": "https://openreview.net/pdf/418e1df8223fbb7bf63768726ff4bb34d0728223.pdf", "updated_at": "2026-10-05T20:45:27.717Z", "categories": [ "hardware", "diffusion" ], "matched_terms": [ "bit-width", "low-bit", "quantization", "quantized", "quantizer", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ww8nQMWeiH", "number": 41948, "title": "LoopQ: Quantization for Looped Language Models", "abstract": "Looped language models (LoopLMs) increase effective computational depth by repeatedly applying shared Transformer blocks without increasing the parameter count. However, this reuse makes them particularly sensitive to post-training quantization (PTQ). We systematically study PTQ in LoopLMs and identify three sources of degradation: distribution shift across roles, state reuse across loop transitions, and recursive error accumulation. We propose LoopQ, a loop-aware PTQ framework that preserves a shared quantized backbone while introducing lightweight adaptations. LoopQ combines loop-aware activation scaling and selective transformations to accommodate changing hidden-state distributions, cross-loop state alignment to stabilize transitions, and trajectory-aware calibration to limit accumulated error. Under W4A4 quantization, LoopQ achieves a 67.8% average relative improvement in five-task mean accuracy across four backbones and an average relative perplexity reduction of at least 75.2% across eight backbone–dataset pairs, compared with the strongest static PTQ baseline for each comparison.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "post-training quantization", "looped language models", "low-bit inference" ], "forum_url": "https://openreview.net/forum?id=Ww8nQMWeiH", "pdf_url": "https://openreview.net/pdf/6cd005d768184d9e6e4cdc71cb3eeb0558d70d2a.pdf", "updated_at": "2026-10-05T20:45:28.070Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MHFRiNs87o", "number": 41969, "title": "SparkDiffusion: A Unified Sparse Warm-Up and Few-Step Distillation Framework for Mitigating the High-Sparsity Trap", "abstract": "Video diffusion transformers are central to visual generation, but inference is dominated by attention over long spatiotemporal sequences. While video attention's sparse-plus-low-rank structure suggests high sparsity is possible, existing models degrade beyond $\\sim90\\%$ sparsity even after step-local training converges. We identify this failure regime as the \\emph{high-sparsity trap}: step-local objectives leave terminal-visible velocity errors that accumulate during sampling without direct gradient signals to correct them. We propose \\method (\\methodfull), a unified framework whose central principle is to first adapt a dense video DiT into a high-sparsity coarse prior through sparse warm-up, and then correct its terminal trajectory through few-step trajectory-mixed distillation; fused FP8 kernels provide the deployment layer. Our idealized analysis isolates the mechanism and limits, while experiments validate the staged framework across multiple backbones, tasks, resolutions, and hardware platforms.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sparse attention", "few-step distillation", "single-GPU acceleration" ], "forum_url": "https://openreview.net/forum?id=MHFRiNs87o", "pdf_url": "https://openreview.net/pdf/5e21f1a996818c72c31e0e86e68126b2f5c646ed.pdf", "updated_at": "2026-10-05T20:45:34.866Z", "categories": [ "training", "hardware", "diffusion" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZqgPzC476U", "number": 41973, "title": "FlashLVSM: A Constant-Time View-Synthesis Transformer", "abstract": "Latent scene representations have emerged as a promising alternative for novel view synthesis, encoding multi-view observations into learned features without relying on predefined geometric primitives or rendering operations. Despite their promise, realizing them with modern neural architectures, such as transformers, often incurs substantial rendering cost. As the number and resolution of source views increase, every target token repeatedly attends to a growing set of source tokens through a deep neural renderer, causing the rendering cost to grow rapidly. We present FlashLVSM a constant-time view-synthesis model designed for scalable, highly efficient rendering from latent scene representations, which is carefully co-designed around four complementary principles, each addressing a distinct source of rendering cost. A richer, more 3D-expressive latent representation enables us to use a lightweight decoder without sacrificing rendering quality. Target-aware view selection restricts computation to source views geometrically relevant to the target, while target-aware token pruning further retrieves only the informative tokens within them, making rendering cost independent of the total number of input views. Finally, mixed-precision inference accelerates the remaining decoder computation.\nTogether, FlashLVSM renders at 142 FPS with 32 input views on a single consumer-grade GPU, and maintains roughly 140 FPS as the number of views increases to 256 while preserving rendering quality. This corresponds to speedups of nearly 12x and 95x, respectively, with 204x fewer FLOPs and 25x lower peak memory at 256 views. Despite this reduction in computation, FlashLVSM remains competitive in rendering quality, matching or exceeding the baseline from 64 input views onward. These results establish efficient latent scene representations as a compelling alternative to conventional explicit representations for efficient novel view synthesis.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Novel view synthesis", "Neural Rendering", "Token Pruning", "Quantization" ], "forum_url": "https://openreview.net/forum?id=ZqgPzC476U", "pdf_url": "https://openreview.net/pdf/e11547b50e7e91ece96ded7c25a2d1b769766ce4.pdf", "updated_at": "2026-10-05T20:45:33.062Z", "categories": [ "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "GXwXAz1F9u", "number": 41993, "title": "Temporal Structure in Contrastive EEG Representation Learning: A Controlled Study for Obstructive Sleep Apnea Detection", "abstract": "Contrastive self-supervised learning for physiological signals typically treats windows as independent apart from augmentation-defined positives, ignoring that windows from the same patient are temporally related. We introduce Temporal NT-Xent, which retains augmentation-defined positives and full-strength cross-patient negatives while continuously down-weighting the repulsive contribution of same-patient negatives according to their temporal distance.\n\nWe evaluate this objective for EEG-based obstructive sleep apnea (OSA) detection along three complementary axes. First, on a canonical patient-held-out split—with patients unseen during pretraining, validation, or test selection—Temporal NT-Xent produces a substantial downstream change relative to vanilla NT-Xent, raising AUROC from 0.5417 to 0.7667 and AUPRC from 0.4013 to 0.6673 (paired permutation p = 0.0455 for AUROC). Second, the same frozen 131K-parameter encoder transfers zero-shot to the independent SHHS-1 cohort without any apnea labels or target-domain fine-tuning, reaching AUROC = 0.724 (95% CI [0.641, 0.801]), compared with 0.718 for a single-channel source-to-target control. This probes cross-cohort transfer separately from the in-domain question addressed by the canonical split.\n\nThird, a five-seed robustness analysis on the source cohort shows that the magnitude of the in-domain downstream gain is not seed-robust (mean ΔAUROC = +0.0350 ± 0.1600, with A1 exceeding A0 in 2 of 5 seeds). We report this stochastic sensitivity as an empirical characterization of small-cohort EEG self-supervised learning rather than as a caveat to be minimized. Controlled comparisons against an explicit temporal-continuity objective, a SoftCLT-compatible soft instance-wise formulation, and three temporal kernels further show that the size and direction of the in-domain effect depend on how temporal information enters the objective. A representation-level diagnostic indicates that the canonical gain is not explained by increased effective rank.\n\nFinally, the compact encoder retains 0.9991 ± 0.0004 cosine similarity between FP32 and INT8 encoder outputs while achieving a 71% reduction in model size, demonstrating a practical consequence of the compact representation for resource-constrained physiological monitoring. Together, these results characterize patient-level temporal structure as a meaningful inductive bias whose in-domain downstream benefit is stochastic in small cohorts, while demonstrating measurable cross-cohort transfer of the learned representation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Self-Supervised Learning", "Contrastive Learning", "Temporal Contrastive Learning", "EEG Representation Learning", "Physiological Time Series", "Obstructive Sleep Apnea", "Cross-Cohort Transfer", "Model Quantization", "Edge Deployment" ], "forum_url": "https://openreview.net/forum?id=GXwXAz1F9u", "pdf_url": "https://openreview.net/pdf/7698efe995585ac6006dd796b9b8e7816dc3c806.pdf", "updated_at": "2026-10-05T20:45:36.830Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Doz8XZY42S", "number": 42097, "title": "EdgeRazor: A Lightweight Framework for Large Language Models via Mixed-Precision Quantization-Aware Distillation", "abstract": "Quantization has emerged as a mainstream approach for deploying Large Language Models (LLMs) on resource-constrained devices, yet reducing precision below 4-bit typically leads to severe performance degradation or prohibitive retraining costs. In this paper, we propose EdgeRazor, a lightweight framework for LLMs via Mixed-Precision Quantization-Aware Distillation. It contains three modules: Structural Quantization with Mixed Precision for fine-grained control over bit-widths, Layer-Adaptive Feature Distillation for dynamically selecting the most informative features for alignment, and Entropy-Aware KL Divergence for balancing forward and reverse KL to handle data from both human annotation and model distillation. With 1.88-bit weights, 8-bit activations, and 8-bit KV cache, EdgeRazor on Qwen3-0.6B outperforms the strongest 2-bit baseline by 11.40 points in average accuracy and the strongest 3-bit baseline by 4.38 points. In terms of training efficiency, EdgeRazor quantizes MobileLLM-350M with 4-10$\\times$ fewer training tokens than the leading quantization-aware training method. On the deployment side, EdgeRazor achieves consistently higher compression ratios than existing methods across all bit-widths. Exported to the TQ2\\_0 quantization type supported by llama.cpp, the 1.58-bit Qwen3-0.6B quantized by EdgeRazor reduces the model size from 1.11 GB to 0.19 GB and decodes 15.16$\\times$ faster than its 16-bit counterpart. Extensive experiments on the MobileLLM and Qwen families confirm the effectiveness and efficiency of EdgeRazor.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Model", "Lightweight", "Quantization-Aware Distillation" ], "forum_url": "https://openreview.net/forum?id=Doz8XZY42S", "pdf_url": "https://openreview.net/pdf/5ca0fa64d20d72fc253507a0e666cce1f416e7f0.pdf", "updated_at": "2026-10-05T20:45:51.095Z", "categories": [ "qat", "kv-cache", "hardware" ], "matched_terms": [ "1.58-bit", "2-bit", "3-bit", "4-bit", "8-bit", "mixed precision", "mixed-precision", "quantization", "quantized", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Y6SR8urjb1", "number": 42148, "title": "MatGPTQ: Efficient and Accurate Inference over Nested Quantized Models", "abstract": "Matryoshka Quantization (MatQuant), Any-Precision-LLM (AP) and AnyBCQ (AB) are recent quantization approaches showing that a single integer-quantized model can be served across multiple precisions. In this paradigm, lower-precision models are extracted from a higher-precision model by simply reading fewer bits of the weights. This enables a single checkpoint to cover a wide range of memory and latency budgets, but makes both quantization and efficient execution substantially harder. Existing methods rely on expensive quantization-aware training (QAT) or gradient-based post-training quantization (PTQ) rather than fast one-shot PTQ, and offer limited system support: dedicated kernels are either missing or restricted to single- or small-batch decoding.\nWe address these limitations with _Post-Training Matryoshka Quantization_ (MatGPTQ), an end-to-end pipeline for nested-model quantization and inference. MatGPTQ casts Matryoshka quantization as multi-precision error compensation, producing a single \"sliceable\" parent model jointly optimized for multiple target precisions in one pass over a small calibration set. We further refine the MatQuant representation so that an $r$-bit model reads exactly $r$ bits, and introduce the first dedicated inference kernels for this format, supporting batch sizes beyond one and integrated into vLLM. Across standard LLMs and benchmarks, MatGPTQ outperforms MatQuant while remaining competitive with AP and AB at the smallest checkpoint size, and our kernels achieve end-to-end speedups of up to 3.5$\\times$ over BF16 at the low-bit regime. Overall, MatGPTQ makes nested quantized models practical to serve from a single, compact checkpoint.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Nested LLMs", "inference", "serving", "gpu", "cuda", "quantization", "compression" ], "forum_url": "https://openreview.net/forum?id=Y6SR8urjb1", "pdf_url": "https://openreview.net/pdf/db57dd44ba1a35190faf2c917f52852bf50e3ec3.pdf", "updated_at": "2026-10-05T20:45:56.714Z", "categories": [ "ptq", "qat", "training", "hardware" ], "matched_terms": [ "low-bit", "ptq", "qat", "quantization", "quantization_", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3XvmQnizjd", "number": 42417, "title": "RETHINKING SUBSPACE SIMILARITY IN LOW-RANK ADAPTATION: WHAT GOVERNS LOCAL UPDATES?", "abstract": "Low-rank adaptation is usually understood as optimization under a low-rank constraint, but its update map is not a projection: it weights a gradient by the spectrum the factors were built from. We reframe it as optimization under a spectrum-weighted map, and identify spectral allocation—how the spectrum is distributed across the two factors—as the variable governing early training, not the rank or the selected subspace. We test this reframing with paired upstream updates, shared-gradient probes, and natural-magnitude component interventions. A modified initialization can preserve its selected subspaces almost exactly while changing the weighted update map appreciably—a divergence we call the subspace–update gap. The distinction reaches training: two factorizations of one weight with identical initial effective weight, identical selected subspaces, and a shared learning rate produce ten-step updates lying 77.5°–88.0° apart, so spectral weighting changes the direction of adaptation, not only its speed. The gap is not an artifact of one spectral choice: cross-subspace coupling reproduces the response under principal-component selection but not under minor-component selection, where changes within the selected block dominate. The criterion is measurable with the same interventions and applies to any modification of the reference weights, including quantization and pruning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "parameter-efficient fine-tuning", "low-rank adaptation", "spectral initialization", "spectral allocation", "spectrum-weighted updates", "subspace-update gap", "optimization geometry" ], "forum_url": "https://openreview.net/forum?id=3XvmQnizjd", "pdf_url": "https://openreview.net/pdf/f817c3ea232bf2a33674a4a9064d395fb143c832.pdf", "updated_at": "2026-10-05T20:46:36.641Z", "categories": [ "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UmfdN8rahl", "number": 42480, "title": "Which Language Pays for Quantisation? The Calibration Corpus under GPTQ", "abstract": "Open language models are evaluated at full precision, but people often run 4-bit copies quantised by third parties. We audit the 10 most-downloaded quantised releases of Qwen3-8B by the shift in the base model's confident answers: the median 4-bit release changes 2.6\\% in English and 6.7\\% to 13.1\\% in six lower-resource languages, and no model card documents the calibration seed. Given that prior work has shown that the calibration language impacts perplexity and accuracy, but that target-language calibration does not reliably restore downstream behaviour after pruning, we investigated whether the calibration corpus is a cause of these changes. We performed 4-bit GPTQ quantisation on Qwen3-8B, varying only the corpus, with 8 to 16 seeds per corpus under pre-registered criteria. On Belebele, averaged across three option orders, English calibration reduces accuracy by 4.0 to 5.8 points in six lower-resource languages; in contrast, target-language calibration maintains 1.9 to 3.3 of these points. Spanish calibration does not benefit Basque, confirming that the effect is language-specific. A 25\\% share of the target language achieves most of the reduction, as the Basque curve predicted in three other languages. At the calibration budget of the most-downloaded GPTQ release (eight times ours), a balanced seven-language corpus keeps 1.5 to 2.2 points across all six languages, English inside a registered margin. The effect holds in Llama-3.1-8B-Instruct and is smaller in Gemma-3-12B-it, which handles these languages better. We release the corpora and recipes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantisation", "calibration data", "GPTQ", "multilingual evaluation", "low-resource languages", "behavioural stability", "model documentation", "pre-registration" ], "forum_url": "https://openreview.net/forum?id=UmfdN8rahl", "pdf_url": "https://openreview.net/pdf/0164feee70c382314516c4310f46437008829a0b.pdf", "updated_at": "2026-10-05T20:46:49.711Z", "categories": [ "ptq" ], "matched_terms": [ "4-bit", "quantisation", "quantised" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "txWaZwg63N", "number": 42491, "title": "Exp-Reduced Two-Level FP8 Attention for Vision Transformers", "abstract": "As GPU compute throughput increases, softmax becomes an increasingly significant bottleneck in attention on modern GPUs, particularly those based on NVIDIA's Blackwell architecture. Existing approaches either sacrifice accuracy, as often observed with sigmoid- and ReLU-based softmax substitutes, or fail to deliver wall-clock speedups in practice. This work proposes the Exp-Reduced Two-Level (ERTL) function, a precision-agnostic alternative to softmax for vision transformers (ViTs). The ERTL function uses block statistics and shifted-ReLU weights for intra-block normalization while retaining exponential normalization across blocks. This design substantially reduces the number of exponential evaluations and maps naturally to tiled attention kernels with an appropriate block size. Furthermore, an FP8 attention kernel is developed by fusing the ERTL function with lazy output rescaling. Extensive experiments on ViTs show that ViTs using the ERTL function recover near-baseline accuracy with only lightweight fine-tuning. Moreover, the FP8 attention kernel using the ERTL function achieves up to a 1.55× speedup over BF16 exact-softmax attention and increases throughput by up to 245.5 TFLOPS compared with FP8 exact-softmax attention.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Efficient attention", "softmax approximation", "vision transformers", "FP8 attention", "GPU kernel optimization" ], "forum_url": "https://openreview.net/forum?id=txWaZwg63N", "pdf_url": "https://openreview.net/pdf/5fc57c9caf95fad32923a850eee06140cd2c31ea.pdf", "updated_at": "2026-10-05T20:46:52.475Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "OBDZfSiQXp", "number": 42561, "title": "STRIDE: An Efficient SNN Training-to-Deployment Framework on Real-World Vision Tasks", "abstract": "Spiking Neural Networks (SNNs) have attracted increasing attention in recent years for their potential for high-efficiency and low-energy computation. The community largely agrees that realizing this potential depends on improvements in SNN training, including low-bit computing and sparse excitation, and adaptation to dedicated hardware platforms. However, evidence connecting SNN training to practical deployment benefits remains fragmented: algorithmic studies often use synaptic-operation-based analytical energy estimates, while hardware-focused studies often provide limited validation of deployed accuracy on practical tasks. In this paper, we bridge this gap with an efficient \\textbf{S}NN \\textbf{TR}a\\textbf{I}ning-to-\\textbf{DE}ployment (STRIDE) framework. STRIDE combines low-bit weights with Integer-LIF neurons to represent activations as compact integer codes, which are unfolded into binary temporal spikes for native, event-driven execution on dedicated SNN hardware. During training, we apply quantization-aware distillation to optimize low-bit weights under task and teacher supervision without backpropagation through time. During deployment, we develop native 1-bit spike $\\times$ 4-bit weight execution on a jointly developed 28-nm SNN accelerator, skipping redundant spike-weight computations when later spikes match the first-timestep reference. This work evaluates the effectiveness of STRIDE on three real-world vision tasks, achieving 61.95\\% mIoU on UDD6 segmentation, 31.71\\%/16.04\\% mAP$_{50}$/mAP$_{50:95}$ on UDD detection, and 67.88\\% top-1 accuracy on ImageNet classification. It outperforms the strongest evaluated compact-code baselines by 3.82, 1.05/0.03, and 1.28 percentage points, respectively, while the evaluated BPTT baseline requires $4.59\\times$, $4.66\\times$, and $3.03\\times$ as much peak training memory. We also evaluate hardware energy consumption by deploying one frozen STRIDE-trained SpikingLETNet checkpoint across the accelerator, GPU, and CPU. Command-block measurements yield an accelerator energy of 2.15\\,mJ per inference, 97.64\\% and 99.96\\% lower than on the evaluated INT8 GPU and CPU, respectively. These results also show that hardware-based energy measurements are far more accurate and reliable than conventional operation-count estimates, which capture some relative energy trends but show significant differences across architectures. Our work provides a feasible SNN implementation for real-world vision tasks and opens the door to developing the training-to-deployment collaboration scheme of SNNs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "spiking neural networks", "quantization-aware distillation", "software–hardware co-design", "SNN accelerator", "event-driven computation", "energy efficiency", "hardware-based energy evaluation" ], "forum_url": "https://openreview.net/forum?id=OBDZfSiQXp", "pdf_url": "https://openreview.net/pdf/ad528888e9e71e114bf7923226bd04f8ebbe1ade.pdf", "updated_at": "2026-10-05T20:47:02.447Z", "categories": [ "qat", "hardware" ], "matched_terms": [ "1-bit", "4-bit", "binary", "int8", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "oNfc4SjU9k", "number": 42698, "title": "Q-PAT: Quantization-Aware Pruning-Aware Tuning for Low-Bit Large Language Models", "abstract": "Structured pruning and low-bit quantization are two key techniques for reducing the deployment cost of large language models. However, structural decisions made during pruning may misalign with the quantization errors introduced later, potentially degrading model performance under low-bit precision. In this work, we propose Q-PAT, a quantization-aware pruning-aware tuning method tailored for low-bit deployment. Building on the Pruning-Aware Tuning (PAT) framework, Q-PAT constructs a channel-wise quantization-sensitivity proxy based on the INT4 fake-quantization error of low-rank weights. The resulting sensitivity signal is incorporated into a Mask-Quant regularization term through a stop-gradient mechanism, thereby constraining the optimization of trainable structural masks at the loss level without directly injecting quantization perturbations into the main forward pass. This design enables mask optimization to account for heterogeneous low-bit quantization sensitivity across channels during task adaptation, better coordinating task adaptation, structural compression, and quantization error. Experimental results show that Q-PAT preserves strong downstream task performance under low-bit quantization and achieves improved INT4 performance on representative Gemma-7B settings. In particular, Q-PAT improves the average score across 14 downstream tasks from 56.69% to 57.38% and increases MMLU accuracy from 24.04% to 33.64%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language models", "structured pruning", "low-bit quantization", "pruning-aware tuning", "quantization sensitivity", "model compression" ], "forum_url": "https://openreview.net/forum?id=oNfc4SjU9k", "pdf_url": "https://openreview.net/pdf/09c798348094eed8837d3aef7b3a0fd33b159e32.pdf", "updated_at": "2026-10-05T20:47:20.745Z", "categories": [ "qat", "training", "hardware" ], "matched_terms": [ "int4", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fbvBYNTB0C", "number": 42709, "title": "NyS-Binary: A Feature-Free, Algebraic Framework for Binary Node Representation Learning", "abstract": "Graph neural networks deliver strong node representations, but their high-dimensional floating-point embeddings are costly to store and slow to serve on resource-constrained devices. Binary graph hashing addresses this by mapping each node to a compact bitcode, yet existing methods force an uncomfortable choice: random-walk and sketching approaches such as node2binary and NodeSketch are slow to construct and discard global structure, while shallow linear hashes such as NodeSig trade away accuracy on harder graphs. We present **NyS-Binary**, a feature-free, algebraic, and sparsely supervised hashing framework that operates purely on graph topology without requiring any node attributes, combines a randomized Nystrom-inspired low-rank sketch of a higher-order graph transition operator with a leakage-safe label-diffusion channel, fuses the two via a convex blend, and quantizes the result with an adaptive per-column median rule.\n\nAcross ten node-classification benchmarks and diverse downstream classifiers, **NyS-Binary** consistently attains the highest average accuracy among all binary hashing methods, substantially outperforming the strongest feature-free binary baselines on both standard neural probes and spiking neuromorphic classifiers, while producing compact bitcodes in fractions of a second, over an order of magnitude faster than NodeSketch and up to three orders of magnitude faster than learned iterative baselines. Crucially, because the landmark structural channel provides a strong topological foundation independent of supervision, **NyS-Binary** operates in a semi-supervised regime and maintains competitive accuracy even under extreme label scarcity.\n\nBecause the representations are strictly binary, they also natively support neuromorphic classifiers built on gradient-free learning and surrogate-gradient spiking networks, which compute via binary spike accumulation rather than floating-point arithmetic. We report where the method is strong, where it is not (notably extreme heterophily), and view it as evidence that principled algebraic sketching remains a competitive, scalable alternative to learned binarization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Binary representation learning", "Gradient free graph learning", "Lightweight Node Embeddings" ], "forum_url": "https://openreview.net/forum?id=fbvBYNTB0C", "pdf_url": "https://openreview.net/pdf/6831795df4416761f6cf3ab756c172c694ba7b2b.pdf", "updated_at": "2026-10-05T20:47:21.848Z", "categories": [ "training", "diffusion" ], "matched_terms": [ "binarization", "binary", "quantizes" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "34MhzBZmew", "number": 42756, "title": "When Quantization Preserves Generation but Degrades Verification: A Frozen-Candidate Study across Model Scale", "abstract": "Quantized language models are commonly evaluated as generators, yet the same\nmodels are increasingly used to judge, rerank, and select candidate answers. These\nroles need not respond identically to reduced precision. We study this distinc-\ntion with a frozen-candidate protocol: for each model, eight candidates are gen-\nerated once in FP16 and then held byte-identical while the verifier is evaluated\nat FP16 and several quantized precisions. Correctness is assigned by determinis-\ntic semantic checking, so verifier comparisons are not confounded by a changing\ncandidate pool. On 300 MATH-500 problems, Qwen2.5-7B-Instruct provides a\nclear example. HQQ 8-bit changes greedy generation accuracy from 0.467 to\n0.473 (∆ = +0.006), while global verifier AUROC falls from 0.825 to 0.544 and\nwithin-problem ranking from 0.698 to 0.411. Best-of-eight selection falls from\n0.500 to 0.430, below the exact random-selection rate of 0.455. The same 7B\ndegradation is observed with bitsandbytes int8 and NF4. In contrast, the HQQ 8-\nbit result does not show a comparable verification drop at Qwen2.5-3B, -14B, or\n-32B; the 3B result is additionally sensitive to the quantizer, with NF4 producing\na separate degradation. A Llama-3.1-8B experiment is not a strong cross-family\ntest because its FP16 verifier is already near chance (AUROC = 0.471). Output-\nlevel analysis of the 7B model shows an approximately 8×reduction in correct-\nversus-incorrect margin separation and a higher decision-flip rate among correct\ncandidates than incorrect candidates. The evidence therefore supports a narrower\nclaim than a universal “quantization breaks verification” hypothesis: generation\naccuracy alone does not certify preservation of verifier ranking, and verification\nshould be evaluated directly for the model, precision, interface, and candidate set\nused downstream.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models (LLMs)", "Model Quantization", "Low-Bit Inference", "Verification", "Verifier Models", "Correctness Verification", "Calibration" ], "forum_url": "https://openreview.net/forum?id=34MhzBZmew", "pdf_url": "https://openreview.net/pdf/37769b1cc556530624a7028a25fb06285de4f78d.pdf", "updated_at": "2026-10-05T20:47:27.123Z", "categories": [ "ptq" ], "matched_terms": [ "8-bit", "int8", "low-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "H32vOZB00e", "number": 42764, "title": "Query-Agnostic Visual KV Compression at Matched Memory Budgets", "abstract": "Prefix caching allows vision-language models (VLMs) to reuse the key-value (KV) cache of an image across questions, which requires compression to be completed before the question is known. Most existing visual KV compression methods are query-aware and therefore cannot be applied in this scenario; for the query-agnostic setting, there is a lack of a systematic and fair comparison under the same memory allocation. In this paper, we investigate query-agnostic visual KV compression under matched memory budgets, across two VLM families and four visual question answering benchmarks. On our fixed paired evaluation manifests, retaining deterministically spaced evicted tokens has higher mean scores than merged tokens in all eight model--task comparisons and matches or exceeds tiled moment summaries in all comparisons. Besides, pre-registered falsification tests and an exploratory mixture study do not establish a summary\nadvantage. Meanwhile, we provide an attention-mass bound on the maximum contribution of an eviction summary and analyze three mechanisms that limit summary methods: dilution of individual value outliers, instability of the dispersion correction, and per-block RoPE interactions in merged keys. Finally, quantization behavior depends on the grouping axis and on metadata cost. Together, these controlled comparisons clarify how byte allocation, retained-token fidelity, and task structure affect query-agnostic visual KV compression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Model", "KV cache" ], "forum_url": "https://openreview.net/forum?id=H32vOZB00e", "pdf_url": "https://openreview.net/pdf/797de50726f7f21cc3d16b545c10d177d46b65e4.pdf", "updated_at": "2026-10-05T20:47:28.537Z", "categories": [ "kv-cache", "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zUiNYteKzS", "number": 42855, "title": "On the Cross-Branch Correlation of Quantization Error in Classifier-Free Guidance", "abstract": "Classifier-free guidance runs a network twice per step, with and without the condition, and extrapolates along the difference of the two predictions. Quantized models are judged per branch, yet the guided error equals $\\Ec+s(\\Ec-\\Eu)+s(1+s)V$, where $\\Ec$ and $\\Eu$ are the branch errors, $s$ is the guidance scale, and $V$, the mean squared difference of the two errors, depends on a cross term that no per-branch metric sees. In six inference stacks, $75.8\\%$ of the activation quantizers give the branches different scales or random numbers, and on these, $19\\%$ to $28\\%$ of the values held by both branches receive different codes. We propose PairRound, which lets the branches share what decides their codes: the scale, and the random numbers where rounding is stochastic. Codes then differ only where values differ, and a quadratic in $s$ states when the coarser grid that sharing gives one branch pays off, a condition our measurements meet. At matched bits and seeds, the shared scale lowers the trajectory distance to the BF16 sampler by $60\\%$ to $68\\%$ on eleven image, video and language backbones, improves FID, VBench or perplexity on every one, and composes with eight published methods and improves each of them.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "bit width" ], "forum_url": "https://openreview.net/forum?id=zUiNYteKzS", "pdf_url": "https://openreview.net/pdf/40e0af78340e2f4ab3a34f29757397f535070f62.pdf", "updated_at": "2026-10-05T20:47:41.532Z", "categories": [ "ptq" ], "matched_terms": [ "bit width", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "T0dqmS8KZ3", "number": 42908, "title": "TurboQuantConv: Channel-Isolated Incoherence Processing for 3D Convolutions in Video Generative AI", "abstract": "Three-dimensional Convolutional Variational Autoencoders (3D-VAEs) serve as the fundamental compression backbone of modern video generative models, yet their deep 3D convolutions (Conv3D) impose prohibitive computational and memory burdens. Pushing post-training quantization (PTQ) down to 4-bit weights and activations (W4A4) typically causes substantial quality degradation due to non-stationary spatiotemporal activation bursts. In this paper, we propose TurboQuantConv, a channel-isolated incoherence processing and quantization framework for deep multi-dimensional convolutions. By harmonizing outlier suppression with convolutional receptive field geometry, TurboQuantConv effectively suppresses spatiotemporal bursts and achieves high-fidelity quantization without requiring fine-tuning or retraining. Under W4A4 compression, TurboQuantConv outperforms prior state-of-the-art PTQ methods by ~4 to 8 dB PSNR across 3D-VAE benchmarks. In downstream 81-frame video generation with the 5B Wan~2.2 Diffusion Transformer, TurboQuantConv achieves the lowest FVD $152.6$ while closely preserving unquantized visual dynamics, avoiding the severe motion freeze and chromatic collapse from which existing PTQ methods suffer. These results establish a practical foundation for efficient edge deployment of video generative models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Video Generative AI", "3D Convolution", "Hardware Optimization", "Edge Deployment", "W4A4" ], "forum_url": "https://openreview.net/forum?id=T0dqmS8KZ3", "pdf_url": "https://openreview.net/pdf/6ca6d515e382436a262558515d900007d37eeabf.pdf", "updated_at": "2026-10-05T20:47:49.391Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "4-bit", "ptq", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HOyAPHHe3q", "number": 42933, "title": "What Does Key Quantization Actually Change in a Delta Rule? The Step Size.", "abstract": "Delta-rule linear recurrent neural networks such as DeltaNet, Gated DeltaNet and DeltaProduct update a matrix state with one least-mean-squares step per key. This step shrinks the residual it corrects only if the effective step size $\\beta s$ lies in $(0,2)$, where $\\beta$ is the step size and $s$ is the squared norm of the key. The models keep $s=1$ with a key normalizer, but a quantizer placed after the normalizer breaks this: with 4-bit integer (INT4) keys, $s$ deviates from one by up to $9.4\\%$. In models with step sizes below $1$, every update then has a relative step-size error of $|s-1|$. In models with step sizes up to $2$, $3.4\\%$ of the steps have $\\beta s>2$, where the update flips and enlarges the residual. If $\\beta$ and $s$ are independent, the share of such steps depends only on the tail of $\\beta$ near $2$ and the distribution of $s$, and measuring the two separately predicts the observed shares within $7\\%$ for INT4 keys. NormFix divides the step size by the squared norm of the key that the kernel receives, $\\eta=\\beta/s$. This restores the trained residual factor $1-\\beta$ for every $s$ and makes the propagation of state errors nonexpansive for any quantizer. With 4-bit weights and INT4 keys, NormFix recovers $57\\%$ of the long-context state-tracking loss of a Gated DeltaNet with negative eigenvalues and $51\\%$ of the loss of a standard one, at less than $4\\%$ extra kernel time. On the negative-eigenvalue model it is better than clipping, skipping, key renormalization and 8-bit keys on every state task and on retrieval. Its gain holds across six models and three scale axes, and where the key is normalized after the quantizer, so that $s=1$, it changes no score by more than $0.01$ points.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization" ], "forum_url": "https://openreview.net/forum?id=HOyAPHHe3q", "pdf_url": "https://openreview.net/pdf/9d2571d7cb6bc9cf1a8ddb19f6abcffe558fbd1b.pdf", "updated_at": "2026-10-05T20:47:51.231Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "8-bit", "int4", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "deHOhhDuWg", "number": 42947, "title": "MAQuant: Modality-Aware Quantization for Omnimodal Large Language Models", "abstract": "Omnimodal large language model (OLLM) deliver strong capabilities but incursubstantial memory and computation costs. Low-bit quantization is a practical remedy, yet existing methods mainly target Large Language Model(LLM) or Large\nVision Language Model(LVLM) with modality-agnostic objectives, failing to handle heterogeneous activation distributions, outlier-dominated degradation, and cross-modal error coupling in OLLM. We propose MAQuant, a modality-aware post-training quantization framework that combines Modality-Specific Learnable Transformation (MSL), Exponent-weighted Activation Penalty (EAP), and\nCross-modal Error Exchange (CME) to learn modality-specific quantization friendly spaces, protect salient activations, and share channel-wise sensitivity across modalities. EAP and CME are calibration-only objectives and add no inference time loss computation. Experiments on diverse benchmarks and low-bit settings show that MAQuant consistently outperforms strong LLM and LVLM quantization baselines. On Qwen2.5-Omni-7B, MAQuant reduces storage by 68.6% under W4A4 while retaining 97.4% of full-precision average performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization ; omnimodel" ], "forum_url": "https://openreview.net/forum?id=deHOhhDuWg", "pdf_url": "https://openreview.net/pdf/9689183be943a5612917558cfd474e787f275336.pdf", "updated_at": "2026-10-05T20:47:53.617Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "low-bit", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8B6NUKPU4m", "number": 43047, "title": "QALF: Deployment Error as a Regularizer for Quantization-Aware Low-Rank Factorization", "abstract": "Low-rank factorization reduces a network's parameter count, while quantization reduces the bits used to store each parameter. Combining them usually means fitting the factors in floating point and rounding them afterwards. Factors that reconstruct a layer accurately can therefore perform poorly at the precision used for storage. We propose QALF, which accounts for quantization during fitting. Under an independent-noise model, we derive each factor's expected contribution to layer-output error in closed form. Adding these contributions to the input-weighted reconstruction loss gives a first-order approximation to the error after quantization. The quantization model determines the regularizer's form, and the target bit-width sets its weight, with no separately tuned trade-off coefficient. On ResNet-18/ImageNet, QALF outperforms dense GPTQ at comparable stored sizes up to the dense 3-bit budget. At 4-bit weights and 8-bit activations, it improves accuracy over the input-weighted reconstruction baseline by 1.8–2.8 percentage points with GPTQ and 4.1–4.7 with round-to-nearest, the rounding rule the objective models. It also exceeds CP-ADMM on its native grid, the closest quantization-aware factorization method, by 0.95–4.43 points at 4 bits. QALF remains more accurate than a tuned component-norm penalty combined with an optimized gauge, despite having higher kernel reconstruction error. The same derivation extends to matrix factorizations: on MLP down-projections in six language models from 410M to 8B parameters, QALF lowers 3-bit perplexity by 1.0–5.4 against SVD-LLM; its factors lose 71–87% less to rounding. For normalized CP factors, we establish conditions under which the regularizer prevents diverging components and the objective attains a minimum.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-rank factorization", "CP decomposition", "quantization-aware optimization", "post-training quantization", "neural network compression", "low-precision inference", "tensor decomposition" ], "forum_url": "https://openreview.net/forum?id=8B6NUKPU4m", "pdf_url": "https://openreview.net/pdf/ba9bcf317670bbaf88da997874d806b90869d555.pdf", "updated_at": "2026-10-05T20:48:05.670Z", "categories": [ "ptq", "qat", "hardware" ], "matched_terms": [ "3-bit", "4-bit", "8-bit", "bit-width", "low-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eKVnluuscF", "number": 43060, "title": "Learning to Execute Tree Search over Continuous Thoughts", "abstract": "Tree-structured reasoning lets a language model explore several intermediate thoughts, but the branching, scoring and backtracking are usually controlled by an algorithm outside the model. We ask whether a Transformer can execute that search internally, as one autoregressive chain of thought over the continuous thoughts it generates and feeds back into its own context, and whether standard training delivers this ability. We study search on depth-$H$ trees with the search history represented as a sequence of vector-valued tokens in the Transformer's context. The Transformer predicts the next token to select, backtrack or terminate, while a fixed front end supplies candidates with noisy rewards. For the search itself, we construct a binary instance in which single-path success has an upper bound that decreases polynomially with depth, while depth-first search succeeds with probability $1-H^{-3}$, so backtracking is essential to the search we execute. On the representation side, we construct a Transformer with one attention layer that executes depth-first search exactly by predicting successive tokens, up to the first reward gap below a threshold, for every fixed branching factor. On the learning side, we give two answers to whether a Transformer trained by empirical risk minimization (ERM) executes it. First, because the correct next action is determined by the prefix, on-policy imitation learning with ERM learns a single executor without quantizing thoughts, with failure $\\widetilde O(D/N)$ for $D$ proportional to the trainable parameter count and $N$ independent instances per round. However, this approach requires repeated collection and labeling of the learner's own rollouts. Second, under teacher forcing with a quantized front end, the guarantee depends on what is written back. With raw feedback, our guarantee can require exponentially small training errors as the horizon grows. Our counterexample class contains near-optimal models that leave the teacher's trace. Writing the quantized state back restores a constant tolerance, under which ERM over the executor's sparse template reaches the goal within $5H$ selections with probability $1-H^{-3}-\\widetilde O(H/N)$. Together, our results show that a Transformer can execute search over its own continuous thoughts and clarify how guarantees of learned execution differ between on-policy imitation learning and teacher forcing, and depend on how predictions are fed back into context.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Transformers", "chain-of-thought", "tree search", "continuous thoughts", "latent reasoning", "expressivity", "learning theory", "imitation learning", "teacher forcing", "backtracking" ], "forum_url": "https://openreview.net/forum?id=eKVnluuscF", "pdf_url": "https://openreview.net/pdf/70df433a3044eec97c5760de4c8feb41dc223201.pdf", "updated_at": "2026-10-05T20:48:07.408Z", "categories": [ "reasoning" ], "matched_terms": [ "binary", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "N06RTPfhCQ", "number": 43088, "title": "Representation Matters in 2-Bit QAT: Finite Barnes–Wall Representations for Large Language Models", "abstract": "Quantization-aware training (QAT) is widely used to recover the quality of large language models (LLMs) at ultra-low precision, yet most existing methods focus on improving how a given low-bit representation is optimized. We study a complementary question: what representation should QAT optimize in the first place? We introduce BW-QAT, a representation-first framework for 2-bit weight-only QAT based on a finite 16-dimensional Barnes–Wall representation. BW-QAT encodes every 16 weights with a fixed 32-bit payload, calibrates the gain alphabet directly against the finite reconstruction space, and initializes trainable proxies from compensated pre-quantization targets. This initialization preserves the exact initial hard-quantized model while retaining the residual information needed for effective discrete updates. The resulting representation can be trained with hard quantized forward passes and a standard straight-through estimator. Across Qwen and Llama models, BW-QAT consistently outperforms matched lattice, scalar-QAT, and vector-QAT baselines in the low-data regime. On Qwen3-1.7B, it improves six-task average accuracy from 52.17% to 58.30% over a matched LC-QAT adaptation and reduces WikiText-2 perplexity from 18.90 to 14.02 across three training seeds. These results show that representation quality is a first-order factor in data-limited 2-bit QAT.