arXiv_CL
-
MemoryTax: The Bitter Lesson of Agent Memory
Abstract
Agent memory systems such as Mem0 and Zep decide what to keep before any question arrives; coding agents keep the raw history and let the model search it with its own tools. We ask whether designed memory still helps once the model is strong enough to search for itself. On one LoCoMo conversation (419 turns, 152 questions), we run four memory methods (full context, BM25, Mem0, Zep) and three coding agents (Pi, Codex, Claude Code), each with GPT-4o-mini, GPT-6 Luna and Claude Opus 5.5, and record what every rollout costs. This is a short empirical study, with one conversation and one run per cell. Agent memory is a costly patch for a weak model: Mem0 beats full context only with GPT-4o-mini, and its build alone costs more than a whole full-context rollout. Agent memory also has a ceiling, set by the model that writes it: a stronger reader does not recover what a weaker writer left out. With Claude Opus 5.5, Pi tops the cost-accuracy frontier, scoring 99% for less than half the cost of full context; its one miss is a wrong answer key. Part of that lead comes from photo captions that only the agents could read: on the 88 questions whose evidence has no photo, Pi and full context differ by one answer.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29823
PDF
-
Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents
Abstract
Literature-searching language agents are becoming a standard component of scientific assistants for the life sciences, yet there is little controlled evidence on which parts of their harness matter. We run a controlled ablation on biomedical claim verification (SciFact) with the live PubMed API as the only tool. Holding the backbone and output format fixed, we compare four harnesses: closed-book answering, single-shot retrieval using the claim as the query, a ReAct-style agent that writes its own PubMed queries, and an oracle given the gold abstracts, each with two efficient frontier backbones. Naive single-shot retrieval finds the gold paper for only 6% of claims and turns overconfident errors into abstentions without improving accuracy. Letting the model formulate its own queries triples gold recall and yields the best verdicts (62% accuracy vs. 49-59% for the other realistic harnesses), although these differences are not significant at our sample size and agents use fewer than half of their search budget. With gold abstracts, the same models reach 79-86% accuracy, a significant 17-24 point gap showing that retrieval, not reasoning, is the main bottleneck. Finally, a citation audit shows that closed-book models cite PubMed IDs that almost always exist (327/335) but are almost never relevant: a cross-family LLM judge rates only 3 of 327 existing citations as relevant to the claim, versus 67% for citations produced by the search agent. Existence checks are therefore insufficient safeguards for scientific agents. We release code, prompts, and all cached tool calls at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29812
PDF
-
MAPS: Macro-Adaptive Parallel Reasoning for Regulating Search Momentum in Retrieval-Augmented Generation
2026-10-21 14:19:17Xinze Li, Dingwen Zhang, Zhenghao Liu, Yukun Yan, Shuo Wang, Yu Gu, Ge Yu, Maosong SunAbstract
Retrieval-Augmented Generation (RAG) improves the performance of Large Language Models (LLMs) by retrieving external knowledge. Recently, some RAG methods interleave reasoning and retrieval, enabling LLMs to progressively generate sub-queries and retrieve external knowledge. However, these methods may suffer from redundant or ineffective sub-queries and insufficiently diverse query exploration. Moreover, due to the tightly interleaved nature of reasoning and retrieval, it remains unclear how the evolving reasoning process influences query intent. To characterize the evolution of query intent during reasoning, we introduce search momentum, which captures the accumulated retrieval tendency induced by historical changes in generated sub-queries along the reasoning trajectory. Through momentum-based analysis, we find that the query generation process tends to continuously evolve along existing retrieval trajectories. Based on this observation, we propose MAPS, a Macro-Adaptive Parallel Reasoning framework for regulating search momentum and mitigating query-intent drift. MAPS first retains the retrieved knowledge and a refined reasoning state, thereby reducing the influence of historical reasoning trajectories on subsequent sub-query generation. Subsequently, MAPS extends the conventional sequential reasoning process into a parallel reasoning paradigm. At each iteration, MAPS identifies complementary retrieval directions, generates corresponding sub-queries for parallel retrieval, and consolidates the retrieved evidence to update the knowledge and reasoning states. Experiments on six question-answering benchmarks demonstrate that MAPS consistently outperforms representative RAG methods across different backbone models. All code and resources are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29792
PDF
-
Where Do LLMs Learn to Recommend AI?
