AI Paper Digest — 2026-10-12

Aug Sep Oct Nov

Highlights

SWE-Duel: Towards a Self-Scaling Arena for Coding Agents

Highlight HF pick · 14▲Agents Wei Minn, Zhensu Sun, Lwin Khin Shar, David Lo Static coding benchmarks leak into training data and saturate as agents improve, so the authors propose a self-scaling arena whose tasks and oracles are created at evaluation time by the participating agents themselves. In SWE-Duel, a Red agent implements a new feature in a real repository, embeds a subtle bug with sealed triggering tests, and the pull request is admitted only if it passes the feature and existing tests, fails exactly the bug tests, and a fresh session of the same agent can repair the bug blind; a Blue agent then reviews the pull request without being told a bug exists, wins only if its revision passes all three suites, and the roles swap. In a tournament of 11 model-harness participants on 15 repositories in five languages over 1,500 turns, GPT-5.5 with Codex CLI ranked first as the strongest solver, repairing 98.3% of the defects it reviewed, while Claude Opus 4.8 with mini-swe-agent ranked second as the strongest task author with a 27.3% Red win rate. Within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% versus 4.0%), preliminary evidence that task difficulty scales with authoring capability.

Static coding benchmarks leak into training data and saturate as agents improve, so SWE-Duel makes two coding agents author tasks for each other at evaluation time, leaving nothing to leak in advance and letting difficulty rise with the participants themselves.

In each Red turn an agent works in three context-isolated sessions on a real repository, implementing a feature with feature tests, then embedding a subtle bug with sealed bug-triggering tests, and the task is admitted only if all visible tests still pass and a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks.
The Blue agent reviews the pull request without being told a bug exists and wins only if its revision passes the existing, feature, and sealed bug test suites, so reverting the feature or shipping the bug both count as losses before the roles swap.
A round-robin tournament of 11 model-harness pairs on 15 repositories in five languages ran 1,500 turns, where GPT-5.5 with Codex CLI ranked first as the strongest solver, repairing 98.3% of reviewed defects, while Claude Opus 4.8 with mini-swe-agent ranked second as the strongest task author with a 27.3% Red win rate, and the same model under Claude Code fell to eighth.
Holding the solver fixed as DeepSeek-V4-Pro and varying only the author, tasks from GLM-5.3 defeated it 42.9% of the time versus 8.3% for GLM-5.1 and 0.0% for GLM-5.2, giving preliminary evidence that difficulty tracks authoring capability, and a late entrant could be placed with only five matches without reordering the incumbents.
The blind self-repair gate caps difficulty at the author's own bug-finding ability, Blue already succeeds on 87.9% of all turns so headroom at the top is thin, a manual audit found a 6.0% false-negative rate from overly specific bug tests, and the GLM ladder rests on only 12 to 14 tasks per version with wide confidence intervals.

arXiv:2610.12569 PDF

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

Highlight HF pick · 77▲Reinforcement Learning Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang, Pengfei Zhou, Wangbo Zhao, Fangyuan Kong et al. Online reinforcement learning (RL) can fine-tune diffusion models from black-box rewards, but reverse-process methods such as Flow-GRPO suffer from poor credit assignment while forward-process methods such as DiffusionNFT are efficient yet unstable without carefully tuned exponential-moving-average (EMA) off-policy training. The authors show that NFT-style objectives contain a hidden forward-reverse consistency force that joint positive-negative training cancels only partially, and that EMA stabilizes training through restoration rather than off-policy rollouts. DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting) aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states, removing the consistency force by construction, and adds a trust-region-masked second update for drift-adaptive restoration without EMA tuning; it needs no differentiable reward model, works with fewer samples per prompt, and supports pairwise reward models. On HunyuanVideo-13B with equally weighted visual and motion quality rewards it reaches the late-stage mean reward levels of both baselines with about 12 times fewer GPU hours while achieving 28.8% and 53.0% higher mean training rewards.

Online reinforcement learning fine-tunes diffusion and flow models with black-box rewards, but reverse-process methods such as Flow-GRPO learn slowly because a trajectory-level advantage is smeared across every denoising transition, while forward-process methods such as DiffusionNFT are efficient yet collapse without an exponential moving average, or EMA, rollout policy whose lag must be tuned per task. The proposed fix aggregates each prompt's rewards into one shared reward-improving direction, regresses every visited reverse-rollout state toward it, and replaces EMA with a second trust-region-masked update that supplies restoration proportional to each state's drift.