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "quantization-aware training", "2-bit quantization", "Barnes–Wall lattice", "model compression" ], "forum_url": "https://openreview.net/forum?id=N06RTPfhCQ", "pdf_url": "https://openreview.net/pdf/c4506b08ffa6862467c1fff0c021ed943465b82e.pdf", "updated_at": "2026-10-05T20:48:10.793Z", "categories": [ "ptq", "qat" ], "matched_terms": [ "2-bit", "low precision", "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Y0lMnZf9XH", "number": 43090, "title": "Shared or Independent? A Controlled Study of Transform Search in LLM Quantization", "abstract": "Low-bit quantization reduces model storage but introduces approximation error. Rotations can reduce this error by changing the coordinates in which weights are quantized. In attention layers, query, key, and value projections can share an input rotation or use independent rotations. Although independent rotations offer greater flexibility, we show that temporarily sharing Q and K during fitting improves the final independent models on OPT and Llama under matched controls. We study this effect in QKV-only ternary post-training quantization. Our construction copies shared rotations into independent variables without changing the quantized model at the transition, then continues fitting independently. This enables a comparison with independent fitting from the outset while holding the final transform family fixed. We evaluate OPT-1.3B, Llama-3.2-3B, Qwen3-4B-Base, and Ouro-1.4B with matched initializations and restart rules. Temporary sharing lowers negative log-likelihood (NLL) on C4 and WikiText-2 in 19 of 24 model–initialization–corpus comparisons, with 18 improvements remaining significant after correction for multiple comparisons. Additional fitting-budget controls retain the gains on OPT and Llama. Across HellaSwag, ARC-Challenge, and PIQA, normalized accuracy improves in 27 of 36 comparisons, including ten significant improvements after correction. On Llama-3.2-3B, HellaSwag accuracy increases by 1.20 and 0.89 percentage points at the two random initializations. Benefits depend on initialization, sharing duration, and evaluation corpus. Lower reconstruction error can still accompany higher NLL even after accounting for model inputs. Interventions on the fitted models further identify the contributions of Q, K, and layer depth to these performance differences. These results establish temporary sharing as a useful fitting strategy in the studied setting and show why reconstruction error, language-model loss, and downstream accuracy must be distinguished when evaluating quantized models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model quantization", "post-training quantization", "ternary weight quantization", "rotation-based quantization", "QKV projections", "transform sharing", "optimization path dependence", "reconstruction error" ], "forum_url": "https://openreview.net/forum?id=Y0lMnZf9XH", "pdf_url": "https://openreview.net/pdf/99f107a35287fae37278eb1ba8bb18ab055b5a76.pdf", "updated_at": "2026-10-05T20:48:11.076Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "quantization", "quantized", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LjikYVIVEu", "number": 43127, "title": "AdaPivot: An Adaptive Core-Tail Representation for Low-Bit LLM Quantization", "abstract": "Post-training quantization (PTQ) of Large Language Models (LLMs) often suffers substantial accuracy degradation because fixed numerical formats poorly accommodate the pronounced outliers in weights and activations. Through extensive profiling across diverse models, we observe a distributional structure: most values concentrate in a compact core, while the outliers, which account for only a small fraction of the values, form a sparse tail, typically separated from the core by a sharp magnitude transition and a nearly unoccupied interval. Building on this observation, we introduce AdaPivot, an adaptive low-bit numerical format that partitions the representable range into distinct core and tail regions. AdaPivot independently adapts the representation granularity of the two regions, together with their boundary and dynamic ranges, to match the underlying tensor distribution within a compact parametric form. We further formulate format fitting as a joint discrete--continuous reconstruction problem and develop a tailored solver that efficiently co-optimizes discrete level selection and continuous format parameters. An integer-constrained variant further reduces metadata overhead while preserving the adaptive structure. Experiments on various models from 7B to 70B parameters show that AdaPivot consistently outperforms representative PTQ baselines. Under W4A4 quantization, it improves average zero-shot accuracy by up to 7.5% over the state-of-the-art baseline, while W4A8 performance closely approaches the FP16 baselines. Our code is \navailable at https://anonymous.4open.science/r/AdaPivot-A.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large Language Models; Post-Training Quantization; Low-Bit Quantization; Adaptive Quantization; Non-uniform Quantization" ], "forum_url": "https://openreview.net/forum?id=LjikYVIVEu", "pdf_url": "https://openreview.net/pdf/ff9aa42aa8f5ea906d2c0ced4fe56533b8810e0c.pdf", "updated_at": "2026-10-05T20:48:16.405Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "w4a4", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "tivPFe5IBr", "number": 43144, "title": "When Does Adaptive LoRA Rank Allocation Help? Signal Reliability and Resource Accounting", "abstract": "Adaptive LoRA redistributes a limited rank budget, yet final accuracy alone does not isolate the benefit of rank placement. Allocation scores depend on existing ranks, while rank transfers can change parameter cost and model state. We analyze the scores and test transfers at fixed active parameter count. For a score based on loss progress, a sign change shared by all modules can reverse their rankings even when the ordering of score magnitudes persists. We separate the immediate loss change caused by a transfer from the loss change during continued training. Experiments cover eight GLUE tasks and quantized Llama commonsense workloads. On commonsense, Spearman correlations between consecutive measurements average 0.944 for score magnitudes and -0.155 for signed scores. At selected RTE checkpoints, online allocation activates 16.4% more parameters than uniform at equal total rank. At a shared RTE checkpoint, one of three module pairs yields lower final loss after moving rank to the higher-scored module, but a less favorable loss change over 20 updates than the reverse transfer. Synthetic controls show that concentrating rank resolves local capacity shortages, while repeated adjustment helps after demand shifts.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "parameter-efficient fine-tuning", "low-rank adaptation", "adaptive rank allocation", "signal reliability", "resource accounting" ], "forum_url": "https://openreview.net/forum?id=tivPFe5IBr", "pdf_url": "https://openreview.net/pdf/3896a6e90fb371bd785f7ff77576b66c3017eef8.pdf", "updated_at": "2026-10-05T20:48:18.935Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xv3yebiu6S", "number": 43163, "title": "InterQ-RL: Interaction-Aware Correction of Execution Mismatch in Quantized Agentic RL", "abstract": "Multi-turn agent reinforcement learning (RL) relies on long closed-loop rollouts, making generation and KV-cache costs a major bottleneck. Quantized rollout reduces these costs, but its low-precision executor also serves as the behavior policy and can induce execution mismatch with the learner. Existing quantized-RL methods are designed primarily for single-turn generation. They align the learner forward with quantized rollout or correct the current-token rollout–learner ratio, without modeling the policy history that produces each agent state. Even after quantization alignment, residual rollout–learner mismatch accumulates along the state-reaching policy history, so a current action may look locally aligned despite a mismatched history. We introduce InterQ-RL (Interaction-aware Quantized Reinforcement Learning) to correct this post-alignment mismatch. It applies bounded execution-prefix correction over the state-reaching policy history, allocates correction strength across action segments by effective support under a fixed displacement budget, and requires no additional model forward pass. We evaluate dense and mixture-of-experts (MoE) backbones from 4B to 30B on ALFWorld, WebShop, and ScienceWorld, with all quantized arms sharing the same rollout with 8-bit weights, activations, and KV cache (W8A8-KV8). Across four environment–backbone settings, InterQ-RL achieves the strongest quantized endpoint result, matching BF16 to within 0.7 points and exceeding it in two settings. It matches or exceeds the learning speed of alignment alone while avoiding its late performance regression.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning", "Quantization", "Agentic RL" ], "forum_url": "https://openreview.net/forum?id=xv3yebiu6S", "pdf_url": "https://openreview.net/pdf/8e7429b344439d5fd33833fed76e35e7d0e1e428.pdf", "updated_at": "2026-10-05T20:48:21.997Z", "categories": [ "kv-cache" ], "matched_terms": [ "8-bit", "low-precision", "quantization", "quantized", "w8a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ICN47EWhiX", "number": 43240, "title": "RetryPrec: Precision as a Retry-Time Resource under an External Verifier", "abstract": "A quantized language model is often served next to its BF16 original, and a task with a verifier lets a failed attempt be tried again. After a low-precision failure, should the next attempt stay in low precision or switch? The usual rule ranks the two copies by average accuracy and repeats the winner. However, a failure is a selection event: the problems it leaves open lean toward those on which the quantized copy is weak, and at our reference setting a BF16 attempt solves more problems per second from the second failure on. We show that exchanging one low-precision attempt for a high-precision one changes the solve rate by the mean gap times the fraction of problems still open, plus a covariance between staying unsolved and the precision gap. We propose RetryPrec, which runs one block of low-precision attempts and then one block of high-precision attempts, with the block lengths chosen on held-out problems by an unbiased estimate of the solve rate. One such schedule is optimal among the policies that react only to failures, and we derive how failures move the ratio of the two hazards. On twelve settings at matched wall clock RetryPrec reaches $78.78$ points, against $75.95$ for the fixed low-precision schedule and $77.60$ for the best of sixteen other strategies. A test registered before the experiments shows that mixing precisions adds little once each problem uses its better one, so most of the attainable gain lies in choosing a precision for each problem, which failures partly reveal. When the budget also charges the memory held by both checkpoints, the schedule stays in low precision.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "calibration" ], "forum_url": "https://openreview.net/forum?id=ICN47EWhiX", "pdf_url": "https://openreview.net/pdf/19b53c2a69268b61606b3178fde3a5de16944bcd.pdf", "updated_at": "2026-10-05T20:48:35.743Z", "categories": [ "ptq" ], "matched_terms": [ "low precision", "low-precision", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "v2s01DQuOl", "number": 43242, "title": "Megrez: Reconstructing Intermediate Supervision for Ternary Looped Transformers", "abstract": "Ternary looped transformers reuse the same quantized weights across recurrent calls, coupling the approximation of intermediate teacher states. We present Megrez, a calibration framework that adapts the supervision targets to the current student. We reconstruct teacher targets to minimize alignment error while requiring the teacher's unchanged native readout to return its original output on those targets. Each token's normalized teacher states are jointly transformed across recurrent calls by a single orthogonal map. The constrained alignment has an exact low-dimensional solution. Megrez alternates target reconstruction with shared ternary weight updates, accommodating gated and terminal readouts without changing the inference architecture. Across Ouro-1.4B, Parcae-1.3B and LoopFormer, reconstruction reduces squared target error by 20.2--55.5\\% on average per window at fixed student weights. On Ouro-1.4B, the complete framework lowers C4 perplexity from 45.44 to 29.54 and raises HellaSwag accuracy from 39.83\\% to 44.92\\% compared with the restricted TernaryLLM adaptation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "ternary quantization", "looped transformers" ], "forum_url": "https://openreview.net/forum?id=v2s01DQuOl", "pdf_url": "https://openreview.net/pdf/05c652df92e0ee56a780e3a5d51cf9b34888cbc3.pdf", "updated_at": "2026-10-05T20:48:35.879Z", "categories": [ "ptq" ], "matched_terms": [ "quantization", "quantized", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "A9scGpnqv6", "number": 43327, "title": "HeteroQuant: Resolving Expert Heterogeneity in MoE Quantization via Targeted Calibration", "abstract": "Mixture-of-Experts (MoE) architectures substantially expand model parameter capacity while maintaining a limited computational cost per token through sparse activation and dynamic routing mechanisms. However, during practical deployment, the complete set of expert parameters usually needs to remain in memory, and the resulting large storage overhead restricts the application of MoE models in resource-constrained environments. Post-training quantization (PTQ) is an important technique for reducing the storage and inference costs of large language models. Nevertheless, due to the significant expert heterogeneity, domain preferences, and imbalanced activation patterns in MoE models, existing quantization methods often fail to sufficiently calibrate critical experts, leading to severe performance degradation. To address these challenges, we propose HeteroQuant, an effective, training-free post-training quantization framework designed to resolve expert heterogeneity through targeted calibration. Specifically, HeteroQuant achieves this by first leveraging the DSPy framework to construct a highly diverse multi-domain candidate pool, thereby comprehensively activating all expert positions to collect their global routing statistics. Based on these statistics, we implement a core two-step selection mechanism to systematically optimize the calibration data. In the first stage, we perform AlphaQ-guided structural-targeted scoring to prioritize candidate samples that precisely target and activate structurally critical experts. In the second stage, we execute activation-redundancy-aware filtering to eliminate samples with overlapping routing footprints, thereby preventing calibration waste while maximizing activation diversity. Experimental results demonstrate that under the W4A4 quantization setting of the Qwen1.5-MoE-A2.7B model, HeteroQuant achieves an average improvement of 5.09 points over the current SOTA method across eight downstream tasks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "Post-Training Quantization", "Large Language Models", "Model Compression", "Calibration", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=A9scGpnqv6", "pdf_url": "https://openreview.net/pdf/d351a891a270101857cb0df2b57ee88447034a61.pdf", "updated_at": "2026-10-05T20:48:51.473Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "ptq", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SI6gzRExOb", "number": 43354, "title": "ShamAN-Q: Shampoo Augmented NanoQuant for Sub-1-bit LLM Weights", "abstract": "We introduce ShamAN-Q, a sub-1-bit post-training quantization method that\nextends NanoQuant (Chong et al., 2026) by replacing its diagonal reconstruction geometry with a\ndense but tractable curvature metric, using a form and computation\nparadigm popularized by the Shampoo optimizer (Gupta et al., 2018). For each linear weight,\nShamAN-Q fits a Kronecker product to the empirical Fisher Information matrix of a small\ncalibration set by Kullback--Leibler minimization and uses the result to form a Mahalanobis\nreconstruction loss. The continuous ADMM updates from NanoQuant become solutions to\nSylvester equations, while its discrete projection and deployment format remain unchanged. Because the curvature is local to\na given set of weights, ShamAN-Q re-measures the input curvature statistic for each layer\nimmediately before layer factorization and periodically refreshes all statistics\non the partially quantized model. ShamAN-Q also redistributes the uniform\nrank from NanoQuant across layers at the same total number of bits. On Qwen3-Base,\nShamAN-Q lowers WikiText-2 perplexity at $\\approx$1\\,bpw from 27.56 to 22.96\n(0.6B), 19.21 to 16.72 (1.7B), and 14.29 to 13.80 (4B) while increasing zero-shot reasoning task performance on the Eleuther LM Evaluation Harness. On 0.6B, ShamAN-Q at $\\approx$0.8\\,bpw matches the published perplexity of NanoQuant at $\\approx$1.0\\,bpw.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Optimization", "Geometry", "LLM", "Hardware" ], "forum_url": "https://openreview.net/forum?id=SI6gzRExOb", "pdf_url": "https://openreview.net/pdf/953a859079d4e071f3c8e3f9199f58b88d8e4e63.pdf", "updated_at": "2026-10-05T20:48:55.999Z", "categories": [ "ptq", "training", "hardware", "reasoning" ], "matched_terms": [ "1-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Y6fASRnvyU", "number": 43362, "title": "RADeMuon: Provable Resource-Aware Quantized Decentralized Muon via Finite-Step Newton--Schulz Orthogonalization", "abstract": "Distributed Muon optimizer enables matrix-aware optimization across multiple devices, but its full-precision local training and model exchange impose substantial resource costs, particularly when resource availability changes throughout decentralized training. We propose RADeMuon, a resource-aware quantized framework for decentralized Muon that supports precision configurations varying with available resources. RADeMuon performs local training with quantized parameters, gradients, and momentum states, applies finite-step Newton--Schulz (NS) orthogonalization to the quantized momentum, and directly communicates the resulting quantized parameter for decentralized aggregation. Theoretically, we trace quantization errors from their accumulation in the momentum state to their transformation by finite-step NS and subsequent propagation through decentralized aggregation. This analysis establishes a convergence guarantee whose dominant stationarity term is $\\mathcal{O}((nKT)^{-1/4})$ under vanishing-error conditions. Experiments on image classification and language-model pretraining show that RADeMuon reduces local computation and communication costs by up to 58.34\\% and 58.67\\%, respectively, relative to full-precision De-EF21-Muon while maintaining comparable model quality.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Decentralized learning", "Muon", "Quantized Training", "Newton-Schulz Iteration", "Convergence Analysis" ], "forum_url": "https://openreview.net/forum?id=Y6fASRnvyU", "pdf_url": "https://openreview.net/pdf/d09fb6758029331efbf4f035d4874faf71290a79.pdf", "updated_at": "2026-10-05T20:48:57.297Z", "categories": [ "training", "theory" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vy1szamWxJ", "number": 43402, "title": "Information Limits and Output Geometry for Quantized Attention", "abstract": "A certificate can reject a quantized attention output even when its actual error is small. We characterize how retained probability and value information determines the tightest attainable error bound. Our exact scalar output envelope combines range, mean, and variance constraints: support bounds clip the classical variance-only optimizer, and two-point laws attain every regime. Joint constraints can certify budgets that separate moment bounds cannot. A distinct construction gives a $\\Theta(n)$ separation from the output conversion of the optimal total-variation (TV) bound. We also derive attainable TV envelopes for selected exact probabilities and lower bounds, with saturation and stopping criteria. Outward moments and a sharp correction for perturbed-distribution value error connect the envelope to quantized attention. On a fixed GPT-2 replay, an exact CPU reference certifies 357 of 1,008 case/budget pairs, versus 249 for midpoint-centered TV and 352 for coordinatewise mixing of variance-only and range/mean bounds. Independent numerical enclosures validate every certified pair. These comparisons distinguish the benefit of value information from the additional gain of joint constraints.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantized attention", "KV cache", "certification", "uncertainty quantification", "output bounds" ], "forum_url": "https://openreview.net/forum?id=vy1szamWxJ", "pdf_url": "https://openreview.net/pdf/498d3ebfce00896ef6c1b74da2103361414ae0fa.pdf", "updated_at": "2026-10-05T20:49:02.740Z", "categories": [ "training", "kv-cache", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3tiANWm9Iv", "number": 43426, "title": "Certified Stopping for Fixed-Codebook Quantization", "abstract": "When pairwise refinement stops in fixed-codebook quantization, how much joint improvement remains possible? We develop a certified stopping framework that separates conditional search, certificate optimization, and relaxation strength, connecting each source of uncertainty to the computation that can reduce it. A budgeted pair oracle maintains a feasible incumbent and a valid conditional lower bound after any number of finite branch scores. A globally valid curvature split extends these bounds to the coupled quadratic. Optimizing signed block curvature and the anchor recovers a block-moment relaxation when $H\\succ0$ and every block codebook has full affine span. Its certified optimization interval decomposes exactly into restricted-master and finite-separation terms. On eight exactly enumerable six-coordinate restrictions of captured GPT-2 reconstruction objectives, one pair-fixed point remains above the finite optimum, while four objectives retain strictly positive block-moment relaxation gaps after certificate optimization. Across all eight stored rational objectives, the optimized certificate is tighter than continuous-box and fixed-curvature controls. The hierarchy distinguishes when to score more candidates, improve the conic state, or change the relaxation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM quantization", "fixed codebooks", "discrete optimization", "optimality certificates", "semidefinite relaxation", "pair search" ], "forum_url": "https://openreview.net/forum?id=3tiANWm9Iv", "pdf_url": "https://openreview.net/pdf/f03ac6ecfcb9d3ce62b1cf2cf942dad246d20343.pdf", "updated_at": "2026-10-05T20:49:06.391Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "z4Yqhv9oj2", "number": 43539, "title": "JAM: Joint Allocation and Matching for Sample Sharing under Emerging Environments", "abstract": "In edge deployments without cloud connectivity, clients encounter different new environments at different times, such as distribution shifts caused by lighting or terrain. Directing information about environments already encountered by some clients to those that have not yet encountered them can support early adaptation. With limited communication, however, each sender must decide whom to send to, how much to send, and what to send. Two challenges arise: class labels do not readily reveal environmental differences within a class, and the same sample has different value to different clients. We propose JAM (Joint Allocation and Matching), which measures differences using features from the common initial model rather than class labels and jointly decides whom to send to, how much to send, and what to send within the budget. Multi-seed experiments on three datasets show that, compared with random sharing, separate allocation and selection, sample selection, and prototype exchange, JAM can significantly improve deployment performance in new environments on datasets with identifiable within-class environmental deviations, without sacrificing average old-environment loss. Quantizing feature averages in JAM control metadata to 4-bit reduces communication while retaining the main adaptation gains.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sample sharing and selection", "collaborative edge learning", "communication-efficient learning", "federated learning" ], "forum_url": "https://openreview.net/forum?id=z4Yqhv9oj2", "pdf_url": "https://openreview.net/pdf/c0850adc00f3cc9386683e9f76c716d54ab7c932.pdf", "updated_at": "2026-10-05T20:49:22.332Z", "categories": [ "hardware" ], "matched_terms": [ "4-bit", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RLrEeIx53i", "number": 43587, "title": "LRSR: Hardware-Friendly Approximation of Per-Element Quantization via Low-Rank Scale Reconstruction", "abstract": "Post-training quantization is a crucial component for deploying Large Language Models (LLMs) in resource-constrained environments. While fine-grained methods (e.g., per-group) offer high accuracy, they heavily rely on hardware-specific implementations and suffer from memory-bound computations, limiting their applicability on standard matrix multiplication accelerators. Conversely, hardware-friendly quantizations (e.g., per-channel) experience severe accuracy degradation at low bit-widths. To address these limitations, we propose Low-Rank Scale Reconstruction (LRSR), a novel quantization method that approximates per-element scale matrices using an upper-bounding rank-1 decomposition, enabling mathematically correct reconstruction via general matrix multiplication (GEMM). Furthermore, we introduce several synergetic clusterization techniques to significantly boost accuracy and allow optimal hardware load balancing. Experimental evaluations on Qwen3 models demonstrate that at INT4, LRSR reduces KL-Divergence by up to 31\\% compared to the per-channel baseline, while utilizing 29$\\times$ less metadata memory compared to per-group (g=128) quantization. This establishes a new Pareto frontier for near-zero overhead, hardware-friendly and calibration-free post-training quantizations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantizations", "model compressions", "hardware-friendly computations", "low-rank approximations" ], "forum_url": "https://openreview.net/forum?id=RLrEeIx53i", "pdf_url": "https://openreview.net/pdf/b5f5d24236110584fd047ad880a78000b25d2c65.pdf", "updated_at": "2026-10-05T20:49:28.904Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "int4", "low bit", "quantization", "quantizations" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "BDngj01uEj", "number": 43595, "title": "Quantized Keys Steal Attention: Bias Correction for KV-Cache Compression in Video Diffusion", "abstract": "Chunk-wise autoregressive video diffusion models rely on a KV cache of previously generated chunks to avoid redundant computation, and this cache quickly becomes a memory bottleneck as videos grow longer. Methods that quantize the KV cache to low bitwidths reduce memory pressure but degrade video quality. We show that a key driver of this degradation is a systematic bias in attention weights: due to the convexity of the exponential in softmax attention, quantization noise inflates the contribution of cached keys, a phenomenon we call the Jensen bias. This effect causes quantized keys to steal attention mass from the unquantized current chunk. We derive a per-attention-score correction that removes this bias in expectation, computed on the fly from the quantization step sizes of the cached keys and the query norm. With a second-order Taylor approximation, the correction adds negligible computational overhead and needs no additional memory alongside the cache. Evaluated on MAGI-1, SkyReels-V2, and HY-WorldPlay at INT2 quantization, our correction recovers most of the quality lost to aggressive quantization, reaching near-BF16 video quality, and can match or exceed INT4 quantization while using 50% less memory. Because it removes the residual Jensen bias rather than the quantization error itself, the correction complements upstream quantizers such as QuantVideoGen.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "autoregressive video diffusion", "efficient multimodal AI systems", "memory-efficient inference", "long-form video generation", "attention bias correction", "post-training quantization", "video generation", "softmax attention" ], "forum_url": "https://openreview.net/forum?id=BDngj01uEj", "pdf_url": "https://openreview.net/pdf/7c3dbff4c3c1cf74231e7ca0d108a682aca08a79.pdf", "updated_at": "2026-10-05T20:49:31.629Z", "categories": [ "ptq", "kv-cache", "diffusion", "multimodal" ], "matched_terms": [ "int2", "int4", "quantization", "quantize", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MuusqOAmAG", "number": 43600, "title": "Post-Training Adaptation of LLMs to Stochastic Weight Perturbations", "abstract": "Analog in-memory computing (AIMC) performs matrix–vector multiplications inside memory arrays that store weights, avoiding costly weight movement during LLM inference. The price it pays is stochastic noise on every weight that is fundamentally different from the deterministic rounding error in digital quantization. Hardware-aware training can mitigate the resulting accuracy loss, but requires a full training pipeline with escalating costs as model sizes grow. Post-training quantization (PTQ) requires only a calibration set, but fails to address stochastic weight noise. We propose a post-training robustification (PTR) method that adapts weights using a noise-aware objective: expected layer-output distortion, decomposed into weight-error and noise-variance terms. Weight-Range Adaptation for Robustness to Perturbations (WARP) combines per-channel input scaling and weight clamping with exactly-solved thresholds for each weight group. WARP+ adds GPTQ-style error redistribution on top. We evaluate 0.6B to 32B models of the Qwen3 family in an AIMC simulator with a hardware-derived programming-noise model. WARP+ improves accuracy across all model sizes and completes in $<$3 hours on 4 A100 GPUs for the 32B model. On Qwen3-14B, WARP+ recovers 40\\% of the accuracy loss, more than 2× the recovery achieved by PTQ, and reaches within 3.5\\% points of noise-free accuracy. On Qwen3-32B, the remaining gap is $<1.5$\\% points. Using WARP+ for PTQ by applying rounding error as weight noise yields a Qwen3-4B model with 3-bit weights at $<$9.2\\% accuracy loss, less than half of OmniQuant (19.3\\%) and GPTQ (22.4\\%).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "analog in-memory computing", "post-training quantization", "post-training robustification", "stochastic weight noise", "noise-aware optimization", "large language models", "hardware-aware deployment", "weight clipping", "model compression", "efficient inference" ], "forum_url": "https://openreview.net/forum?id=MuusqOAmAG", "pdf_url": "https://openreview.net/pdf/42655f9cc5aaa3b4f9047c97b670002488508c55.pdf", "updated_at": "2026-10-05T20:49:31.826Z", "categories": [ "ptq", "hardware", "theory" ], "matched_terms": [ "3-bit", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wTyHpsny23", "number": 43896, "title": "$R^2$-VLM: Detecting and Correcting Vision-Language Hallucination via Active Representational Perturbation and Residual Spectral Grounding", "abstract": "Vision-Language Models (VLMs) can confidently generate text that contradicts the visual input, a failure commonly referred to as $perception$ $hallucination$. Existing mitigation methods predominantly analyze unperturbed forward passes, limiting their ability to assess the stability of internal representations under perturbations. We propose $R^2$-VLM, a training-free, single-pass framework for hallucination detection and correction based on active representational perturbation. Our detection module applies structured latent perturbations to the hidden states of a frozen VLM across visual, linguistic, and cross-modal subspaces. These perturbations yield a calibrated hallucination risk score and reveal that joint cross-modal reliability provides a stronger indicator of hallucination than unimodal signals alone. To correct detected failures, we introduce $Residual Spectral Grounding (RSG)$, which restores visual evidence by additively injecting a per-sample spectral decomposition of the visual representation into the decoder hidden states. We further establish theoretical guarantees showing that the risk score is bounded within $[0,1]$, the adaptive threshold converges, and RSG strictly improves representational energy stability. Empirically, across Qwen2.5-VL (3B/7B) and LLaVA-1.5 (7B) on ColorBench, POPE, and CableDetection, our risk score exhibits a strong correlation with ground-truth correctness, achieving a point-biserial correlation of up to $r=0.54$ and Hedges' effect size of up to $|g|=1.50$ ($p<0.001$). Furthermore, RSG nearly doubles cross-modal representational alignment from $0.35$ to $0.66$ and achieves relative average accuracy improvements of $+2.4%$ on color-attribute tasks and $+1.84%$ on CableDetection, while avoiding the performance degradation observed with competing subtractive correction baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Models (VLMs)", "Multimodal Large Language Models (MLLMs)", "perception hallucination", "hallucination detection", "hallucination mitigation", "hallucination correction", "R²-VLM", "Residual Spectral Grounding (RSG)", "active representational perturbation", "representational resilience", "hidden-state analysis", "latent-space perturbation", "structured latent perturbation", "visual subspace", "linguistic subspace", "cross-modal subspace", "cross-modal reliability", "visual grounding", "cross-modal grounding", "representational fragility", "structural stability", "hallucination risk score", "adaptive risk scoring", "adaptive flagging", "dynamic thresholds", "sliding-window calibration", "LOW/MEDIUM/HIGH risk tiers", "single-pass inference", "training-free inference", "test-time mitigation", "decoder hidden states", "visual encoder representations", "language decoder representations", "spectral decomposition", "spectral grounding", "high-frequency visual information", "Fourier transform", "discrete Fourier transform (DFT)", "frequency-domain analysis", "content centroid", "complexity signal", "spectral grounding token", "residual injection", "additive correction", "residual correction", "skip connections", "representation alignment", "Linear CKA", "cross-modal alignment", "Frobenius energy ratio", "directional centroid alignment", "KL divergence", "output divergence", "energy stability", "subspace erasure", "SVD", "singular vectors", "rank collapse", "token permutation", "token ablation", "semantic drift", "anisotropic spectral scaling", "adversarial robustness", "joint manifold projection", "energy equalisation", "perturbation severity", "fine-grained attribute grounding", "visual-language grounding", "uncertainty", "confidence margin", "calibration", "statistical significance", "point-biserial correlation", "Spearman correlation", "Hedges’ g", "ColorBench", "POPE", "CableDetection", "Qwen2.5-VL", "Qwen-VL", "LLaVA-1.5", "inference efficiency", "quantized inference", "safety-critical VLMs", "medical imaging", "industrial inspection", "autonomous driving", "hallucination robustness", "representation engineering" ], "forum_url": "https://openreview.net/forum?id=wTyHpsny23", "pdf_url": "https://openreview.net/pdf/ac8b9fc8371aaae81b24c6b6e2e9cc45da6ee755.pdf", "updated_at": "2026-10-05T20:50:32.887Z", "categories": [ "ptq", "multimodal", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "U3tUQMooJO", "number": 43950, "title": "Non-vacuous Generalization Bounds for Reinforcement Learning with Verifiable Rewards", "abstract": "While reinforcement learning with verifiable rewards (RLVR) is widely used to improve the reasoning capabilities of large language models (LLMs), the generalizability of the resulting models remains poorly understood. In this work, we establish the first non-vacuous generalization bounds for parameter-efficient RLVR fine-tuning at the billion-parameter scale. Our approach adapts PAC-Bayes compression bounds to this setting and addresses the inherent stochasticity of token generation by applying the Gumbel-max reparameterization trick. To obtain non-vacuous bounds, we develop Progressive RLVR, a pipeline combining RLVR with on-policy distillation, TinyLoRA, and quantization, which achieves high empirical accuracy and high compressibility. Progressive RLVR retains 84–97% performance of standard LoRA fine-tuning while producing learned weight updates that are 14,796× more compressible. We show that this framework yields non-vacuous generalization bounds in four domains: math, programming, general-knowledge reasoning, and text-to-SQL. Our bounds exceed the accuracy of the base model by 9–51% and lie within 6–11% of the accuracy of the fine-tuned models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "generalization bound", "reinforcement learning with verifiable rewards" ], "forum_url": "https://openreview.net/forum?id=U3tUQMooJO", "pdf_url": "https://openreview.net/pdf/77e3397680d055c0f53802e6279b0b65dbc1357d.pdf", "updated_at": "2026-10-05T20:50:40.085Z", "categories": [ "reasoning", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zJAORUCoh5", "number": 43970, "title": "Activation Denoising: A Robustness View on Parallel vs Sequential LLM Quantization", "abstract": "Post-training quantization is a powerful tool for compressing large language models. The most scalable methods quantize every layer in parallel, but quantization errors then compound through the residual stream, as no layer corrects for the errors of the layers before it. Sequential quantization accounts for this error compounding by re-calibrating each layer on the already-quantized outputs of its predecessors, yielding stronger results but at the cost of a serial schedule that becomes a bottleneck at scale. As a solution, we propose parallel quantization with activation denoising, which recovers much of the sequential benefit while keeping quantization fully parallel. Rather than re-calibrating layer-by-layer, we take a robustness perspective and model the upstream error as noise, regularizing to be robust to it through a preprocessing step followed by metric-weighted rounding. Applied at every layer, this regularization forms a depth-compounding smoothness penalty that dampens how strongly quantization errors amplify through the model. Unlike orthogonal rotations commonly used in quantization, which must preserve the model's function, we multiply the weights by a more general linear transformation. We find that the two are complementary and their effects compound. Empirically, our robustness regularization recovers a significant part of sequential quantization's benefit in a single parallel pass, at a fraction of its cost. Overall, by treating compounding quantization errors as a robustness problem, we offer a principled foundation for more efficient and accurate LLM quantization at scale.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "robustness regularization", "large language models" ], "forum_url": "https://openreview.net/forum?id=zJAORUCoh5", "pdf_url": "https://openreview.net/pdf/72ff018c277371e2f97afd8e34dfcc23f48fc9ea.pdf", "updated_at": "2026-10-05T20:50:43.400Z", "categories": [ "ptq", "diffusion", "theory" ], "matched_terms": [ "quantization", "quantize", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "t7pXlmF518", "number": 44044, "title": "Layer Collapse in Diffusion Language Models", "abstract": "Diffusion language models (DLMs) have recently emerged as competitive alternatives to autoregressive (AR) language models, yet the differences in their activation dynamics remain poorly understood. We characterise these dynamics in LLaDA-8B and identify a striking layer-collapse property: early layers exhibit highly similar, collapsed activation patterns, and a single large super-outlier channel dominates the activations at every token position in layers 1–24 (of 32). Despite its apparent redundancy, we have found that this outlier is critical: pruning it causes the model's outputs to degrade into repetitive random token loops. Paradoxically, besides this outlier, layers in LLaDA contain more redundant representations, where redundancy is particularly pronounced in earlier layers. This pattern is the reverse of what is commonly observed in AR language models, where deeper layers tend to become more redundant due to undertraining, as measured by representation similarity. Our analysis further indicates that layer collapse in DLMs is not driven by undertraining. Rather, it appears to arise from overtraining: a dominant outlier becomes an indispensable carrier of information, while the remaining representations collapse into redundant structure. These observations have strong practical implications, as we verify through controlled pre-training experiments. First, DLMs are surprisingly robust to compression: performance of LLaDA under 3-bit GPTQ quantization drops by only 1.8% on GSM8K, whereas Llama-3.1-8B under the same settings drops by 64.7%. Second, optimal non-uniform sparsity allocation reverses between the two model families: under an average budget of 50% sparsity, allocating more sparsity to early layers in LLaDA yields +8.4% over the reverse strategy, while for Llama the same early-layer-sparse allocation incurs -8.4%. Our findings reveal that the DLM training objective fundamentally reshapes layer dynamics relative to AR models, with direct consequences for how such models should be compressed and deployed. Our code will be made available upon publication.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Deep learning", "Diffusion", "Curse of depth", "learning dynamics" ], "forum_url": "https://openreview.net/forum?id=t7pXlmF518", "pdf_url": "https://openreview.net/pdf/52691efa24a256cb5888117975590038e8018c54.pdf", "updated_at": "2026-10-05T20:50:54.612Z", "categories": [ "diffusion" ], "matched_terms": [ "3-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "D0VShthCdd", "number": 44076, "title": "AdaSC-Muon: Adaptive Subspace Compression for Communication-Efficient Muon Training", "abstract": "Muon has emerged as a promising optimizer for large-scale neural network training, but communication and orthogonalization costs can limit its efficiency in distributed settings.\nEF21-Muon by Gruntkowska et al. reduces communication by compressing momentum differences. Its best-performing compressor still requires workers to transmit local singular bases at every iteration, so the worker-to-server communication cost scales with the original matrix dimensions.\nWe introduce adaptive subspace-compressed Muon (AdaSC-Muon), which uses shared subspaces for both momentum compression and Muon computation. The key insight is that aggregating projected momentum before orthogonalization enables a global update to be computed directly in a reduced matrix space. Workers reuse the shared bases and quantize projected momentum differences with error feedback. Periodic full-momentum corrections update the subspaces as momentum changes during training. We establish convergence guarantees and communication complexity bounds for AdaSC-Muon.\nLanguage model pretraining experiments with NanoGPT on FineWeb show that, compared with EF21-Muon, AdaSC-Muon and its diagonal variant reduce worker-to-server communication cost by up to 47.8\\% and reach the target validation loss in up to 53.8\\% less wall-clock time.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Muon", "compression", "error feedback", "distributed training" ], "forum_url": "https://openreview.net/forum?id=D0VShthCdd", "pdf_url": "https://openreview.net/pdf/8660a5bc55d14b43087c8ae11b1cba4f1f768242.pdf", "updated_at": "2026-10-05T20:50:59.531Z", "categories": [ "training", "theory" ], "matched_terms": [ "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "91Q3RaHTYz", "number": 44078, "title": "OPQ: On-Policy Quantization for Low-Bit Large Language Models", "abstract": "Low-bit quantization reduces the memory required to deploy large language models. However, learning on fixed sequences can miss the contexts reached by a quantized model's own predictions. We propose on-policy quantization (OPQ), a framework that learns directly on student-generated trajectories. On-policy quantization faces two main challenges: correcting output discrepancies introduced by low precision and maintaining consistent internal attention representations. To address the first, Dynamic Kullback--Leibler (KL) divergence adjusts output supervision using floating and quantized views of the student. To address the second, key/value (K/V) alignment supervises their attention memory. Our theoretical analysis explains the corrective gradients of the output objective and how K/V discrepancies affect attention outputs. With 2-bit and 3-bit weights, OPQ achieves the highest average task accuracy among the evaluated low-bit methods across two model families. Its gains are largest at 2-bit precision, reaching 9.49 percentage points over the strongest baseline. Ablations further show how student trajectories, Dynamic KL and K/V alignment contribute to capability preservation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Bit Quantization", "Post-Training Quantization", "Large Language Models" ], "forum_url": "https://openreview.net/forum?id=91Q3RaHTYz", "pdf_url": "https://openreview.net/pdf/211e1b4e3bf23bb09e6116b57e846cfefd718ff3.pdf", "updated_at": "2026-10-05T20:51:00.020Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "2-bit", "3-bit", "low precision", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AXq5Lx2rOG", "number": 44125, "title": "ACU-RL: Adaptive Correction and Update Control for Low-Precision LLM Reinforcement Learning", "abstract": "Low-precision rollout can reduce the dominant generation cost of reinforcement learning for large language models, but it also creates a policy mismatch between the quantized behavior policy and the BF16 learner. Existing truncated importance sampling (TIS) mitigates this mismatch by reweighting sampled tokens, yet it can strongly attenuate tokens favored by the quantized rollout policy. We find that relaxing this lower-side attenuation improves early learning but can amplify late-stage policy drift, revealing that sample correction and actor stabilization must be controlled jointly. To address this trade-off, we propose Adaptive Correction and Update Control for RL (ACU-RL), which couples Adaptive Two-Sided Truncated Importance Sampling (AT-TIS) with risk-gated actor-update control. AT-TIS adapts the correction strength while retaining more of the gradient contributions from tokens favored by the quantized rollout policy. Complementing this sample-side correction, risk-gated actor-update control increases drift regularization when persistent excess mismatch coincides with degradation signals and reduces the policy-gradient multiplier when the mismatch becomes more severe. ACU-RL requires no additional sampling or model forward passes. We evaluate ACU-RL on Qwen2.5-Omni-3B and Qwen3-8B-Instruct across six multimodal and four mathematical reasoning benchmarks, respectively. It achieves average accuracies of 45.26% and 60.82%, compared with 44.72% and 58.47% for BF16 RL. NVFP4 rollout further provides a $1.17\\times$ end-to-end training speedup.