2026-10-21 14:13:52Benaya Trabelsi, Jonathan Shaki, Sarit KrausAbstract
Asked what to study, which career to pursue, or what product to build, large language models (LLMs) often recommend AI-related options. We ask where this tendency comes from by comparing models before and after fine-tuning and by varying how questions are asked. First, pretrained models already include explicit AI/ML terms in the recommendations of 27% of their responses, and their released instruction-tuned versions do so in 59%. Second, fine-tuning models ourselves shows that this rise is not inevitable: general instruction tuning leaves the rate unchanged, but removing the 7% of instruction examples that concern AI cuts it by 14 percentage points, below the level of the pretrained models it starts from. Data rewarding helpfulness raise it by 9 points, while data ranked by the preferences of diverse simulated users lower it by 9. Finally, with models held fixed, asking for the "best" option or for what to do "right now" raises the rate by 8 and 18 points, whereas telling the model it is an AI advisor adds 4. Overall, the tendency starts in pretraining, and what models learn afterward, together with how a question is phrased, decides how strongly it appears.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29785
PDF
-
Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting
2026-10-21 14:09:57Zhipeng Chen, Yunsen Liang, Yanxi Chen, Hongkang Zhao, Lingyun Feng, Ruiyang Ren, Guoxin Chen, Jia Deng, Feng Chang, Chuan Hao, Ran Tao, Jian Yang, Ji-Rong Wen, Bryan Dai, Wayne Xin ZhaoAbstract
Most contemporary artificial intelligence (AI) systems are released with frozen parameters and improve through external updates, whereas an ideal system would keep improving from its own experience, a capacity we call recursive self-improvement (RSI). Yet studying RSI requires an environment that is continuously changing, objectively verifiable, and naturally grounded, posing a fundamental challenge for agent evolution. Here we introduce DeepEvolve, a sustainable evolution agent system grounded in future event prediction, in which agents continuously interact with a dynamically evolving environment and improve by internalizing and externalizing the experience gained through these interactions. A living environment synthesizes verifiable tasks, and a recursive-evolution engine turns the reasoning trajectory behind each prediction back into capability: decoupling search from prediction adapts parameter-level evolution to the contamination-prone live environment, while distilling reusable skills lets the agent evolve from the limited live trajectories such an environment affords. On the external FutureX leaderboard, which evaluates predictions against real-world outcomes and is updated weekly, our self-improved forecaster ranked first across all systems over three consecutive weeks, achieving up to a 195% improvement in rating. On our synthesized evaluation, it ranked among the top three across four base LLMs, outperforming the strongest baseline by up to 10.7 Bradley--Terry points. More broadly, these results suggest that a system built for self-improvement is best understood as an agent that co-evolves with a genuinely living, open-ended environment, pointing to a shift in AI from training static models on fixed data toward cultivating agents that grow stronger from their own experience, with such an open world as the indispensable basis of that growth.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29784
PDF
-
ALIAS: Evaluating Performance and Memorization in LLM-Written TLA+ Specifications
2026-10-21 14:02:02Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria, Brian Ortiz, Eric Spencer, Beatriz Santos, TaiNing Wang, George K. Thiruvathukal, Konstantin Laufer, Mohammed AbuhamadAbstract
Memorization inflates the benchmark scores of large language models (LLMs), since a model can reproduce answers it saw in training. Researchers test for it with consistent identifier renaming (CIR), which gives every identifier in a task a new name, and treat a lower score on the renamed task as evidence of recall. We test this method on the Temporal Logic of Actions Plus (TLA+) language, whose model checker acts as an oracle by confirming whether a renamed task preserves the exact behaviors of the original. Across 111 TLA+-Bench tasks, renaming lowers the correctness of Claude Opus 4.8 from 28% to 18% and of GPT-5 from 9% to 7%. Most of this decrease comes from how the test grades, which requires exact names that most descriptions never state. Grading by behavior shrinks the decrease to about 3 percentage points for Opus 4.8 and to 0 for GPT-5. In two model families fine-tuned on disjoint halves of the tasks, renaming removes about half of the training gain in likelihood and costs 4.9 percentage points (Qwen) and 1.3 percentage points (Gemma) on unseen tasks. A lower score after renaming therefore mixes memorization, name-based grading, and the cost of new names.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29767
PDF
-
Bridging Foundational and Agentic Capabilities of LLMs
2026-10-21 13:50:43Wei Guo, Yuanjian Xu, Jianing Hao, Yongyao Duan, Ran Chen, Zhong Li, Guang ZhangAbstract