The diagnosis is that DiffusionNFT's forward-noised targets carry a hidden forward-reverse consistency force that joint positive and negative training cancels only across samples rather than at each state, and that EMA stabilizes training through restoration toward the lagged reference rather than through off-policy data, since keeping EMA rollouts but dropping the old-policy anchor still collapses.
For each prompt the method computes the covariance between rewards and clean endpoints over the K rollout samples, RMS-normalizes it into a guidance vector, and regresses each visited state's velocity toward the cached rollout prediction plus that vector, applying two updates on disjoint halves of the batch so the second update's gradient becomes guidance minus current drift, with a KL-proxy trust-region mask at one fixed threshold blocking excessive outward drift and a pairwise variant accepting winner/loser labels in place of scalar rewards.
On SD3.5-M it matches DiffusionNFT with EMA in training efficiency without any EMA tuning and reaches Flow-GRPO's step-500 reward with 10.0×, 5.2× and 9.6× fewer steps on GenEval, OCR and PickScore, and on FLUX.1-dev it scores 0.9684 on GenEval versus 0.9183 for Flow-DPPO and 0.8490 for DiffusionNFT.
On HunyuanVideo-13B with equal visual-quality and motion-quality weights it reaches the late-stage mean reward of DiffusionNFT and Flow-GRPO-Fast with 12.02× and 11.72× fewer GPU hours while finishing 28.8% and 53.0% higher, and for image-to-video on Wan2.1-I2V-14B it reaches motion quality 6 in 45.8 GPU hours versus 220.3 for Flow-GRPO-Fast, a 4.8× reduction.
Every video gain is measured against the learned VideoReward model with no human evaluation or over-optimization check, the trust-region mask that prevents a mid-training drop on video slightly slows image training, and the FLUX.1-dev comparison runs DiffusionNFT with hyperparameters transferred untuned from SD3.5-M, so that baseline was never tuned at scale.

arXiv:2610.12572 PDF

Recurrent Harness for Agentic Continual Learning

Highlight HF pick · 11▲Agents Yuxin Xiao, Zhiyun Lu, Jianyu Wang, Yangtian Zhang, Yang Song, Qifan Wang et al. Language-model agents operating over long interaction streams need to adapt online to distribution shift, but existing harnesses either update their state only through forward information flow or require extra replay or validation data. ReHarness is a recurrent harness that adds backward revision to forward read and write under a strictly online streaming protocol: after each instance, the resulting trajectory serves as feedback to revise the reflections that were retrieved for that instance, propagating the signal backward through the textual harness state. Across three continual learning benchmarks and three language-model backbones, ReHarness consistently outperforms raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches. Analysis shows the revisions are source-grounded, correct overgeneralized or outdated experience under shift, and improve both stability and plasticity.

Language-model agents that learn from a stream of tasks typically store memories or reflections and only ever read them forward, so an overgeneralized or outdated reflection persists and keeps misleading later decisions after a distribution shift. ReHarness adds a backward revision step: after every instance, a revision agent rewrites the reflections that were retrieved for it, using the new trajectory as feedback while keeping any new content grounded in each reflection's original source trajectory, all under a strictly online protocol with no replay, validation set, or gradient updates.

Each step runs forward read (retrieve the top-5 reflections with Qwen3-Embedding-8B), backward revision (independently rewrite each retrieved reflection given its source trajectory and the current outcome), and forward write (a memory agent produces a new reflection using the revised ones as demonstrations).
Across CL-Bench, MemoryBench, and MemoryArena with Claude Opus 4.8, GPT-5.5, and Gemini 3.7 Flash, ReHarness reaches a mean normalized gain of 0.240 versus 0.168 for Top-K Raw and 0.164 for Full Raw, ranks first or tied in 8 of 9 benchmark-backbone combinations, and beats every baseline including Mem0, RBank, DoT, ACE, and GEPA-Online under Bonferroni-corrected Wilcoxon tests (p_adj 0.039 vs Top-K Raw, 0.043 vs Full Raw).
A claim-level audit finds over 95% of revised-reflection claims stay grounded in the source trajectory with only 0.8–2.1% leakage from the downstream instance, and on CL-Bench the method posts the highest average stability (0.015) and plasticity (0.241) while all other structured memory and reflection baselines show negative stability on average.
Dropping revision entirely cuts the CL-Bench/Claude gain from 0.217 to 0.144, and a diagnostic counterfactual replay that reapplies the revised state to the same instance (which violates the streaming protocol) lifts it to 0.335, confirming the revisions carry real information.
On CL-Bench with Claude, ReHarness consumes ~41.6M total tokens versus 52.7M for Top-K Raw and 142.1M for Full Raw, but its heavier output-token use puts its cost at about $254, cheaper than Full Raw ($314) yet far above Top-K Raw ($85), and gains are marginal on exploitable poker where noisy hand-level rewards give the revision agent little diagnostic signal.