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning", "Low-Precision Rollouts", "Rollout–Training Mismatch", "Adaptive Importance Sampling", "Training Stability" ], "forum_url": "https://openreview.net/forum?id=AXq5Lx2rOG", "pdf_url": "https://openreview.net/pdf/e50f89a61e36c038e2d511058f3c38392e38d763.pdf", "updated_at": "2026-10-05T20:51:07.913Z", "categories": [ "training", "reasoning", "multimodal" ], "matched_terms": [ "low-precision", "nvfp4", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jdbTopkRnn", "number": 44145, "title": "BitLadder: Bit-Level Selective Loading of the KV Cache for Efficient Long-Context LLM Inference", "abstract": "Long-context large language models (LLMs) repeatedly load a growing key-value\n(KV) cache during autoregressive decoding, making attention memory-bandwidth\nbound. Existing methods reduce this traffic through query-aware sparse attention\nor KV-cache quantization, but neither aligns loading with current token importance: sparse attention uses a coarse keep-or-skip mask that cannot differentiate retained contributions and may discard the aggregate contribution of long-tail\ntokens, whereas quantization uses static bit-width assignments that cannot track\nquery-dependent importance. Our analysis shows that token importance is graded\nand changes across decoding steps. We present BitLadder, an algorithm-system\nco-designed framework for bit-level selective loading that dynamically adapts\nloaded KV-cache bit widths to token importance at each decoding step. To support\ndynamic bit widths without storing multiple quantized copies or performing per-step requantization, we propose truncation-consistent nested quantization (TCNQuant), which encodes the KV-cache once and derives lower-bit representations\nby truncating the same code. To translate fewer loaded bits into practical speedup,\nwe further design PlaneWeave, a memory layout that stores this nested representation so that a GPU kernel loads only the bits each token needs and supports efficient mixed-precision execution. We evaluate BitLadder on long-context understanding (LongBenchV2, MRCR) and reasoning (GPQA, MATH500, AIME25)\nbenchmarks. At matched accuracy, BitLadder achieves 3.0× and 2.3× speedups\nover full attention and the state-of-the-art baseline, respectively.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KVCache Quantization", "Sparse Attention", "LLM Inference" ], "forum_url": "https://openreview.net/forum?id=jdbTopkRnn", "pdf_url": "https://openreview.net/pdf/f4c36acb7d1052d1fb7bb4d15a1f065fb0a2b8bc.pdf", "updated_at": "2026-10-05T20:51:11.772Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "bit-width", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "T2AxHRI9eT", "number": 44199, "title": "Q3.6: Metadata‑Conditioned Lossless Compression for Quantized LLMs", "abstract": "Low-bit quantization has become a popular way to deploy large language models (LLMs) on resource-constrained devices. However, after a model has already been quantized, residual redundancy in the quantized weights remains largely unexploited, while further quantizing to even lower bit widths can cause unacceptable generation-quality loss. This motivates dedicated lossless compressors for quantized LLMs: they reduce storage after quantization while preserving the exact quantized weights. Existing lossless LLM compressors target floating-point formats such as bfloat16 (BF16), re-quantize before coding, or treat weights as generic byte streams; none conditions on the native metadata of a fixed low-bit object. We propose Q3.6, a lossless compressor for already-quantized LLMs with bit-exact decoding. Q3.6 builds on the observation that stored metadata, such as the scales, is predictive of the symbol distribution of the weight blocks it describes. It derives deterministic entropy buckets from the native metadata and encodes each bucket with a matched table-based entropy model, so blocks with similar metadata share a bucket and a bitstream. Q3.6 stores less than a strong Zstandard (zstd) baseline in every evaluated setting, covering three models, five quantized formats, and three bit widths, and saves about 10% on average relative to the uncompressed target tensors. In storage-offloaded serving at 1.5 GB/s, Q3.6 gives the smallest stream and achieves the fastest decoding among the evaluated GPU codecs; fed to a fused decode-plus-GEMM kernel, the entropy-coded stream lowers the time per output token by 7 to 10%, and the weight buffers the kernel frees yield 16% to 90% more KV cache and up to 2.1× the throughput in vLLM on 8 to 24 GB GPU budgets, with tokens identical to the stock engines. Code is available at https://anonymous.4open.science/r/iclr2027-code-4B00.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "lossless compression", "entropy coding", "quantization metadata", "memory-efficient inference" ], "forum_url": "https://openreview.net/forum?id=T2AxHRI9eT", "pdf_url": "https://openreview.net/pdf/cd5614be3209f2ab15e7536bc8797acd37601341.pdf", "updated_at": "2026-10-05T20:51:17.127Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantize", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sYlzZaMPnS", "number": 44203, "title": "SparseMark: Routing-Aware Differential Parameter Watermarking for Mixture-of-Experts LLMs", "abstract": "Parameter watermarking for sparse mixture-of-experts (MoE) language models faces a placement challenge: evidence concentrated in a few experts is vulnerable to localized damage, while embedding must preserve model utility. We introduce SparseMark, a post-training multi-bit watermark that distributes evidence across experts and limits the influence of individual parameter pairs during recovery. SparseMark combines balanced expert allocation with response-based carrier selection, encodes a secret codeword through antisymmetric parameter updates, and spreads each bit across experts using keyed block ordering. Verification recovers the codeword from stored baseline contrasts through unit sign voting, without model forward passes. Controlled experiments show that balanced allocation improves recovery under targeted expert damage at comparable utility cost, while unit voting outperforms mean aggregation under pruning at fixed embedding energy. SparseMark achieves exact clean recovery of 256-bit codewords across three MoE backbones with small measured changes in perplexity. Detailed OLMoE evaluations demonstrate watermark retention under the evaluated pruning, quantization, and fine-tuning settings, while identifying limited recovery margin under feedback-guided attacks after quantization. These findings support distributed carrier placement and bounded pairwise influence as design principles for parameter watermarking in sparse MoE models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts", "Model Watermarking", "Ownership Verification" ], "forum_url": "https://openreview.net/forum?id=sYlzZaMPnS", "pdf_url": "https://openreview.net/pdf/9fabf90e435490e60ff8ed8c2a38185e443245e8.pdf", "updated_at": "2026-10-05T20:51:18.206Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ORg442EApD", "number": 44367, "title": "CARE-Attn: Component-Aware Residual Error Correction for FP4 Attention QAT", "abstract": "Recent advances in attention quantization-aware training (QAT), notably Attn-QAT, have substantially narrowed the quality gap of FP4 attention. Nevertheless, a measurable gap to higher-precision attention remains, and end-to-end recovery alone does not reveal what QAT actually adapts to inside the attention operator. We revisit FP4 attention from a component-wise perspective and find that Attn-QAT adapts highly unevenly: it substantially reduces the error from quantized queries and keys ($QK$), while errors from attention probabilities ($P$) and values ($V$) barely change and a non-negligible $QK$ residual remains. These residuals also have distinct structures: the $QK$ residual is sparse and heterogeneous across layers and heads, the $P$ error is limited by the quantizer, and the $V$ error is associated with channel-wise offsets. We therefore propose CARE-Attn (Component-Aware Residual Error Correction for Attention), which corrects each residual separately: it recomputes predicted high-error $QK$ tiles in higher precision under adaptive budgets, decouples training and inference precision through $P$-exact QAT, and removes channel offsets from $V$ with exactly compensated mean-centering. On video diffusion transformers, CARE-Attn improves generation quality over Attn-QAT, produces outputs closer to those of the BF16 fine-tuned model, and reduces QAT overhead; on language models, it consistently lowers perplexity and substantially improves long-context retrieval. These results show that attention QAT alone is not enough: explicitly correcting the residuals it leaves is key to accurate FP4 attention. Code is available at https://anonymous.4open.science/r/CARE-Attn.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "QAT", "attention quantization", "quantization-aware training" ], "forum_url": "https://openreview.net/forum?id=ORg442EApD", "pdf_url": "https://openreview.net/pdf/5c2411bc27e1250339be5f4f85d3f432ffe29e8d.pdf", "updated_at": "2026-10-05T20:51:46.701Z", "categories": [ "qat", "training", "diffusion" ], "matched_terms": [ "fp4", "qat", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hmcSnEcYXB", "number": 44392, "title": "PRESTO: PeRspective-based Entropy-Sparsity Training Optimization", "abstract": "Pruning and quantization make a neural network smaller, and when the goal is storage size, a lossless compressor is eventually applied (Han et al., 2016) to a model that is usually not trained to be compressible.\n\nWe introduce PRESTO, a quantization-aware training method that adds to the loss a proxy for the compressed size of the weights, leaving the activations in floating point. This proxy is the value of a convex inner problem over the quantization grid learned by LSQ (Esser et al., 2020), written in perspective form, i.e. the tightest convex relaxation of the decision to keep a weight. A Lagrangian relaxation makes that problem separable, and an envelope argument gives a (sub)gradient of its value. To our knowledge, PRESTO is the first method to jointly balance sparsity and the codebook symbol distribution arising from quantization within a single convex inner problem, rather than through separate terms that act independently, as in HEMP+LOBSTER (Tartaglione et al., 2021) and LilNetX (Girish et al., 2023).\n\nOn ImageNet, with sizes measured on the whole checkpoint after lossless coding, PRESTO reaches $20.4\\times$ on ResNet-18, $19.5\\times$ on ResNet-50, $22.6\\times$ on ViT-B/16, $11.0\\times$ on DeiT-Small and $9.9\\times$ on EfficientNet-B0, above full-precision accuracy on the two ResNets and less than $0.75$ accuracy points below it on the other three; on five of the six networks, we hold an operating point that no published method we compare against betters on both axes. On AlexNet, prune-first methods remain ahead, a deficit we attribute to sparsity allocation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Neural Network Compression", "Lagrangian Relaxation", "Quantization-Aware Training", "Convex Optimization", "Pruning", "Entropy Coding", "Perspective Reformulation" ], "forum_url": "https://openreview.net/forum?id=hmcSnEcYXB", "pdf_url": "https://openreview.net/pdf/29b3ce1d10ef305cd0efcc14c3bd10c0149e8d07.pdf", "updated_at": "2026-10-05T20:51:49.739Z", "categories": [ "qat", "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ihD01i4wdi", "number": 44455, "title": "QKSieve: Query-Key Balanced Mixed-Bit Retrieval for Efficient Long-Context Attention", "abstract": "Autoregressive decoding repeatedly scans the complete key-value (KV) history. Query-aware sparse attention helps only when selection is cheaper than dense QK scoring and faithful to joint Query-Key geometry. We present QKSieve, a training-free, GPU-resident exact-KV retrieval method. Request-local Query and Key second moments define biorthogonal coordinates that preserve every unquantized dot product and order dimensions by joint score energy. A per-layer, per-head allocator assigns 0/1/2/4/8 bits to eight 16-D bands under a 240-bit token/head budget by minimizing a separable logit-MSE surrogate. The packed index targets B(N) = min{N, 1280, max(256, ceil(0.06N))} positions, after which the original 16-bit K/V receive exact sparse attention. No learned router, exact-QK reranking, task rule, or dense fallback is used.\n\nQKSieve-Robust adds a current-Query rank-16 INT4 estimate of the omitted softmax partition and Value numerator, bringing total auxiliary storage to 7.47%. On complete LongBench, Robust retains 99.98%, 99.73%, and 100.00% of Full macro score on Llama-3.1-8B, Qwen3-4B, and Mistral-7B while targeting 6.85-7.43% of history. Across 13 RULER tasks from 4K to 128K, it retains 100.33% on average with 4.97% measured active attention. In a controlled native-MHA RTX 3090 implementation, the complete Robust attention path is 2.09-4.12x faster at 32K-128K, and real steady decode is 1.32-3.98x faster. We separately report index construction and persistent-cache break-even rather than hiding fixed cost in steady-state timing. Analysis connects representation error to ranking margins, omitted probability mass, and attention-output error.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "sparse attention", "KV cache retrieval", "mixed-bit quantization" ], "forum_url": "https://openreview.net/forum?id=ihD01i4wdi", "pdf_url": "https://openreview.net/pdf/a40ae643f3324d50373d62960d01a01e7a37bac7.pdf", "updated_at": "2026-10-05T20:52:01.521Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "mYDDvIglKc", "number": 44526, "title": "Continuous Variational Synthesis", "abstract": "Biological machine learning was long bottlenecked by our ability to synthesize designed DNA. Variational synthesis models control chemical reactions to physically manufacture quadrillions of designed sequences in DNA. However, training these variational synthesis models is challenging: constraints on chemical synthesis can force many parameters into a discrete space, limiting the ability to pre-train and fine-tune. In this article we train \"free\" variational synthesis models using stochastic gradient descent in continuous space, and then discretize with post-training quantization. This enables variational synthesis models to satisfy stringent reward criteria, while still producing diverse designs, achieving a strictly dominating quality-diversity Pareto frontier. We demonstrate by training variational synthesis models of peptides, antibody CDRH3s, regulatory DNA elements, and enzymes. *In silico* performance is maintained *in vitro*.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "chemical synthesis", "quantization", "protein design", "DNA design" ], "forum_url": "https://openreview.net/forum?id=mYDDvIglKc", "pdf_url": "https://openreview.net/pdf/1d63d90199932677374d16adcfe9bd46d60071c6.pdf", "updated_at": "2026-10-05T20:52:11.558Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "u5d3kM6ahC", "number": 44592, "title": "MiLD: Mining Informative Low-Bit Cues for Training-Free AI-Generated Image Detection", "abstract": "Low-bit planes have been shown to contain discriminative information for\nAI-generated image detection. We further observe that the Hamming states\ninduced by low-bit morphological residuals provide useful real/fake cues.\nHowever, the Hamming-state occupancy is a static image-level statistic and is\nsensitive to JPEG compression and Gaussian noise. This motivates us to use\nHamming states as spatial cues rather than as the final detection score, and to\nconstruct a stable, robust, and efficient training-free detector. We propose\nMiLD, which selects channel-disagreement locations from low-bit morphological \nresiduals and\nbuilds a masked low-bit reversal. MiLD then compares the frozen encoder\nfeatures of the original and reversed images, using their feature discrepancy as\nthe detection score. Across five benchmarks, MiLD improves the\nmean AUROC over RIGID by \\textbf{2.78\\%} while achieving the highest efficiency\nscore. Extensive ablations further demonstrate the stability, transferability,\nand robustness of MiLD.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Training-Free", "AI-Generated Image Detection" ], "forum_url": "https://openreview.net/forum?id=u5d3kM6ahC", "pdf_url": "https://openreview.net/pdf/3ea147f68bc21a47a1ada3d17c4e26e993dda578.pdf", "updated_at": "2026-10-05T20:52:21.101Z", "categories": [ "theory" ], "matched_terms": [ "low-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "AenXyGuGX8", "number": 44616, "title": "DyQ-VLA: Dynamic Quantization and Compensation for VLA models with Runtime Error Awareness", "abstract": "Vision-language-action (VLA) models achieve favorable task performance, yet runtime errors in closed-loop execution evolve with alternating updates of actions and observations. Existing VLA quantization methods mainly trade off inference speed and model performance while rarely investigating how quantization alters runtime errors. By comparing error distributions between full-precision and quantized models, we identify distinct evolutionary patterns for the two types of errors: quantization enlarges the variance of translation error distributions and increases their dispersion, whereas the distribution center of rotation errors gradually shifts across execution steps, demonstrating cumulative drift. Motivated by this observation, we rethink the optimal quantization strategy for VLA models and propose \\textit{DyQ-VLA}, a runtime-error-aware quantization framework, which dynamically selects activation precision according to execution steps and tracks as well as compensates rotation errors via accumulated quantization residuals. Corresponding operators and runtime adaptation strategies are devised within the framework to enable dynamic-precision execution. Experiments show that \\textit{DyQ-VLA} achieves a 1.88$\\times$ to 1.93$\\times$ inference speedups while maintaining comparable or higher average task success rates. Moreover, it reduces mean execution steps by 17.7\\% to 19.8\\%. Our code is here: https://anonymous.4open.science/r/DyQ-VLA-7F51/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Dynamic Quantization，Error Control，VLA Models，Compensation" ], "forum_url": "https://openreview.net/forum?id=AenXyGuGX8", "pdf_url": "https://openreview.net/pdf/3df0339a55c3db6281b65d7b2b016165e17f5cae.pdf", "updated_at": "2026-10-05T20:52:29.394Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "iNzzmPFGg4", "number": 44738, "title": "REGAIN: Restoring Token Distributions in Low-Bit Language Models via Truncated-Support Distillation", "abstract": "Low-bit weight quantization substantially mitigates the memory footprint and latency bottlenecks of large language models, yet aggressive compression regimes incur non-negligible capability degradation. While low-rank quantization error compensation effectively recovers coarse benchmark accuracy with minimal parameter overhead, it fails to restore the nuanced output distributions of the full-precision teacher. Specifically, even when Top-1 predictions appear intact, quantization systematically distorts the probability margins and relative rankings among competing tokens, precipitating severe compounding errors in sequential and structured generation tasks. To resolve this discrepancy, we propose \\textsc{Regain}, a parameter-efficient recovery framework based on truncated-support distillation. Instead of optimizing over the entire vocabulary or naively mirroring teacher targets, \\textsc{Regain} dynamically constructs a compact token support from the union of the Top-$k$ candidates from both the full-precision teacher and the quantized student. This design preserves teacher-favored modes while directly penalizing quantization-induced spurious competitors, using an auxiliary Ghost token to aggregate the residual probability mass for bounded computation. With the low-bit backbone frozen, \\textsc{Regain} exclusively trains low-rank adapters under this compact distributional objective. Extensive evaluations across Qwen3 architectures from 8B to 32B, multiple post-training quantization schemes including GPTQ and QuIP\\#, and aggressive 2 to 4-bit regimes demonstrate that \\textsc{Regain} substantially improves distributional fidelity and downstream performance across question answering, tool invocation, and structured reasoning, setting a new Pareto frontier for low-rank recovery of ultra-low-bit language models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Bit Quantization; Post-Quantization Recovery; Knowledge Distillation; Large Language Models; Parameter-Efficient Fine-Tuning" ], "forum_url": "https://openreview.net/forum?id=iNzzmPFGg4", "pdf_url": "https://openreview.net/pdf/5663bd95bf2d40696b0205bc21f7e6afecd42176.pdf", "updated_at": "2026-10-05T20:52:53.220Z", "categories": [ "ptq", "hardware", "reasoning" ], "matched_terms": [ "4-bit", "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lIct6Odo0j", "number": 44880, "title": "Budgets Transfer, Hardness Labels Do Not: Auditing What Survives Quantization of Reasoning Models", "abstract": "Reasoning models are usually quantized just before deployment, but two settings that guide their use are measured earlier, on the full-precision (BF16) model: the decode budget, which caps how many tokens the model may generate, and prompt hardness labels, which decide which prompts receive extra training. We ask whether these BF16 measurements can be reused after quantization. A direct comparison is misleading, because a longer budget can look better only because fewer answers are cut off, and because two runs of the same model already disagree on which prompts are hard. We therefore check truncation before comparing budgets, and compare every cross-precision change in the hard set with an independent rerun of the same model. On Qwen3.5-4B with 4-bit GPTQ, the two settings behave differently. The budget can be reused: serving the 4-bit model at the budget chosen on BF16 changes GPQA accuracy by −0.59 points on average over six seeds (95% interval [−1.94, +0.76]). Hardness labels cannot. The BF16 and 4-bit hard sets overlap by 0.414 (Jaccard), well below the 0.593 between two BF16 runs. The lower overall accuracy of the 4-bit model does not explain the gap: shifting and rescaling the BF16 scores to match it still predicts an overlap of 0.523. Quantization changes which prompts are hard, not only how many. A second quantizer, round-to-nearest, shows the same mismatch. The practical rule is to inherit the budget from BF16 but remeasure hardness on the deployed model, and to test any hardness-based training against a random control, since in our runs it gives no detectable gain.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Test-Time Scaling", "Metadata Transfer" ], "forum_url": "https://openreview.net/forum?id=lIct6Odo0j", "pdf_url": "https://openreview.net/pdf/6499c9b46f1ad551a2e3fce88371d90e26894793.pdf", "updated_at": "2026-10-05T20:53:17.050Z", "categories": [ "hardware", "reasoning" ], "matched_terms": [ "4-bit", "quantization", "quantized", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QsD8PLNCMA", "number": 44929, "title": "RoleQ: Role-Aligned Scales for Converting Language Models to Ternary", "abstract": "Weight quantization below four bits shrinks memory and bandwidth at a perplexity that looks acceptable, but often at the cost of factual knowledge and task accuracy. Ternary weights have mostly been obtained by training from scratch or long fine-tuning. We introduce RoleQ, a quantization framework that places scales by the functional roles of Transformer units rather than by the layout of the weight matrix, as conventional formats do. Vectors that read from the residual stream take one scale per output unit; those that write to it, one per input unit. The rule is structural, needs no search, calibration or stored state, and holds on integer and ternary grids. In a rotated basis it matches a per-layer axis search at half the solver work, and at ternary it cuts perplexity by up to half against layout-based placement in two families at every size from 0.6B to 14B, where per-row scales leave MMLU at chance. Our lightweight post-training pipeline converts an existing checkpoint to 1.59 bits per weight in a few GPU-hours, a fraction of the cost of learned ternarization, and leads every post-training ternary method at matched or lower memory-resident bits on perplexity, zero-shot accuracy and MMLU, most widely on knowledge. A sparse second ternary plane placed by a per-unit Fisher saliency keeps the matmul multiplication-free and recovers what ternary loses first: at 8B, a fraction of a bit returns about half the MMLU that conversion costs, well ahead of random or Hessian-based placement at equal bits.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "ternary quantization", "low-bit large language models", "weight-only quantization", "model compression", "efficient inference", "quantization scales", "rotation-based quantization", "GPTQ", "sub-2-bit" ], "forum_url": "https://openreview.net/forum?id=QsD8PLNCMA", "pdf_url": "https://openreview.net/pdf/2cbda5a5a02ce4a7a99eafee011dbafb020bbdbc.pdf", "updated_at": "2026-10-05T20:53:23.708Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "low-bit", "quantization", "ternary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "y6dMoHxLTD", "number": 44944, "title": "Investigating Sample-Level Sharpness as a Curriculum Scoring Function", "abstract": "Curriculum learning presents training examples to a model in order of increas-\ning difficulty, with sample-level difficulty estimations typically relying on metrics\nthat are implicitly deduced from the model’s loss landscape. Yet, the connection to\nminima sharpness, a key characteristic of the local loss landscape, which is often\ninvestigated in the context of generalisation on a model or dataset level, has barely\nbeen made in the literature. In this work, we introduce per-sample sharpness as a\ndifficulty estimation that measures how quickly the loss on a single example rises\nunder random, filter-normalized weight perturbations of a trained model, along\nwith a Gradient-Guided variant that perturbs along the example’s own loss gra-\ndient. We evaluate both on CIFAR-10, CIFAR-100, and the Minimum Viewing\nTime dataset, comparing sharpness against five established scoring functions and\nhuman difficulty labels.We find that (i) sharper minima on a sample-level min-\nima correlate with higher difficulty according to other difficulty metrics, imply-\ning a related but distinct notion of difficulty, which, however, does not correlate\nwith human perception, (ii) random effects w. r. t. to underlying filter-normalized\ndirections can be mitigated through ensembles, which also increase correlation\nwith gradient-guided sharpness, (iii) sharpness-based curriculum learning outper-\nforms a random-order baseline but not standard training (iv) under post-training\nweight quantization, model accuracy collapses at the same perturbation magnitude\nat which the Sharpness Score ordering itself becomes unreliable, consistent with\na transition from a locally convex to a chaotic region of the loss landscape. We\nshare our code on https://anonymous.4open.science/r/curriculum-learning-7302/", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Sharpness", "Curriculum Learning", "SAM", "Sample Difficulty" ], "forum_url": "https://openreview.net/forum?id=y6dMoHxLTD", "pdf_url": "https://openreview.net/pdf/232aae406d2d3573e239a5e6b81a794b50399de8.pdf", "updated_at": "2026-10-05T20:53:28.747Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KMXlKKogU0", "number": 44985, "title": "REAL-TIME JOINT AUDIO–VIDEO GENERATION BY PARALLEL ADAPTER COMPOSITION", "abstract": "Deploying a joint audio--video diffusion transformer for real-time, interactive\ngeneration normally requires two essential modifications: block-autoregressive\nattention, so frames can be emitted before the whole clip is finished, and few-step\nsampling, so each block is cheap.\nConventionally, the streaming video literature obtains both capabilities\nfrom a \\emph{chained} pipeline. It first distills a bidirectional teacher\ninto a causal student, then into a few-step one, or proceeds in reverse order.\nEach stage of such a chain fine-tunes the weights the previous one produced,\nso a later objective can undo an earlier capability.\nFollowing the idea of model merging, we show that on a packed audio--video\nbackbone the two capabilities can be\nacquired \\emph{in parallel}. A causal\nadapter is trained against the \\emph{frozen} backbone, and an off-the-shelf few-step\nadapter provides the few-step capability. As the two edit different\nfunctional axes, we predict, and then verify, that their weight-update\ndirections are near-orthogonal, without any explicit orthogonality\nconstraint during training. Orthogonal updates should combine without\ninterfering, so parallel composition is a direct sum. The two adapters are\nsimply \\emph{added} at inference, with no joint training, yielding few-step,\nstreaming audio--video whose image quality tracks the bidirectional teacher. \nCompared to the chained baselines, the composed model\nmatches or beats them on most metrics, making parallel composition a\npractical approach. The resulting streaming system generates joint audio--video in\nreal time, ${\\approx}26$\\,fps at $480{\\times}832$ without quantization, and sustains $30$\\,s of\ncontinuous generation with stable image quality. Demo Page: \\url{https://pac-demo-2027.github.io/demo/}", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Joint audio-video generation", "Real-time streaming generation", "Model merging" ], "forum_url": "https://openreview.net/forum?id=KMXlKKogU0", "pdf_url": "https://openreview.net/pdf/ad42551206888e80a92dd7338ef163b3ed7565be.pdf", "updated_at": "2026-10-05T20:53:37.231Z", "categories": [ "diffusion" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "y5hvGgjglI", "number": 44994, "title": "CentriQ: Calibration-Free Quantization of Diffusion Transformers via Exact Mean Centering", "abstract": "Diffusion transformers (DiTs) achieve state-of-the-art image generation, but their sampling cost limits deployment. Quantizing both weights and activations to 4 bits reduces this cost, yet existing methods fall short in one of two ways. Calibration-based methods are tied to a specific checkpoint and prompt distribution, whereas data-free Hadamard rotation, effective for LLMs, loses quality on DiTs. We show that this loss has a structural cause. Adaptive layer-norm conditioning adds a per-token mean to the activations, and at the widths of the evaluated DiTs, the Hadamard rotations used by data-free methods cannot spread this mean uniformly across coordinates. A single dominant direction therefore survives the rotation and sets the quantization range. We introduce CentriQ, a calibration-free quantizer that centers each token before rotation and restores the mean exactly through a rank-1 full-precision branch, so that per-token scales follow in closed form without data. Weights are fitted under a robust $\\ell_p$ objective that tracks the dense mode of each group and discounts heavy tails. Across three DiTs, CentriQ matches the quality of calibrated SVDQuant at 4 bits, whereas calibration-free weight quantizers with plain per-token activation quantization collapse or degrade substantially. CentriQ outperforms the strongest calibration-free method reported to date at 2-bit weights. It is also the first calibration-free method to retain usable image quality at 2-bit activations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Diffusion Models", "Quantization", "Calibration-free", "DiTs", "Image Generation" ], "forum_url": "https://openreview.net/forum?id=y5hvGgjglI", "pdf_url": "https://openreview.net/pdf/2b2068657f5572653fae521571380994ee9ee90d.pdf", "updated_at": "2026-10-05T20:53:39.936Z", "categories": [ "ptq", "hardware", "diffusion" ], "matched_terms": [ "2-bit", "quantization", "quantizer", "quantizers", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rwRyrGOG1L", "number": 45085, "title": "NormSub: Normalization-Affine Subspace Adaptation for Quantization Recovery", "abstract": "Low-bit quantization can substantially degrade model accuracy. Beyond improving quantizers and weight reconstruction, recovery requires understanding which parameters are useful to adapt under a limited budget. We investigate normalization-affine parameters as a structured, low-dimensional update subspace. In a fixed-graph study with matched parameter counts, data, and update budgets, affine updates achieve lower perplexity and higher mean accuracy than the tested weight-coordinate controls, with a small advantage over gradient-selected coordinates. Functional diagnostics further show that task-level recovery need not reduce every local reconstruction error. These findings motivate NormSub, a two-stage affine adaptation framework around a target post-training quantization (PTQ) backend. The pre-stage adapts only normalization-affine parameters on a disposable backend-aligned quantized proxy, then discards the proxy, restores the original floating-point weights, and transfers only the backend-compatible affine state for fresh PTQ reconstruction. The post-stage refines the affine coordinates exposed by the reconstructed graph while freezing its non-affine persistent state. Across the evaluated dense language models and PTQ backends, NormSub improves perplexity and mean zero-shot accuracy. On Qwen3-8B W2A4KV4 with QuaRot+ResComp, the unweighted six-task mean accuracy increases from 49.6\\% to 58.1\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Model Quantization", "Post-Training Quantization", "Large Language Models", "Low-Bit Quantization", "Model Compression", "Normalization-Affine Adaptation" ], "forum_url": "https://openreview.net/forum?id=rwRyrGOG1L", "pdf_url": "https://openreview.net/pdf/9ba09d4ae2da9ffe548f4d27b3e025d2e1769154.pdf", "updated_at": "2026-10-05T20:53:49.873Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "quantizers", "w2a4kv4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ruFeSThcNq", "number": 45147, "title": "Memory Has an Interface: From Facts to Accepted Actions", "abstract": "The same remembered facts can lead to different task outcomes depending on how they enter an agent's context and how its responses reach execution. We study this interface by removing reference-action facts from user requests and restoring byte-identical marked JSON either inline or through a historical assistant–tool exchange. Across 146 selected tasks and three models at one primary seed, historical delivery achieves 29.68% success versus 23.52% for inline delivery, a gain of 6.16 percentage points. A separate 59-task retail study with quantized Qwen3.6-27B reveals a distinction between proposing an action and having it accepted: all 25 first-response pairs with different native acceptance outcomes propose identical tool names and arguments. We then cross the delivery interface with the execution rule in 236 complete interactions. Native rules yield 1/59 inline and 6/59 historical successes; admitting schema-valid text-plus-tool responses to execution raises both totals to 56/59. Content controls further establish the value of correct facts through both interfaces, with model-dependent delivery contrasts. Together, these measurements identify message construction and response acceptance as consequential stages between selected facts and completed tasks, and provide a concrete framework for evaluating how agent memory becomes action.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "agent memory", "large language models", "context construction", "tool-use agents", "agent evaluation" ], "forum_url": "https://openreview.net/forum?id=ruFeSThcNq", "pdf_url": "https://openreview.net/pdf/cef22dec8b0085155eb148afa3252a07a8cf0492.pdf", "updated_at": "2026-10-05T20:54:00.377Z", "categories": [], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zaMfAiz3m5", "number": 45350, "title": "Quantization Beyond Matrix Multiplication for Efficient LLM Inference", "abstract": "Quantization is widely used to accelerate LLMs inference, but existing approaches primarily optimize matrix multiplication while leaving surrounding operators and data movement in high precision.\nThis limits end-to-end benefits of quantization, particularly for memory-bound operators.\nIn this paper, we proposed an INT8 pipeline that extends quantization beyond matrix multiplication.\nOur design preserves intermediate tensors in INT8 across both matmul and non-matmul operators, reducing global-memory traffic.\nWe further introduce a hierarchical warp--block reduction strategy that effectively utilize on-chip resources and maintains high GPU occupancy during compute deep learning operators.\nWe evaluate our approach on Qwen3 and Llama-3.2 models across multiple tasks and workloads.\nCompared with BF16 inference, our approach achieves up to\n$1.6\\times$ inference speedup and reduces energy consumption by up to 39\\%, while maintaining comparable accuracy.\nThe benefits remain consistent across GPUs, decoding workloads, and batch sizes, demonstrating that quantization beyond matmul is important for efficient end-to-end LLM inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM Quantization", "GPU Kernel Optimization" ], "forum_url": "https://openreview.net/forum?id=zaMfAiz3m5", "pdf_url": "https://openreview.net/pdf/cf9e2f9f0eff6fccd99b746a2f9a0ac6cb477304.pdf", "updated_at": "2026-10-05T20:54:31.676Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sDQ8S6CKCh", "number": 45374, "title": "Open-MMUnlearning: Unifying Methods and Evaluation for MLLM Unlearning", "abstract": "As multimodal large language models (MLLMs) become more capable and widely deployed, concerns about privacy and safety have become increasingly pressing. Machine unlearning offers one approach to addressing these concerns by removing designated information from trained models while preserving unrelated capabilities. However, fragmented implementations and evaluation protocols, incomplete robustness testing, and limited understanding of metric reliability make progress in MLLM unlearning difficult to assess systematically. We introduce **Open-MMUnlearning**, an open-source, extensible framework that integrates target-model preparation, multimodal data processing, unlearning, and evaluation through shared interfaces and structured configurations. The framework supports five benchmarks spanning privacy, safety, and copyright, eight MLLMs from four model families, and twelve unlearning methods. Its evaluation suite jointly assesses forgetting effectiveness, retained utility, and robustness to model interventions, adversarial inputs, and membership inference attacks. Using a common evaluation protocol, we compare ten representative unlearning methods. In this comparison, GD and MIP-Editor tie for the highest overall score: GD achieves the highest Forget Quality, while MIP-Editor preserves more Model Utility. We further introduce a metric meta-evaluation protocol that tests faithfulness using models with controlled exposure to target knowledge and robustness under quantization and relearning. Among the thirteen evaluated metrics, BLEU achieves the highest aggregate reliability score. KS-Test attains the highest faithfulness AUC but performs less well on robustness. Together, the framework and these findings support reproducible comparison of MLLM unlearning methods and systematic assessment of evaluation reliability. We release the code and model checkpoints at [this anonymous repository](https://anonymous.4open.science/r/Open-MMunlearning-4E20).", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Machine Unlearning", "Multimodal LLM" ], "forum_url": "https://openreview.net/forum?id=sDQ8S6CKCh", "pdf_url": "https://openreview.net/pdf/553d3b84b03764f28ad7bb56888273ebc5b605bf.pdf", "updated_at": "2026-10-05T20:54:35.575Z", "categories": [ "multimodal", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "YvrtNbSjDZ", "number": 45405, "title": "Transformer-FSQ: Discrete Latents for Extreme KV-Cache Compression", "abstract": "Private, personalised language models must run on local consumer devices, where frontier models cannot fit and capability depends on what the model retains in context. However, context memory---the key-value (KV) cache---grows linearly with retained cache entries, quickly dwarfing the model itself: for Qwen3-1.7B-Base, 200,000 retained entries require 21 GiB, 6.6 times more than the model weights. While high token capacity is urgently needed on edge hardware, the KV-cache compression methods we measured fail beyond modest compression ratios. We introduce *Transformer-FSQ*, addressing extreme KV-cache compression by replacing standard key and value projections with a finite scalar quantisation bottleneck. Rather than storing large vectors, the cache retains a packed vector of discrete integer coordinates per retained token per layer, requiring no per-token quantization metadata. Converting a pretrained transformer requires training these projections on 0.25 tokens per original parameter.We evaluate Transformer-FSQ against five baseline families: token eviction, scalar and vector quantisation, low-rank projection, and latent attention on models from 0.6B to 8B parameters. On long-context retrieval (RULER), the methods fall into two regimes. Up to about 5$\\times$, vector quantisation and low-rank projection stay close to the uncompressed model; from 12$\\times$ to 97.5$\\times$, Transformer-FSQ is the most accurate method among our primary evaluated implementations, and leads at extreme compression across scales. Extreme KV-cache compression is therefore a practical way to extend how much context a model on a fixed device can hold.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache compression", "finite scalar quantisation", "discrete latents", "long-context inference", "memory-efficient inference", "on-device language models" ], "forum_url": "https://openreview.net/forum?id=YvrtNbSjDZ", "pdf_url": "https://openreview.net/pdf/10c8756fe3a936a686bf85321f268b00971f2d98.pdf", "updated_at": "2026-10-05T20:54:39.552Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantisation", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "vTzM83OUJn", "number": 45419, "title": "THE LOW-BIT MEMORY LAW: PRECISION BUYS PROPAGATION, DIMENSION BUYS RETRIEVAL, AND GATING MUST BE STABLE", "abstract": "Recurrent-state architectures (linear RNNs, SSMs, Gated DeltaNet, Kimi Delta Attention) replace the KV cache of most layers with a fixed-size state, making state quantization a primary deployment bottleneck: recent work reports uniform INT8 already degrades complex reasoning and INT4 collapses to near zero. No quantitative law, however, says how long a b-bit recurrent state can remember—expressivity analyses ask whether such states can compute a function, and deployment heuristics give no account of why persistence sets risk or when the heuristic breaks. We supply the law. Under finite precision, memory is governed by three resources with closed-form bounds: (i) precision buys propagation—reach is k_max = (2√3 / eC) * 2^b steps, each bit doubling it, at a unique self-consistent forgetting rate α_*(b); (ii) dimension buys retrieval—read accuracy saturates beyond b_sat(α) = 1/2 log2( C² / (3(1−α²)) ) bits, above which 8 bits is indistinguishable from fp32 at every load while 4 bits suffices only up to M ≈ d_k/4 (the retrieval bit requirement grows with load); (iii) gating noise is exponentially amplified—one standard deviation of gate-logit noise costs 0.72 state bits. A fourth result repairs current practice: the correct per-channel persistence measure is P = 1/(1 − E[a_t²]), and the geometric-mean version used by DAMP (Zhang et al., 2026) probably underestimates risk (Jensen), by 2.7–5.7× under the occasional-reset gates that selective architectures learn; correcting it removes 99.7 ± 0.6% of the excess risk of Top-K protection across five independently trained models at an identical bit budget. That gain is a statement about gate variability, not a universal one: where the gates are nearly deterministic the correction is inert. The resulting bit-requirement log-law b_req(P) = 1/2 log2(C² P /3) explains why protected channels must be FP16, why INT4 variants cannot be repaired by layout changes, and why DAMP’s Hadamard transform works (it lowers C by up to Δb ≈ 1.24 bits on heavy-tailed blocks). Beyond closed-form verification we measure a real pretrained Mamba-1 (130M, 1536 channels): its channel statistics transfer (23× energy concentration above uniform; material damage from 4–5 bits, INT8 costing 0.12% perplexity) while the ranking correction does not, because that model’s gates are nearly deterministic—mask-to-mask variation (0.44 PPL s.d.) exceeds the correction (0.19 PPL)—so its scope is gates with non-negligible Var(ln a). All results are reproducible from released scripts.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "low-bit memory law", "state quantization", "SSM", "Mamba", "gating noise", "finite precision", "quantization" ], "forum_url": "https://openreview.net/forum?id=vTzM83OUJn", "pdf_url": "https://openreview.net/pdf/47d084395ee5238cdb35fcc95a17afb5f794d6d5.pdf", "updated_at": "2026-10-05T20:54:41.909Z", "categories": [ "kv-cache", "hardware", "reasoning" ], "matched_terms": [ "int4", "int8", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gKdwQoLM75", "number": 45540, "title": "SpecGuard: What Should LLMs See to Generate Assertions That Catch RTL Bugs?", "abstract": "Hardware bugs that escape the rigorous verification and validation process are costly to fix, often requiring recalls, software patches that mask the error rather than fix it, and similar remedial actions. This motivates rigorous verification of each implementation against its intended behavior. Assertion-based verification captures this intent using SystemVerilog Assertions (SVA), which can be used in both formal and simulation-based verification. However, composing SVAs remains largely manual and time-consuming, causing a bottleneck in time-to-market. LLMs show promise in automating this by translating a design’s natural-language specification into SVA without manual formalization. In practice, SVA generation through LLM is impeded by restating shown implementation behavior, placing claims in the wrong clock cycle, and naming signals outside the checked module. Recent studies show SVAs generated by frontier models detect only 5.1–11.7% of injected bugs. In this paper, we analyze assertion suites generated by existing LLM-based SVA generation methods. We evaluate them against injected bugs and identify four failure modes, along with two attributes that place the resulting assertions beyond repair after generation, rendering them ineffective for actual bug detection. These failure modes and attributes are not captured by the compile and proof rates typically used to evaluate generators. While a stronger model improves these rates, its generated assertions do not necessarily detect more bugs. We subsequently develop SpecGuard, an SVA generation framework that addresses each failure by separating what an assertion must draw from a model versus what can be fixed with design’s structure. SpecGuard was evaluated on the 83-design AssertLLM2 benchmark, using a local 14.8-B model with int4 quantization. Its assertions detected up to 66% of the AssertLLM2 bugs in a single run, compared with 11.7% for the best-performing of six frontier models. Furthermore, on more complex hierarchical designs from NVDLA and Google Coral NPU, it achieves up to a 51% bug-detection rate, compared with 14% for existing state-of-the-art assertion-generation techniques.