Large language models (LLMs) are increasingly deployed as agents, yet it remains unclear how their foundational capabilities translate into agentic capabilities. End-to-end scores leave unresolved whether low benchmark capability or weak deployment drives a failure, limiting actionable diagnosis. We bridge foundational and agentic capabilities through an operation-level measurement framework. We represent each model with a foundational profile derived from capability benchmarks and an operation profile derived from interaction trajectories, and define capability realization gaps as deviations of observed operation performance from predictions based on foundational capabilities. Experiments on fifteen model services in $\tau^2$-Bench show that foundational profiles only partially predict agentic operations. Models with comparable profiles can differ substantially in execution, and these gaps remain consistent across task subsets. Operation profiles add information and substantially improve task-score prediction. The framework locates execution departures at the operation level to inform model development and agent evaluation. We will release our code and analysis artifacts.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29751
PDF
-
Accept or Repair: Preserving Passage-Supported Answers in Contrastive Decoding
2026-10-21 13:32:29Jungseob Lee, Seungyoon Lee, Hyeonseok Moon, Seongtae Hong, Suhyune Son, Sugyeong Eo, Minhyuk Kim, Chanjun Park, Jaehyung SeoAbstract
Large language models often answer from knowledge learned during training, even when a provided passage contradicts that knowledge. To address this problem, contrastive decoding contrasts the predictions under the full and a modified prompt at every token, but it also alters correct answers that the passage supports. We show that checking answers against the passage removes this trade-off. Based on this observation, we propose Accept-or-Repair (AoR), a training-free procedure. AoR keeps a sampled answer if the passage supports it. Otherwise, it tries up to two replacements with contrastive decoding, returning one as soon as it passes this check. It rejects replacements that repeat the answer generated without the passage when the passage favors another answer. Our experiments show that AoR raises the score over sampling by 12.5 points on counterfactual passages and 2.3 on ordinary ones, surpassing eight contrastive and prompting baselines on both, including when each receives the same rejection rule. Replacing only initial answers that the passage does not support improves the average score of every decoder. These results suggest that models can follow passages that contradict their knowledge while preserving accuracy on passages that agree with it. Code will be released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29727
PDF
-
Retrieval Augmented Generation Compositionality
2026-10-21 13:30:29Ido Levy, Segev Shlomov, Leshem ChoshenAbstract
Retrieval Augmented Generation (RAG) grounds large language models in external knowledge. Modern pipelines combine embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification, yet these techniques are typically evaluated one at a time. It remains unclear whether their isolated gains survive composition, whether retrieval quality and answer faithfulness favor the same design choices, and whether conclusions from small corpora hold at large scale. We pre-declare three hypotheses and introduce LASAGNA, an open-source framework for controlled RAG evaluation. We test 450 configurations across eight benchmarks and corpora ranging from 5K to 200K documents. Three findings emerge. First, individually strong techniques from different phases of a RAG pipeline do not necessarily compose: three retrieval techniques predict a +47% combined improvement but deliver only +17% when used together. Second, component gains stack differently for retrieval and faithfulness, and optimizing one can hurt the other. Third, the best composition changes as the corpus grows: the combination that wins on a small corpus may not remain best at large scale. These findings show that gains measured in isolation are not reliable predictors of realistic performance, and that new RAG techniques should be tested with the full surrounding stack. We call on researchers to do so and supply LASAGNA to simplify these experiments.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29722
PDF
-
Scaling Multilingual Pretraining of Pixel Language Models
2026-10-21 12:57:41Ilker Kesen, Jenny Kunz, Kasra Malihi, Desmond ElliottAbstract
Pixel language models process text as rendered images instead of a finite vocabulary, which gives them strong cross-lingual transfer abilities. However, not much is known about how these abilities are influenced by larger-scale pretraining. In this paper, we train a family of MPIXEL models along three axes of variation: more languages, more scripts, and larger model capacity under fixed per-language compute. In evaluations on five tasks covering 83 languages, we find that scaling acts locally: each axis improves the added languages and scripts, while increased capacity mitigates the curse of multilinguality from pretraining on more languages. Transfer to unseen scripts improves with multilingual pretraining, but does not further improve with scaling. On typologically distant low-resource languages, our models outperform subword and character-level baselines across many writing systems. We also show that larger pixel models reserve more layers for visual processing, and that cross-lingual alignment weakens as more languages are added. Overall, our findings characterize how multilingual scaling improves pixel language models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29681