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Assertion-based verification", "LLM-based hardware verification", "Design specification translation", "Hardware bug detection" ], "forum_url": "https://openreview.net/forum?id=gKdwQoLM75", "pdf_url": "https://openreview.net/pdf/7a6ddc91a6f7d16399da450811646d4b21e6a162.pdf", "updated_at": "2026-10-05T20:55:01.948Z", "categories": [ "hardware" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PXGDPZB3Av", "number": 45559, "title": "ModTrain: Training Modulated Diffusion Models for Extreme Low-Bit Quantization", "abstract": "Diffusion models have redefined generative AI, but their multi-step inference remains a critical bottleneck for efficient deployment. Quantization offers a powerful remedy, yet extremely low-bit activation quantization remains challenging even for state-of-the-art quantization-aware training (QAT), as its errors are difficult to absorb during training. Modulated diffusion models offer a promising path by exploiting temporal redundancy across denoising steps, but have so far been limited to post-training quantization. We present ModTrain, a training framework for modulated diffusion models. Training these models is non-trivial because historical states couple consecutive denoising steps, complicating data construction, gradient computation, and activation alignment. ModTrain addresses these issues with sequential forward sampling for efficient data construction, history-aware backpropagation for correct gradient flow, and full-precision alignment to mitigate activation distribution shift. ModTrain is orthogonal to existing QAT methods and integrates seamlessly with them. Experiments on LSUN and ImageNet show that ModTrain outperforms post-training MoDiff, reducing activation precision from 4 to 2 bits, and consistently improves QAT baselines at extremely low precision, with up to a 10.90 FID improvement. ModTrain advances extremely low-bit inference and may further motivate hardware support for such computation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantized-Aware Training", "Diffusion Models", "Efficiency" ], "forum_url": "https://openreview.net/forum?id=PXGDPZB3Av", "pdf_url": "https://openreview.net/pdf/61c040192f4fb1a794d5faf5e979085f6bbb47e5.pdf", "updated_at": "2026-10-05T20:55:04.881Z", "categories": [ "ptq", "qat", "training", "hardware", "diffusion" ], "matched_terms": [ "low precision", "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "CpPnm0EePQ", "number": 45598, "title": "Rollout-Guided QAT for FP4 Reinforcement Learning of MoE Language Models", "abstract": "Reinforcement learning (RL) for post-training large language models (LLMs)\nincurs substantial computation and memory overhead during rollout generation,\nwhich motivates low-precision rollout for efficient RL training. However, existing\nFP4 RL methods suffer from a key limitation: they primarily optimize quantiza-\ntion accuracy on the training and rollout paths independently rather than directly\nreducing the discrepancy between the two quantized execution paths. In this work,\nwe propose NTR, an FP4 quantization framework for RL training of Mixture-of-\nExperts (MoE) language models that addresses the limitation of existing FP4 RL\nmethods. NTR incorporates rollout-guided quantization-aware training that uses\nrollout-side quantization outcomes to guide training-side FP4 rounding decisions,\ndirectly reducing train-rollout discrepancy. Moreover, NTR adopts an efficient\nquantization-information caching scheme that selectively retains mantissa and\nscale information from deeper layers to reduce the storage and communication\noverhead introduced by rollout guidance. We evaluate NTR on four large-scale\nMoE language models across reasoning, coding, and long-horizon RL tasks. Our\nresults demonstrate that NTR enables joint FP4 weight/activation and FP4 KV-\ncache rollout with RL performance comparable to BF16 rollout, while achieving up\nto 5.4× rollout speedup and strong final FP4 performance compared with post-hoc\nFP4 quantization of BF16-trained policies.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "RL Training", "AI Infrastructure", "LLM Quantization" ], "forum_url": "https://openreview.net/forum?id=CpPnm0EePQ", "pdf_url": "https://openreview.net/pdf/254728e56e44f75a7c8f362e1a547afa64e7b1c1.pdf", "updated_at": "2026-10-05T20:55:12.170Z", "categories": [ "ptq", "qat", "training", "reasoning" ], "matched_terms": [ "fp4", "low-precision", "qat", "quantiza", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "8LqHKWRyvw", "number": 45644, "title": "ThriftAttention: Selective Mixed Precision for Long-Context FP4 Attention", "abstract": "Efficient attention algorithms are critical to mitigate the quadratic cost of attention in long-context workloads. Such methods often rely on low-bit quantisation and/or hard sparsity to accelerate inference. FP4 attention on Blackwell GPUs offers large speedups but degrades sharply relative to FP16 as context length grows while sparsity methods discard query-key interactions outright which can degrade model quality. We address both failure modes through ThriftAttention, a mixed-precision attention mechanism that uses a heuristic to promote only the most important interactions to FP16 while computing the remainder in FP4. ThriftAttention relies on our finding that the output impact of quantisation error is highly non-uniform and increases with the importance of each query-key interaction, concentrating functionally relevant error in a small number of attention blocks that contain the most important tokens. Our experiments show across long-context benchmarks and model families that by computing only 5% of the attention blocks in FP16, ThriftAttention recovers on average 92.3% of the FP4$\\to$FP16 performance gap. We demonstrate ThriftAttention's advantage as context length grows, the regime where FP4 attention performs worst.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Training-free", "FP4", "attention", "efficient", "sparsity", "mixed-precision", "inference", "llm" ], "forum_url": "https://openreview.net/forum?id=8LqHKWRyvw", "pdf_url": "https://openreview.net/pdf/d3269dc995a38aebea7d1ff319c79c29df81aef0.pdf", "updated_at": "2026-10-05T20:55:21.032Z", "categories": [ "training", "hardware" ], "matched_terms": [ "fp4", "low-bit", "mixed precision", "mixed-precision", "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZkBCD3Iati", "number": 45668, "title": "Topological Path Signatures under Physical Micro-Fractures: Pareto-Optimal Battery Degradation Estimation on Edge Microcontrollers", "abstract": "Sequence learning paradigms demonstrate low empirical error on curated laboratory battery benchmarks, yet rely upon an unstated operational prior: continuous, uniformly sampled charging trajectories globally aligned to a standardized temporal axis. In commercial vehicular deployment, this prior is violated by physical micro-fractures, including unanchored opportunistic charging slices ($5\\%$-$25\\%$ State-of-Charge intervals), sensor quantization floors ($10\\,\\text{mV}$ ADC steps), and stochastic communication dropouts. Under this manifold, continuous spline-based ODE solvers encounter numerical stiffness deadlocks, self-attention mechanisms undergo key-padding dilution, and unrolled hidden states exceed edge microcontroller (MCU) memory bounds. In this work, we present RIIS-MicroBurst, an edge-native framework utilizing a parameter-free feature extraction operator formulated in free Lie algebra. By replacing numerical integration with deterministic tensor-vector operations via Chen's identity, RIIS maps arbitrary variable-length inputs into a static, invariant geometric representation. Combined with thermodynamic phase-plane scaling and a closed-form Geometry-Oracle, RIIS establishes a new Pareto efficiency frontier across simulation, dynamic stress, synthetic fault, and commercial fleet benchmarks. Operating strictly within automotive MCU hardware ceilings ($\\le 1\\,\\text{MB}$ SRAM, $\\le 10^5$ FLOPs), it achieves state-of-the-art accuracy while delivering $4,900\\times$ to $57,000\\times$ lower computational operations compared to recurrent and continuous baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Physics-Informed Machine Learning", "Path Signatures", "Deterministic Routing", "TinyML", "Resource-Constrained AI", "Battery Degradation" ], "forum_url": "https://openreview.net/forum?id=ZkBCD3Iati", "pdf_url": "https://openreview.net/pdf/4dde518e6b55e5735e4c8ea8a44d020ee38fd398.pdf", "updated_at": "2026-10-05T20:55:26.087Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rj4obAa7ez", "number": 45727, "title": "SP-QAT: Spectral-Preserving Quantization-Aware Training for Extreme Low-Bit Vision-Language-Action Models", "abstract": "Vision-language-action (VLA) models are rapidly advancing and becoming central to embodied AI, yet their deployment on edge devices remains challenging. Since in VLA model, LLM backbone dominates memory consumption and inference latency, aggressive quantization of this LLM module presents a straightforward solution. Unfortunately, existing VLA quantization techniques mostly operate at 4-bit or higher bit widths and avoid fixed sub-4-bit quantization because of its catastrophic performance collapse, leaving it largely underexplored. To bridge this gap, we propose SP-QAT, an efficient quantization-aware training framework (QAT) that enables fixed 2-bit quantization of the Large Language Model (LLM) backbone in VLA, to our knowledge, for the first time.\nThrough theoretical and experimental analysis, we identify that preserving the cognitive feature spectral subspace is the key to extreme low-bit quantization for the LLM backbone of VLAs. Therefore, we first constrain the weight-update subspace to preserve the pretrained model priors. Second, we propose Soft-Mask Gradients with Severity-Aware Scale Regulation to mitigate outlier-driven step size expansion from causing quantized weights to collapse into noise. Finally, we propose Difficulty-Aware Curriculum Learning with SNR-Guided Static Reweighting to emphasize hard samples and low-noise timesteps, thereby preserving the action-sensitive energy in the cognition feature spectrum. Extensive experiments on SIMPLER with CogACT and $\\pi0$ show SP-QAT only results in a small drop of $\\leq2.6\\%$ compared to FP16 baselines, with 75%–84\\% memory savings and $1.6\\times$ inference speedup. Code is available as supplementary material.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "VLA", "Quantization" ], "forum_url": "https://openreview.net/forum?id=rj4obAa7ez", "pdf_url": "https://openreview.net/pdf/322d8fbfeddae3ed03a93bdf772ed247f3a452b5.pdf", "updated_at": "2026-10-05T20:55:34.183Z", "categories": [ "qat", "training", "hardware", "multimodal", "theory" ], "matched_terms": [ "2-bit", "4-bit", "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "gcXYUSLhNC", "number": 45763, "title": "SparseDiag: State-Dependent Diagnosis and Repair for Low-Bit Large Language Models", "abstract": "Post-training quantization (PTQ) methods for large language models (LLMs) have largely advanced by changing how quantized layers are optimized, through improved objectives, learnable parameterizations, reconstruction granularities, and layer- or window-wise schedules. However, existing methods generally follow optimization schedules that are specified independently of the residual state that emerges as quantization progresses, leaving a complementary question underexplored: once a model reaches a particular quantized state, is every scheduled layer still worth optimizing?\nThrough controlled layer-wise probing, we find that the marginal utility of further optimization is highly non-uniform and strongly state-dependent. In relatively weak quantized states, repairable errors can be broadly distributed across the network. After stronger PTQ optimization, however, many layers provide little additional benefit, while the remaining headroom often concentrates in a small set of non-contiguous residual bottlenecks. Moreover, repairing one bottleneck reshapes the subsequent repair landscape, suggesting that the layers worth optimizing cannot, in general, be determined reliably from the initial state alone.\nMotivated by these observations, we propose Sparse Diagnosis (SparseDiag), an iterative framework that explicitly separates diagnosis from correction. At each round, SparseDiag constructs lightweight layer-wise repair candidates using calibration data, ranks them on a disjoint calibration split, verifies the selected candidate on held-out calibration samples, and updates only a validated residual bottleneck. It then re-diagnoses the updated model and automatically stops when no reproducible repair opportunity remains. Consequently, the optimization schedule is discovered from the evolving quantized state rather than prescribed in advance.\nExperiments on Llama and Qwen models under W4A16 and W2A16 show that SparseDiag consistently exposes additional repairable headroom across different quantized starting states, while stronger PTQ states exhibit substantially sparser residual repair landscapes than weaker ones. Repeated diagnosis further reveals an evolving progression from broadly distributed repairability to sparse residual bottlenecks and eventually to headroom exhaustion. These results recast low-bit PTQ refinement from merely asking how to optimize each layer to deciding whether, where, and when further optimization is still beneficial.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "low-bit quantization", "efficient inference", "adaptive optimization" ], "forum_url": "https://openreview.net/forum?id=gcXYUSLhNC", "pdf_url": "https://openreview.net/pdf/872e02e1085d870c8eb4ac04635779c253b8e47c.pdf", "updated_at": "2026-10-05T20:55:38.930Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized", "w2a16", "w4a16" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "KzeQR3Rqgi", "number": 45991, "title": "RESCOM: ACTIVATION RESIDUAL COMPENSATION FOR LARGE LANGUAGE MODEL QUANTIZATION", "abstract": "Low-bit weight–activation quantization reduces the memory and computation needed to deploy large language models, but activation rounding can still degrade accuracy after calibration. Reconstruction from rounded inputs cannot distinguish activations within the same quantization bin. We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the result\u0002ing layer-output error. An output-aware calibration procedure selects residual di\u0002rections by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights. On Llama-3-8B with four-bit weights and activations, rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 per\u0002centage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting. On B200, the evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "weight-activation quantization", "activation quantization residuals", "low-rank compensation", "reduced-rank regression" ], "forum_url": "https://openreview.net/forum?id=KzeQR3Rqgi", "pdf_url": "https://openreview.net/pdf/fa7ef68e7dcbe35207424f0c466f637e0509cc6c.pdf", "updated_at": "2026-10-05T20:56:11.237Z", "categories": [ "ptq", "training" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "oIaznhm0uU", "number": 46017, "title": "LatticeLoop: Learning Loop-Specific Weights from Shared Low-Bit Codes", "abstract": "Looped Transformers apply one stack of core layers several times. Full weight tying keeps the model small, but every loop must apply the same weights. Independent weights for each loop remove this restriction, but the core storage grows with the number of loops. Partial untying and per-loop LoRA add less storage, but they limit the change between loops to a few rows or to a low-rank update. Block-scaled low-bit formats such as NVFP4 already split every weight tensor into two parts: 4-bit codes set the pattern of each block of 16 weights, and 8-bit block scales set its magnitude. LatticeLoop shares the codes across loops and learns one table of block scales for each loop. Each additional loop costs 0.5 bits per core weight, one ninth of an independent copy, and every loop runs as one standard block-scaled matrix multiplication. In language model pretraining at matched storage, LatticeLoop closes 92.5% of the loss gap between full tying and independent weights, while partial untying and per-loop LoRA close less than a third of the gap. The advantage holds up to 12 loops and 80B training tokens, and decoding speed stays close to full tying.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "loop", "nvfp4" ], "forum_url": "https://openreview.net/forum?id=oIaznhm0uU", "pdf_url": "https://openreview.net/pdf/d5a1278eeb25b18cc92bfa5f4e96a572041b1915.pdf", "updated_at": "2026-10-05T20:56:14.993Z", "categories": [], "matched_terms": [ "4-bit", "8-bit", "low-bit", "nvfp4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "o2BJPdnrm0", "number": 46026, "title": "Forward Similarity Does Not Guarantee Training Equivalence: Precision Placement in Differentiable Gram Solvers", "abstract": "Changing the numerical precision of a differentiable Gram solver may leave predictions nearly unchanged while altering the computation used to learn its parameters. We study this distinction in controlled Spectral Koopman Attention (SKA) and an independent episodic ridge classifier, assigning FP32 or FP64 to separate solver stages without changing their exact-arithmetic definitions. Matched comparisons fix the model, input, and optimizer state; separately trained endpoints are evaluated through a common precision route. For SKA, we compare $\\mathsf q_{110}$ with the higher-precision solver route $\\mathsf q_{111}$, a numerical reference rather than mathematical ground truth. Across 18 matched SKA validation comparisons from three training seeds, median relative discrepancies in logits, full-model gradients, and restored AdamW updates are $0.962\\%$, $8.61\\%$, and $9.88\\%$, respectively, each normalized by its corresponding $\\mathsf q_{111}$ reference norm. Prediction agreement can therefore miss differences in the learning computation. In a seven-seed matched classifier comparison, imposing externally prescribed ridge schedules shared across routes within each seed during training reduces the absolute between-route spread in post-training regularized-Gram conditioning by $97.4\\%$ relative to learning ridge separately under each route. This shows that the observed conditioning separation depends on the training-time ridge policy. Precision choices in these solvers therefore warrant checks of parameter updates alongside predictions, with endpoint conditioning and task outcomes assessed separately.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "differentiable solvers", "numerical precision", "mixed precision", "numerical linear algebra", "optimization", "training dynamics" ], "forum_url": "https://openreview.net/forum?id=o2BJPdnrm0", "pdf_url": "https://openreview.net/pdf/cf4494f6dd788eb2bdaaad2fcb73bf6987b8ea3d.pdf", "updated_at": "2026-10-05T20:56:16.029Z", "categories": [ "ptq", "training", "reasoning" ], "matched_terms": [ "mixed precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "x2je9XgN06", "number": 46039, "title": "SignMuon: Communication-Efficient Distributed Muon for LLM Pretraining", "abstract": "Full-precision gradient exchange can limit the scaling of data-parallel training. We propose SignMuon, a distributed optimizer that combines Muon's matrix update directions with signSGD's majority-vote communication. Each worker approximates the polar factor of its local momentum matrix and communicates only its entrywise signs. Workers aggregate these signs using int8 AllReduce or packed 1-bit AllGather and apply the same update, keeping the polar computation local. We give conditional stationarity bounds for the actual momentum-polar vote and distinguish polar-solver error from the loss introduced by signing. On large language models (LLMs) up to 1.5B parameters trained on eight GPUs, SignMuon achieves lower perplexity at matched token budgets than the tested sign-based optimizers, while also showing favorable log-perplexity trajectories in wall-clock comparisons. We further evaluate optimizer memory usage and distributed communication efficiency in multi-GPU training. Together, these results show that SignMuon combines Muon's local matrix transformation with highly compressed global communication, delivering strong token efficiency and lower log perplexity for distributed LLM pretraining.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "SignMuon", "Muon", "distributed optimization", "communication-efficient training", "gradient compression", "sign-based optimization", "majority-vote aggregation", "matrix-aware optimization", "LLM pretraining", "data-parallel training", "distributed deep learning" ], "forum_url": "https://openreview.net/forum?id=x2je9XgN06", "pdf_url": "https://openreview.net/pdf/122ba65369de4e3d26c94facd3ff24e870a0cc90.pdf", "updated_at": "2026-10-05T20:56:17.531Z", "categories": [ "training", "hardware" ], "matched_terms": [ "1-bit", "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "u2AWpCOTtb", "number": 46141, "title": "The Vulnerability Window: Decision Instability in Quantized Reward Models", "abstract": "Reward models and LLM judges are routinely quantized to 4 bits so that scoring many candidates stays affordable, and the usual check is that benchmark accuracy barely moves. That check is close to blind. Across eight reward models from two families (0.6B–8B), three generative judges and four 4-bit quantizers, quantization rewrites 1.1–12.3% of pairwise preference decisions on RewardBench while aggregate accuracy moves by as little as 0.03 points ($p = 1.00$), because the errors it fixes and introduces cancel. Flips concentrate where the full-precision margin is small relative to the quantization noise scale $\\sigma$: a Gaussian approximation $\\Phi(-|m|/\\sigma)$ describes the aggregate rate, over-predicting it in all 30 scalar settings and failing for generative judges, and occupancy of the resulting vulnerability window falls from 25% to 7% with model size, which is what makes small judges fragile. The churn stays within noise under benign best-of-16 and becomes clear under adversarial pressure for some quantizers (146:54 harmful:helpful flips on one family, reversing sign on another). Downstream it changes what ships: across six (model, quantizer) pairs from two families, 14.6–65.7% of problems get a different best-of-64 selection while only 2.6–9.6% get a different final answer, with selections changing 5–8$\\times$ more often than answers in every pair, and the pair the benchmark calls identical (0.03 points, $p = 1.00$) still changes 34.5% of them. Benchmark churn ranks that downstream change ($\\rho = 0.89$); benchmark accuracy does not order it reliably. Held-out audits then locate the useful diagnostic in the margin rather than the Gaussian: every margin-based predictor orders flips identically (AUC 0.927), and on the level a directly measured churn rate wins in-distribution. As a routing signal it is matched by plain top-$k$. The contribution is the characterization, not the model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "reward models", "LLM-as-a-judge", "quantization", "post-training quantization", "RLHF", "best-of-N", "reward hacking", "robustness", "evaluation", "calibration" ], "forum_url": "https://openreview.net/forum?id=u2AWpCOTtb", "pdf_url": "https://openreview.net/pdf/4227352f1ae58bca1485c652d381b076fcbdcf73.pdf", "updated_at": "2026-10-05T20:56:33.106Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "4-bit", "quantization", "quantized", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hJ9kQATIq3", "number": 46179, "title": "HYBRIDQUANT: CALIBRATION-FREE ENTROPYCODED E8 LATTICE QUANTIZATION OF THE KV CACHE", "abstract": "Compressing the key–value (KV) cache is central to long-context inference: for\nthe compact and distilled models deployed where memory is scarce, the cache, not\nthe weights, sets the limit. Established quantizers force a trade-off: calibration-\nfree methods use fixed-rate scalar codes that leave a large gap to the rate–distortion\nlimit, while entropy-coded compressors close the gap but fit their models to cal-\nibration data. An entropy code needs no calibration if the vectors it sees are\nGaussian, isotropic and nearly independent across coordinates. HybridQuant\nachieves this with three rounds of the randomized fast Walsh–Hadamard trans-\nform, 3d log2 d operations per vector instead of the d^2 of a dense random rotation,\nthen rounds to the E8 lattice and codes with one static rANS code built for a Gaus-\nsian source; the step follows from the target rate alone, with every side bit counted.\nWe prove how much rotation is needed: one round leaves a spike far from Gaus-\nsian (Kolmogorov distance 0.34), two bring every input close but leave a floor of\norder d\n−1/2\n, and three lower it to order 1/d up to a logarithm (0.035 versus 0.0097\nat d = 128, computed exactly). A dithered variant has input-independent errors,\nunbiased attention logits and an attention concentration bound that does not grow\nwith context length, and an exact, calibration-free centering removes the collapse\nthat the projection biases of Qwen2 models cause in rotation-based quantizers. On\nQwen2.5-7B, Mistral-7B-v0.3 and Llama-3.1-8B the rate is within 0.003 bits of\nits target, and at 2.125 bits TurboQuant-MSE has 8.8%, 13.3% and 31.3% higher\nperplexity than HybridQuant, and a rotated entropy-coded scalar quantizer 2.3%,\n3.0% and 6.5% higher, all significant under a paired bootstrap.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-Cache Compression", "Large Language Models (LLMs)", "Vector Quantization", "Fast Walsh-Hadamard Transform (FWHT)", "Wasserstein-2 Optimization", "Cornish-Fisher Expansion", "Talagrand Concentration", "E8 Lattice", "Bit Rate Distortion", "Sub-Gaussian Tail Bounds", "Long-Context Inference." ], "forum_url": "https://openreview.net/forum?id=hJ9kQATIq3", "pdf_url": "https://openreview.net/pdf/86656a23b847674cb9d8703d54398351b4a01e86.pdf", "updated_at": "2026-10-05T20:56:36.865Z", "categories": [ "kv-cache", "theory" ], "matched_terms": [ "quantization", "quantizer", "quantizers" ], "review_status": "abstract", "summary": "HybridQuant 用三轮快速 Walsh–Hadamard 变换把 KV 向量变得更接近高斯和各向同性，再舍入到 E8 晶格，以预先为高斯源构造的静态 rANS 编码，无需校准语料。它还给出旋转轮数的分布逼近分析，并提供 dither 版本和无需校准的中心修正以处理注意力误差与投影偏置。", "contribution": "在 Qwen2.5-7B、Mistral-7B-v0.3、Llama-3.1-8B 上，编码率距目标不超过 0.003 bits；在 2.125 bits，TurboQuant-MSE 的困惑度分别比 HybridQuant 高 8.8%、13.3%、31.3%，摘要称配对 bootstrap 均显著。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。困惑度百分比以 HybridQuant 为比较基准，不能直接改写成相同百分比的下降；摘要未报告端到端吞吐或总显存收益。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "LLM KV 缓存的 E8 lattice 量化与熵编码", "bit_width": "报告 KV 编码率 2.125 bits，计入 side bits", "evidence": [ { "location": "官方摘要", "claim": "三种所列 7B/8B 模型的编码率距目标在 0.003 bits 内。" }, { "location": "官方摘要", "claim": "2.125 bits 下 TurboQuant-MSE 困惑度分别高 8.8%、13.3%、31.3%。" } ] }, { "id": "xQ57THZa4v", "number": 46229, "title": "Quantizing Diffusion LLMs Under Parallel Decoding: Commitment Risk and the Parallelism Tax", "abstract": "Diffusion large language models (dLLMs) are served with accelerated decoders that cache context and commit, in one step, every masked position whose confidence clears a threshold. Yet post-training quantization of dLLMs has so far been calibrated and evaluated under fixed-step samplers without caching. We study W4A4 quantization under the deployed decoder and find that the decoder determines both where quantization harms accuracy and what it costs. Irreversible errors enter chiefly through forced commitments, made when no position clears the threshold: they are a minority of commitments but carry most quantization-induced flips. Quantization also levies a parallelism tax, extra forward passes per response: by compressing top logit gaps, it pushes positions below the threshold while leaving predictions largely intact. We propose DECAL, which reweights layer-wise calibration by a commitment-risk prior measured on the deployed decoder, and an argmax-preserving commit temperature that recovers the tax. On LLaDA and Dream, DECAL attains the best average accuracy among W4A4 methods, with lower run-to-run variance than prior calibration.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language models; post-training quantization" ], "forum_url": "https://openreview.net/forum?id=xQ57THZa4v", "pdf_url": "https://openreview.net/pdf/c8f463e8e50127b4b9ce50b765038df8e3446617.pdf", "updated_at": "2026-10-05T20:56:43.349Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantization", "quantizing", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FThYK1pTZc", "number": 46246, "title": "Don't Drop Tokens, Dim Them: Bit Allocation for Multi-Turn KV-Cache Compression", "abstract": "Agentic language models spend substantial wall-clock time paused while waiting\nfor tool responses or user input. During these pauses, their KV caches either re-\nmain resident in GPU memory or are managed by a memory store and transferred\nbetween storage and the GPU. How to compress the paused cache, and which\ncompression best preserves answer quality over the turns that follow, has no set-\ntled answer: the two training-free families, token selection and quantization, are\nboth built and evaluated for a single prompt answered immediately. We show that\nboth fail in the paused regime, for opposite reasons. Token selection commits\na keep-set before the queries arrive, so methods perfectly recover a planted fact\non a single-turn needle collapse on SCBench key–value retrieval. Quantization\nmakes no such commitment but spends its budget uniformly, and on multi-turn\nretrieval it collapses at 2 bits, which is also the most it can compress. We trace\nboth failures to one cause: precision is uniform and token selection is binary. We\ntherefore propose SOFT-BIT KV. Rather than choosing which tokens survive, it\ngives every token its own bit-width, fixed once at the pause by rate–distortion\nallocation under a measured-byte budget. A small MLP scores each cached to-\nken from the resuming query’s own per-head attention, and reverse water-filling\nspends a byte budget over those scores. Average across four models, SOFT-BIT\nKV retains 92% of full-cache accuracy on multi-turn SCBench at 5.2× compres-\nsion, compared with 85% for TurboQuant and 72% for the best token-dropping\nmethod. At ≈ 11× compression, SOFT-BIT KV retains 99% on single-turn needle\nretrieval and 65% on long-context RULER, compared with 61% and 56% for the\nbest token-dropping method, and 38% and 23% for TurboQuant. At the system\nlevel, it reduces the paused cache of a 98k-token Llama-3.1-8B trajectory by 81%\nand accelerates host-to-device transfer upon resumption by 5.2×.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM KV Cache", "Agentic LLM", "multi-turn conversation" ], "forum_url": "https://openreview.net/forum?id=FThYK1pTZc", "pdf_url": "https://openreview.net/pdf/2fa193ee9e753ad9bcc8db742bccf3e0de607e20.pdf", "updated_at": "2026-10-05T20:56:45.496Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "binary", "bit-width", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UCSm45sWOk", "number": 46281, "title": "Receiver-Conditioned Selection and Reconstruction for Latent Communication Between LLM Agents", "abstract": "Latent communication through key-value (KV) caches can avoid natural-language decoding and repeated prefilling by transferring internal states between large language model (LLM) agents. Existing approaches fix transmission choices before the current query is known or reuse latent states formed outside the receiver context. We find that the content a sender ranks first can hurt the receiver and that different receivers need different content. We therefore condition what to retrieve and how to read it on the receiver. Each document is written once, before any receiver or query exists, as an INT8 snapshot of its early-layer state. At read time, the receiver uses its state to select a document and reruns the upper layers under its private context, so document states are formed for that receiver rather than transplanted. Across Qwen3-4B, Qwen3-8B, and Llama-3.1-8B-Instruct, receiver-aware selection with BGE-M3 improves F1 by 15.5--17.2 points over query-only selection for both text and checkpoint delivery. With selection fixed, read-time reconstruction improves F1 by 5.1--13.5 points over full static KV reuse with payloads $32\\times$ to $57\\times$ smaller. Under controlled TCP networking, the complete system reduces time to first token (TTFT) by 6.91\\%, 5.81\\%, and 4.23\\% on Qwen3-4B, Qwen3-8B, and Llama-3.1-8B-Instruct, respectively. In the end-to-end evaluation, answer quality remains comparable to native text, with F1 differences ranging from $-0.82$ to $+1.88$ points. Code is available at anonymous.4open.science/r/rcsr-latent-comm-B184.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-Cache Communication; Multi-Agent LLM Systems; KV-Cache Reuse; Receiver-Aware Selection;" ], "forum_url": "https://openreview.net/forum?id=UCSm45sWOk", "pdf_url": "https://openreview.net/pdf/0b4570a543dab5ba2e22415c0cfc4a287170d639.pdf", "updated_at": "2026-10-05T20:56:50.137Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TP87m7HjKC", "number": 46321, "title": "TIGER: Inverting Multimodal Transformer Gradients via Embedding-Subspace Distance Optimization", "abstract": "Federated learning (FL) allows multiple clients to jointly train a shared model by sending gradient updates to a central server while keeping their raw inputs local. Gradient inversion attacks, however, have shown that these updates can leak enough information to reconstruct client inputs.\nFor transformers, though, existing attacks fall short in some of the following ways: they scale poorly to realistic model sizes, larger inputs, or non-causal attention; they require an enumerable vocabulary of token embeddings, preventing the recovery of continuous non-textual inputs; or they are brittle under numerical noise, e.g., from quantization or Differential Privacy (DP).\nIn this paper, we introduce TIGER, a continuous gradient inversion attack that exploits the natural low-rank structure of attention gradients. Its novel objective aligns the features of dummy inputs across layers with the low-rank gradient subspaces, resulting in a single framework that recovers text from both causal and bidirectional attention models and, for the first time, multimodal inputs. \nOn settings featuring bidirectional attention, TIGER substantially improves reconstruction quality over prior work and can recover images at roughly 720p resolution in the multimodal setting. On causal models, TIGER is significantly more robust than prior attacks, enabling accurate reconstructions even under quantization and DP-style defenses. Overall, TIGER uncovers further privacy vulnerabilities in LLMs, underscoring the need for stronger defenses in deployed FL systems.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "federated learning", "gradient inversion", "gradient leakage", "llm", "vllm", "multimodal", "privacy" ], "forum_url": "https://openreview.net/forum?id=TP87m7HjKC", "pdf_url": "https://openreview.net/pdf/efa7168972015c68dfce68af6cf5ea42b980c46e.pdf", "updated_at": "2026-10-05T20:56:54.812Z", "categories": [ "ptq", "training", "multimodal" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lNkc5LGSt4", "number": 46483, "title": "QuMix: Query-Guided Mixed-Precision KV-Cache Quantization for Block Diffusion Language Models", "abstract": "Block diffusion lets diffusion language models (dLLMs) reuse completed blocks through a key–value (KV) cache, but storing and repeatedly reading this growing cache limits context length, batch size, and decoding speed. KV-cache quantization reduces these costs, yet has been studied almost entirely in autoregressive (AR) models. Applying an existing AR quantizer to a block-diffusion model, we find that two-bit quantization preserves nearly all of its original TriviaQA score, whereas reducing the precision to one bit causes a large drop. This motivates mixed precision, which lowers average storage by retaining more bits for sensitive components and using fewer bits elsewhere. Such allocation must preserve information that attention relies on, including the key of the attention sink—a token that many heads repeatedly attend to. AR models typically place this sink at the first token, making it possible to protect by position, but sink positions can vary in dLLMs. We therefore introduce QuMix, which uses queries to guide precision allocation without assuming a sink position. From calibration queries and keys, QuMix constructs a query-weighted basis and assigns more bits to directions whose errors have a larger effect on attention scores, without model training. To our knowledge, this is the first study of KV-cache quantization for block diffusion. QuMix retains 98–99% of the 16-bit LongBench score across three models at an average of 1.99 bits per cache element. A custom CUDA kernel that attends directly to the compressed cache supports a $10.5\\times$ larger maximum batch and $2.3\\times$ faster denoising steps at the same batch size (batch 15) than BF16 FlashAttention4 on an RTX PRO 6000 GPU.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "diffusion language models", "block diffusion", "KV-cache quantization", "mixed precision", "efficient AI" ], "forum_url": "https://openreview.net/forum?id=lNkc5LGSt4", "pdf_url": "https://openreview.net/pdf/9a48702d9f1430cb5c580c233d9c1547e9e6ac81.pdf", "updated_at": "2026-10-05T20:57:15.415Z", "categories": [ "ptq", "kv-cache", "hardware", "diffusion" ], "matched_terms": [ "mixed precision", "mixed-precision", "quantization", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "PMMJ2mpa1m", "number": 46709, "title": "Generalization behavior of OPTQ and the role of regularization", "abstract": "Large neural networks can be compressed by rounding or ``quantizing'' their weights to numbers that admit representations with fewer bits. One algorithm for quantization, OPTQ, progressively quantizes the weights of a neural network so that the squared quantization error on a specified calibration dataset is as small as possible. We study the performance of OPTQ and a variant algorithm, stochastic OPTQ, in a generalization setting and derive bounds for the expected squared error accrued by the algorithm when a test point is drawn from a fixed distribution. We prove two results. One result relates the generalization error to the error on a calibration dataset comprising independent samples from the same distribution as the test distribution. The other result bounds the generalization error of stochastic OPTQ for all sufficiently nice distributions, regardless of the calibration dataset. In both of these results, the regularization term $\\lambda$ plays an important role. We use insights from these results to make a new recommendation for the choice of $\\lambda$ and see that this choice of $\\lambda$ preforms favorably in experiments when compared to prior recommendations in the literature.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "OPTQ", "quantization", "generalization", "regularization" ], "forum_url": "https://openreview.net/forum?id=PMMJ2mpa1m", "pdf_url": "https://openreview.net/pdf/6c5aaf2b8397418b2b657d8a666925e15f67fb1b.pdf", "updated_at": "2026-10-05T20:57:46.647Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "quantization", "quantizes", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dMu09wVQ9Y", "number": 46774, "title": "BitLift: Unified Mixed-Precision Quantization and Caching for Efficient MoE Inference", "abstract": "Mixture-of-Experts (MoE) models reduce computation through sparse activation, but moving expert weights can dominate decoding when the model exceeds device memory.\nDynamic mixed precision and DRAM-caching can reduce this cost, yet their effectiveness depends on how weight representation, precision allocation, and cache residency interact.\nWe introduce $\\mathrm{BitLift}$, a training-free framework that co-designs these components.\nA router-rank policy dynamically allocates a fixed nominal precision budget across selected experts.\nEach expert is encoded as a 2-bit base and two 1-bit refinements, jointly optimized for the target policy through weighted scale fitting and error compensation.\nA two-pool cache separates base and refinement planes and applies cache-resident precision refinement without fetching additional weights for an upgrade.\nAcross three MoE model families, rank-assigned execution approaches independently optimized 3-bit quality at nominal budgets of 2.333--2.375 bits per routing slot, and outperforms the evaluated 2.5-bit baselines. \nOn Qwen3-30B-A3B, cache-resident refinement matches uniform 4-bit WikiText-2 perplexity with a cache holding 50\\% of the 4-bit routed-expert footprint.\nAt smaller matched capacities of 6.25--25\\%, it improves on uniform 3-bit perplexity.\nIn this flash-limited regime, our analytical hardware model estimates a throughput gain of $1.17$--$1.26\\times$ over uniform 3-bit and $2.06$--$2.49\\times$ over uniform 4-bit inference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-experts", "large language models", "post-training quantization", "dynamic mixed precision", "progressive quantization", "expert caching", "memory-constrained inference" ], "forum_url": "https://openreview.net/forum?id=dMu09wVQ9Y", "pdf_url": "https://openreview.net/pdf/575c8ec484a9f150c16dfdf793c4bb079aa70289.pdf", "updated_at": "2026-10-05T20:57:54.802Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "1-bit", "2-bit", "3-bit", "4-bit", "5-bit", "mixed precision", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "UpR5uTu7JV", "number": 46794, "title": "CISA: Retrofitting Dense Attention Models with Chunk-Indexed Sparse Attention", "abstract": "As context becomes grows, LLM inference becomes increasingly expensive because attention repeatedly reads a growing key-value (KV) history.\nWe introduce Chunk-Indexed Sparse Attention (CISA), a method for retrofitting pretrained dense attention models without gradient updates.\nCISA selects fixed-size chunks of the KV history using compact summaries derived from the model's existing keys, avoiding token-level indexing and a separate full-length index cache.\nFor indexing, CISA uses a single query per KV group and shares the selected chunks across the query in that group.\nDuring prefill, we reuse chunk selections across adjacent queries to reduce indexing cost and improve KV reuse and GPU utilization.\nWe call this technique *QShare*.\n\nWe evaluate CISA on the NVIDIA Nemotron~3 Nano and Ultra models, as well as on Inkling-Small.\nIn one-layer BF16 benchmarks at batch size 32 using Ultra's per-device attention dimensions, CISA accelerates core attention by $11.7\\times$ for prefill queries near the end of a 1M-token prompt and $13.0\\times$ during decode.\nIncluding chunk scoring, selection, and all other measured GPU work, the corresponding speedups are $8.6\\times$ and $3.0\\times$.\nOn Ultra NVFP4, at $k_{\\mathrm{chunks}}=4{,}096$, CISA keeps aggregate scores within 1.5 percentage points of dense attention on MRCR, LongBench v2, and GraphWalks.\nRULER sweeps on Nano and Ultra BF16 up to one million tokens show scores within 2.05 percentage points of dense attention with 8--16$\\times$ fewer KV reads.\n\nWhen serving million-token prompts with 32 concurrent requests and 4K output tokens per request, CISA achieves $2.29\\times$ the end-to-end throughput of dense attention using Nemotron~3 Ultra NVFP4 on four NVIDIA B300 GPUs.\nSelecting fewer chunks increases throughput to $2.61\\times$ compared with dense attention; in separate GraphWalks evaluations, the smaller budget widens the aggregate score gap from 1.49 to 3.59 percentage points. These results expose a practical quality--speed tradeoff without requiring sparse training.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "attention", "sparse attention", "efficiency" ], "forum_url": "https://openreview.net/forum?id=UpR5uTu7JV", "pdf_url": "https://openreview.net/pdf/6cf0fe5491b0a76c335ea23d0d36ef6e8a513594.pdf", "updated_at": "2026-10-05T20:57:57.512Z", "categories": [ "training", "kv-cache", "hardware" ], "matched_terms": [ "nvfp4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "O2IGIHRtWc", "number": 46860, "title": "Spurious states are labels: A multi-label code in the Hopfield-RBM duality", "abstract": "For forty years, the superpositions of stored patterns known as spurious states have been treated as the defining defect of the Hopfield model. We read them as an answer: a mixture of $L$ patterns is a composite signal, and identifying its constituents is multi-label classification. A Hopfield configuration cannot represent a set of labels; its dual restricted Boltzmann machine (RBM) can. We show that the Hopfield-RBM duality analytically fixes this representation: across $N$ visible neurons, active labels sit on the hidden layer at scale $\\sqrt{N}$ with an amplitude matching the archetype-mixture overlap. Under this label assignment, the class averages of noisy training examples form an exact fixed point in mean of one-step supervised contrastive divergence. We prove a finite-size drift bound that decomposes into four vanishing error channels: thermal sampling, dataset noise, mini-batch consensus, and finite-dimensional fluctuations. For continuous mixtures combined prior to quantization (such as musical chords and hyperspectral pixels), this overlap amplitude transitions to an arcsine law via one-bit random projections. Across four benchmarks on three public datasets, the resulting configuration acts as an optimal analytical warm start: computed in a single pass without gradient updates, it surpasses random initializations by $0.08$ to $0.43$ in score, saving thousands of training steps. Finally, the induced parameter-free read-out rivals computationally intensive one-bit compressed sensing baselines at a fraction of their inference cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "statistical physics", "associative memories", "Hopfield networks", "restricted Boltzmann machines", "multi-label classification", "contrastive divergence", "energy-based models" ], "forum_url": "https://openreview.net/forum?id=O2IGIHRtWc", "pdf_url": "https://openreview.net/pdf/ff467ff32969818be3be768d4171bd6f1490ea19.pdf", "updated_at": "2026-10-05T20:58:04.577Z", "categories": [ "training" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Cb794G8ZzP", "number": 47014, "title": "Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization", "abstract": "Quantizing AdamW's optimizer states reduces persistent storage, but quantization errors propagate through the moment recurrences and perturb subsequent adaptive updates. We redesign 4-bit optimizer-state quantization for AdamW from the perspective of \\emph{rounding space}: the coordinate in which a quantizer chooses between adjacent reconstruction levels. For the second-moment, a local analysis of the quantization cell adjacent to zero shows that small mean state error does not ensure a small mean error in the next-step preconditioner. A one-dimensional quadratic construction further shows qualitatively different optimization dynamics under state-space and preconditioner-space rounding. These results motivate Zero-Inclusive Preconditioner-space Stochastic Rounding (\\textbf{ZIP-SR}), which retains zero in the second-moment codebook and computes stochastic-rounding probabilities in preconditioner space. As a complementary route, Zero-Excluding EDEN calibration (\\textbf{ZE-EDEN}) uses a zero-excluding second-moment codebook and rescales the quantized second-moment block to mitigate the preconditioner distortion caused by the positive quantization floor. Both configurations use 4-bit NormalFloat (NF4) for the first moment, with targeted stochastic rounding of the LM-head first moment during the final 10\\% of training. Across GPT- and Llama-style pretraining experiments ranging from \\textbf{130M} to \\textbf{2.7B} parameters, both methods reduce TorchAO 4-bit AdamW's validation-loss gap to 32-bit AdamW at every evaluated scale, with the largest reported gap reduction reaching \\textbf{70\\%}. In full-parameter supervised fine-tuning, both recipes achieve lower validation loss than TorchAO while remaining close to 32-bit AdamW on downstream tasks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Optimizer-state quantization", "AdamW", "memory-efficient optimizers", "low-precision training", "large language models" ], "forum_url": "https://openreview.net/forum?id=Cb794G8ZzP", "pdf_url": "https://openreview.net/pdf/1ac1996f2a1fe57074dc18a35452e49c83530ad4.pdf", "updated_at": "2026-10-05T20:58:25.214Z", "categories": [ "ptq", "training" ], "matched_terms": [ "4-bit", "low-precision", "quantization", "quantized", "quantizer", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ECMhMhrx4X", "number": 47054, "title": "SAGE: Efficient Protected Vision-Language Inference on Smart Glasses", "abstract": "Smart glasses are a promising platform for vision-language assistants, but private\nsplit inference remains expensive: protecting visual representations before\ntransmission does not remove the cost of on-device vision encoding, wide\nprotected tokens, or hardware-unaware execution. We present SAGE\nSmart-glasses-Aware Guarded Encoding), a\nhardware-aware framework for efficient protected VLM inference on smart glasses.\nSAGE keeps raw images and unprotected visual tokens on the trusted device,\nreduces visual input cost, executes shape-specialized INT8 graphs, and compresses\nprotected representations into compact 256-dimensional INT8 packets before\ntransmission. Unlike efficiency methods optimized only for nominal model size,\nSAGE selects its operating point using downstream answer quality together with\nmeasured latency, memory, and thermal behavior on real Mentra Live glasses.\nOn a matched Qwen workload, SAGE reduces glasses inference latency by\n39.18\\%, device-sensor temperature rise by 21.91\\%, and protected packet size\nby 96.56\\%, while retaining 85.62\\% of the protected reference's relaxed\nutility and comparable prompted containment. Direct glasses-to-phone tests\ndeliver all 192 packets and reduce median packet-plus-acknowledgment time from\n71.3 to 10.0\\,ms. The compact protected interface also transfers across four\nadditional VLM families. These results show that private wearable VLM inference\nmust be optimized jointly across representation protection, communication, and\nreal-device execution rather than through model compression alone.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "smart glasses", "vision-language models", "privacy-preserving inference", "split inference", "on-device inference", "efficient multimodal systems" ], "forum_url": "https://openreview.net/forum?id=ECMhMhrx4X", "pdf_url": "https://openreview.net/pdf/7c9675c5b841d9e3937bb67dc43bfb198d18bc18.pdf", "updated_at": "2026-10-05T20:58:30.342Z", "categories": [ "hardware", "multimodal" ], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "jZIAMQEVr7", "number": 47240, "title": "ResQAdam: Residual-Preserving Adaptive Optimization for Memory-Efficient Quantization Aware Training", "abstract": "Quantization-aware training (QAT) is the most reliable way to obtain accurate low-bit large language models (LLMs), but training with AdamW keeps two dense FP32 states for every weight matrix. We find that low-rank memory-efficient optimizers such as GaLore and LDAdam degrade substantially in QAT. A considerable part of the QAT momentum lies outside its top-$r$ singular subspace, and projection-based methods discard this part at every step. We further observe that this residual remains useful when only its sign is kept. Based on these observations, we propose ResQAdam. It keeps a single BF16 momentum, applies Adam only to the $r\\times r$ coordinates of the principal subspace, and adds a normalized sign update computed from the residual. We evaluate ResQAdam with two from-scratch QAT methods and three QAT fine-tuning pipelines, with weights quantized down to 2 bits. It reduces optimizer-state memory by 71\\% to 74\\% relative to AdamW, stays close to AdamW in most settings, and lowers training throughput by less than 2\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization-Aware Training", "Model Compression", "Efficient Training" ], "forum_url": "https://openreview.net/forum?id=jZIAMQEVr7", "pdf_url": "https://openreview.net/pdf/60c43f04d4a16259133abdaf375d8779d6f46553.pdf", "updated_at": "2026-10-05T20:58:52.238Z", "categories": [ "qat", "training", "hardware" ], "matched_terms": [ "low-bit", "qat", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HALd2iUuls", "number": 47366, "title": "TAP-MoE: Task-Aware Profiling for Memory-Efficient Mixture-of-Experts Deployment", "abstract": "Mixture-of-Experts (MoE) models scale capacity through sparse activation, yet their large expert pools impose substantial storage and weight-transfer costs that hinder efficient edge deployment. When experts cannot all reside in accelerator memory, deployment must address both compression-induced quality loss and expert-loading overhead. We propose TAP-MoE, a task-aware framework that connects fine-grained expert behavior to task-level deployment decisions. It captures expert activation and sensitivity patterns through shared class-conditional statistics. Reweighting these statistics by the target task's token-class distribution yields a workload-specific profile. This profile guides sensitivity-aware mixed-precision allocation, task-conditioned expert prefetching and residency, and risk-aware output compensation. Precision allocation determines both expert sizes and quantization residuals, linking memory planning and error correction to the same task profile. The framework separates offline calibration from online execution. It preserves the pretrained routing rule and shared experts without model retraining. Our analysis bounds profile-transfer error and derives optimality conditions for a continuous precision-allocation surrogate. We evaluate TAP-MoE on OLMoE-1B-7B, Qwen1.5-MoE-A2.7B, and DeepSeek-MoE-16B-Base across three task domains under matched storage and runtime memory budgets. Compared with BF16, TAP-MoE reduces routed-expert weight storage by 75% while maintaining near-BF16 quality. Against the strongest evaluated SOTA baseline for each metric, TAP-MoE achieves 1.0-5.0% relative reductions in perplexity and 1.25-1.75× decoding speedups.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture-of-Experts; Task-Aware Profiling; Mixed-Precision Quantization; Expert Offloading; Memory-Efficient Deployment" ], "forum_url": "https://openreview.net/forum?id=HALd2iUuls", "pdf_url": "https://openreview.net/pdf/0d3fd4d6556a6686b394ce719fc55042cee1e5f8.pdf", "updated_at": "2026-10-05T20:59:06.973Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Qb1S3aJ2Ir", "number": 47399, "title": "Towards Verifying Neural Networks Against Multi-Parameter Bit-Flip Perturbations", "abstract": "Hardware faults can flip bits in the stored weights of a quantized neural network, potentially compromising its predictions. While such faults typically affect multiple parameters simultaneously, existing verifiers are limited to single-parameter perturbations due to the combinatorial explosion of possible flip locations in large networks. We present mBFV ($m$-BitFlip Verifier), an efficient verification framework that proves robustness against simultaneous bit flips across multiple parameters without explicitly enumerating these combinations. mBFV achieves this via a novel multi-parameter bound propagation technique that directly aggregates the $m$ worst-case contributions. To further tighten these bounds, mBFV employs a branch-and-bound mechanism over perturbation locations, partitioning the potential flips to smaller groups of neurons. Evaluated on 625 instances, mBFV successfully verifies 293, significantly outperforming a prior single-parameter verifier (38 verified instances) and an exact mixed-integer linear programming baseline (0 verified instances). Notably, while these baselines are restricted to single-parameter flips on small networks (up to 13k parameters), mBFV scales to verify networks with up to 1.15M parameters against up to four simultaneous parameter flips.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "dnn verification", "bit flip attack" ], "forum_url": "https://openreview.net/forum?id=Qb1S3aJ2Ir", "pdf_url": "https://openreview.net/pdf/6632d749c81ea900b242a2525456a0884eaeb5bd.pdf", "updated_at": "2026-10-05T20:59:10.587Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "dCi1PXcCQT", "number": 47422, "title": "QAttrib-MoE: Runtime-Controlled Attribution of Cross-Precision Trajectory Divergence in Mixture-of-Experts Inference", "abstract": "Low-precision mixture-of-experts inference is usually compared by endpoint accuracy, yet similar accuracy can conceal substantial changes in the autoregressive trajectory, while runtime variability can confound attribution to numerical precision. We introduce QAttrib-MoE, a pre-frozen, runtime-controlled, noise-floor-aware protocol for attributing cross-precision trajectory divergence. Same-path fresh-server variability defines an empirical control floor, cross-path divergence is the observed contrast, and excess divergence measures the contrast remaining above the two path-specific floors. The protocol freezes the evaluation cohort and decision rule before generation, counterbalances execution order, and reconstructs multiple decoding budgets by censoring a single stored trajectory rather than regenerating requests. We instantiate QAttrib-MoE on a shared GLM-5.2 checkpoint executed through FP8 and online-NVFP4 routed-expert paths. At the pre-frozen 256-token prefix, both paths exhibit 0.00%/0.00% within-path mismatch on 64 controls, while cross-path mismatch and excess divergence reach 98.44% (95% CI: 95.31-100.00%); the effect is positive on both GSM8K and MMLU-Pro and grows from 50.00% at 32 tokens to 98.44% at 256 tokens. In contrast, strict 1024-token accuracy is 64.45% versus 66.41% (difference 1.95 pp, 95% CI: -1.95 to 5.86 pp; McNemar p = 0.442). A gated teacher-forced diagnostic preserves token-ID alignment yet finds top-1 prediction flips in all 48 anchored inputs, with median first flip position 2. The results show that individually reproducible precision paths can be strongly non-equivalent at the reasoning-trajectory level without a statistically significant endpoint-accuracy shift in the evaluated setting.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language model inference", "quantization", "mixture-of-experts", "reasoning trajectories", "inference reproducibility", "FP8/NVFP4" ], "forum_url": "https://openreview.net/forum?id=dCi1PXcCQT", "pdf_url": "https://openreview.net/pdf/3f50b0f444320900535b6f64fa56cc151a5533db.pdf", "updated_at": "2026-10-05T20:59:13.373Z", "categories": [ "training", "reasoning" ], "matched_terms": [ "fp8", "low-precision", "nvfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "QztV2AB7oQ", "number": 47479, "title": "Fidelity Is Not Selection: Low-Bit Optimizer States Across Codebooks and Bit Widths", "abstract": "Quantizing Adam’s persistent moment states reduces optimizer memory, but representations with the same nominal bit width can train differently, making exhaustive comparison expensive. A cheap selector is the error a representation introduces into the next Adam update. However, this score reads a reference state once, whereas training repeatedly applies the codec to states it has already encoded; whether its ranking remains valid for selection is unclear. We test this prospectively by locking each choice before candidate training across codebooks, bit widths, training working points, and an external low-bit optimizer. At a common 130M condition across 12 independent training units, the error retained the same ordering of our three representations, yet its choice was best by point estimate under a uniform map in all 12 units, intermediate under a floating-point map in all 12, and never best under a dynamic-exponent map, where it was worst in 11. In one paired 4+4-bit condition the ordering persisted and the mismatch did not recur. Similar fidelity values also accompanied sharply different stability outcomes at larger working points. In the external optimizer, the error rule beat the published default in all six units at the longer budget but missed the lowest-loss setting in five. Thus, the failure mode is not an unstable fidelity ranking, but a stable one whose training meaning changes across settings.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "optimizer-state quantization", "low-bit optimizers", "update fidelity", "selection validity", "AdamW", "training stability" ], "forum_url": "https://openreview.net/forum?id=QztV2AB7oQ", "pdf_url": "https://openreview.net/pdf/2442ed78fc4798518079c68220308cbfbf683133.pdf", "updated_at": "2026-10-05T20:59:20.634Z", "categories": [ "training" ], "matched_terms": [ "4-bit", "bit width", "low-bit", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "3zKL0tfzM3", "number": 47545, "title": "KLIQ: KL-Informed Low-Bit Quantization via Attention-Distribution-Aware Fisher-Kronecker Curvature", "abstract": "Post-training quantization (PTQ) reduces the memory and computation costs of large language models (LLMs), but preserving accuracy at very low bit-widths remains challenging. Existing reconstruction-based curvature approximations in second-order PTQ can overlook the distribution-dependent sensitivity of attention probabilities to query and key quantization errors. We introduce KLIQ, a backpropagation-free PTQ method that uses KL divergence between full-precision and quantized attention distributions to guide curvature estimation, quantization parameter selection, and error compensation. KLIQ introduces three key innovations: (1) a Fisher-Kronecker curvature surrogate that maps a diagonal approximation of the attention-score categorical Fisher to query and key weight spaces; (2) staged functional calibration that conditions key quantization parameter selection on already quantized and compensated queries to account for query-key error interactions; and (3) KL-derived compensation statistics for compensating quantization errors propagated from preceding layers. Extensive experiments across multiple model families show that, when combined with conventional outlier-suppression techniques, KLIQ achieves state-of-the-art performance on language modeling and downstream tasks under low-bit weight-only and weight-activation quantization. The source code is available at https://anonymous.4open.science/r/KLIQ-7440/.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "large language models", "attention-distribution-aware quantization", "KL divergence", "Fisher information" ], "forum_url": "https://openreview.net/forum?id=3zKL0tfzM3", "pdf_url": "https://openreview.net/pdf/6cff50edfc6104aceb02f3687f17b02a30c49045.pdf", "updated_at": "2026-10-05T20:59:30.191Z", "categories": [ "ptq" ], "matched_terms": [ "low bit", "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RnqxAT7Dnp", "number": 47712, "title": "On Learning Expand-and-Sparsify Representations", "abstract": "Expand-and-sparsify (EaS) representations lift an input to a much higher dimension through a linear map and keep only the k largest coordinates. Existing theory builds that map at random and needs the expanded dimension to grow exponentially in the input dimension, which puts the construction out of practical reach. We ask what is gained by learning the map by gradient descent instead, and how that gain depends on how wide the expansion is. We sweep 73 OpenML classification tasks and 48 regression tasks under a protocol that holds the datasets, expanded dimensions, sparsity levels and splits fixed across every method, and we summarise each comparison by a median over datasets and a Wilcoxon signed rank test. We report four findings. First, learning the expansion helps most when the expansion is narrowest, and its advantage over a random map decays to nothing by the widest expansions we run, so a learned EaS layer can be made small. Second, with the map learned, a top-k representation that keeps at most a tenth of its coordinates is indistinguishable from a dense ReLU layer of the same expanded dimension at every expansion factor we test, while GELU and SiLU are not. Third, when the representation is forced to be binary the sparse code is the better of the two, and the reason is that binarization costs it nothing while it costs the dense layer accuracy. Fourth, how sparse the representation can be is governed by how wide it is: the sparsity level that is free rises as the expansion grows, so there is no fixed limit on how sparse it can be. The regression tasks reproduce the first finding and qualify the second and third.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "expand-and-sparsify", "sparse representations", "top-k activation", "learned projections", "random projections", "winner-take-all", "binary codes", "empirical study" ], "forum_url": "https://openreview.net/forum?id=RnqxAT7Dnp", "pdf_url": "https://openreview.net/pdf/b780b5cb452f3096ed3bc79d51645a47c137613b.pdf", "updated_at": "2026-10-05T20:59:52.547Z", "categories": [ "training" ], "matched_terms": [ "binarization", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6g7ZKY3FFp", "number": 47714, "title": "Interventional Commutativity: Effect Preservation and Reference Selection Under Weight-Only Quantisation", "abstract": "Can an intervention discovered before quantisation be reused after deployment? This question has two targets: preserving its original effect and matching an intervention re-discovered on the deployed model. We formalise the latter as interventional commutativity and audit both on behaviourally matched inputs. A round-to-nearest construction proves that exact functional equivalence can coexist with complete source-effect loss; an effect-mass bound separates mean sign agreement from magnitude retention. Empirically, a corrected 504-cell FP16/INT8 audit finds median native-relative retention 0.999 across 245 informative INT8 cells, with eight severe gaps in four configurations. However, the 243 cells whose source-effect magnitudes exceed the same descriptive guard have source-fidelity point estimates between 0.792 and 1.186. Crossed execution explains the distinction: on 117 matched template-held-out Qwen prompts, 4.481 logits of a 4.497-logit native-reference gap already exist before quantisation; only 0.016 comes from different responses to deployment. Thus a large native gap can diagnose reference selection while the transported effect's point estimate stays near its source value. Control-calibrated floors, paired gaps and cluster uncertainty make the comparison measurable; sequence checks show that its verdict remains endpoint-specific. The resulting audit quantifies preservation against the chosen reference and clarifies the target of repair.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mechanistic interpretability", "post-training quantization", "model compression", "causal interventions", "activation patching", "steering vectors", "circuit ablation", "intervention transfer", "evaluation methodology", "reliability of explanations" ], "forum_url": "https://openreview.net/forum?id=6g7ZKY3FFp", "pdf_url": "https://openreview.net/pdf/a707f12e27668a064c488d129d5bc7253c6e8b08.pdf", "updated_at": "2026-10-05T20:59:52.862Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "int8", "quantisation", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "xvJv0Fh1FD", "number": 48042, "title": "The Arithmetic Error Bar: A First-Passage Law for the Numerical Fragility of Language Models", "abstract": "A benchmark score carries an uncertainty nobody reports: change the arithmetic\nthat produced it — a numerical precision, a machine — and the number moves. On\na 1000-question set the aggregate accuracy is unchanged, 66.2% against 66.1%,\nwhile 53 answers differ and 29 items change verdict; the cancellation that keeps\nthe total still is an accident, and on a second machine of a different architecture the\nsame kind of flips cost a full point. We call the resulting band the arithmetic error\nbar — here ±1.06 points — and we show that it is predictable in advance, item by\nitem, from a single run, and that it can be paid down at a stated compute price. The\nmechanism is not the one usually assumed. Chaos predicts that reducing a pertur-\nbation makes the divergence arrive later; measurement shows that it removes it,\nand a state perturbation decays by an order of magnitude within ten tokens and\nnever grows. Generation is a hybrid system — a contracting flow interrupted by a\nquantisation — and all of its sensitivity sits in the quantisation. The distance of a\ndecision from its boundary is the geometric margin ρj = (ℓ1−ℓj)/∥w1−wj∥;\nnormalised by the state norm it defines a threshold landscape along a generation,\nand the probability that the text diverges is a first-passage problem whose only\nparameter is the amplitude of the numerical noise, measured rather than fitted. It\npredicts which answers change (AUROC 0.960 against 0.797 for entropy, on held-\nout questions), places all 29 observed correctness flips in the half of the items\nit had flagged, and predicts 43 changed answers against the 53 observed; car-\nried unchanged to the second machine it predicts exact agreement in float32,\nconfirmed on all 1500 questions, and 71 divergences in bfloat16 against 66\nobserved. Run on two further models the law ranks well on one and not on the\nother. Rather than record this as a limitation we bound it: if the law holds, many\nquestions are genuine coin tosses that no score can separate, so achievable rank-\ning accuracy has a ceiling computable from the landscape and the amplitude alone,\nbefore any comparison of precisions is run. It depends on essentially one number,\nhow often the arithmetic decides the answer, and against it all three models land at\nor near the best any score could achieve. Finally, re-running in float32 only the\ntenth of a corpus the law flags recovers 72.4% of the correctness flips and halves\nthe error bar, with the residual predicted at every budget.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mechanistic Interpretability", "First-Passage Problem", "Stability Radius", "Unembedding Geometry", "Autoregressive Generation", "Numerical Precision Sensitivity" ], "forum_url": "https://openreview.net/forum?id=xvJv0Fh1FD", "pdf_url": "https://openreview.net/pdf/c2c2347c1874b474241f5044443a24b437aec0f8.pdf", "updated_at": "2026-10-05T21:00:55.119Z", "categories": [], "matched_terms": [ "quantisation" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "wcQVK7UR0H", "number": 48169, "title": "ULMoE: Algorithm-System Co-design for Efficient Ultra-Low-Bit MoE Compression and Serving", "abstract": "Mixture-of-Experts (MoE) models now dominate frontier LLMs. However, the massive parameter count of experts makes them expensive to fit into GPU memory, and the traffic of reading expert weights on every decode step makes them expensive to serve. We show that ultra-low-bit weight quantization addresses both challenges, and that it pays off more in MoE than in dense models, because sparse routing makes the memory-bound region of grouped GEMM wider than dense GEMM, covering the practical serving batch sizes.\nAchieving practical serving speedup requires the quantization format and the kernel to be designed together. Methods such as vector quantization and expert-level mixed precision are designed for accuracy alone, and although they reach two bits, no grouped GEMM kernel can execute them efficiently. Methods designed around kernels achieve efficient serving performance but stop at four bits.\nWe present \\emph{ULMoE}, an algorithm-system co-design for accurate and efficient ultra-low-bit MoE serving. Its format spends 2.5 bits per weight and places the quantization group scales on two integer levels. The metadata for 64 weights fits in one aligned 32-bit word for efficient memory access. On top of the format we build a fast W2A16 grouped GEMM kernel and integrate it into SGLang.\nAcross seven MoE models from three families, ULMoE compresses expert weights ${\\sim}6\\times$ at $2.6$ to $2.8$ effective bits and loses $1.8$ average accuracy points. At matched interactivity it delivers $5.7\\times$ the serving throughput of BF16 on Qwen3.5-35B and $1.7\\times$ that of Int4 on Qwen3.5-122B.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture of Experts", "Large Language Models", "Model Compression" ], "forum_url": "https://openreview.net/forum?id=wcQVK7UR0H", "pdf_url": "https://openreview.net/pdf/3d74914e18923bb15b9beae8785c1f1d28529050.pdf", "updated_at": "2026-10-05T21:01:12.724Z", "categories": [ "hardware" ], "matched_terms": [ "int4", "low-bit", "mixed precision", "quantization", "w2a16" ], "review_status": "abstract", "summary": "ULMoE 针对 MoE 的 expert 存储与每个 decode 步的读权重流量，同时设计超低位格式和执行内核。其 group scale 使用两级整数编码，让 64 个权重的元数据装入一个对齐的 32-bit word，并实现集成到 SGLang 的 W2A16 grouped GEMM。", "contribution": "在三个家族的七个 MoE 模型上，expert 权重约压缩 6 倍，有效位宽为 2.6–2.8 bits，平均准确率损失 1.8 个百分点；在匹配交互性条件下，Qwen3.5-35B 吞吐为 BF16 的 5.7 倍，Qwen3.5-122B 为 Int4 的 1.7 倍。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。压缩倍数针对 expert 权重，吞吐数字分别对应不同模型/基线且要求匹配交互性；不能视为统一的整模型压缩或任意 batch 加速。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "MoE expert 权重格式与 W2A16 grouped GEMM 服务内核", "bit_width": "格式 2.5 bpw；计入开销的有效位宽 2.6–2.8 bits；W2A16 内核", "evidence": [ { "location": "官方摘要", "claim": "七个 MoE 模型的 expert 权重约压缩 6 倍，平均准确率损失 1.8 个百分点。" }, { "location": "官方摘要", "claim": "匹配交互性下吞吐提升：Qwen3.5-35B 对 BF16 为 5.7 倍，Qwen3.5-122B 对 Int4 为 1.7 倍。" } ] }, { "id": "klXcnEEMjl", "number": 48302, "title": "TTQ: Activation-Aware Test-Time Quantization to Accelerate LLM Inference On The Fly", "abstract": "To tackle the high computational cost of large foundation models, post-training activation-aware quantization methods have been widely used. However, because these methods rely heavily on calibration data, calibration–deployment mismatch may arise on unseen downstream tasks. We propose test-time quantization (TTQ), which constructs prompt-specific low-bit weights online from the current prompt and reuses them during decoding. TTQ removes the need for a predetermined calibration dataset when constructing these weights and targets decode-time weight traffic and integer execution rather than resident model-memory reduction. Experiments across language, vision-language, and vision-language-action models show improved low-bit accuracy over offline baselines in several settings while retaining most of AWQ’s decode throughput at a measurable prefill cost.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Activation-Aware Compression", "Test-Time Compute", "LLM" ], "forum_url": "https://openreview.net/forum?id=klXcnEEMjl", "pdf_url": "https://openreview.net/pdf/b5fc5020738228ed93e1ed0f3184605a820c896a.pdf", "updated_at": "2026-10-05T21:01:32.366Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "5TBeW5XhGB", "number": 48349, "title": "PROBE: Frontier Coding Agents Find Different Bugs Than Maintainers Fix", "abstract": "Proactive bug discovery is an important software-engineering problem. Even though coding agents are already widely deployed in this domain, their capabilities remain understudied. In particular, the challenging task of finding all bugs that human maintainers would repair across an entire repository has not been studied yet. To close this gap, we introduce PROBE, a benchmark consisting of all 2,457 bugs fixed over the course of 18 months across 20 diverse GitHub repositories, together with a method for matching candidate findings to this partial reference set.\n\nWe evaluate coding agents across frontier models and find they all achieve a recall of below 21\\% and a precision of less than 9\\% with respect to the reference set. While stronger models increase recall substantially from 8\\% to 17\\%, they do not improve precision. Surprisingly, despite the low precision, LLM and human evaluators still judge 85\\% of findings valid. Crucially, we find that found bug distributions are similar across agents but differ substantially from those of bugs fixed by maintainers. The practical relevance of bug discovery and the low recall of frontier models make this benchmark a valuable resource for both model users and developers.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Bug Finding", "Coding Agents", "LLMs", "Benchmarking" ], "forum_url": "https://openreview.net/forum?id=5TBeW5XhGB", "pdf_url": "https://openreview.net/pdf/e749d1a6fedd290798f78b489045c9d15850d8d2.pdf", "updated_at": "2026-10-05T21:01:39.484Z", "categories": [], "matched_terms": [ "low precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1dlWJpwRNd", "number": 48392, "title": "CodedNN: Coding-Theoretic Redundancy for Robust Neural Classification", "abstract": "Neural classifiers can degrade substantially under input corruptions and perturbations to learned features. We investigate a simple way to improve reliability: encode labels into redundant binary outputs and train a shared-backbone network to predict the resulting codeword. Coded Neural Networks (CodedNN) preserve the original label coordinates in multilabel classification (MLC), while redundant parity coordinates provide additional supervision and can optionally be exploited by belief-propagation (BP) decoding; in single-label classification (SLC), classes are represented by redundant binary codewords and prediction is performed by soft codeword matching. A central difficulty is that learned output errors are strongly dependent, so the independent-channel abstraction underlying classical error-correcting codes does not accurately describe neural predictions. We therefore develop a dependence-aware analysis of aggregate codeword errors, including a distribution-free correction-radius bound and a Gaussian-margin reference model. Empirically, redundant outputs provide gains in both clean and corrupted regimes. Across 13 clean MLC benchmarks, raw CodedNN improves micro-F1 over a matched binary-cross-entropy predictor by 1.5 percentage points on average, while explicit decoding provides additional robustness in some corrupted regimes. On CIFAR-10-C, redundant 8-bit output representations improve mean corruption accuracy by up to 1.67 percentage points over softmax with essentially unchanged clean accuracy, and on ImageNet the coded model exceeds softmax by 12.2 points under the strongest tested feature perturbation. These results show that structured output redundancy can improve neural reliability even when classical coding assumptions fail, providing a framework for exploiting redundancy through both representation learning and decoding.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Robust neural classification", "error-correcting codes", "structured output redundancy", "covariance conditioning", "correlated prediction errors", "multilabel classification", "reliable AI" ], "forum_url": "https://openreview.net/forum?id=1dlWJpwRNd", "pdf_url": "https://openreview.net/pdf/17fdf179ffd9cf0961228654e3a9f2c3872a212c.pdf", "updated_at": "2026-10-05T21:01:47.224Z", "categories": [ "theory" ], "matched_terms": [ "8-bit", "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZR9OJLbqQN", "number": 48417, "title": "QSLR: A Quantized + Sparse + Low-Rank Decomposition of LLM Weights", "abstract": "Post-training Quantization (PTQ) of Large Language Models (LLMs) is a popular approach to reduce inference costs but can be restrictive under aggressive compression constraints. Hence, recent methods augment quantized weights with (i) sparse corrections $\\mathbf{Q}+\\mathbf{S}$ to capture outliers (e.g., SpQR is a specialized solver for the quantized-plus-sparse format) or (ii) low-rank corrections $\\mathbf{Q}+\\mathbf{L}$ to capture shared structure. \nWe combine both in a unified quantized-plus-sparse-plus-low-rank representation, $\\mathbf{W}\\approx\\mathbf{Q}+\\mathbf{S}+\\mathbf{L}$,and propose an ADMM-based algorithm $\\texttt{QSLR}$ that fits all three components jointly. Our novel ADMM framework builds upon computationally efficient subroutines and is agnostic to the quantization format. Extensive experiments demonstrate our joint optimization improves upon sequential fitting procedures. $\\texttt{QSLR}$ reduces the WikiText-2 perplexity of a Llama-3.2-1B from $16.79$ to $14.82$ compared to fitting SpQR followed by a rank-16 correction, at the same compression budget. Additionally, we show that splitting the allocation between sparsity and low rank can improve performance over using either correction alone. For Qwen3-4B with (sym 2-bit) quantization at 2.47 bits per weight, using $\\texttt{QSLR}$ to fit all three allocations, the mixed representation achieves a WikiText-2 perplexity of 9.76, reducing the perplexity gap to the dense model by 30.9% relative to $\\mathbf{Q} + \\mathbf{S}$ and 35.2% relative to $\\mathbf{Q} + \\mathbf{L}$. Finally, we design GPU kernels that support $\\mathbf{Q}+\\mathbf{S}+\\mathbf{L}$ and establish speedups over FP16 dense computation in matrix-multiplication and decoding benchmarks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Pruning", "Low-Rank", "Efficient ML", "Model Compression", "Non-Uniform Compression", "ADMM" ], "forum_url": "https://openreview.net/forum?id=ZR9OJLbqQN", "pdf_url": "https://openreview.net/pdf/d6017a168b10c46628b2a68cf7497bf954df3592.pdf", "updated_at": "2026-10-05T21:01:52.848Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "2-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "U97XcRRIm1", "number": 48454, "title": "Causal Models of Convolutional Binary Spiking Neural Networks for Explanation", "abstract": "We propose a novel logic-based causal approach to explaining convolutional binary spiking neural networks (C-BSNNs). Specifically, we provide an SMT (Satisfiability Modulo Theories) encoding of the binary causal model (BCM) underlying a C-BSNN and use it to compute formally correct abductive explanations for its outputs. We demonstrate that our SMT-based causal analysis extends beyond static inputs, preserving the recurrent temporal dependencies of spiking dynamics to support formal explanations of complex event-based DVS-Gesture data alongside static MNIST. To enable training of our networks with Batch Normalization while preserving the Boolean activations of neuronal units required for mapping them to BCMs and subsequent SMT encoding, we introduce a novel Batch Normalization absorption procedure. Finally, we compare our logic-based causal approach with SHAP with respect to the relevance of the elements included in an explanation. While our approach provides a formal guarantee that all such elements are relevant, SHAP provides no such guarantee.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Formal explainability", "Abductive explanations", "Spiking neural networks", "Binary neural networks", "SMT solver", "Causal models", "Neuromorphic computing" ], "forum_url": "https://openreview.net/forum?id=U97XcRRIm1", "pdf_url": "https://openreview.net/pdf/568e1f416fb6026fa6ef52ae5a9f9e6df1888f9e.pdf", "updated_at": "2026-10-05T21:01:56.403Z", "categories": [], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "pMEd8dqKuA", "number": 48502, "title": "GenomeOcean Anywhere: Private WebGPU Inference for Genome MoEs", "abstract": "Genome foundation models are most useful where sequences are generated, yet the largest models need datacenter accelerators and a place to send private DNA. We ask whether a 15-billion-parameter genome mixture-of-experts (MoE) model can instead run on volunteers' web browsers, with the experts spread across many untrusted devices, without changing its predictions and without revealing the sequence to any single device. We build a system in which a trusted coordinator runs attention and routing while browser workers run every expert feed-forward network through hand-written WebGPU kernels, and we protect the expert inputs with real-valued Lagrange coded computing: each worker receives only a Gaussian-padded share, computes the expert's linear maps, and the coordinator decodes from any two of three workers. On GenomeOcean-MoE (8 experts, top-2 routing, 24 layers), the browser path matches native llama.cpp at every quantization level, the distributed path stays at the BF16 numerical noise floor (KL 0.0036 nats per token), and an unfitted latency model predicts decode time within 0.74% (median) under emulated wide-area links. We first show that plaintext expert inputs are not private: a probe recovers the token from a single vector at every depth, and one worker can identify the source genome from 300 unordered tokens with 92% accuracy. With coded experts, an adaptive attacker trained on shares falls to the most-frequent-token baseline, one worker's information about each token is bounded below one bit per forward pass, and the fidelity cost stays below the BF16 noise floor; in Chrome, coded decoding runs at 220 to 376 ms per token, depending on how much of the routing is hidden, and continues without replicas when a worker fails.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "genome foundation models", "WebGPU", "decentralized inference", "edge AI", "efficient inference", "computational genomics" ], "forum_url": "https://openreview.net/forum?id=pMEd8dqKuA", "pdf_url": "https://openreview.net/pdf/e435696d0f7b53a0269da183bd9cdfc1b2ba2a87.pdf", "updated_at": "2026-10-05T21:02:03.091Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "FjGiCaa2p1", "number": 48577, "title": "Inference-Time Machine Unlearning via Gated Activation Redirection", "abstract": "Large Language Models (LLMs) memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. \nMachine unlearning seeks to remove the influence of a targeted forget set \\mathcal{D}_f while preserving model performance, ideally approximating a model retrained from scratch without \\mathcal{D}_f. Once an LLM is in use, every new request to make it forget specific content demands updating its weights. However, unlearning through parameter updates is expensive, hard to audit, and can be undone by quantization.\nWe show that unlearning can be enforced entirely at inference time, without training, gradients, or weight changes. We introduce Inference-Time Unlearning via Gated Activation Redirection (GUARD-IT), a training- and gradient-free method that unlearns via input-dependent activation steering at inference time.\n\\gls{guard} stores the content to be forgotten as a small library of activation directions, and during inference, it routes each query through a similarity gate that activates only for relevant directions and applies them as a norm-preserving rotation of the residual stream, while unrelated queries pass through the unmodified model.\nThe same design carries across three model families and nine checkpoints from 0.8B to 8B parameters, and new forget requests are absorbed by one offline pass of forward passes.\nOn TOFU, against 16 gradient-based and inference-time baselines, and on MUSE and WMDP, GUARD-IT forgets without breaking the model, and on TOFU it is the only method that suppresses memorization in every Llama configuration without collapsing: it keeps utility and fluent generation in every configuration, moves the model's output distribution closest to a model that never saw the forgotten data on the \\texttt{forget01} split, survives 4- and 8-bit quantization and ten sequential forget requests, and holds under jailbreak attacks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "machine unlearning", "inference-time adaptation", "activation steering", "representation editing and engineering", "residual stream intervention", "gated networks", "model privacy", "data deletion", "LLM safety", "interpretability", "activation space geometry", "activation engineering", "parameter-efficient methods", "quantization robustness" ], "forum_url": "https://openreview.net/forum?id=FjGiCaa2p1", "pdf_url": "https://openreview.net/pdf/bc366e8908d141cdcfcbd7123b136f3879b2adac.pdf", "updated_at": "2026-10-05T21:02:17.530Z", "categories": [ "ptq", "training", "theory" ], "matched_terms": [ "8-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ELBs8YHQwb", "number": 48607, "title": "How You Build Superintelligence: From Scaling Laws to Construction Laws", "abstract": "A capability target does not specify the system that should be built to attain it. Scaling laws predict performance along a chosen development path, yet even exact knowledge of a scaling trajectory can leave the optimal resource allocation unidentified. We formulate construction laws: target-conditioned maps from identified mechanisms and physical resource constraints to a complete executable recipe, a forecast, and an auditable qualification test. We formalize superintelligence as an expert-relative capability specification that fixes the task population, within-task reliability, across-task breadth, and common information conditions. Our main inverse theorem converts uncertainty in forward predictions, resource costs, and task breadth into a guarantee on a qualifying construction; the controlling quantity is the price of target tightening, not prediction error alone. We then derive explicit construction laws for resource allocation, numerical precision and training data, reusable information, and conditional specialization, including an exact rule for jointly choosing sharing strength and specialist data. Prospectively frozen neural constructions reduce packed-weight storage by 1.33–1.59× while satisfying a predeclared 0.010-nat noninferiority criterion in all six target–seed comparisons. In a controlled hidden-support neural task, reusable structure reaches the same 80% reliability target with 32 rather than 256 task queries, while task-private structure removes the advantage. In external multi-hop retrieval, inherited transition structure consistently improves low-budget recovery, but a preregistered 2× budget-crossing forecast misses validation by 0.66 percentage points, illustrating the instability of discrete construction decisions near a target boundary. Construction laws turn capability scaling from a forward forecast into an inverse question: which system reaches the target, at what resource cost, and with what evidence?", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "scaling laws", "construction laws", "learning theory", "resource allocation", "inverse design", "model scaling", "quantization", "conditional computation", "computational inheritance", "superintelligence" ], "forum_url": "https://openreview.net/forum?id=ELBs8YHQwb", "pdf_url": "https://openreview.net/pdf/294bea1f6b917f855967ccb0373df12f85dd247a.pdf", "updated_at": "2026-10-05T21:02:21.173Z", "categories": [ "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nVZXNXJGIv", "number": 48652, "title": "Timescale Identification from $k$-Step Transition Operators: The Structural Entropy Spectrum", "abstract": "Agents and world models discretise experience into a transition graph at one fixed step, yet behaviour has structure at several timescales. Which step should the graph be built at, and which timescales are present? Multiscale tools (Markov stability, the map equation at Markov time, implied timescales of Markov state models) sweep a time parameter but attach no significance statement to the scales they report. We make the step count k an argument of a clustering objective on the k-step operator P^k and calibrate it against a matched null: chains whose stationary flow matrix has identical margins and dwell (self-loop) mass. Our theory uses structural entropy (SE), a two-level code-length objective: the structural entropy spectrum SE(k), its null SE_null(k) and the calibrated spectrum ΔSE(k) = SE_null(k) − SE(k). We prove that the k-dependence factors through an exact identity, SI(k;C) = Σ_a F_aa(k) log2(1/V_a) (k-step module persistence × module code length), prove that when metastable blocks have stationary volume < 1/2 the metastable partition is the unique global optimizer of the idealized in-window operator, and sketch a Simon–Ando extension to nearly uncoupled chains on a window whose endpoints scale as 1/(1−λ) for the fast and slow eigenvalue clusters, so the optimal partition path coarsens through a planted hierarchy in order. On a 36-state three-level chain, the path recovers both planted levels in order at calibrated effect sizes up to 1.8 bits, although at this fixture's modest scale separation ΔSE(k) shows no plateaus; a single-scale expander control is featureless; the spectrum is estimable from a single trajectory. On latent transition graphs from a pixel-based JEPA world model, the one-step graph's calibrated effect falls below the preregistered 0.1-bit floor, while the spectrum peaks reproducibly at k* = 11–16 with 1.6–1.9 bits across five training seeds; a self-loop control shows that this structure survives removal of dwell but that the peak location is a dwell-inflated step count (two state changes). The calibration is not SE-specific: wrapped around the map equation it passes the planted-hierarchy, single-scale and structureless controls, and around Markov stability all but the expander control; both recover two balanced blocks, SE's boundary case, and the scale boundaries move with the objective. All decisive experiments were preregistered with pass/fail gates and are reported verbatim.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "structural entropy", "Markov chains", "timescale identification", "metastability", "spectral methods", "world models" ], "forum_url": "https://openreview.net/forum?id=nVZXNXJGIv", "pdf_url": "https://openreview.net/pdf/ee2065ec7babc00fba0aeaba0eacc48d8d918338.pdf", "updated_at": "2026-10-05T21:02:27.205Z", "categories": [ "ptq", "training" ], "matched_terms": [ "1-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "R0kHnQCcBJ", "number": 48726, "title": "From Hidden States to Probability Measures: Measuring Quantization-Induced Deformation in Large Language Models", "abstract": "Attention rows and next-token output distributions are probability measures. Prior work evaluates compressed models with accuracy, output KL, and answer flips; we extend this output-level evaluation to attention routing and treat both distributions as a measure-valued diagnostic layer. We compare discrepancies with and without ground geometry to characterize quantization-induced redistribution over positions and vocabulary outcomes. Using same-prefix, teacher-forced comparisons isolates local deformation from free-running history divergence. In the tested, comparatively mild INT8-versus-FP8 weights-only comparison, attention KL and \\(W_1\\) distinguish all six model-domain pairs with a consistent direction; perplexity is significant in only \\(3/6\\) and disagrees in direction even among significant pairs. Output KL has one directional exception. Under stronger INT4 weights-only quantization, perplexity also detects degradation. Adding activation quantization reverses the INT8/FP8 ordering in every tested pair, with model- and diagnostic-dependent magnitude and architecture-dependent layerwise shape. These diagnostics measure local redistribution and are not downstream-accuracy surrogates or full path-law comparisons.