PDF
-
Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization
2026-10-21 12:48:10Zhiqian Xia, Jielong Tang, Ling-An Zeng, Shixiong Kai, Rong Yao, Hao Kong, Xiao Dong, Wei Wang, Mingxuan Yuan, Siyu Xia, Haifeng XiaAbstract
Long-term agent memory is shifting from predefined workflows toward learnable policies for memory construction and retrieval. Existing learning-based methods either optimize construction alone or combine retrieval-side online reinforcement learning (RL) with offline construction supervision, limiting co-adaptation between the two stages. In this paper, we explore end-to-end online RL for a unified memory manager (Uni-Mem) that learns both capabilities under a shared policy. However, joint optimization with existing RL algorithms such as GRPO faces confounded group comparisons and sparse feedback over long memory-management trajectories. To this end, we propose Stage-Decoupled GRPO, which alternates construction and retrieval optimization while controlling variation from the other stage within each rollout group, yielding stage-specific learning signals. We further introduce Event-Wise Credit Assignment (ECA), which localizes feedback to responsible events and propagates event-level advantages to their constituent operations. Experiments on LoCoMo and LongMemEval show that Uni-Mem outperforms the strongest RL-based memory, demonstrating the effectiveness of jointly learning memory construction and retrieval via Stage-Decoupled GRPO and ECA.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29671
PDF
-
Conflicted Records: Typed Adversarial Contracts for Auditable LLM Output Quality Enforcement
2026-10-21 12:43:35Kardelen Subasi, Jaros{\l}aw A. ChudziakAbstract
LLM-generated records increasingly populate production information systems where they are treated as data assets subject to the same quality expectations as records from operational databases. A single such record can simultaneously satisfy and violate a quality rule: a grounded fact adjacent to a hallucination, a useful answer adjacent to a stereotype. Classical data cleaning cannot resolve this because the two judgments are incompatible over the same span; a single LLM judge cannot either, since it shares the training biases that produced the conflict. Existing interventions such as prompt engineering, chain-of-thought, and RLHF reduce surface error rates but provide no audit trail and cannot distinguish genuine remediation from blanket suppression. The question is whether a deterministic check over structured inter-agent records can catch the defects that referential checks alone miss. We present BiasGuard-D: two adversarial roles drawn from distinct model families exchange typed claim documents, a three-model quorum adjudicates by majority vote, and a combined referential-plus-semantic invariant validates the verdict. The semantic component fires on verdicts referential checks miss; removing the iterative loop drops invariant satisfaction sharply. The deliberation layer transfers from stereotype to hallucination remediation on HaluEval QA with only role-framing changes, raising right-answer recovery by 49 percentage points; the triage gate required a cross-family second auditor to maintain recall on objective hallucination. Every claim, vote, and instruction is stored as a typed record giving operators a provenance trail auditable under emerging AI regulations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29669
PDF
-
SPARE: Sparsity Allocation via Routing Entropy for Efficient MoEs
2026-10-21 12:40:00Yifan Yang, Hanling Yi, Dongquan Yang, Feng Lin, Rong Xiao, Aijun Li, Xiaotian YuAbstract
Mixture-of-Experts (MoE) architectures scale model capacity efficiently, yet their substantial memory footprint poses a major barrier to deployment. Expert pruning reduces this footprint, but most methods focus on which experts to remove and prune every layer equally. We propose SPARE (Sparsity Allocation via Routing Entropy), a training-free, selector-agnostic method that decides how many experts to remove from each layer using routing entropy, the entropy of tokens' normalized Top-K gate weights on calibration data. When a token's gate weight is spread evenly, removing one active expert discards little of it, so SPARE prunes layers with more diffuse routing more aggressively. A perturbation bound under random pruning formalizes this intuition, and controlled interventions with a fixed selector link concentrated routing to larger output changes. With the same selector and budget, SPARE improves average accuracy over uniform allocation on Mixtral, Qwen3, and DeepSeek models, especially at high pruning ratios. At 50% expert pruning, it raises the eight-task average of Qwen3-30B-A3B from 57.2% to 63.8%, and the math, code, and science average of DeepSeek-R1 from 80.1 to 82.8.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29662
PDF
-
Effects of Model Scale and Quantization on Social Bias and Abstention in Question Answering
2026-10-21 12:38:04Kuniko Paxton, Ifeoluwa Wuraola, Dhavalkumar ThakkerAbstract