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "transformers", "attention", "Wasserstein distance", "KL divergence", "language models" ], "forum_url": "https://openreview.net/forum?id=R0kHnQCcBJ", "pdf_url": "https://openreview.net/pdf/983010805c03345f79fe54d0d1bef8736d8a753a.pdf", "updated_at": "2026-10-05T21:02:37.759Z", "categories": [ "training" ], "matched_terms": [ "fp8", "int4", "int8", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1Wi94G6EF8", "number": 48772, "title": "Binary Expansion Group Intersection Network", "abstract": "Conditional independence is fundamental to modern statistics and machine learning, yet beyond special parametric families it rarely admits an exact covariance-based characterization. We introduce the binary expansion group intersection network (BEGIN), a distribution-free graphical framework for representing conditional independence in multivariate binary data and bit-encoded multinomial variables. We show that for arbitrary binary random vectors, conditional independence is equivalent to a zero cross block of a generalized Schur complement. Its entries are probability-weighted sums of conditional covariances, with signs determined by conditioning interactions. The representation encompasses Ising models, higher-order binary models, and categorical graphical models, including distributions with structural zeros. We characterize the rank of the interaction covariance and establish Markov properties and conditions for concatenating BEGIN graphs. We test conditional independence using the maximum absolute value of the estimated entries in the BEGIN cross block, and we show that this test attains the optimal sample order under sparse interaction alternatives. Simulations show the power gain under sparsity and the computational efficiency of the interaction scan. An analysis of categorical Mushroom records shows how the maximizing binary interactions explain detected conditional associations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "conditional independence", "graphical models", "binary data", "Schur complement", "distribution-free methods", "group theory" ], "forum_url": "https://openreview.net/forum?id=1Wi94G6EF8", "pdf_url": "https://openreview.net/pdf/a8cbad883d0d89d12a74fa81e76074d9ad44a581.pdf", "updated_at": "2026-10-05T21:02:42.865Z", "categories": [], "matched_terms": [ "binary" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7KCUTpkPaS", "number": 48882, "title": "CipherGenome: Homomorphic Inference for Genomic Mixture-of-Experts", "abstract": "Genome foundation models are growing into sparse mixture-of-experts (MoE) networks whose expert weights no longer fit on the machines that hold the sequences, yet sending a private genome to rented accelerators exposes it: we show that a single server hosting one expert recovers the input nucleotides with 99.8% top-1 accuracy. We present CipherGenome, a protocol that keeps the embedding, attention and router of a 15.1B-parameter MoE genome model on a trusted thin client and outsources every expert projection, 95.8% of the parameters, to untrusted and possibly colluding GPU servers under module-LWE encryption. The design exploits three structural facts: expert layers are linear between two SwiGLU gates, expert weights are public, and GPU integer tensor cores can evaluate a ciphertext–weight product exactly modulo $2^{48}$ in a single GEMM. The client evaluates the nonlinearity exactly and re-encrypts with fresh secrets, so no polynomial approximation or bootstrapping is ever needed. On 72 windows from 12 bacterial genomes, encryption adds $2.54 \\times 10^{-4}$ nats per token of KL divergence (95% CI upper bound $3.95 \\times 10^{-4}$), below a pre-registered non-inferiority margin and indistinguishable from bf16 inference, while the same inversion attack falls to chance level. A reusable public hint cuts end-to-end latency by 3.54 times, wire compression reduces traffic 6.8 times, per-layer padding reduces routing leakage from 54.9% to 8.9% accuracy, and HE-compatible int4 experts remain non-inferior to their plaintext counterparts. Per expert and token, the server-side cost is more than six orders of magnitude below a CKKS baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "genome foundation models", "homomorphic encryption", "mixture of experts", "privacy-preserving inference", "distributed inference", "confidential computing" ], "forum_url": "https://openreview.net/forum?id=7KCUTpkPaS", "pdf_url": "https://openreview.net/pdf/10fc047edcf6947fcb1e357500cb16876c9ea799.pdf", "updated_at": "2026-10-05T21:02:57.116Z", "categories": [ "hardware" ], "matched_terms": [ "int4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "78Tj0cLhki", "number": 48917, "title": "Federated Fine-Tuning of Code Language Models: A Comparison of FedAvg, FedProx, and FedAdam UnderLanguage-Heterogeneous Clients", "abstract": "Large language models specialized for source code excel at program synthesis, but further domain gains require fine-tuning on codebases that are often proprietary and cannot be pooled centrally. Federated Learning (FL) offers a path to collaboratively fine-tune a shared model without exchanging raw data, yet the comparative behavior of FL aggregation algorithms on code generation—particularly under realistic, language-based client heterogeneity—has not been measured systematically. We present a controlled empirical study that fine-tunes Qwen2.5-Coder-1.5B with Quantized Low-Rank Adaptation (QLoRA) inside the Flower framework, comparing three aggregation algorithms (FedAvg, FedProx, FedAdam) across three data-partitioning strategies (random IID, Dirichlet α=0.5, and a realistic non-IID split where each client specializes in one programming language). The resulting global models are benchmarked against a centralized upper bound on HumanEval+, MBPP+, and MultiPL-E using greedy pass@1, with three seeds per condition and paired significance testing. Two findings emerge. First, on algorithm choice, FedAvg consistently matches or exceeds FedProx and substantially outperforms FedAdam in every partition (HumanEval+ pass@1 of 40.0, 36.4, 29.1 for FedAvg across IID, Dirichlet, and by-language), whereas FedAdam is unstable and collapses to 11.6±5.8 under quantity skew; contrary to the expectation that server-adaptive optimization best handles heterogeneity, the simplest aggregator is the most accurate and robust. Second, on the cost of federation, the best federated model is statistically indistinguishable from the centralized reference (38.2) under IID and Dirichlet but lags by ≈9 points under the by-language split. A per-language MultiPL-E analysis reveals a stable competence gradient from Python and JavaScript down to Go, without collapse toward a dominant language. We release adapters and configurations for reproducibility.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Federated Learning", "Code Generation", "Large Language Models", "QLoRA", "Parameter-Efficient Fine-Tuning", "Non-IID Data", "HumanEval", "Qwen2.5-Coder." ], "forum_url": "https://openreview.net/forum?id=78Tj0cLhki", "pdf_url": "https://openreview.net/pdf/95ff0e956721d11b02e5dca1bc56f5ff433a8525.pdf", "updated_at": "2026-10-05T21:03:01.525Z", "categories": [ "training", "reasoning" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LfMs6G4Duf", "number": 49034, "title": "Tail-W4: Predictive Precision Scheduling for the Long Tail of RL Rollouts", "abstract": "Reinforcement learning (RL) post-training can be dominated by rollout, where synchronous execution waits for a few long responses to finish. Weight-only 4-bit quantization (W4) is promising for addressing those small-batch tails because it reduces the cost of loading model weights. However, naïvely applying W4 in RL post-training does not guarantee faster RL steps: quantized sampling can generate more tokens, increasing both decoding and downstream RL work. To address this challenge, we propose TAIL-W4, a rollout precision scheduler that chooses when to switch from BF16 to W4 by jointly predicting response lengths and execution costs. Reserving W4 for the tail can mitigate token inflation, but switching too early risks extra work and switching too late misses acceleration. Fixed schedules require repeated trials to tune and cannot fully adapt to changing training workloads. TAIL-W4 calibrates estimates of token inflation, updates them from training rollouts, and periodically predicts the benefit of switching to W4. On Qwen3.5-4B, TAIL-W4 achieves 1.38× rollout and 1.26× measured RL-step speedup over BF16, versus 1.12× and 1.03× for uniform W4 with the same quantized weights.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning Post Training; LLM Reasoning; Rollout Efficiency; Weight-Only Quantization LoRA; Long-tail Generation; Systems for Machien Learning" ], "forum_url": "https://openreview.net/forum?id=LfMs6G4Duf", "pdf_url": "https://openreview.net/pdf/7754d1495cadf3ebdfcaf9ec4dd70a9c80edcfe8.pdf", "updated_at": "2026-10-05T21:03:17.893Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "4-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zOQ0HAXFKi", "number": 49185, "title": "FastVLM2: Efficient On-Device Dense Grounding for Generalist VLMs", "abstract": "Generalist vision-language models (VLMs) with dense grounding capabilities can perform a broad range of classical vision tasks through a single model and natural-language interface. However, efficient on-device inference with such models faces two complementary bottlenecks: costly high-resolution visual encoding and lengthy autoregressive decoding of numerous spatial locations. We introduce FastVLM2, a unified framework that addresses both. For input-side efficiency, we introduce a native-resolution strategy that adapts pretrained convolutional and hybrid convolution-transformer backbones (ConvNeXt and FastViTHD, respectively) through VLM instruction tuning, enabling arbitrary image resolutions and aspect ratios without additional vision pretraining, image tiling, or explicit layout tokens. On a MacBook with an M1 Max, FastVLM2-4B achieves VQA accuracy comparable to Qwen3-VL-4B with $4.6\\times$ lower prefill latency. For output-side efficiency, we use quantized location tokens to compactly represent grounding outputs, reducing the output-token budget by over $60\\%$ relative to textual coordinate representations. To recover the localization accuracy lost to quantization, we introduce a simple Gaussian distance-aware soft cross-entropy objective that brings quantized grounding close to textual-coordinate performance while preserving this decoding advantage. At higher grounding accuracy, FastVLM2-1.8B achieves $7.3\\times$ lower end-to-end latency than Rex-Omni. We further introduce a training-free, geometry-aware box confidence estimator that directly exploits the predictive distributions of quantized location tokens. We show that this scoring rule generalizes across three independently trained grounding models and consistently improves COCO mAP by more than $8$ points over the score-free ranking protocol used in Rex-Omni. Together, FastVLM2 makes accurate, confidence-aware dense grounding practical for on-device generalist VLMs.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vision-Language Models", "Efficient Inference", "Visual Grounding", "Native-Resolution Vision", "Structured Prediction", "Confidence Estimation" ], "forum_url": "https://openreview.net/forum?id=zOQ0HAXFKi", "pdf_url": "https://openreview.net/pdf/af7683fbfd415d8049641f0fb9e16ba0db2b313c.pdf", "updated_at": "2026-10-05T21:03:42.527Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "fhLTsfDLtr", "number": 49212, "title": "MORSE: Mixture of Representations from Sparsified Experts", "abstract": "Recent work on quantization (e.g., TurboQuant) and dimensionality reduction (e.g., Matryoshka Representation Learning; MRL) demonstrates that text embeddings contain substantial representational redundancy. However, enforcing extreme sparsity (retaining only a small number of active dimensions) can reduce embedding generality and favor task-specific performance, as observed in work on\nContrastive Sparse Representations (CSR) and CSRv2. We show that such task-specific, low-active dimension representations can be learned by training a light-weight adapter (namely, Sparsified Expert) over any base model using a novel weighted-top-k objective. Our approach departs from prior methods that either (1) concentrate information in predetermined leading dimensions (i.e., prefix-based),\nas in MRL and Matryoshka Adapters, or (2) project embeddings into a higher-dimensional space to identify sparse representational axes, as in CSR and CSRv2. We then introduce Mixture of Representations from Sparsified Experts (MORSE), which combines representations from multiple SEs to recover task-agnostic performance while maintaining sparsity in representations. MORSE occupies the previously underexplored regime between moderately sparse, task-agnostic representations and extremely sparse, task-specific ones, revealing a Pareto frontier between representational sparsity and generality. Across three base models, our SE outperforms other methods at ultra-low dimensions and MORSE improves over others on cross-lingual generality.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "efficient representation learning", "text embeddings", "information retrieval", "embedding compression", "sparse representations" ], "forum_url": "https://openreview.net/forum?id=fhLTsfDLtr", "pdf_url": "https://openreview.net/pdf/f932f2f30925df00c25c8ed936ece3e0280db56c.pdf", "updated_at": "2026-10-05T21:03:47.241Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DR0qHuYqlL", "number": 49216, "title": "Unmasking the Hidden Bias, Fairness, and Safety Costs of Compression with Mixture-of-Expert Models", "abstract": "Mixture-of-Experts (MoE) language models are often compressed for deployment, yet the fairness, bias, and safety effects of these interventions remain poorly understood, particularly for Expert Compression (EC) algorithms and their required calibration data. We investigate how compression algorithms and calibration data mixtures affect fairness and safety behaviour in MoE language models. Across three diverse MoE architectures, we find that EC is highly sensitive to calibration data composition. Expert merging produces the largest regressions on average, although the relative behaviour of merging algorithms varies substantially across architectures. Expert pruning is generally less disruptive, while quantization and weight sparsity are more stable and less sensitive to calibration data. We show that benchmark level improvements can be misleading: some aggressive compression settings appear more stable on selected fairness benchmarks while degrading task accuracy and instruction following, and prompt matched generation diagnostics reveal greater behavioural drift from the uncompressed baseline. Targeted interventions show that pruning just three of the 6,144 experts in Qwen3-30B-A3B can turn a refusal of a harmful AdvBench request into compliance, whereas quantizing the same experts to 2 bits preserves the refusal. These results demonstrate that the fairness and safety effects of MoE compression are neither simple nor monotonic and highlight calibration data, model architecture, and algorithm selection as key considerations when compressing MoEs for deployment.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "bias", "fairness", "moe", "mixture-of-experts", "compression", "quantization", "sparsity", "pruning", "expert merging" ], "forum_url": "https://openreview.net/forum?id=DR0qHuYqlL", "pdf_url": "https://openreview.net/pdf/d1143bf4c62bfc3f9c85e9a24e6a7c17cdbc764d.pdf", "updated_at": "2026-10-05T21:03:47.423Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MLdEdeBNWU", "number": 49294, "title": "RACE: RoPE-Aligned Attention Compression Based On Query-Key Activation Energies", "abstract": "Head-dimension compression can reduce the memory required by the key-value (KV) cache, but existing methods often reconstruct full-dimensional representations during inference, adding significant decoding FLOPs overhead. We formalize the design considerations for converting pretrained attention into a natively reduced-dimensional form that reduces both KV-cache storage and attention computation using a calibration set. Specifically, we consider reductions that (1) apply a shared projection to queries and keys so that they remain in a common reduced feature space, (2) derive this projection jointly from query and key activations to better preserve their interaction, and (3) are compatible with the pretrained RoPE transformation. To obtain a reconstruction-free realization of this reduction, we restrict it to coordinate selection and theoretically show that RoPE compatibility then holds if and only if complete rotary pairs are retained. Building on this characterization, we introduce RACE (RoPE-Aligned Attention Compression based on activation Energy), a post-training method that selects rotary pairs using their joint query-key activation energy. We theoretically show that this criterion minimizes a separable, calibration-dependent upper bound on post-RoPE query-key product approximation error. The resulting reductions are folded into the model weights and require neither runtime reconstruction nor specialized kernels. Across three model families and different compression levels, RACE outperforms structured pruning and remains competitive with reconstruction-based methods on language-modeling, zero-shot, and long-context tasks. At 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8%, yielding 1.79× higher throughput, 43.2% lower time-to-first-token (TTFT), and 28.1% lower allocated GPU memory in comparison to an uncompressed model. We also show that RACE is compatible with token eviction or compaction as well as quantization based KV-cache compression methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-cache compression", "Efficient LLM inference", "Attention compression", "Rotary positional embeddings" ], "forum_url": "https://openreview.net/forum?id=MLdEdeBNWU", "pdf_url": "https://openreview.net/pdf/1b20353a32e224e0ba98b85040195698db4678cd.pdf", "updated_at": "2026-10-05T21:03:57.278Z", "categories": [ "ptq", "kv-cache", "hardware", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IlNqX0B95x", "number": 49407, "title": "Offline Hidden-State Distillation from Compact Teacher Caches", "abstract": "For language models, logit distillation is now routinely performed offline: the teacher's next-token distributions are computed once, ahead of training, and then cached for use in subsequent training runs.\nStoring the full distributions would be prohibitive, but truncation or sampling compresses the cache by orders of magnitude without significantly impacting performance.\nHidden-state distillation, which additionally uses the teacher's hidden states to supervise the student's intermediate computation, has no compact offline counterpart.\nExisting methods either perform teacher inference within the training loop, slowing down training significantly, or require enormous amounts of storage.\nWe show that a per-token random projection of the teacher's hidden states, coarsely quantized, lets hidden-state distillation run from a cache of the same order of size as the logit cache, at the training throughput of logit distillation.\nWith the hidden states compressed 32$\\times$, students distilled from four teachers of 1.5B to 8B parameters need at most about 2\\% more training tokens than students trained on the full hidden states to reach the same held-out loss at 1B tokens.\nOver training runs of up to 15B tokens, they need about 1\\% more tokens than the full-state students and match their downstream accuracy, while a student distilled from logits alone needs 1.4 to 1.7 times as many.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Knowledge distillation", "hidden-state distillation", "feature distillation", "offline distillation", "teacher caching", "random projections", "activation compression", "quantization", "unbiased estimation", "language model pre-training", "efficient training" ], "forum_url": "https://openreview.net/forum?id=IlNqX0B95x", "pdf_url": "https://openreview.net/pdf/7432f5a4b4cadec2f0e9266206d69e49ede3e699.pdf", "updated_at": "2026-10-05T21:04:13.255Z", "categories": [ "hardware" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "TZJ8Z6Dl0M", "number": 49413, "title": "Quantar: Quantizing All-Reduce for Efficient Tensor-Parallel Serving", "abstract": "Quantization has reached nearly every tensor a language model stores or computes, yet tensor-parallel all-reduce still communicates in BF16. We find that this communication accounts for as much as 37% of prefill device time across models ranging from an 8B dense transformer to a 671B mixture-of-experts model, and its share grows as lower-precision computation makes the rest of inference faster. We introduce Quantar to quantize the reduction itself. Existing dynamic quantization, if applied to all-reduce, would require an extra communication round for scale agreement; without it, the reduced values cannot be mapped back to the correct sum. Quantar instead freezes communication scales calibrated once and shares them across ranks and requests. We show that transformer normalization places an input-independent bound on all-reduce activations, making one frozen scale practical across ranks. On eight B300 GPUs, Quantar speeds up the complete communication boundary by up to 1.81× and prefill by up to 1.28×, and on eight RTX PRO 6000 GPUs, performance measurements reach 1.92× and 1.61×. At matched settings, Quantar outperforms per-hop FP8 by up to 1.24× at the communication boundary and 1.10× in prefill, while avoiding its 1.54× error growth from two to eight ranks. We also show that quality remains at BF16 parity: full 12,032-question MMLU-Pro scores change by at most 0.42 percentage points across five models, and all nine benchmarks and 33 subtasks evaluated across three models remain within 2σ. Quantar also matches BF16 on LongBench-E, GSM8K-CoT, and RULER from 8k through 128k, including 85.4 versus 85.0 at 128k. Quantar's advantage grows as models become larger, more sharded, and more heavily quantized, extending low-bit gains across the inference path.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Inference", "Quantization", "Tensor Parallel", "Communication Collectives" ], "forum_url": "https://openreview.net/forum?id=TZJ8Z6Dl0M", "pdf_url": "https://openreview.net/pdf/ef15e414b213e0f6852c199b7cc379aafc68b396.pdf", "updated_at": "2026-10-05T21:04:14.010Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp8", "low-bit", "quantization", "quantize", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Nzz5ro1Y1A", "number": 49586, "title": "Verifiable Robust Federated Inference of LLMs", "abstract": "Organisations increasingly favour open-weight large language models (LLMs) because on-premise deployment avoids exposing proprietary assets and private data to third-party APIs. This choice trades one risk for another: open weights can be fine-tuned, quantised, or backdoored by anyone in the supply chain, and a single deployed model becomes a single point of failure whose faults are silent. Federated inference, which queries a diverse ensemble of models and aggregates their outputs, is a natural mitigation, but it only helps if aggregation itself tolerates a subset of faulty or adversarial models.\n\nIn this work, we study Byzantine-robust federated inference of LLMs, and unlike prior work, we consider proper generation tasks, where the output space is not restricted to single tokens drawn from a small label set for classification tasks.\nOur key observation is that such tasks supply two resources classification lacks: a partial semantic-equivalence oracle over outputs, and an external verifier induced by executability or formal properties.\nWe combine them in a unified framework based on plurality voting with abstention, which returns an answer only when its support exceeds a consensus threshold $\\tau$, and abstains otherwise. \nOur theoretical analysis bounds error and abstention for any instantiation of the framework in terms of the Byzantine coincidence rate, honest competence, equivalence oracle completeness, and verifier strength. The consensus threshold $\\tau$ then controls the trade-off between error and coverage. We quantify how verifier strength buys plurality margin, and raise fault tolerance from $n/2$ to $n-1$ in ideal conditions.\nOn HumanEval+, our method achieves 0.6\\% error at 89\\% coverage, versus 7.3\\% error for the best base model, and reaches zero error at higher $\\tau$. On MMLU, an adaptive adversary that endorses the largest incorrect cluster reduces accuracy by only 1\\%.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "collaborative inference", "robust inference", "LLMs" ], "forum_url": "https://openreview.net/forum?id=Nzz5ro1Y1A", "pdf_url": "https://openreview.net/pdf/0708ec45405fd7811f05155b720b91b768a2b7dd.pdf", "updated_at": "2026-10-05T21:04:37.409Z", "categories": [ "hardware", "theory" ], "matched_terms": [ "quantised" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Hxnudbtet7", "number": 49640, "title": "Lossy Compressive Text Autoencoders", "abstract": "Our work explores learning a compressed latent representation of text, at the intersection of data compression and representation learning. We propose an autoencoder architecture that performs residual downscaling and upscaling of hidden representations along the time axis, \nwith a residual low-dimension discrete bottleneck. We analyze our approach for different quantization methods, training objectives, and datasets. For different levels of compression, \nwe evaluate the similarity between the original and reconstructed text\nboth at the surface-level (BLEU) and at the semantic-level (LLM-based judge). Additionally, we evaluate our models on downstream question-answering and semantic text similarity benchmarks. Our approach results in compressed representations which are on par with lossless text compression algorithms at 2.24 bits per byte on web text data, while having good reconstruction and downstream task performance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "text autoencoder", "text compression", "lossy compression", "embedding model" ], "forum_url": "https://openreview.net/forum?id=Hxnudbtet7", "pdf_url": "https://openreview.net/pdf/65a9f542a67ca34d4f4e592e1bb8df1f9b8397d3.pdf", "updated_at": "2026-10-05T21:04:45.196Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "nsIAFF5JI9", "number": 49687, "title": "One Token Owns the Calibration Hessian: Why Layer-wise Quantization Can Silently Break Instruction Following", "abstract": "Quantized instruction-tuned models are usually validated with perplexity and multiple-choice accuracy, which can miss a collapse of instruction following. At 3 bits, reference GPTQ with act-order, calibrated on short documents as widely used libraries sample them by default, lowers the IFEval score of Llama-3.1-8B-Instruct from 0.768 to 0.155 and that of Qwen2.5-14B-Instruct from 0.820 to 0.426, below round-to-nearest (0.565, 0.697), while MMLU stays in the ordinary 3-bit range; GPTQModel's own column order avoids the collapse. We trace the failure to a blind spot of the layer-wise objective. At the down-projection that forms the attention sink, the first token owns the calibration Hessian, yet the loss is almost insensitive to its output. GPTQ's compensation, which we derive in closed form, therefore protects that token and writes its rounding error along the sink pattern into one matrix: a lesion that the chat-template tokens read out. Excising the lesion restores both models. The blind spot is present in all twenty-six models whose first token dominates a down-projection, and of the seventeen of these that we quantized, fourteen carry the lesion, yet only two collapse. Transplants indicate that the other matrices of a healthy network cancel the lesion, which would explain why the failure is rare and depends on how the calibration text is cut and which documents are drawn. Weighting each calibration token by its output-side sensitivity divided by its input energy removes the blind spot and the four collapses it was run on, at a cost of at most 5.6 points on healthy models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "GPTQ", "attention sink", "instruction following", "calibration data", "layer-wise reconstruction", "large language models", "mechanistic interpretability" ], "forum_url": "https://openreview.net/forum?id=nsIAFF5JI9", "pdf_url": "https://openreview.net/pdf/5d3d4e0c2d95b54dfc6c2435f802c8d8fd244730.pdf", "updated_at": "2026-10-05T21:04:52.591Z", "categories": [ "ptq" ], "matched_terms": [ "3-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VG6BlmCGWO", "number": 49706, "title": "Reparametrizing Shampoo and SOAP", "abstract": "Shampoo-based methods, such as KL-Shampoo and SOAP, have demonstrated strong performance in neural network training, but their QR-based updates become costly for large preconditioning matrices. To address this, we propose a preconditioner reparametrization that accelerates full-basis QR computation and enables faster partial basis updates. By maintaining preconditioning factors in the current orthogonal basis, our approach exposes useful input structure to facilitate QR decomposition and allows dynamically selected basis vectors to be cheaply updated via subspace QR while preserving the full-basis orthogonality. To further accelerate QR computation, we provide a customized implementation supporting mixed-precision computation and batched execution. The resulting framework unifies full-basis and subspace updates and applies broadly to QR-based updates, including KL-Shampoo, SOAP, and KL-SOAP. Empirically, our approach reduces computational overhead and improves the training performance of SOAP-type methods in mixed precision, enabling KL-SOAP to match or exceed KL-Shampoo. Overall, our approach improves the time efficiency of Shampoo-based methods.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "optimization", "deep learning" ], "forum_url": "https://openreview.net/forum?id=VG6BlmCGWO", "pdf_url": "https://openreview.net/pdf/a368ae1a74ccba5ccdcbcd4cf6b88d699c154c35.pdf", "updated_at": "2026-10-05T21:04:54.160Z", "categories": [], "matched_terms": [ "mixed precision", "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "0ucx7j6l6g", "number": 49724, "title": "LSMXQ: Layerwise Salience-Aware Mixed-Precision Quantization for Efficient LLM Inference", "abstract": "Post-training quantization (PTQ) is an effective approach for reducing the memory footprint of large language models (LLMs), enabling larger models to fit within limited GPU memory and improving inference efficiency under memory-bound workloads. While recent sub-4-bit quantization methods can achieve high accuracy compared to FP16 baselines, they do not necessarily leverage modern hardware-supported FP4 Tensor Core architectures, limiting their ability to fully utilize the computational throughput of modern GPUs, particularly at larger batch sizes. In this work, we introduce LSMXQ, a hardware-aware mixed-precision post-training quantization approach that allocates low-bit representations under an average bit budget while minimizing layer-wise propagated quantization error. LSMXQ is designed for coalesced memory access and efficient NVFP4 Tensor Core execution, supporting both memory-bound low-batch inference and high-throughput large-batch workloads. Across a broad evaluation of LLMs and zero-shot benchmarks, LSMXQ retains up to 98.5% of FP16 accuracy at 2-bit precision, while matching FP16 accuracy at 2.5- and 3-bit precision. For inference efficiency, LSMXQ achieves up to 3.71× the throughput of FP16 in batch-1 inference, while delivering substantial throughput improvements at larger batch sizes in vLLM.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "post-training quantization", "mixed-precision quantization", "hardware-aware", "tensor core" ], "forum_url": "https://openreview.net/forum?id=0ucx7j6l6g", "pdf_url": "https://openreview.net/pdf/f7e40ae278358d5fc3de893d308e63a00bac151f.pdf", "updated_at": "2026-10-05T21:04:55.909Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "2-bit", "3-bit", "4-bit", "fp4", "low-bit", "mixed-precision", "nvfp4", "ptq", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ICmOwVeKGi", "number": 49797, "title": "Analysis of MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible error", "abstract": "MXFP4 arithmetic can dramatically accelerate reinforcement learning (RL) post-training of large language models (LLMs), yet the quantization error introduces severe accuracy degradation. Existing work treats the quantization error as a monolithic noise term, missing the distinct mechanisms upon interpreting how quantization error damages training. We prove that this error is not monolithic but admits an exact, \\textit{privileged} decomposition into three additive components: \\textit{scale bias} from power-of-two rounding, \\textit{deadzone truncation} from zeroing small values, and \\textit{grid noise} from rounding to the 4-bit grid.\nThe deadzone component is exactly orthogonal to the other two, and, crucially, the grid component is \\textit{invariant to scale precision}: refining the block scale reduces scale bias without touching grid noise, driving the total error toward an irreducible floor set by the grid and deadzone alone.\nEach component dominates a distinct RL failure mode: scale bias accumulates multiplicatively through the backward pass, affecting gradient accuracy; deadzone truncation degrades rollout quality; and grid noise raises the policy's entropy. Our contribution is this diagnostic framework rather than a new correction: to validate it, we use three \\textit{existing}, mechanism-targeted techniques as controlled interventions (Macro-block scaling for scale bias, Outlier Fallback for deadzone, and Adaptive Quantization Noise for grid noise/entropy), each removing one component and recovering its predicted RL pathway.\n On Qwen2.5-3B (dense) this recovers BF16 accuracy to within 0.7 pp, on Qwen3-30B-A3B-Base (MoE) to parity, and on Llama-3.1-8B (a different family) to within 3.2 pp. The same component-to-pathway mapping holds across all three, and the results generalize from GSM8K to MATH500.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization error", "low-precision training", "large language models", "reinforcement learning" ], "forum_url": "https://openreview.net/forum?id=ICmOwVeKGi", "pdf_url": "https://openreview.net/pdf/c11d4246e5e78465dfce44d18bf3bb0f0bdb850a.pdf", "updated_at": "2026-10-05T21:05:04.090Z", "categories": [ "ptq", "training" ], "matched_terms": [ "4-bit", "low-precision", "mxfp4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DuDwRhNU6j", "number": 49818, "title": "DUALQUANT: Dual-Scale Calibration-Free Weight-and-Activation Quantization for LLMs", "abstract": "Low-bit inference on accelerators with native 4-bit support requires quantizing both weights and activations to realize the full throughput gains these formats provide. While post-training weight quantization has been extensively studied, activation quantization remains considerably more challenging and typically requires calibration data. In this work, we introduce DUALQUANT, a dual-scale,\ncalibration-free post-training quantization framework that directly minimizes the weight-reconstruction Mean Squared Error (MSE) and, as an immediate by-product, yields per-channel activation rescalings prior to activation quantization. Our key observation is that the scaling factors obtained by minimizing weight-reconstruction error also provide effective activation rescalings, enabling competitive activation quantization without calibration data. We formulate DUALQUANT as a non-differentiable weight-reconstruction problem and develop an efficient solver based on variable splitting and block coordinate descent, where each subproblem admits a closed-form update. By treating the quantizer as a black box and projecting scales onto their format-specific admissible set after each update, DUALQUANT provides a unified algorithm supporting a wide range of formats, including INT4, INT8, the full MX family, and NVFP4. To our knowledge, DUALQUANT is the first dual-scaling method to formally support MX and NV block formats. Across Llama and Qwen models evaluated on WikiText-2, C4, and zero-shot downstream tasks, DUALQUANT matches or exceeds the accuracy of calibration-based SmoothQuant across most models and formats, and substantially outperforms calibration-free baselines, with particularly large gains for MX and NV block formats. We further validate these gains with a real packed W4A4 inference path for Llama-3.1-8B on an RTX 4090. The resulting implementation closely matches the accuracy of fake quantization while delivering up to 2.61× memory reduction and up to 2.46× throughput speedup over BF16. Our anonymized codebase is available https://anonymous.4open.science/r/dualquant-69B3/README.md.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "dual-scale quantization", "calibration-free", "low-bit inference", "efficient inference", "block coordinate descent", "weight-reconstruction error", "block data format" ], "forum_url": "https://openreview.net/forum?id=DuDwRhNU6j", "pdf_url": "https://openreview.net/pdf/9b1d9dc7856e733bb721dc3a59137530bbbd29ca.pdf", "updated_at": "2026-10-05T21:05:07.288Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "4-bit", "int4", "int8", "low-bit", "nvfp4", "quantization", "quantizer", "quantizing", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "hsO0iX5XH3", "number": 49863, "title": "Quantizing Looped Transformers: Feedback Exposure and Calibration Blindness", "abstract": "Looped transformers reuse weights across recurrence steps, making low-bit quantization especially attractive. We identify two distinct failure modes of standard post-training quantization. On Huginn-3.5B, per-channel INT4 fails primarily at the non-residual loop-entry adapter, while quantizing the residual core is much less damaging. We call this *feedback exposure*: a quantized layer perturbs the recurrent state without an identity path, and the resulting error is fed back at later steps. Controlled experiments on linear filters and Mamba state-space models show that feedback exposure also occurs outside transformers. Grouped INT4 reveals a separate failure, *calibration blindness*: our one-step GPTQ baseline builds its Hessian from step-0 activations, leaving input directions used later in the recurrence nearly unweighted. Across nine checkpoints from seven looped architectures, one-step GPTQ is worse than round-to-nearest (RTN) on the primary task metric for five checkpoints. Accumulating the GPTQ Hessian across recurrence steps outperforms both one-step GPTQ and RTN on all nine checkpoints and recovers bf16-level accuracy on Huginn. These results separate two questions for PTQ on looped models: where quantization error enters the recurrence, and which states calibration sees.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "looped transformers", "post-training quantization", "weight-only quantization", "GPTQ", "recurrent inference", "low-bit inference", "model compression" ], "forum_url": "https://openreview.net/forum?id=hsO0iX5XH3", "pdf_url": "https://openreview.net/pdf/0445c02f9167655ecb3dafc10194acc6cfa9c5e6.pdf", "updated_at": "2026-10-05T21:05:14.227Z", "categories": [ "ptq" ], "matched_terms": [ "int4", "low-bit", "ptq", "quantization", "quantized", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Eg1GBLzKmB", "number": 49923, "title": "QReS: Quantizing Recurrent Language Models Is Easier Than You Think", "abstract": "Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models, motivating quantization schemes tailored to recurrent structure. We study how quantizing linear recurrences differs from quantizing softmax attention along two axes. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B models, we find that decayed recurrences develop much weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates and recurrent decay create a different numerical problem from KV cache quantization. We use these insights to develop QReS, a pipeline that combines value-axis scale grouping, decay-based mixed-precision allocation and stochastic rounding for recurrent state quantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, QReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement QReS in vLLM and demonstrate end-to-end decode speedups.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "language modeling", "linear attention", "quantization", "efficiency" ], "forum_url": "https://openreview.net/forum?id=Eg1GBLzKmB", "pdf_url": "https://openreview.net/pdf/77928ae3140d2b0da5ac334cceb3b9fd5c453819.pdf", "updated_at": "2026-10-05T21:05:20.274Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "8-bit", "mixed-precision", "quantization", "quantize", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SQ048EUIXv", "number": 49941, "title": "Does Fine-Tuning Still Beat In-Context Learning Under LoRA?", "abstract": "Does the advantage of fine-tuning (FT) over in-context learning (ICL) on synthetic formal languages persist under parameter-efficient adaptation? We compare full FT, LoRA and 4-bit QLoRA with single-pass ICL on Qwen2.5 base models from 0.5B to 3B parameters, using matched sets of 32 training strings and a published discriminative ROC-AUC evaluation. Under the published protocol, which selects each run's best epoch using test scores, all tested fine-tuning methods achieve higher mean AUC than ICL at every evaluated model size. However, at 3B, this ranking reverses when the same runs are evaluated after 50 epochs. To test whether the advantage survives validation-based checkpoint selection, we repeat the 3B comparison across four languages, choosing epochs on fresh validation strings and evaluating on separate held-out test strings. Full FT, LoRA and QLoRA achieve mean test AUCs of 0.942, 0.921 and 0.919, respectively, versus 0.873 for ICL; only the full-FT advantage remains statistically significant after correction for multiple comparisons. Selecting epochs on the test strings instead raises mean AUC only slightly, while all three fine-tuning methods again fall below ICL on average at epoch 50. Under the published protocol, adapter-versus-full-FT comparisons also depend on the adapter learning rate. Thus, in this 32-example setting at 3B, validation-based stopping preserves full FT's advantage over single-pass ICL, with weaker evidence for the adapters, whereas training for 50 epochs reverses the comparison.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "parameter-efficient fine-tuning", "LoRA", "QLoRA", "in-context learning", "formal languages", "evaluation methodology", "checkpoint selection", "reproducibility" ], "forum_url": "https://openreview.net/forum?id=SQ048EUIXv", "pdf_url": "https://openreview.net/pdf/ff5886bd7ed1711c4733a3f29a409e309eb102c2.pdf", "updated_at": "2026-10-05T21:05:22.736Z", "categories": [], "matched_terms": [ "4-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "m6tf9gModK", "number": 50118, "title": "TReQ: Refitting Output Transforms Without Requantizing Language Models", "abstract": "Low-bit weight--activation quantization introduces errors that can\naccumulate across language-model layers. We present TReQ, a post-training\nquantization method with block-diagonal transformations before and after\nlow-precision matrix multiplication. The output transformations mix\nchannels within small blocks, providing additional reconstruction freedom\nwhile keeping quantized weights and scales fixed during each refit.\nWe combine two-sided transformation construction with reconstruction of\nattention scores, attention context, and downstream projection outputs.\nThe output transformations can be fused into the GEMM epilogue, avoiding\na separate kernel and intermediate-memory round trip.\nAcross Ministral-3-8B, Llama-3.1-8B, and Qwen3-8B, TReQ reduces block\nreconstruction error relative to matched WUSH baselines in INT4, MXFP4,\nand NVFP4. Under INT4, it lowers task Kullback--Leibler (KL) divergence\nand improves accuracy on all five benchmarks for every model, while\nreducing WikiText-2 perplexity. End-to-end improvements in the FP4 formats\nare less consistent. A fixed-checkpoint Ministral INT4 comparison shows\nthat output refitting alone reduces held-out KL by 8.63\\%.\nOn NVIDIA B200, fused output transformations add only 0.1--1.5\\% to MXFP4\nfull-model prefill latency at the tested workloads with at least 2,048 total tokens.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Post-training quantization", "Large language models", "Weight–activation quantization", "Output transformation refitting", "Quantization error correction" ], "forum_url": "https://openreview.net/forum?id=m6tf9gModK", "pdf_url": "https://openreview.net/pdf/6543e63b75381130e16893d852a0c8564fa9c809.pdf", "updated_at": "2026-10-05T21:05:53.495Z", "categories": [ "ptq", "training", "hardware" ], "matched_terms": [ "fp4", "int4", "low-bit", "low-precision", "mxfp4", "nvfp4", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "9KMvoYdxb9", "number": 50160, "title": "Efficient Multimodal Inference through Adaptive Acquisition and Sequential Fusion", "abstract": "Multimodal systems often encode every available input, even when a subset suffices for prediction. Adaptive acquisition can reduce this cost by using predictions from incrementally fused evidence to *decide which modality to encode next and when to stop*. However, sequential fusion makes these predictions order-dependent, so decisions based on them may need to distinguish factorially many histories of the same acquired set. We introduce SeMARC, which couples a Sequential Modality Aggregator (SeMA) with an Adaptive Runtime Controller (ARC) and uses acquired evidence to select each modality *before* its encoder runs. SeMA executes only selected encoder and fusion branches, updates a fixed-size state, and predicts after each acquisition without recomputing earlier branches. We supervise every acquisition prefix under randomized modality subsets and orders to encourage consistent predictions across acquisition orders. ARC combines a set-dependent marginal-utility prior with residual fitted-Q learning to select the next available modality or stop, without inspecting unacquired inputs or retaining acquisition order. Across six multimodal classification datasets and eleven baselines, SeMARC achieves 3.2% higher macro-F1 and 61.4% lower total inference GFLOPs on average relative to each dataset's most accurate baseline. End-to-end latency falls by 44.0% across GPU and CPU and by 47.2% on Android INT8 relative to the fastest measured baseline, on average. Under varying runtime modality missingness, SeMARC still skips available modalities, matching or exceeding the best baseline macro-F1 in 21 of 24 conditions with 14.8% lower total GFLOPs on average. SeMARC thus offers a practical path toward efficient multimodal inference across heterogeneous devices.