The deployment of small and quantized language models in resource-constrained settings raises an important fairness question: when a compressed model produces fewer biased responses, does this reflect fairer behavior or greater abstention? We study this problem in Question Answering (QA) using the BBQ benchmark across multiple model sizes and quantization settings from Qwen3, Gemma-3, and SmolLM2. We also examine whether the observed trends remain stable across alternative prompt formulations and when the ``Unknown'' option is removed in a forced-choice setting. Our results show that larger models achieved higher QA accuracy, but bias showed no simple scaling trend and quantization effects varied across configurations. Notably, 20 of the 26 configurations with fewer biased responses also show increased over-abstention, indicating that lower observed bias can partly reflect greater abstention. These findings show that lower observed bias can partly reflect greater abstention, and that fairness evaluation should consider biased responses, abstention, prompt formulation, and answer-option design jointly.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29657
PDF
-
LLMs as survey respondents: Comparison to simple semantic baselines
2026-10-21 12:26:20Esko Lehtonen, Tommi Buder-Gr\"ondahlAbstract
Large language models (LLMs) have been explored as surrogates for human survey respondents. A more promising application is data augmentation, where LLMs predict new responses from partial human data. However, the semantic similarity between questionnaire items is a known source of correlation between answers. It is thus unclear whether LLMs actually infer meaningful relationships between question-response pairs, or use surface-level comparison between questions as a heuristic. To assess this, we evaluated how three LLMs can predict responses to held-out items in two surveys measuring attitudes toward automated or light-electric vehicles, based on varying proportions of known responses. Performance was compared to two baselines, where new responses were derived by maximizing embedding-similarity to a prior response. While LLMs mostly predicted human responses better than baselines, their performance consistently declined with a lower share of known responses, even aligning with the baseline in one survey. Our results show that LLMs' capacities for augmenting replies varied across surveys, and can revert to the same levels as simpler embedding similarity baselines when less data is available.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29645
PDF
-
Overview of CHIP 2026 Shared Task 1: Patient-Specific Phenotype Concept Recognition in Full-Text Biomedical Literature
2026-10-21 11:45:35Xuanlyu Zhu, Feng Tian, Yang Zhong, Pengqi Zheng, Xiwen Bi, Ling Luo, Yuanyuan Sun, Jian Wang, Lingfei LinAbstract
Automatic phenotype information extraction plays an important role in rare disease diagnosis, genotype-phenotype analysis, and precision medicine. However, accurately identifying phenotype concepts and determining their associations with specific patients remain challenging, particularly in full-text biomedical articles involving multiple patients and complex clinical descriptions. This paper presents an overview of CHIP 2026 Shared Task 1, which aims to promote the development of methods for patient-specific phenotype concept recognition from full-text biomedical literature. The task consists of two subtasks: full-text phenotype concept recognition and patient-specific phenotype association. For this evaluation, we used 200 annotated full-text biomedical articles from the PatientPheX dataset, containing 506 patients, 15,193 phenotype mentions, and 5,656 patient-phenotype associations. A total of 455 teams registered for the competition, with 189 and 103 valid submissions received on the Phase A and Phase B leaderboards, respectively. The best-performing system achieved an overall score of 0.7389 on the final test set, outperforming the baseline by 8.37 percentage points. The top-performing teams explored diverse approaches combining biomedical language models, large language models (LLMs), and ensemble strategies. The evaluation results highlight both the potential of these approaches and the remaining challenges in patient attribution and phenotype normalization from full-text biomedical literature. The post-evaluation phase remains open at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29606
PDF
-
A Diagnostic Study of Faithfulness in Diffusion Language Models for Table-to-Text Generation
2026-10-21 11:42:19Zaynab Merimi, Noussayba El Marrakchi, Arthur SatoufAbstract
Diffusion language models generate by iteratively denoising a masked sequence under bidirectional attention, without the strict left-to-right commitment of autoregressive decoding. Whether this difference affects how reliably they remain grounded in structured inputs is largely unexplored. We study this question in table-to-text generation using ToTTo, where highlighted source cells make factual grounding directly testable. We compare the masked diffusion models MDLM and LLaDA with GPT-2 and GPT-Neo under an identical retrieval-based few-shot protocol, together with a QLoRA-adapted LLaDA, and we score reference similarity and source grounding as separate axes. The two pull apart. Within LLaDA, additional demonstrations raise BLEU from 11.3 to 22.0 while NLI entailment against the source table falls from 0.78 to 0.52; QLoRA reaches the best BLEU of the study, 24.15, and grounds no better for it. Manual annotation shows the failure modes differ in kind: the autoregressive baselines tend to wander off the table, while LLaDA-based models more often stay on topic and swap out individual entities. Reference-based metrics alone are insufficient to assess faithfulness in table-to-text generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29602
PDF
-
WRDO: Weighted Refusal Direction Optimization in Aligned Language Models
2026-10-21 10:42:25Grigoriy Evseev, Vladimir Solodkin, Dmitry Bylinkin, Daniil Medyakov, Vladimir Aletov, Aleksandr BeznosikovAbstract
Aligned large language models are trained to comply with benign requests while refusing harmful ones. However, recent work shows that this refusal behavior is fragile and can often be disrupted by simple activation- or weight-space interventions. In particular, refusal has been shown to be largely mediated by a single residual-stream direction, typically computed as a difference-of-means vector between harmful and harmless prompts. However, this standard construction aggregates heterogeneous harmful prompts into one uniform average, potentially conflating distinct refusal-related phenomena. In this work, we study whether refusal interventions can be strengthened by learning how to combine prompt-specific refusal directions. We propose Weighted Refusal Direction Optimization (WRDO), a procedure that constructs a composite refusal direction from a bank of per-prompt directions using learned aggregation weights. This approach allows different harmful prompts to contribute unequally to the final intervention direction. We evaluate WRDO against prior single-direction methods on three open-weight language models using three guard judges and HarmBench as an external evaluator. We further measure the effect of the intervention on general model capability through standard utility benchmarks. Finally, we investigate a defense-oriented setting in which the same optimization machinery is used in reverse to restore refusal behavior in an abliterated model checkpoint. Our results show that weighted aggregation provides a more effective and flexible mechanism for manipulating refusal representations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29555
PDF
-
PACE: Periodic and Channel-Aligned Encoding of Four-Bit Recurrent States
2026-10-21 10:31:52Patrick Haller, Filipe Laitenberger, Alan AkbikAbstract
Linear-attention language models keep their context in a recurrent state of fixed size, where an inference system stores one copy of this state for every sequence it decodes at the same time. Storing the state at lower precision therefore lets more sequences share an accelerator. Existing state quantizers round the state after every generated token and report that four bits lose accuracy unless the bit width is raised or the model is fine-tuned.\par We introduce PACE, a post-hoc four-bit state quantizer. PACE rounds the state only once every few generated tokens, as chunked kernels already do once per chunk during prefill, and keeps the tokens generated in between in a small buffer. Rounding less often reduces the error that accumulates in the state. Additionally, we find that the large entries of the gated-delta state gather in columns, because some channels of the update written at each token remain large across many tokens. PACE therefore shares each quantization scale within one column. It also keeps the heads that retain rounding error longest at eight bits. On four hybrid models from 4B to 35B parameters, PACE with mixed four- and eight-bit heads is competitive with eight-bit DAMP on the three larger models at about two thirds of its memory, PACE with four-bit heads throughout scores higher than Q-Mamba at equal memory, and rounding every few tokens preserves the decoding speedup of the buffer.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29545
PDF
-
RCSD: Reinforcement and Continual Learning for Domain-Specific Speculative Decoding
2026-10-21 10:16:23Mingkang Liu, Weiming Meng, Peipei Jiao, Jiawei Zhang, Tianyuan Liu, Junyan Li, Lichen PangAbstract
Speculative decoding accelerates large language model inference by employing a compact draft model to propose candidate tokens that are verified in parallel by a target model. Conventional training relies on soft-label distillation, which forces the draft model to mimic the target's full output distribution, including high-entropy regions that degrade acceptance rates and increase early rejection risk. We propose RCSD (Reinforcement and Continual Learning for Speculative Decoding), a training framework that aligns draft model optimization with the system-level objective of maximizing acceptance length. RCSD leverages naturally occurring accepted and rejected sequences during speculative inference to construct implicit preference pairs, with masked loss to avoid penalizing target-aligned tokens. Evaluated across multiple benchmarks, RCSD improves average acceptance length by 35% over HASS and achieves 1.35x faster end-to-end inference than HASS on practical workloads, without compromising generation quality.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29527
PDF