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multimodal learning; adaptive modality acquisition; sequential fusion; efficient inference; reinforcement learning; missing-modality robustness" ], "forum_url": "https://openreview.net/forum?id=9KMvoYdxb9", "pdf_url": "https://openreview.net/pdf/2b9c70cc6645292167a45a36b832aa06e59342a0.pdf", "updated_at": "2026-10-05T21:05:59.656Z", "categories": [ "hardware", "multimodal", "theory" ], "matched_terms": [ "int8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "eScfcPjl2U", "number": 50268, "title": "EVERY BYTE MUST RANK: A RETRIEVAL-AWARE CODEC FOR FROZEN MULTIMODAL EMBEDDINGS", "abstract": "Multimodal retrieval systems increasingly store every candidate as a high-dimensional embedding, and the bill is paid per vector: one billion 2,048-dimensional FP32 embeddings occupy 8.2 TB before any index is built. This scale makes embedding storage a critical systems problem. Yet the tools used to compress these vectors were designed to reconstruct them. Product quantizers minimize reconstruction error even though retrieval depends on the resulting rankings, and fixed-width codes charge every symbol equally regardless of predictability. We introduce EmbedCodec, a codec trained for ranking rather than reconstruction. A lightweight adapter shared by queries and candidates is optimized through the hard product-quantizer assignments used at deployment, reshaping the frozen encoder’s geometry for post-quantization retrieval. A task-conditioned Gaussian mixture models the resulting indices through codeword geometry; rANS then entropy-codes those indices, and rate is measured from serialized bytes rather than nominal bit width. Results on five retrieval tasks spanning text, cross-modal, and visual-document search with two frozen encoders show that EmbedCodec uses about 125 B/vector, a roughly 65× storage reduction relative to FP32, while retaining 99.6–99.8% of the corresponding five-task average NDCG@10. At about 500 B/vector, it exceeds the FP32 average on both encoders. Every persisted stream is exactly decodable back to its input index tuples, and each of the ten encoder–task panels contains at least one Pareto-nondominated EmbedCodec operating point. Our findings suggest that retrieval-aware source and entropy coding can substantially reduce embedding storage without requiring faithful coordinate reconstruction.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "multimodal embeddings", "representation compression", "cross-modal retrieval" ], "forum_url": "https://openreview.net/forum?id=eScfcPjl2U", "pdf_url": "https://openreview.net/pdf/bbeb90eb8e0750f6ee493165c02006c16ca6a30b.pdf", "updated_at": "2026-10-05T21:06:13.759Z", "categories": [ "ptq", "hardware", "multimodal" ], "matched_terms": [ "bit width", "quantization", "quantizer", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DdeRtXKXHr", "number": 50366, "title": "All-in-One: A Composable Attack on LLM Transformations", "abstract": "Quantization, pruning, and model merging are popular transformations applied to deployment-ready LLMs to improve efficiency and downstream performance. Recent work has shown that these transformations can introduce a security threat: an adversary can train a model that passes safety checks in its pre-transformed state, only turning malicious once it is transformed for deployment. However, prior attacks target one specific transformation (e.g., quantization) and generally rely on transformation-specific techniques, making them difficult to combine in practice. In this work, we introduce a simple, composable attack framework that targets multiple transformations simultaneously and triggers when any one of them is applied by a user. Our key insight is that the effect of each transformation can be approximated as a perturbation to the model weights, allowing us to jointly train the model with a benign loss on the original model and a malicious loss on each simulated transformed model. These perturbations are simple enough to simulate at every training iteration across diverse transformations, making them practically combinable within a single training objective, enabling a unified attack. Through extensive experiments across attack scenarios, model families, and a wide range of deployment-time model transformations, we show that our attack achieves high success rates. This demonstrates that model transformations may be utilized to trigger significantly different behavior from the original model and reinforces the associated security concerns by demonstrating that a single attack can jointly target diverse transformations.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "quantization", "pruning", "model merge", "transformation-conditioned attack" ], "forum_url": "https://openreview.net/forum?id=DdeRtXKXHr", "pdf_url": "https://openreview.net/pdf/3887ef5a9f2b9c28afdcd75033f0795206ef687a.pdf", "updated_at": "2026-10-05T21:06:30.125Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ISh3jyOh0e", "number": 50379, "title": "Level of Detail Attention: Allocating Resolution Instead of Sparsity", "abstract": "Most sparse attention methods attend to a subset of keys, ignoring the rest and thereby amplifying any mistakes while entirely removing the ignored set from softmax normalization. We introduce Level of Detail (LoD) Attention, a training-free approximation to dense softmax attention that represents less important keys and values at reduced resolution instead of discarding them. For language inference, LoD requires no changes to pretrained model weights and achieves $O(N^{3/2})$ prefill work and $O(N^{1/2})$ work per decode step. Experiments on Qwen3.8 and K2 Horizon show that LoD largely preserves aggregate quality. On Qwen3.8, two-tier LoD accelerates attention computation by over 6x during both prefill and decode in our 128K prompt batch size 8 benchmarks. Its similarity-based grouping scheme also permits the reduction of the LLM key-value cache to a compact INT4 representation, saving VRAM with little additional aggregate quality loss. To show the versatility of our method across modalities we also demonstrate LoD for MiniMax-H3 video generation. We will release our code under the Apache 2 license.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "attention", "LLM", "Video", "sparsity", "resolution", "multiresolution", "generative", "attention", "KV", "KV-Cache", "vector quantization", "transformer", "compressive state" ], "forum_url": "https://openreview.net/forum?id=ISh3jyOh0e", "pdf_url": "https://openreview.net/pdf/dddcb08ba51f94a3f049ebf98dc45e7a03bdb1c2.pdf", "updated_at": "2026-10-05T21:06:31.093Z", "categories": [ "kv-cache" ], "matched_terms": [ "int4", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "DMUfRers4X", "number": 50405, "title": "AutoCache: Agentic Search for LLM Prefix Cache Compression", "abstract": "Storing and reusing computed states allows LLM serving systems to avoid redundant prefill computation for shared prefixes, reducing time to first token. As context lengths and the number of reusable prefixes grow, however, limited cache capacity forces evictions and expensive recomputation, making storage efficiency critical. This challenge is further complicated by modern hybrid architectures that maintain recurrent and sliding-window states in addition to KV caches, requiring compression strategies that account for model architecture, workload characteristics, and reconstruction cost. We present AutoCache, an agent-driven framework that automatically discovers compression and reconstruction programs for prefix caches. Given a model, dataset, and accuracy target, AutoCache explores a large combinatorial space of compression techniques to compose effective cache-compression recipes, refine existing methods, discover data and model architecture-specific compression opportunities. On the RULER benchmark with 32K context, AutoCache reach 291× KV-cache compression on Qwen3.5-9B, compared with 16.0× for the strongest evaluated baseline. It does so by composing factorization, quantization, entropy encoding, and protection of high-saliency cache entries in ways that would be difficult to design manually. More importantly, AutoCache can seamlessly extend compression beyond KV caches to recurrent states in hybrid models. By discovering and exploiting a previously unexploited low-rank structure in the recurrent state of linear-attention layers, AutoCache achieves 222× compression of the total cache memory footprint. These results demonstrate the potential of automatically discovering compression algorithms that are tailored to the model architecture and workload, therefore enabling substantially more storage-efficient prefix caching for LLM serving.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV-Cache", "Hybrid Model", "Agentic Search" ], "forum_url": "https://openreview.net/forum?id=DMUfRers4X", "pdf_url": "https://openreview.net/pdf/d6e9a332349c66dc81a0d5f6812b4cd8cb7c079c.pdf", "updated_at": "2026-10-05T21:06:35.138Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sL7JFiA4c0", "number": 50441, "title": "Read What Matters: Query-Adaptive Quantization for KV Caches", "abstract": "KV-cache entries are stored before their future queries are known, but each de-\ncoding query needs precision in different places. We study this mismatch using\nseparate budgets for retained bits and bits fetched per query. READKV stores each\nkey and value in a progressive code whose prefixes support different reconstruction\nprecisions. For each query, it allocates key-channel prefixes using the query, com-\nputes attention from the reconstructed keys, and then allocates value-token prefixes\nusing that attention. Stored entries remain unchanged. Each stage optimizes a\ncalibrated distortion objective under a fixed budget; we prove exact allocation\nunder diminishing refinement gains and relate these objectives to attention-output\nerror. We also exhibit a finite-dimensional attention family where query-dependent\naccess strictly outperforms every query-independent reader at the same read budget,\neven with unrestricted competing encoders and decoders.\n\nAcross six base models, reading four bits on average from an eight-bit cache\nincreases C4 perplexity by at most 0.66%, using about one quarter of the logical\nreads and half the retained capacity of a 16-bit cache. It is consistently more\naccurate than storing and fully reading four bits at the same payload-read budget.\nLong-context question answering and retrieval on two instruction-tuned models\nprovide additional quality evidence. On the tested 8K-token, batch-one, single-layer\nworkload on an NVIDIA A10G, a restricted eight-bit READKV reader with a two-\nbit mean payload-read budget has 39% lower latency than the tested TurboQuant\ncodec.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Query-Adaptivity", "Asymmetric read and write" ], "forum_url": "https://openreview.net/forum?id=sL7JFiA4c0", "pdf_url": "https://openreview.net/pdf/970167b3de3bad62f86a5c9b886a212837314f1d.pdf", "updated_at": "2026-10-05T21:06:41.826Z", "categories": [ "ptq", "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ir9shPsHPh", "number": 50504, "title": "Clean: Second-order LLM Training at Linear Memory Cost via Nyström Sketching", "abstract": "Training large language models (LLMs) entails a fundamental trade-off: memory-efficient optimizers such as Adam discard cross-parameter curvature, whereas full-curvature methods such as SOAP can accelerate convergence at prohibitive memory costs.\nWe introduce **Clean**, a memory-efficient and full-curvature optimizer designed to resolve this bottleneck. Clean leverages the randomized Nyström method to accurately approximate the left and right preconditioners in SOAP, and to reduce the optimizer's memory complexity from quadratic to linear in terms of model dimension.\nWe subsequently reintegrate the off-subspace components to capture curvature information beyond the low-rank approximation, preserving rich curvature at minimal memory cost. \nWe further propose **Q-Clean**, a low-precision variant that aggressively compresses optimizer states. Q-Clean reduces optimizer memory consumption by **over 50%** compared to Muon when pre-training a LLaMA-1.3B architecture, all while maintaining strong and competitive predictive performance. Notably, Clean operates with a smaller optimizer-state footprint than standard AdamW while reaching AdamW's final performance **26% faster** in wall-clock time. Furthermore, our methods uniquely enable the pre-training of a 13B-parameter model on a single 80GB GPU, providing a scalable, efficient, and accessible approach to large-scale model optimization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "memory-efficient optimization", "low-rank gradient", "second-order information", "subspace tracking", "Nystrom sketching" ], "forum_url": "https://openreview.net/forum?id=Ir9shPsHPh", "pdf_url": "https://openreview.net/pdf/66298bcc304b3f8bdc34a4c83a5a7b33c200dbaf.pdf", "updated_at": "2026-10-05T21:06:52.812Z", "categories": [ "training", "hardware", "theory" ], "matched_terms": [ "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "RiqkjlPZhc", "number": 50532, "title": "RAPID: Extreme KV Cache Compression At Near-Full Quality from the Context Alone", "abstract": "The size of KV caches is a dominant bottleneck in long-context model serving. Modern KV compression algorithms fall short on at least one of the following objectives. 1. To maintain uncompressed generation quality. 2. To compress fast, without heavy training, data generation, or model use (beyond context itself). 3. To achieve best possible compression: SOTA as of mid-2026 is over 10× rate as compared to full KV cache storage at half-precision floating points.\n\nWe introduce RAPID algorithm (Rank And Precision allocation with Independent per-layer Distillation) which our evaluations show to be the closest to satisfying all three objectives. RAPID combines layer-wise latent representations, cross layer budget allocation, per token dimension reduction, per dimension quantization, and intra-layer optimization using prefill query vectors only. RAPID doesn't need training data and access to the model beyond each layer's attention projections. A variant of our algorithm, RAPID++, is always within 2.5 points of the full cache quality at 20× compression, on every evaluated benchmark and model, including Qwen3-4B, a 117B-parameter mixture-of-experts GPT-OSS-120B, and an MLA model whose cache is already compressed.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "long-context inference", "efficient LLM serving", "rate–distortion optimization", "low-rank approximation", "quantization", "knowledge distillation" ], "forum_url": "https://openreview.net/forum?id=RiqkjlPZhc", "pdf_url": "https://openreview.net/pdf/049a22be691b4f6d94392c94458c748342dee8ad.pdf", "updated_at": "2026-10-05T21:06:55.680Z", "categories": [ "kv-cache" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "zaBYeZd24c", "number": 50547, "title": "A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models", "abstract": "Looped models reason by applying the same block of weights many times, so compressing that block saves memory traffic\non every loop. Compressed looped models, however, often collapse, and the collapse is usually blamed on rounding error\nthat accumulates from loop to loop. In this work we test that account on more than 30 models from five families and find, to our surprise, that it\nholds only for loops that never settle. When a loop settles, a fixed rounding error does not accumulate. It moves the\npoint where the loop settles, much as tilting a bowl moves where a ball comes to rest, and the answer is lost only when\nthe shift is larger than the readout tolerates. This picture lets us predict which models fail from a single label-free\nmeasurement, and it tells us why failed models recover: their loops still settle, so a few final loops with 8-bit\nweights bring the answer back. Motivated by these findings, we build a controller that stops when the model's halting head fires and then finishes\nwith 8-bit loops. On Sudoku-Extreme it beats fixed-depth inference by 15 points at under a third of the weight traffic.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "looped models; recursive reasoning; quantization; model compression; fixed points; test-time compute; early exit; deep equilibrium models; on-device inference" ], "forum_url": "https://openreview.net/forum?id=zaBYeZd24c", "pdf_url": "https://openreview.net/pdf/0691254ff6600a20042bb99b7042d7f35cc4ade7.pdf", "updated_at": "2026-10-05T21:06:57.042Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "8-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "yP6MSpYNAi", "number": 50574, "title": "COMPACT-Bench: One Budget Axis for Memory Compaction in LLMs and Agents", "abstract": "Methods that reduce how much of its context a language model keeps in memory\nare developed in four separate literatures, and each reports results in its own\nunit. KV-cache compression reports memory saved at fixed perplexity, prompt\ncompression reports tokens saved at fixed accuracy, sub-quadratic architectures\nreport recall at a fixed state size, and agent-memory systems report benchmark\nscores without stating how much they store. All of these methods spend the same\nresource, yet their numbers cannot be compared. We introduce COMPACT-Bench,\nwhich measures every method on one axis, the bytes of retained state per token\nof history (BPT), and asks four questions at matched budgets: how much accuracy\na method keeps, whether exact or lossy retention does better, whether the\nsystem can tell what it discarded, and whether its stated confidence reflects\nwhat its memory still supports. Every self-report is checked against\nuncompacted controls in which the fact is present or absent. On 21 settings of\n16 open models from 0.5B to 24B parameters we find four results. First, the\nliteratures disagree once every stored byte is counted. On Qwen3-1.7B a 2-bit\nKV cache keeps 0.98 accuracy at 22\\% of the full cache, where eviction, prompt\ncompression, and summarization keep at most 0.56 within 30\\%, while on\nQwen2.5-1.5B the same quantizer fails below 8 bits. Second, tolerance to\neviction depends on model generation and not on size. In five same-family pairs\nfrom 0.5B to 14B, and across all 21 settings, every model released from\nNovember 2024 onwards loses half its accuracy at 33--69\\% of its budget,\nagainst 76--86\\% for every earlier model. Third, content-based scorers fall\nbelow random eviction only because they cannot see the question, and letting\nthem see it recovers up to 0.98 in accuracy at the same budget. Fourth, eviction\nthat scatters its losses leaves fragments of a fact that the model takes for the\nfact itself. On the models whose yes/no self-report passes both controls,\nSnapKV's failures are reported as still answerable in up to 74\\% of cases,\nagainst at most 11\\% for StreamingLLM, which removes facts whole, and on the\nmodels whose stated confidence falls for an absent fact, SnapKV's failures\ncarry higher confidence than a fact that was never there. We release\nthe benchmark and a reference implementation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "memory compaction", "KV cache compression", "long-context LLMs", "prompt compression", "calibration", "benchmark" ], "forum_url": "https://openreview.net/forum?id=yP6MSpYNAi", "pdf_url": "https://openreview.net/pdf/c508821b62af0911299fa898d4723677771c2703.pdf", "updated_at": "2026-10-05T21:07:00.373Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "2-bit", "quantizer" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "HWCHKGep3x", "number": 50625, "title": "Disentangled Representations for Zero-shot Behavioral Style Transfer", "abstract": "Human behavior is exhibited through bodily motion that not only achieves its intent-driven or communicative goals, but also demonstrates stylistic and idiosyncratic patterns. Existing approaches for generative motion synthesis learn to transfer these styles to novel motions by learning adapters or designing task-specific guidance objectives. However, disentangling style and content within the core motion representation remains an open problem. We present DisCodec, a general framework for building disentangled representations of style and content in human motion, which can be used for zero-shot style transfer at the decoding stage. Our approach learns these representations through a neural motion codec trained with a Variational Information Bottleneck and SNR-adaptive Transform Quantization to disentangle style from the motion's action content. Two streams of encoders decompose any reference motion into style and content latents, which are then combined to reconstruct the motion, while also ensuring that these latents represent mutually distinct information about the motion. Therefore, any generative motion synthesis task learned using the proposed representation not only inherits zero-shot style transfer capability but can also sample styles from the learned style subspace. We benchmark the style transfer performance against existing approaches and independently demonstrate applications of DisCodec representation in text-to-motion synthesis and co-speech gesture generation tasks. Reader is urged to watch accompanying video demo.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Human Motion Generation", "Style Transfer", "Gesture Synthesis" ], "forum_url": "https://openreview.net/forum?id=HWCHKGep3x", "pdf_url": "https://openreview.net/pdf/fc033dd0e2efac988d669831ac222a59ccab6c01.pdf", "updated_at": "2026-10-05T21:07:08.263Z", "categories": [], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cqXDGJfGuu", "number": 50668, "title": "Architecture-Conditioned KV Compressibility", "abstract": "KV-cache compression is usually evaluated as a deployment method, but a prior question is often left implicit: which native cache representations contain structure worth compressing? We study KV compressibility as a conditional property of representation, architecture, context length, domain, and tensor folding. The main audited screen contains nine completed instruction-tuned models from Qwen2.5, Llama, and Mistral, spanning 0.5–22.2B parameters, four domains, four contexts from 512 to 32K, and three seeds: 432 valid cells and 1,296 representation rows. A separate exact-length matrix contains 420 cells from three models, four domains, seven contexts from 512 to 32K, and five seeds. Three results emerge. First, representation ordering is universal in both protocols: pre-RoPE keys beat post-RoPE keys, which beat values, in 432/432 and 420/420 paired cells. Second, quantized tensor-train (QTT) compatibility is architecture-conditioned and non-monotone: Qwen2.5-7B reaches a 7.32× median pre-RoPE-key QTT ratio in the main screen, while neighboring Qwen2.5-3B and 14B are near 1×; the protocol-separated Qwen2.5-1.5B exact-context matrix reaches 26.82×. Third, long context amplifies representation mismatch: from 512 to 32K, rank-16 error rises by 0.0246 for pre-RoPE K, 0.1448 for post-RoPE K, and 0.0819 for V. The paper provides a structural screening prior, not a claim of resident-memory or latency improvement.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache compression", "Efficient transformer inference", "Long-context language models", "Tensor-train decomposition", "Architecture-conditioned compression", "Attention geometry", "Tensor trains" ], "forum_url": "https://openreview.net/forum?id=cqXDGJfGuu", "pdf_url": "https://openreview.net/pdf/b1948285a6619ee07494cbb85bb342e82d3d35d9.pdf", "updated_at": "2026-10-05T21:07:14.964Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "7jOGNBSsvG", "number": 50685, "title": "Know Your Tokens: Structure-Aware Gradient Allocation for Agentic Search via Reinforcement Learning", "abstract": "Reinforcement Learning (RL) from verifiable rewards is a powerful paradigm for training Large Language Model (LLM)-based search agents. Applying RL to agentic search models faces a core technical bottleneck: (a) standard Group Relative Policy Optimization (GRPO) causes gradient instability under parameter-efficient training (LoRA), and (b) existing fixes that subsample tokens by position (Stochastic GRPO, S-GRPO) restore convergence but discard gradient signal from late-position decisions that are equally critical. We show that position is not a well suited proxy for importance in multi-step search trajectories where critical decisions can occur at any position. We propose SA-GRPO (Structure-Aware S-GRPO), a general methodology for training LLM-based search agents via RL that addresses both problems: each model-generated token receives an inclusion probability in the policy gradient loss based on its functional role in the trajectory; search decisions and queries always receive gradient; answer tokens at high probability; reasoning is subsampled; and redundant re-searches are rarely included. Role assignment requires only deterministic parsing of tool-use tags, adding zero compute overhead. We evaluate on the HotpotQA and MuSiQue multi-hop Question Answering (QA) benchmarks using Mistral-7B-Instruct and Qwen2.5-7B-Instruct with QLoRA (NF4 quantization with LoRA adapters), comparing the QLoRA baseline against four RL-trained arms: GRPO, S-GRPO, SA-GRPO, entropy-weighted baseline under identical training conditions. Using a normalized answer-quality score (0-1) on Mistral/QLoRA, all subsampled arms improve over the baseline (0.620) while full GRPO barely moves (0.642). On HotpotQA, S-GRPO leads (0.703) with SA-GRPO close behind (0.684, 50% fully correct vs 40% baseline), under the same expected token budget ($k = 100$). On genuine 2-search trajectories under Qwen, where role-based selection provides maximum benefit, SA-GRPO scores 0.820 versus S-GRPO's 0.777. On the harder compositional questions of MuSiQue, SA-GRPO leads in both regimes: under the resource-constrained Mistral setting it is the only arm to improve over the baseline (0.416 vs 0.349, McNemar $p = 0.001$ vs base), while GRPO falls below base and the entropy baseline collapses; under the stronger Qwen base all arms improve and SA-GRPO scores highest (0.590 vs 0.517) though within noise of GRPO (0.579). Beyond the method, we identify a key empirical finding: role-based selection requires multi-hop trajectory diversity to differentiate from position-based selection, a design guideline with implications for evaluation design in agentic RL.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "GRPO", "RL", "Multihop", "SGRPO" ], "forum_url": "https://openreview.net/forum?id=7jOGNBSsvG", "pdf_url": "https://openreview.net/pdf/0971b35e3aeeea5c990417c85517522c07fd1bb5.pdf", "updated_at": "2026-10-05T21:07:18.449Z", "categories": [ "training", "reasoning", "theory" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "lL0RG6hq8D", "number": 50697, "title": "The Calibration Set Is in the Checkpoint: How Release Geometry Governs Membership Leakage in Post-Training Quantization", "abstract": "Quantizing a model often uses a small calibration set after training is complete. This set can remain encoded in the released checkpoint. In SmoothQuant-style methods, comparing the released checkpoint with its public base model recovers activation statistics computed from the calibration data. A maximum can copy a calibration record's activation exactly into the checkpoint, while averaging removes this exact constraint but can keep other membership signal. We show that leakage grows on a common scale, the number of released channels per calibration record ($d/k$), while the released statistic sets its shape. On OPT-125M, the reference SmoothQuant implementation and NVIDIA Model Optimizer's stock exporter each identify about 68% of calibration records at measured false-positive rates below $10^{-4}$. Changing only the statistic in a Qwen2.5-7B exporter changes detection at low false-positive rates by an order of magnitude at essentially unchanged perplexity, and the same effect appears in isotonic probability calibration. Post-training calibration can therefore itself be a privacy-relevant release mechanism.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "membership inference", "post-training quantization", "calibration data", "privacy", "SmoothQuant", "model compression", "extreme value theory", "probability calibration" ], "forum_url": "https://openreview.net/forum?id=lL0RG6hq8D", "pdf_url": "https://openreview.net/pdf/d6d2a9281bd10ada240b6cfe38a0c29d4310ddb8.pdf", "updated_at": "2026-10-05T21:07:18.888Z", "categories": [ "ptq", "training" ], "matched_terms": [ "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "VuzgEFCuwx", "number": 50717, "title": "Measure First: The Noise Floor of an Autonomous Research Loop Dominates Its Accept Rule", "abstract": "Autonomous research agents change their code, retrain, evaluate, and keep the change if the score went up. Recent work proposes better statistical tests for this keep-or-discard decision. We argue that the first problem is that nobody measures how noisy the score is, and we measure it in four ways. First, we re-run one production LLM distillation-and-quantization pipeline, unchanged, thirty times in each of sixteen settings (480 runs): run-to-run spread varies more than 13-fold across settings, and the improvement the pipeline's own authors were chasing cannot be distinguished from this noise (its 95% upper bound is 0.54 of it). Second, we read the source code of 26 systems: sixteen of eighteen deployed agent loops keep a change after a single run, and none lets a noise estimate reach the decision. Third, we replay eleven decision rules on the measured runs. The anytime-valid e-process, written against a noisy baseline as is natural, is not valid (false-accept 0.038, 0.148 and 0.278 at 8, 32 and 128 runs); corrected, its best variant reaches 0.048 power at half a standard deviation, where a five-run Welch t-test reaches 0.170 and needs no external noise estimate. Taking that estimate from three runs breaks the corrected rule in all sixteen settings. Fourth, we evaluate the same frozen weights 3,627 times on seven NVIDIA GPU models. Scores move in whole questions; 13 of 39 model, task and GPU combinations return the identical score every time while others drift by up to thirty questions; two GPUs that are each perfectly repeatable still disagree by twenty; and running inputs one at a time instead of batched removed the drift in every case we tested. We recommend two things: before trusting any comparison, re-run the unchanged pipeline several times on the hardware the loop will use, to learn how much the score moves on its own; and decide with a test that estimates the noise from the runs themselves, such as a Welch t-test, rather than one handed a noise level from outside.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "evaluation; reproducibility; autonomous research; agents; measurement; anytime-valid inference; nondeterminism" ], "forum_url": "https://openreview.net/forum?id=VuzgEFCuwx", "pdf_url": "https://openreview.net/pdf/3ee084a0605dd2bc864069a942c64f3111f9c405.pdf", "updated_at": "2026-10-05T21:07:21.575Z", "categories": [ "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "NR8lYUNnka", "number": 50775, "title": "Query Complexity of Forward-Only Hebbian Learning: Theory and a Native Fit for Diffusion Models", "abstract": "We study the query complexity of forward-only, gradient-free learning rules for training neural networks, motivated by settings where backpropagation is structurally unavailable: activation-tape memory that grows with depth and context, and objectives (quantized weights, non-differentiable rewards) whose exact gradient is zero almost everywhere. We analyze a three-factor Hebbian rule, an antithetic zeroth-order gradient estimator that perturbs parameters along a random probe and broadcasts a single scalar score to every synapse, and derive an exact, non-asymptotic identity for its gradient alignment as a function of probe count and problem dimension, yielding a closed-form query budget $M^* \\approx (P+1)/3$. We relate this upper bound to known minimax lower bounds for zeroth-order stochastic convex optimization, characterizing how close the rule's query complexity comes to the best possible rate for any gradient-free method. We further identify an architecture-dependent separability criterion, a measurable coupling scope $S$ and feedback bandwidth $K$, that predicts when this query cost stays small, and confirm it empirically on a diffusion denoiser, where the per-noise-level residual gives $S=1$: the rule closes most of the gap to backpropagation from scratch while eliminating its activation-memory overhead entirely, and continues to descend the loss through a quantized codebook argmin where the exact gradient is identically zero.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "query complexity", "zeroth-order optimization", "lower bounds", "Hebbian learning", "gradient-free training", "diffusion models", "evolution strategies", "learning theory", "memory-efficient training" ], "forum_url": "https://openreview.net/forum?id=NR8lYUNnka", "pdf_url": "https://openreview.net/pdf/8ee83b160a35153787765ef33ad137fdc20e72ca.pdf", "updated_at": "2026-10-05T21:07:28.343Z", "categories": [ "training", "diffusion" ], "matched_terms": [ "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LVON04OMU2", "number": 50779, "title": "Task-Aware QUBO Allocation for Mixed-Precision Quantization", "abstract": "Mixed-precision quantization requires discrete allocation of weight and activation bit-widths, followed by recovery of the selected network. We develop a task-aware quadratic unconstrained binary optimization (QUBO) surrogate with separate weight and activation profiles, a bit-operation (BOP) cost, and selected structural priors, then refine the resulting route through direct validation-based PROTES search. Across three public pretrained SIDD checkpoints with source-image-disjoint calibration and evaluation, the production simulated-annealing QUBO routes are within 0.01 dB of the strongest SFC routes in a retrospective 19-solver comparison. Refinement gives single-run static gains of 0.27–1.01 dB over QUBO and 0.39–1.34 dB over the HAWQ-style allocator at approximately 4.1% routed-layer BOPs, while remaining 0.39–0.87 dB below FP32. On a compact NAFBlock-based denoiser evaluated on a held-out patch split with shared source images, the refined route improves static PSNR over HAWQ-style by 0.21–2.46 dB at nearby costs across the three tightest requested targets. Recovery narrows these differences and can change route ordering, so we evaluate allocation quality separately before and after LSQ+. We report achieved cost, search variation, and optimization expense; BOP reductions are analytical, and the reference inference path remains floating point.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "mixed-precision quantization", "neural network compression", "QUBO", "combinatorial optimization", "PROTES", "image denoising" ], "forum_url": "https://openreview.net/forum?id=LVON04OMU2", "pdf_url": "https://openreview.net/pdf/687fce641bad67f85c1a20e1bc1a3acc2bb49740.pdf", "updated_at": "2026-10-05T21:07:28.807Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "binary", "mixed-precision", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "kY0EvTgJsL", "number": 50820, "title": "Key-Value Means: Transformers with Expandable Block-Recurrent Compressed Memory", "abstract": "Recall presents a difficult choice: transformers have a linearly growing memory that slows each successive token, while linear RNNs typically have fixed costs but limited recall. We present Key-Value Means (\"KVM\"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth.\nKVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. We will release our code and trained models under the Apache 2.0 license.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "attention", "sparsity", "RNN", "LRNN", "recurrent", "KV", "KV-Cache", "vector quantization", "transformer", "TTT", "test time training", "linear attention", "rwkv", "gdn", "kda", "compressive state" ], "forum_url": "https://openreview.net/forum?id=kY0EvTgJsL", "pdf_url": "https://openreview.net/pdf/1356542b3bf64ef58ba22f5293245319b0509c87.pdf", "updated_at": "2026-10-05T21:07:33.646Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "b2gtUEJtve", "number": 50846, "title": "MoE-PACT: Confidential MoE Inference via Searched Bit Partitioning and Optimal Clustering of Expert Tensors", "abstract": "Mixture-of-Experts (MoE) architectures are widely deployed thanks to their low training and inference compute cost, but their large parameter counts impose severe memory and host-to-device (H2D) transfer bottlenecks. These bottlenecks are most severe on cloud server nodes whose H2D bandwidth is restricted by NVIDIA Confidential Computing (CC) mode. Existing solutions for memory and bandwidth overhead come at the price of degraded accuracy, or of expensive retraining or auxiliary memory for mixed-precision weight storage.\n\nWe present MoE-PACT, an MoE inference system that reduces memory footprint and H2D traffic under Confidential Computing. We first systematically analyze NVIDIA CC mode to characterize the PCIe bottlenecks introduced by encryption and decryption, achieving a $5.3\\times$ bandwidth gain through vector AES and multiprocess transfers. We then propose an exhaustive per-tensor search over base-delta bit partitions with optimal 1D $k$-means clustering, and a fused GPU kernel that performs bit-unpacking, lookup-table (LUT) base-delta reconstruction, and dequantization on the fly. Against the strongest baselines under CC, MoE-PACT stores 25% fewer expert bytes and cuts H2D traffic by up to half, accelerating decoding by up to $1.65\\times$ at batch size one and $2.01\\times$ at large batches, with negligible accuracy degradation.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Mixture of Experts", "Confidential Computing", "Offloading" ], "forum_url": "https://openreview.net/forum?id=b2gtUEJtve", "pdf_url": "https://openreview.net/pdf/a8b41a89c7ca4737c6544ffc49d684562433f408.pdf", "updated_at": "2026-10-05T21:07:40.618Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "mixed-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "N3DesjnN82", "number": 50887, "title": "Privacy-Aware Post-Training Quantization for Mitigating Membership Inference in Diffusion Models", "abstract": "Diffusion models are vulnerable to membership inference attacks that can reveal whether a sample was used during training. \nExisting defenses typically rely on privacy-aware training, which adds computational cost and can degrade generation quality. \nMeanwhile, model quantization provides an efficient way to compress AI models and can affect their membership privacy, yet existing quantization methods are not specified to optimize diffusion models for membership leakage. \nIn this work, we propose PAQ-Diffusion, a privacy-aware post-training quantization framework that incorporates membership privacy into quantization optimization without requiring privacy-aware retraining. \nPAQ-Diffusion first performs layer-wise privacy-aware reconstruction by aligning the quantization residual distributions of member and non-member samples while preserving reconstruction fidelity. \nIt then introduces timestep-specific activation quantization scales and optimizes them with a denoising-based privacy objective, allowing privacy optimization to be performed independently across timesteps. \nWe evaluate the proposed PAQ-Diffusion on different diffusion models against representative membership inference attacks.\nExperimental results demonstrate that PAQ-Diffusion reduces membership leakage (e.g., reducing SecMI-NN attack AUC by up to 20.62\\%) while maintaining superior generation quality, providing an effective privacy-utility trade-off for quantized diffusion models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post training quantization", "diffusion model", "membership inference attack", "privacy and security" ], "forum_url": "https://openreview.net/forum?id=N3DesjnN82", "pdf_url": "https://openreview.net/pdf/91eee5d0436e750c32f6b9b0d26ba4d2d1e6dd53.pdf", "updated_at": "2026-10-05T21:07:44.519Z", "categories": [ "ptq", "diffusion" ], "matched_terms": [ "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "T2sw3QfYe5", "number": 51004, "title": "Bits Under ZK-LLM: Evaluating Zero-Knowledge-Friendly Quantization for Verifiable Private LLM Inference", "abstract": "Zero-knowledge (ZK) proofs are emerging as a promising approach for enabling private, verifiable LLM governance and auditing, where regulators, users, and auditors need to verify claims about training-data usage or LLM inference-time behavior, while model providers must protect proprietary model parameters. However, despite the growing interest in ZK-LLMs, the understanding of ZK-friendly quantization remains limited. \nThis gap matters because in the ZK setting, quantization directly shapes the arithmetic structure, constraint complexity, and proving cost of ZK inference. ZK protocols operate over finite fields and incur costs that depend heavily on the number and type of arithmetic operations, nonlinearities, and lookup constraints. Understanding ZK-friendly quantization is therefore essential for making ZK-LLMs practical. In this work, we present the first systematic study of ZK-friendly quantization for LLMs. \nWe first formalize the definition of ZK-friendly quantization, capturing the properties required for ZK proof generation. \nWe then evaluate nine language models, including Qwen2.5-14B and the mixture-of-experts model Qwen3-30B-A3B, across a broad design space of weight, activation, and nonlinear lookup table precision. Our results show that activation precision is substantially more sensitive than weight precision, while nonlinear lookup approximations can become the dominant source of utility degradation. Also, we identify RMSNorm inverse-square-root lookups as a recurring bottleneck in several large models and recover near-baseline utility by selectively increasing precision only at the bottleneck.\nFinally, we show that reducing bit-width or lookup-table size does not necessarily yield proportional end-to-end proving savings, showing that conventional low-bit quantization heuristics do not directly translate to ZK proving efficiency and motivating operator-aware precision selection.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM", "Quantization", "Zero-knowledge proof" ], "forum_url": "https://openreview.net/forum?id=T2sw3QfYe5", "pdf_url": "https://openreview.net/pdf/1d84dda70beba5a4686412e268fb0ffb77645dc5.pdf", "updated_at": "2026-10-05T21:07:59.075Z", "categories": [], "matched_terms": [ "bit-width", "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "iMKJmbumOJ", "number": 51020, "title": "RateQuant: Attention-Aware Rate–Distortion Allocation for KV Cache Quantization", "abstract": "Effective KV cache compression combines accurate quantization with precision allocation to preserve model quality within a memory budget. Allocation is coupled through existing attention error and inputs from earlier quantized layers. We introduce RateQuant, a rate--distortion framework allocating KV precision across layers and heads through measured attention responses. These responses capture how adjacent bit-width changes reinforce or cancel the existing attention error. Dynamic programming combines them into budget-preserving joint updates, and remeasurement refreshes the responses for the next round. Full-model cross-entropy on separate calibration data then selects the final allocation. Our analysis connects normalized attention distortion to task loss under bounded suffix sensitivity and establishes exact optimization of the local additive objective. Across three models and two quantizers, RateQuant achieves the lowest perplexity in all 18 settings spanning 2--3-bit budgets, with reductions of 12.6--94.5\\% at 2 bits and 3.2--10.3\\% at 2.5 bits versus each setting's strongest tested baseline. Frozen allocations achieve the highest three-model mean scores in all 30 LongBench evaluation settings (five tasks, two quantizers, and three budgets), without task-specific recalibration.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization", "mixed-precision quantization", "rate-distortion theory", "large language model inference", "post-training quantization", "model compression", "memory-efficient inference" ], "forum_url": "https://openreview.net/forum?id=iMKJmbumOJ", "pdf_url": "https://openreview.net/pdf/d4032e451d652232b0aa287f4b7511e0a4d990df.pdf", "updated_at": "2026-10-05T21:08:00.694Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "3-bit", "bit-width", "mixed-precision", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "e49RXNVfRz", "number": 51039, "title": "Compressed Adaptation of Diffusion Transformers via Globally-Sensitive Sketch", "abstract": "Adapting diffusion transformers under tight memory budgets requires both compression and parameter-efficient fine-tuning.\nExisting sketch-based compression methods use locally sensitive sketches, where each lookup table is tied to a predetermined spatial region and cannot be shared by statistically similar weight groups at different locations.\nWe propose \\textit{Globally-Sensitive Sketch (GSS)}, a unified framework for compression and parameter-efficient fine-tuning. GSS clusters weight groups within each weight matrix by normalized distributional shape, allowing similar groups to share learnable sketches across arbitrary matrix positions. Each weight stores a 4-bit index into its group's assigned sketch, and only the shared sketch entries are updated during fine-tuning.\nGSS is extensively evaluated on six diffusion transformers spanning 2B to 20B parameters. We consider two settings: adaptation for subject-driven personalization and paired image editing, and standalone text-to-image generation.\nGSS has the lowest mean weight quantization error among all evaluated 4-bit baselines and preserves text-to-image generation quality across all six backbones.\nIn addition, GSS achieves competitive quality on both subject-driven personalization and paired image editing compared to full fine-tuning while using over 100$\\times$ fewer trainable parameters.\nWe provide our code at \\url{https://anonymous.4open.science/r/gss-anonymization/}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Compression", "PEFT", "Sketching" ], "forum_url": "https://openreview.net/forum?id=e49RXNVfRz", "pdf_url": "https://openreview.net/pdf/c057d2f572c4add346c3edcc7ac3c8e0c955acc7.pdf", "updated_at": "2026-10-05T21:08:03.742Z", "categories": [ "diffusion" ], "matched_terms": [ "4-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MfGbNWGAaO", "number": 51068, "title": "QUENCH: Prune Iteratively and Quantise Once for Training-Free LLM Compression", "abstract": "Large language models require substantial memory to store their weights. Pruning removes selected weights and quantisation represents them with fewer bits, but both can reduce accuracy. A small low-rank correction can recover some of the lost accuracy without fine-tuning. We introduce QUENCH, a training-free pipeline that designs pruning and quantisation around such a correction. We target the linear layers of the transformer architecture. QUENCH alternates pruning and activation-aware correction, enabling the sparse weights to respond to the previous correction. It then quantises the surviving weights once with elastic-bucket quantisation (EBQ), which learns representative weight values (codebooks) for each output channel, and refits the correction. On LLaMA-3-8B with 2:4 sparsity and 3-bit values, QUENCH reaches WikiText2 perplexity 12.01 and 34.22% ARC-Challenge accuracy, against 12.60 and 31.22% reported for EoRA with 2:4 sparsity and 4-bit weights. We also weigh the improvement from activation weighting against low-rank correction. It improves standalone codebooks by 12.29 points, but provides no consistent additional benefit after a rank-128 EoRA correction.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "large language models", "model compression", "post-training quantization", "pruning", "low-rank approximation", "quantization error compensation", "semi-structured sparsity" ], "forum_url": "https://openreview.net/forum?id=MfGbNWGAaO", "pdf_url": "https://openreview.net/pdf/e944bf5d51c51bcc1ae170eddf0873116a8ba75c.pdf", "updated_at": "2026-10-05T21:08:06.658Z", "categories": [ "ptq" ], "matched_terms": [ "3-bit", "4-bit", "quantisation", "quantise", "quantises", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "JzBNtG1tg6", "number": 51093, "title": "Data-Free Transformer Quantization Using Parameter-Space Symmetry", "abstract": "Transformer models are widely used in many learning tasks but incur large memory and compute costs, limiting their deployability. Post-Training Quantization (PTQ) is a promising solution but can lead to significant performance degradation. Many PTQ methods estimate weight and activation distributions with calibration data to account for outliers and maintain quantized performance. We propose a data-free approach to improve quantization by exploiting parameter space symmetries. We address outliers and high variability in weights by finding a transformation of the model weights that minimizes quantization error variance. Our approach is light-weight, data-free, and can be integrated as a pre-processing step within other PTQ methods. We evaluate our approach by testing quantized large language models on several benchmark tasks.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "parameter-symmetry" ], "forum_url": "https://openreview.net/forum?id=JzBNtG1tg6", "pdf_url": "https://openreview.net/pdf/ddb2fa7bc6c7e9f61f8aa87088e0d29b77bb13fa.pdf", "updated_at": "2026-10-05T21:08:08.328Z", "categories": [ "ptq" ], "matched_terms": [ "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "IeYyISsoBe", "number": 51189, "title": "Quantizing Rhythm with Music Foundation Models", "abstract": "Rhythm quantization, mapping the expressive timings of a musical performance into a score's precise note values, must be learned from fewer than a thousand paired performances and fewer than two hundred scores. We show that a pretrained symbolic-music foundation model can be adapted into a rhythm quantizer with no new parameters: fed the performance as its anticipated control stream, the Anticipatory Music Transformer becomes a fixed-lookahead transcriber of entire songs, fine-tuned on interleaved pairs of events from a performance and its corresponding score. Musical pretraining substantially improves this adaptation. Under an identical training recipe and checkpoint-selection rule, the pretrained model more than doubles whole-song strict note F1 relative to random initialization (8.7\\% vs. 3.2\\%). We further find that teacher-forced fine-tuning in this data-limited regime is susceptible to compounding errors during whole-song autoregressive inference. To address this, we introduce a three-stage training curriculum that progressively bridges the distribution shift between teacher-forced training and model-generated inference. On the 59-performance ASAP test set, our model reaches 9.33\\% strict note F1 and 77.93\\% inter-onset-interval F1, compared with 8.96\\% and 67.81\\%, respectively, for the strongest task-specific baseline. These gains are accompanied by a substantially higher rate of producing outputs at the reference score's time scale.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "rhythm quantization", "symbolic music", "exposure bias", "low-resource learning" ], "forum_url": "https://openreview.net/forum?id=IeYyISsoBe", "pdf_url": "https://openreview.net/pdf/a9728c069bcf9963aa2d5450360fb387d11897ee.pdf", "updated_at": "2026-10-05T21:08:21.422Z", "categories": [], "matched_terms": [ "quantization", "quantizer", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "1mmsBjYOlx", "number": 51366, "title": "IDEAL-KV: Inference-Aligned Codebook Learning and Calibrated Stage Allocation for KV-Cache Compression", "abstract": "Residual-codebook quantization reduces key--value (KV) cache storage, but preserving prediction quality at high compression ratios depends on both the learned codewords and the allocation of residual stages. Reconstruction error alone does not fully characterize the predictive effects of these decisions. We propose IDEAL-KV, which combines inference-aligned codebook learning with continuation-calibrated global allocation. With the language-model backbone frozen, IDEAL-KV learns structured Key and Value codebooks by aligning attention maps, attention outputs, and continuation distributions with the dense model. It then measures the continuation KL induced by each layer--component--depth choice in an otherwise dense cache and uses the resulting additive scores to allocate residual stages under a global KV-payload budget. The selected offline policy retains all historical tokens and supports packed-code attention without materializing the full historical KV cache during decoding. Across three LLM backbones, IDEAL-KV achieves near-BF16 performance on LongBench, RULER-32K, and GSM8K at approximately $8\\times$ historical-KV payload compression. At approximately $16\\times$, LongBench and RULER-32K averages remain within 0.67 and 0.80 points of BF16, respectively. Code is available at \\url{https://anonymous.4open.science/status/IDEAL-KV-7562}.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Large language models", "KV-cache compression", "residual vector quantization", "multi-rate codebooks", "rate allocation" ], "forum_url": "https://openreview.net/forum?id=1mmsBjYOlx", "pdf_url": "https://openreview.net/pdf/29cc83a4dd1aeac19d6758efea986b63939a5b6b.pdf", "updated_at": "2026-10-05T21:08:46.756Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "D61YRmQOq6", "number": 51397, "title": "SOPQ: Simple Optimized Product Quantization via Corrected Reverse Water Filling", "abstract": "This paper presents SOPQ. \nIt is a provably correct and one-shot optimization algorithm for optimizing Product Quantization (PQ).\nThe main combinatorial challenge is that of optimally assigning coordinates to *variable-width* PQ segments.\nOPQ suggests a good fixed-segment-width heuristic.\nSOPQ, on the other hand, provably approximates the optimal solution.\nThe crux of the algorithm is a novel Corrected Reverse Water Filling (CRWF) formula that accounts for dimension-dependent effects in $k$-means clustering. \nThereby, converting the coupled combinatorial problem to standard bin-packing of coordinates to segments by their CRWF allocated bits.\nWe then prove that the SVD rotation globally minimizes the CRWF waterline while simultaneously decorrelating the coordinates.\nSOPQ outperforms EDEN, E-RaBitQ, TurboQuant, and PQ at low bit rates across five \\vqb datasets.\nIt matches OPQ's compression quality while offering strong guarantees, faster and predictable training, and a smaller memory footprint.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Vector Quantization", "SVD", "Reverse Water Filling", "Product Quantization", "KV Compression" ], "forum_url": "https://openreview.net/forum?id=D61YRmQOq6", "pdf_url": "https://openreview.net/pdf/49fd638020b8fd34a40dee10b1d74e88786faec0.pdf", "updated_at": "2026-10-05T21:08:51.876Z", "categories": [ "ptq" ], "matched_terms": [ "low bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "SGGmsRFBIS", "number": 51410, "title": "Positional Specialization for Autoregressive Video KV-Cache Compression", "abstract": "KV-cache compression for autoregressive video is usually framed as deciding how much of the past to preserve. Using persistent, position-matched interventions, we show instead that different temporal positions require different kinds of information. Distant states, including pinned and mid-lag positions, can be spatially collapsed from 1,560 tokens to a single token with negligible loss, whereas the same collapse on recent states causes severe frame-to-frame flicker. Spatial compressibility, however, does not imply dispensability: evicting distant states degrades image quality and nearly halves motion, whereas retaining each as a single token does not, and replacing them with another video's cache shifts output semantics increasingly with semantic distance. These findings motivate AxisKV, a two-axis allocation that spatially compresses distant states while quantizing recent states. At $8\\times$ cache compression on Self-Forcing, AxisKV outperforms matched-memory uniform pooling and, against uniform quantization, trades a loss in aesthetic quality for higher image quality; it also remains usable at $12\\times$, where uniform quantization has no valid operating point. The positional specialization transfers across related autoregressive video generators.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Video generation", "diffusion model" ], "forum_url": "https://openreview.net/forum?id=SGGmsRFBIS", "pdf_url": "https://openreview.net/pdf/6b78a35910cc79506a667c4da675f0202a1d3e74.pdf", "updated_at": "2026-10-05T21:08:54.274Z", "categories": [ "kv-cache", "diffusion" ], "matched_terms": [ "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "egWrgyZaVz", "number": 51672, "title": "KV-Fold: Beyond Chunked Prefill with Manipulable KV State", "abstract": "Long-context inference in pretrained transformers relies on a growing key--value (KV) cache that is usually treated as an internal artifact of attention. We instead view the KV cache as an explicit \\emph{computational state} that can be carried, modified, and transferred between computations. Standard chunked prefill is the identity case of this view: carrying the KV state unchanged across chunks is exactly equivalent, in exact arithmetic, to a monolithic causal forward pass. We prove this equivalence and verify it across several pretrained transformer families, with next-token loss matching full attention to within $0.002$ nats.\n\nThis exact identity provides a controlled reference point for studying what happens when the carried state is changed. Without modifying model parameters, we quantize, selectively retain, re-address, and transfer KV state. The unmodified state preserves exact factual retrieval across contexts up to $128$K tokens and $511$ chunk transitions while reducing peak working memory enough to process a context for which the corresponding monolithic forward does not fit on the same hardware. Under lossy transformations, per-step quantization reveals a graded memory--fidelity trade-off, whereas position removal is substantially less forgiving.\n\nMost strikingly, KV state can serve as a sparse message between otherwise separate inference processes. At $32$K tokens, a consumer that never reads the original document recovers a designated fact in all $20$ trials from a shared subset of only $36$ of $32{,}768$ cache positions; matched-bandwidth random and recency controls recover none. A per-head question-aware selector succeeds even at $20$ retained positions, and RoPE-correct re-addressing allows selected state to be moved into a new positional coordinate frame without loss of retrieval. These results suggest that the KV cache is not merely stored context, but a manipulable state interface through which pretrained transformers can preserve, compress, address, and communicate information without architectural changes or fine-tuning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Persistent KV State", "Stateful Inference", "Long-Context Inference" ], "forum_url": "https://openreview.net/forum?id=egWrgyZaVz", "pdf_url": "https://openreview.net/pdf/a0a1116c30c422080aece0ee9b2670641a128987.pdf", "updated_at": "2026-10-05T21:09:26.445Z", "categories": [ "kv-cache", "hardware" ], "matched_terms": [ "quantization", "quantize" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "ZQepYgzoGR", "number": 51715, "title": "An $\\alpha$-Law for Scale Multipliers in Low-Bit Post-Training Quantization", "abstract": "Preserving language-model accuracy below four bits per weight\nmotivates understanding how quantizers allocate their limited\nrepresentable values.\nWe ask whether post-training quantizers with different optimization\nprocedures choose scales that follow a shared, predictive relationship.\nWe compare their choices through a per-group multiplier $\\alpha$\nthat measures scale contraction relative to the group's largest weight\nmagnitude.\nFor AutoRound, GPTQ, and AWQ, its median follows\n$\\alpha^* = q_{\\max}/(q_{\\max}+c)$, where $q_{\\max}$ is the largest\npositive integer quantization code and $c$ is fitted per method and\nconfiguration.\nA symmetric clipped mean-squared-error approximation interprets\nthis contraction through the trade-off between clipping and\nrounding error.\nAcross seven models from six families at two, three, and four bits\nper weight, the law achieves in-sample $R^2 > 0.97$.\nConstants fitted on models with at most nine billion parameters\npredict the sampled median $\\alpha$ of Llama-3.1-70B with absolute\nerrors below 0.016 for AutoRound and 0.008 for GPTQ at two and\nthree bits per weight.\nAmong the four tested methods, larger fitted $c$ corresponds to\nhigher two-bit accuracy averaged across the seven models.\nFor Llama-3.1-70B quantized to two bits per weight, AutoRound and\nits ablation without rounding iterations have similar sampled\nmedian scale multipliers yet retain 91\\% and 45\\% of the\nhalf-precision reference model's ten-task mean accuracy.\nThe $\\alpha$-law reveals shared structure in method-specific scale\nchoices and supports prediction across bit widths and, for AutoRound\nand GPTQ, to a held-out 70B model.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "calibration data", "low-bit quantization", "large language models" ], "forum_url": "https://openreview.net/forum?id=ZQepYgzoGR", "pdf_url": "https://openreview.net/pdf/05350b23b2db6de94432d4e03afd9d4b1f373cc5.pdf", "updated_at": "2026-10-05T21:09:31.762Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "LoTA2X1BKF", "number": 51837, "title": "FLASHLight: Preserving Accepted Prefixes in Low-Bit Speculative Decoding", "abstract": "Target quantization accelerates speculative verification, making draft execution cost increasingly important. W4A4 drafting reduces this cost, but lower verifier agreement can offset the gain. On DFlash and DSpark, the conditional-acceptance gap to a 16-bit draft widens toward later positions even under basis-consistent rotation-based PTQ (Aligned PTQ). We call this phenomenon quantization-induced suffix decay. We propose FLASHLight, which reframes W4A4 draft quantization as selecting a low-bit representation for the deployed verifier rather than reconstructing the high-precision draft. With target and draft weights frozen, FLASHLight calibrates the quantization-sensitive interface, then selects an independent draft basis using verifier guidance and a prefix-coupled overlap objective for accepted-prefix preservation. On DFlash and DSpark, FLASHLight retains 97.21-97.68% of the greedy-decoding AL of the 16-bit draft paired with the same W4A4 target, improving average AL over Aligned PTQ by approximately 11.8-22.8%. Repeated measurements on a single RTX 4090 show 28.5-29.8% higher mean end-to-end throughput at batch size 1 than that reference.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "speculative decoding", "large language models", "post-training quantization", "low-bit inference", "efficient LLM inference", "model compression" ], "forum_url": "https://openreview.net/forum?id=LoTA2X1BKF", "pdf_url": "https://openreview.net/pdf/ed9f84da4ea94223792448f4366e778aab1124da.pdf", "updated_at": "2026-10-05T21:09:47.672Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "ptq", "quantization", "w4a4" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "sUzaZTsvLQ", "number": 51906, "title": "ROKI: RoPE-Aware KV Cache Quantization via Isofrequency Grouping", "abstract": "Low-bit KV-cache quantization reduces the memory and bandwidth costs of long-context LLM inference, but post-RoPE per-token quantization is sensitive to large activations that dominate the shared quantization range. Cross-channel rotations can reduce these outliers, yet general cross-pair rotations do not commute with Rotary Position Embedding (RoPE) and therefore cannot be folded into Q/K projection weights. We propose ROKI, a RoPE-aware KV-cache quantization framework based on Isofrequency Commutant Expansion (ICE). Rather than restricting transformations to the fixed RoPE operator, ICE selectively ties nearby low-frequency RoPE pairs, creating isofrequency groups that enlarge the exact RoPE-commuting transformation space from pair-wise to cross-pair rotations. ROKI uses this expanded space to redistribute Q/K activations and applies transformation-aware mixed-precision key quantization based on post-transformation QK contributions. All Q/K transformations are optimized offline and folded into the projection weights, requiring no additional online transformation. At an average 3-bit KV-cache budget, ROKI achieves the highest average LongBench-E score among the evaluated quantized baselines on three Llama/Qwen models while remaining close to FP16. Our ROKI\nimplementation also achieves up to 4.22$\\times$ decode speedup over the\nevaluated low-bit baselines at long contexts.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "KV cache quantization" ], "forum_url": "https://openreview.net/forum?id=sUzaZTsvLQ", "pdf_url": "https://openreview.net/pdf/2ab4c960783dafa6dab073f44c9db5c0857c95a9.pdf", "updated_at": "2026-10-05T21:09:56.203Z", "categories": [ "ptq", "kv-cache" ], "matched_terms": [ "3-bit", "low-bit", "mixed-precision", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "x9agewUkzQ", "number": 52039, "title": "CRePE: Convolution-aware Relative Importance in Post-training Pruning with Efficient Search", "abstract": "Post-Training Pruning (PTP) reduces the memory and computational costs of Large Language Models (LLMs) by removing weights without retraining. RIA, a state-of-the-art scoring method, normalizes each weight by its row and column sums, but uses only 1D cross-shaped information and weights the two directions equally. We propose CRePE (Convolution-aware Relative Importance in Post-training Pruning with Efficient Search), which adds a 2D local neighborhood term and adaptive per-term coefficients to Relative Importance scoring. Since searching these coefficients by perplexity (PPL)-based hill climbing takes about 11 hours on LLaMA-2-7B, we further propose PHO (Proxy-based Hyperparameter Optimization), which uses the Gini coefficient of pruned importance scores as a surrogate objective and reduces the search to about 20 minutes, a 30$\\times$ speedup with nearly the same PPL; the resulting coefficients transfer from LLaMA-2-7B to LLaMA-2-13B without re-tuning. Across LLaMA-1/2/3 under 2:4 and 4:8 sparsity and LLaMA-1/2 under unstructured sparsity, CRePE consistently improves over scoring-only baselines, with margins that widen under tighter structural constraints or higher sparsity. Under 2:4 sparsity, CRePE lowers PPL by 0.43 on average relative to RIA. CRePE also combines with channel permutation, non-uniform sparsity allocation, quantization, and re-pruning.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training pruning", "large language models", "semi-structured sparsity", "N:M sparsity", "pruning metric", "hyperparameter optimization", "proxy objective" ], "forum_url": "https://openreview.net/forum?id=x9agewUkzQ", "pdf_url": "https://openreview.net/pdf/ba93b7c734efe29f5c679308645c0c1a4d83d598.pdf", "updated_at": "2026-10-05T21:10:10.978Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Mt7JawGQgl", "number": 52096, "title": "HEPH: 3-Bit Grid-Aligned Trellis Quantization of MXFP4 Mixture-of-Experts for Accurate, Hadamard-Free Serving", "abstract": "A growing class of mixture-of-experts models, including gpt-oss, Kimi K3, and DeepSeek-V4, ships routed experts natively in the MXFP4 hardware microscaling format after quantization-aware post-training. HEPH is a trellis-coded codec for these weights. Existing methods such as QTIP and QuIP# achieve strong rate–distortion performance but decode to arbitrary FP16 values in a randomized-Hadamard-rotated basis. Serving them requires a Hadamard transform on every forward pass and cannot use the native FP4 tensor-core kernel. HEPH instead targets the hardware grid: it reuses each checkpoint block’s UE8M0 exponent as its exact scale and restricts the trellis codebook to the 15 distinct E2M1 values, making every decoded weight a valid MXFP4 element.\n\nThis design improves both accuracy and serving speed. On gpt-oss-20b, HEPH achieves **80.03** mean accuracy across five zero-shot benchmarks at 3.04 bits per weight, versus **77.72** for a rate-matched QTIP baseline and within 0.76 points of the 4.25-bpw MXFP4 checkpoint. When both compressed streams remain resident and decode weights in-kernel on each forward pass, HEPH reaches **96.1** tokens/s for batch-1 decode on a Grace Blackwell GPU in SGLang, versus **74.2** for QTIP ($1.30\\times$); it remains $1.1$–$1.5\\times$ faster across batches of 1–32. HEPH also supports a second serving mode: its approximately 3-bpw storage representation can be expanded into a resident 4.25-bpw MXFP4 tensor and executed by the unmodified production fused grouped-MoE kernel. This trades resident memory for throughput while retaining the storage saving. In this mode, HEPH reaches **382** tokens/s at batch 1, versus **72.8** for QTIP ($5.25\\times$), with speedups of $5$–$29\\times$ across batches of 1–32. Aligning the codec with the native FP4 grid thus yields both higher accuracy at a matched compression rate and faster mixture-of-experts serving.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "weight quantization", "post-training quantization", "trellis-coded quantization", "mixture-of-experts", "MXFP4", "microscaling formats", "LLM serving", "inference throughput", "hardware-aligned quantization", "Blackwell GPU" ], "forum_url": "https://openreview.net/forum?id=Mt7JawGQgl", "pdf_url": "https://openreview.net/pdf/93095cf18bee0909caea7829308556569f0575af.pdf", "updated_at": "2026-10-05T21:10:17.348Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "3-bit", "fp4", "mxfp4", "quantization" ], "review_status": "abstract", "summary": "HEPH 压缩已有 MXFP4 expert 权重，同时保持解码值可直接进入原生 FP4 执行网格。它复用 checkpoint block 的 UE8M0 exponent，并把 trellis 码本限制为 15 个 E2M1 值，避免每次前向做 Hadamard 变换；部署可选择内核内解码压缩权重或先展开为驻留 MXFP4。", "contribution": "gpt-oss-20b 在 3.04 bpw 的五项零样本平均准确率为 80.03，率匹配 QTIP 为 77.72；Grace Blackwell、SGLang 的 batch-1 内核解码模式为 96.1 对 74.2 tokens/s（1.30 倍）。先展开驻留 4.25-bpw 模式为 382 对 72.8 tokens/s（5.25 倍），但需要更多驻留内存。", "limitations": "仅依据官网摘要，未读官方 PDF，实验设置未核实。两种 serving 模式的内存条件不同：5.25 倍结果以展开驻留为代价，不能当作同样约 3-bpw 驻留条件的加速；摘要数值也限定于所述模型、硬件和 batch。", "review_basis": "official_abstract", "official_pdf_verified": false, "quantization_target": "原生 MXFP4 MoE routed expert 权重的 grid-aligned trellis codec", "bit_width": "报告 3.04 bpw 存储；解码为 MXFP4，另有 4.25 bpw 驻留展开模式", "evidence": [ { "location": "官方摘要", "claim": "3.04 bpw 下 gpt-oss-20b 五项平均准确率 80.03，率匹配 QTIP 为 77.72。" }, { "location": "官方摘要", "claim": "batch-1：压缩流内核解码 1.30 倍加速；展开驻留 MXFP4 模式 5.25 倍加速，后者增加驻留内存。" } ] }, { "id": "lmBI0rWxso", "number": 52240, "title": "OrBIT: Structure-Guided Embedding Compression", "abstract": "Embedding tables are among the largest components of modern language models. Most compression methods fix a coding geometry such as coordinate blocks, low-rank subspaces, or unrestricted codebooks, and optimize within it. We instead ask whether the coding geometry can itself be discovered. We introduce OrBIT, a structure-guided embedding compression framework that learns reusable local geometry from orbit dynamics and uses it to constrain a small set of shared codewords. The global reconstruction residual then decides where the fixed coding budget is spent, while redundant overlapping charts let local errors compensate one another after gluing. Our theory shows how tight-chart geometry controls distortion, how the global residual directs sequential allocation, and how data-geometry-guided refinement improves the codec. The resulting orbit machinery is compiled away, leaving a compact decoder in which the learned structure governs what is stored, where capacity is allocated, and how local information is assembled globally. Across four LLM embedding tables, OrBIT achieves $37.9\\times$ compression on GPT-2 and over $23\\times$ on each 7B table relative to 16-bit storage, while delivering competitive rate-distortion performance against established quantization and low-rank baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "learned geometry", "structured codebooks", "orbit dynamics", "representation compression", "local-to-global learning" ], "forum_url": "https://openreview.net/forum?id=lmBI0rWxso", "pdf_url": "https://openreview.net/pdf/1f7c7b355b6e4cae840aa72fc7fc3613c70fecc1.pdf", "updated_at": "2026-10-05T21:10:34.314Z", "categories": [ "ptq" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MmjYzay4iJ", "number": 52297, "title": "QuantThink: Quantized Reasoning Models Can Learn When to Commit", "abstract": "Large reasoning models (LRMs) achieve strong performance through extended chains of thought, but their long outputs impose substantial KV-cache and inference overhead. Although KV-cache quantization effectively reduces memory cost, we find that it can paradoxically prolong reasoning by inducing repeated post-solution verification and delaying answer generation even after a correct solution has already been reached. To address this issue, we introduce QuantThink, which learns when reasoning is sufficient to terminate. QuantThink combines (i) a sufficiency-aware trajectory formulation that identifies checkpoints for verification continuation or reasoning termination, and identifies redundant post-solution verification, and (ii) a parameter-efficient reinforcement learning method Q-GRPO that optimizes termination decisions under quantized inference. Across DeepSeek-R1-Distill-Qwen, QwQ, and Phi-4 models, five reasoning benchmarks, and three KV-cache quantizers---KVQuant, KIVI, and RotateKV---QuantThink improves pass@1 by up to 4.32 percentage points while reducing generated tokens by up to 63.1\\% compared with the KIVI K2V2 baseline.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reasoning Model", "Quantization", "Chain-of-Thought", "Reinforcement Learning", "Early Stopping" ], "forum_url": "https://openreview.net/forum?id=MmjYzay4iJ", "pdf_url": "https://openreview.net/pdf/cdb945573c11a6bf127c53902711dc58602c2be7.pdf", "updated_at": "2026-10-05T21:10:42.848Z", "categories": [ "kv-cache", "reasoning" ], "matched_terms": [ "quantization", "quantized", "quantizers" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "MjqcLg39np", "number": 52306, "title": "Beyond a Single Deterministic Teacher: Invariant Synthesis for Data-Free Quantization of Vision Transformers", "abstract": "Data-Free Quantization (DFQ) of Vision Transformers (ViTs) enables post-training quantization without real calibration data by synthesizing surrogate calibration samples from a pre-trained full-precision teacher model. However, existing DFQ methods typically optimize synthetic samples against a single deterministic teacher, which overlooks their sensitivity to internal structural variations, such as attention-head responses and spatial token dependencies. As a result, synthesized samples may exhibit structural brittleness, leading to unreliable calibration statistics and degraded quantization performance. To address these issues, we revisit data-free synthesis for ViTs from the perspective of invariant synthesis over structurally perturbed teachers. We theoretically show that: 1) structural sensitivity of synthetic samples contributes to the discrepancy between real and synthetic calibration statistics; and 2) structurally perturbed teachers can expose response variations hidden by deterministic supervision. Motivated by these insights, we propose DropIS, a novel and effective Dropout-Induced Invariant Synthesis approach for DFQ of ViTs. Unlike conventional dropout regularization, DropIS employs structured dropout as a teacher probing mechanism during synthesis: attention head dropout suppresses excessive reliance on specific head responses, while spatial token dropout mitigates dependence on narrow local token regions. Furthermore, DropIS incorporates auxiliary attention-space regularizers to stabilize perturbed responses and synthesize semantically consistent calibration samples. Extensive experiments across diverse backbones and downstream tasks verify the advantages of DropIS over the state-of-the-art methods, achieving up to 8.46% Top-1 accuracy improvements on ImageNet under challenging low-bit regimes. Our code is available in supplementary material package.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Data-Free Quantization", "Vision Transformers", "Invariant Synthesis", "Structural Probing" ], "forum_url": "https://openreview.net/forum?id=MjqcLg39np", "pdf_url": "https://openreview.net/pdf/5b7373cda9f752a4d39ae9423e6f7b3502fe26ab.pdf", "updated_at": "2026-10-05T21:10:43.893Z", "categories": [ "ptq", "theory" ], "matched_terms": [ "low-bit", "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "iFMhFS2r1E", "number": 52322, "title": "NibbleMLA: Dual-Axis FP4 Attention over a Single Latent Cache", "abstract": "Multi-head latent attention (MLA) reduces the key-value cache to a single latent per token that serves as both key and value. What remains is numeric: how many bits store that latent, and in what precision the two attention GEMMs read it. FP4 latent caches already exist, but they dequantize to BF16 before attention. We show that this is forced by a structural mismatch rather than by accuracy: block-scaled FP4 tensor cores accept scale factors only along the contraction dimension, whereas the shared latent is contracted along channels by QK and along tokens by PV. We prove that a scale written once per token can be consumed by both GEMMs only if it is a per-token scalar times a static channel vector—precisely the layout that suits raw latents worst, because their channel magnitudes are highly uneven. NibbleMLA resolves this tension by changing the basis instead of the scale: a fixed $512$-dimensional Hadamard co-rotation, which costs no storage, equalizes the channels so that one E4M3 scalar per token suffices. The resulting $257$-byte-per-token cache ($0.44\\times$ FP8) feeds both attention GEMMs on FP4 tensor cores without dequantization. On DeepSeek-V4-Flash, FP4 attention matches BF16 attention computed on the same cache (output cosine $\\ge 0.9999$ in all $43$ layers); end-task accuracy is statistically indistinguishable from the FP8 cache on GSM8K, AIME, and LongBench-v2 and within $0.7$ points on $13$-task RULER at $128$K; and the decode-attention kernel is $1.10$–$1.27\\times$ faster than the FP8 kernel and, on average, faster than the official FP4 kernel, which dequantizes.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Multi-head Latent Attention", "FP4 attention", "KV cache compression", "paged attention", "long-context inference", "low-precision inference" ], "forum_url": "https://openreview.net/forum?id=iFMhFS2r1E", "pdf_url": "https://openreview.net/pdf/2ae25101c7823127ffbf19bc3a1c89bc20ebf615.pdf", "updated_at": "2026-10-05T21:10:44.839Z", "categories": [ "ptq", "training", "kv-cache", "hardware" ], "matched_terms": [ "fp4", "fp8", "low-precision" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "cNE7lJxw6i", "number": 52378, "title": "Beyond Local Error: Functional Reconstruction for Vision-Language Model Quantization", "abstract": "Post-training quantization (PTQ) aims to reduce the cost of vision-language models\nwhile retaining their capabilities. Reconstruction-based methods pursue this goal\nby matching quantized intermediate outputs to their full-precision counterparts,\nwith modality-aware approaches further balancing visual and textual errors. Yet an\nerror’s impact depends on what happens next: subsequent operations can respond\ndifferently to perturbations of similar magnitude. A smaller local error therefore\nneed not better preserve the computation that follows. We introduce Modality-\nNative Functional Equalization (MNFE), a PTQ method based on multimodal\nfunctional reconstruction. MNFE evaluates quantization errors at module-specific\nstates formed by downstream operations and balances their contributions across\nvisual evidence, textual context, and answer information. A structured search\nrefines channel scales through bounded group-wise adjustments. Extensive ex-\nperiments on Qwen2.5-VL-7B and InternVL2.5-8B show that MNFE achieves\nthe highest nine-benchmark average among evaluated quantized methods under\nW3A16, W4A8, and W4A6. Notably, under W4A6, MNFE retains 94.8% of the\nfull-precision average score on InternVL2.5-8B. Our code will be made publicly\navailable upon acceptance.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "post-training quantization", "vision-language models", "multimodal large language models", "model compression" ], "forum_url": "https://openreview.net/forum?id=cNE7lJxw6i", "pdf_url": "https://openreview.net/pdf/598691077fd4c29f792f2c46a8ed4a5cbcf4263d.pdf", "updated_at": "2026-10-05T21:10:51.997Z", "categories": [ "ptq", "multimodal" ], "matched_terms": [ "ptq", "quantization", "quantized", "w3a16", "w4a6", "w4a8" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "rPwh8vaDni", "number": 52413, "title": "PipeSearch: Buying Composition Labels Only Where They Change the Pipeline", "abstract": "Serving a large language model on a single accelerator asks quantization, pruning,\nlow-rank factorization and cache compression to share one memory and throughput\nbudget, so what is deployed is an ordered composition of methods rather than a\nsingle choice. Choosing that composition needs the pairwise interactions of the\nmethods, yet they have to be bought: a pairwise label costs a full compression pass,\nreversing two methods yields a different model, and some pairings refuse rather than\ndegrade. We propose PipeSearch, which buys a pairwise label only while it can still\nchange the returned pipeline, that is, while the pipelines of its optimistic and\npessimistic completions disagree. Decision-relevant acquisition observes the cheapest\nunobserved pair either completion applies, order-aware prefix search carries which\nmethod acted last and keeps a sequence apart from its reverse, and feasibility\nmaintenance excludes a refused pair at expansion time. The key insight is that the\nobservations which reveal the interaction structure are part of the decision rather\nthan a preliminary to it. Across catalogues of four to twelve atoms, PipeSearch\nreturns the pipeline of complete offline labelling while building 0.201 to 0.500 of\nthe $K(K-1)$ ordered pairs, at 75.1 accelerator-hours against 367.4 at $K{=}12$.\nPaying for every candidate pair instead leaves perplexity, zero-shot average, peak\nmemory and decode rate unchanged at both memory budgets, at 15.0 and 3.6 times the\nsearch cost. Under a cap of $0.14\\times$ the fp16 footprint, where no single\nmethod fits, PipeSearch fits Llama-3.1-8B at a predicted 8.79 GB and 9.59 perplexity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "LLM compression pipelines", "composed model compression", "order-aware pipeline search", "cost-aware label acquisition", "memory-constrained LLM deployment", "post-training quantization and pruning" ], "forum_url": "https://openreview.net/forum?id=rPwh8vaDni", "pdf_url": "https://openreview.net/pdf/4aa459d04b5c765e8f70398bf19082321ce0da2f.pdf", "updated_at": "2026-10-05T21:10:57.729Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "quantization" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "6egpMpGwZE", "number": 52450, "title": "LinQuant: Functional Invariances for Low-Precision Linear Recurrences", "abstract": "Linear recurrences are emerging as practical alternatives to softmax attention in large language models, appearing in both standalone and hybrid architectures. However, despite their rapid emergence, low-precision inference remains much less developed for these models than for softmax attention. During prefill, chunkwise-parallel algorithms cast much of the recurrence as dense matrix multiplications, suggesting a natural path to low-precision execution. Simply quantizing these operations in their native form, however, can be fragile. In this work, we identify functional invariances that preserve the full-precision computation while making the operands more amenable to quantization. Building on this observation, we introduce _LinQuant_, a framework for low-precision recurrent prefill based on seven such invariances, including rotations, diagonal rescalings, and alternative placements of recurrent decay and write factors. We derive when these transformations remain valid across common recurrence families spanning linear, gated, and delta-rule updates, and show that they provide complementary degrees of freedom for improving low-precision accuracy. Finally, we translate this framework to production by augmenting SSDi8, an existing low-precision prefill pipeline for Mamba-2, with the applicable invariances. _LinQuant_ improves perplexity at matched latency on Mamba-2 1.3B and 8B and Nemotron-H 8B, demonstrating the practical utility of invariance-based quantization for recurrent models.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "quantization", "linear recurrence", "SSM", "efficiency", "kernels" ], "forum_url": "https://openreview.net/forum?id=6egpMpGwZE", "pdf_url": "https://openreview.net/pdf/7004871577bf637b9cf2d2ea2250e7d1374311ca.pdf", "updated_at": "2026-10-05T21:11:02.146Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-precision", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Q4gS70AH9E", "number": 52494, "title": "A FEW GOOD WEIGHTS: SPARSE SUBNETWORKS IN PREFERENCE-OPTIMIZED LLMS", "abstract": "Reinforcement learning fine-tuning of Large Language Models (LLMs) induces sparse, full-rank weight updates concentrated subnetworks containing fewer than $2.5\\%$ of parameters. We show the structure of these subnetworks is determined primarily by the learning objective, rather than the training data. Subnetworks produced by DPO and GRPO differ substantially in parameter selection and internal representation, even on similar data, while subnetworks produced by the same objective on different datasets remain closely aligned. To establish this result rigorously, we develop a theoretical framework for identifying the sparse subnetwork by scoring each parameter's importance. We derive an optimal Oracle scoring function with guarantees on loss retention, prove the existence of an optimal scoring function at any sparsity level, establish a certifiability condition for evaluating practical scoring functions, and develop a scalable threshold estimation algorithm for constructing subnetworks in multi-billion-parameter LLMs. We propose warm-start magnitude scoring, which recovers dense DPO and GRPO training dynamics at or beyond $97.5\\%$ sparsity across mathematics, coding, syntax, and semantics. We characterize objective-specific subnetwork divergence through Jaccard similarity, Centered Kernel Alignment, and linear probing. Finally, we introduce a sparsity-aware AdamW optimizer kernel implemented in Triton that reduces optimizer step complexity by the factor $(1-\\rho)$, achieving $\\sim1.7\\times$ and $\\sim3.4\\times$ speedups over PyTorch and HuggingFace 8-bit AdamW respectively at $97.5\\%$ sparsity.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Reinforcement Learning", "Sparse Reinforcement Learning", "Sparse Optimization", "Preference Optimization" ], "forum_url": "https://openreview.net/forum?id=Q4gS70AH9E", "pdf_url": "https://openreview.net/pdf/d18e334da350b487ef343ff5978071233322368a.pdf", "updated_at": "2026-10-05T21:11:07.495Z", "categories": [ "training", "hardware", "theory" ], "matched_terms": [ "8-bit" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "Ydv8uEEheu", "number": 52594, "title": "Layer-wise Sensitivity-aware Sparsity Allocation for Efficient LLM Inference", "abstract": "Large Language Model (LLM) inference presents substantial computational challenges when executed on commodity hardware, thereby necessitating the development of efficient acceleration techniques. While existing approaches predominantly focus on uniform compression strategies, they neglect the heterogeneous sensitivity patterns exhibited across different transformer layers. In this paper, we introduce the Adaptive Sparsity Allocation Framework (ASAF), which integrates rotation-based low-bit quantization with layer-wise adaptive sparsity allocation. The framework comprises two sequential phases with a dynamic programming strategy. In Phase 1, coarse-grained optimization determines the optimal number of layer groups and narrows sparsity rate search intervals. In Phase 2, fine-grained optimization determines precise consecutive layer allocation and exact sparsity rates within each group. The joint optimization of layer grouping decisions and sparsity rate assignments creates a combinatorial explosion in the solution space, rendering brute-force approaches computationally prohibitive; we employ a dynamic programming strategy that decomposes the exponential search space into manageable subproblems across both phases, achieving practical computational efficiency while guaranteeing optimality within the discretized two-phase search space. Experiments on the Llama-2 family, with additional memory and zero-shot results on Llama-3, reveal that our proposed framework sustains average benchmark accuracy degradation within 1\\%, concurrently achieving up to 3.63$\\times$ prefill acceleration over an FP16 baseline and 12.63\\% memory reduction over the quantized baseline, on a single NVIDIA RTX 3090 GPU.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "efficient LLM inference", "layer-wise sparsity allocation", "rotation-based quantization", "joint sparsity and quantization", "layer sensitivity", "dynamic programming" ], "forum_url": "https://openreview.net/forum?id=Ydv8uEEheu", "pdf_url": "https://openreview.net/pdf/6462419f20cbf7cfb8486a3742803f6f5d7900ca.pdf", "updated_at": "2026-10-05T21:11:23.988Z", "categories": [ "ptq", "hardware" ], "matched_terms": [ "low-bit", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "4P0gseVIi1", "number": 52635, "title": "Feature Correlation Preserving Quantization for Analog Compute-in-Memory", "abstract": "Analog compute-in-memory (ACIM) enables efficient neural network inference by performing matrix operations within memory arrays, but its limited numerical precision requires low-bit quantization. Existing post-training quantization (PTQ) methods mainly minimize reconstruction error between full-precision and quantized representations. This objective does not explicitly preserve feature relations, which can be important when multiple features share the same quantization configuration within an ACIM mapping group. We propose Feature Correlation Preserving Quantization (FCPQ), a groupwise PTQ framework that preserves feature relations while optimizing reconstruction. FCPQ consists of two components. First, Feature Correlation Calibration (FCC) standardizes features and aligns the correlation matrices of full-precision and quantized representations within each mapping group. Second, Correlation Aware Quantization Optimization (CQO) estimates the contribution and distortion of feature relations and uses them to assign adaptive reconstruction weights. Together, these components guide quantization toward preserving both numerical values and feature relations. Experiments across multiple datasets and quantization settings show that FCPQ improves model accuracy and maintains the feature structure of full-precision models under low-bit quantization.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Low-Bit Quantization", "Feature Correlation", "Analog Compute-in-Memory", "Efficient Inference" ], "forum_url": "https://openreview.net/forum?id=4P0gseVIi1", "pdf_url": "https://openreview.net/pdf/e95dddfeb72ce10438bc2c9d72db21aff0fdee4c.pdf", "updated_at": "2026-10-05T21:11:29.443Z", "categories": [ "ptq" ], "matched_terms": [ "low-bit", "ptq", "quantization", "quantized" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "bheeBphxZf", "number": 52651, "title": "QuantLRM: Quantization of Large Reasoning Models via Fine-Tuning Signals", "abstract": "Weight-only quantization is important for compressing Large Reasoning Models (LRMs). However, existing post-training quantization (PTQ) methods are suboptimal for LRMs that are typically fine-tuned from base models, because they leave their fine-tuning traces wasted during PTQ. Inspired by the spirit of classical magnitude pruning, we study whether the magnitude of weight updates during reasoning-incentivized fine-tuning can provide valuable signals for quantizing LRMs. We hypothesize that the smallest and largest weight updates are more important than those of intermediate magnitude, a phenomenon we term \"protecting both ends\". Upon hypothesis validation, we introduce QuantLRM, the first method for weight quantization of LRMs that explicitly leverages fine-tuning signals. We fit simple restricted quadratic functions on weight updates to protect both ends. By multiplying the average quadratic values with the count of zero weight updates of channels, we compute channel importance that is more effective than using activation or second-order information, which guides quantization by protecting important weights. To demonstrate the effectiveness and applicability of QuantLRM, we evaluate sub-4-bit quantization performance on various fine-tuning strategies (SFT, DPO, and RLVR) over different model families (Llama, Qwen, and Olmo) on a range of reasoning benchmarks (AIME-120, FOLIO, temporal sequences, and GPQA-Diamond). QuantLRM delivers consistent gains in LRMs quantization, improving average accuracy by up to 6.55% over the best baselines.", "source_kind": "openreview", "active_submission_verified": true, "keywords": [ "Quantization", "Weight-Only Quantization", "LLMs Compression", "LRMs Compression" ], "forum_url": "https://openreview.net/forum?id=bheeBphxZf", "pdf_url": "https://openreview.net/pdf/3c98269aab41466b1984a92075f002066238e342.pdf", "updated_at": "2026-10-05T21:11:32.293Z", "categories": [ "ptq", "reasoning" ], "matched_terms": [ "4-bit", "ptq", "quantization", "quantizing" ], "review_status": "candidate", "summary": null, "contribution": null, "limitations": null }, { "id": "EQuFNpm6gc", "number": 52688, "title": "ThinQuant: Scalable Rotation Learning For Weight And Activation Quantization", "abstract": "Learned rotations play an important role in enabling low-bit weight and activation quantization of\nlarge language models by smoothing outliers in the activation distribution. State-of-the-art approaches include gradient-based procedures such as SpinQuant and computationally friendlier gradient-free\napproaches such as DartQuant, but both remain hard to scale to the largest architectures.\nTo address the computational bottlenecks in gradient-free rotation learning, we introduce two ideas for efficiency, (i) a data selection procedure which reduces the required number of calibration data points, and (ii) an exact reduction of the associated optimization on this reduced calibration set. Our data selection procedure exploits the geometric structure of the convex hull of the activations. Using this idea, we show that a carefully selected calibration set with several orders of magnitude fewer activations than state-of-the-art rotation-based methods can match their performance in low-bit quantization settings. Under this extreme data efficiency, the selected activations span an\n$r$-dimensional subspace with $r