arXiv:2610.12668 PDF

Surfing the Activation Wave: Analyzing the Information and Geometry of LLM Hidden States for Speculative Decoding

Highlight HF pick · 10▲Large Language Models Raghavv Goel, Sudhanshu Agrawal, Mukul Gagrani, Risheek Garrepalli, Mingu Lee Speculative decoding drafters such as EAGLE3 draw on a few target-model hidden layers chosen by fixed position heuristics, even though picking the best subset is a combinatorial problem that cannot be searched exhaustively. The authors score candidate layer sets on two criteria, informativeness measured by matrix entropy and diversity measured by centered kernel alignment (CKA), and study layer-wise entropy, entropy under concatenation, and pairwise CKA dynamics across depth, revealing distinct informational compartments in decoder representations; a token-entropy filter drops hidden states for low-predictive-entropy tokens to clean up the selection statistics. In EAGLE3-style speculative decoding, the entropy- and CKA-selected layer sets outperform fixed defaults, including the standard three-layer rule, while also explaining why some defaults remain competitive.

Speculative decoding drafters such as EAGLE3 feed on hidden states from a handful of target-model layers chosen by fixed position heuristics, even though picking the best subset is a combinatorial problem that grows infeasible for deep models. The paper replaces the heuristic with a principled selection framework that scores candidate layer sets on two complementary criteria: informativeness, measured with matrix entropy, and diversity, measured with centered kernel alignment (CKA).

The analysis tracks layer-wise Shannon entropy, aggregate entropy when hidden states are concatenated, and pairwise CKA dynamics across depth, revealing distinct geometric and informational compartments inside the decoder representations.
Building on these findings, the entropy and CKA-based selection algorithms favor layer sets that are highly informative yet minimally redundant, and a token-entropy filter discards hidden states tied to low predictive-entropy tokens so the selection statistics are computed on more discriminative positions.
In EAGLE3-style speculative decoding, the selected layer sets outperform fixed baselines, including the default [2, N//2, N-3] rule (with N the number of decoder layers), while the entropy and CKA profiles also explain why that default remains competitive.
The abstract reports no quantitative speedups, acceptance lengths, or model sizes, and the evaluation is confined to a single EAGLE3-style setup, so the practical margin over the default rule and its transfer to other speculative-decoding architectures remain to be shown.

arXiv:2610.12939 PDF

Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction

Highlight HF pick · 31▲Agents Yuqi Huang, Hui Li, Yifei Cao, Zikun Cai, Tianci Xie, Yangfan Gao et al. Computer-use agents still struggle with unfamiliar software and workflows that require human domain knowledge, and although human demonstrations convey that knowledge through the same GUI, reusing them means interpreting actions in context and separating general workflow rules from task-specific details. Work2Skill provides WorkRecorder, which aligns screen recordings, mouse and keyboard events, and accessibility-tree data, and uses it to build WorkDemo-350, a synchronized multimodal dataset of 350 screened human desktop demonstrations; an agentic skill-induction method then inspects this evidence on demand to write reusable skills capturing workflow rules, step dependencies, and completion checks. With four execution models equipped with induced skills under shared settings, Work2Skill scores highest, raising mean scores over a no-skill baseline by 4.67 to 10.54 points on MyPCBench Bounded Action and 3.11 to 7.50 points on a 286-task subset of OSWorld-Verified.

Computer-use agents handle increasingly complex tasks but still fail on unfamiliar software and organization-specific workflows whose procedures are not spelled out in the request. The core idea is to record human demonstrations with aligned screen video, input events, and accessibility-tree evidence, then let a tool-using agent inspect that evidence on demand and write reusable text skills that an executor can follow without ever seeing the recording.

WorkRecorder groups timestamped mouse and keyboard events into actions, links each to keyframes and video intervals, and attaches accessibility-tree names, roles, and bounds for click targets, so an induction agent can read what a control means rather than infer it from pixel coordinates.
WorkDemo-350 packages 350 screened, successful human demonstrations covering 64 MyPCBench Bounded Action tasks and a 286-task OSWorld-Verified subset, about 8.48 hours of recorded experience with roughly 33 actions and 24 evidence images per episode.
The induction agent interleaves evidence inspection and skill revision in one conversation, sampling video coarsely and then zooming into ambiguous steps, and writes SKILL.md files that generalize demonstration-specific inputs into task-dependent descriptions while preserving step dependencies, selection rules, and completion checks, with no execution feedback during induction.
Across four executors (GPT-5.5, Opus-4.8, Qwen3.5-35B-A3B, Qwen3.5-9B) with skills injected as a text block, Work2Skill beats the No Skill baseline by 4.67 to 10.54 score points on MyPCBench and 3.11 to 7.50 on the OSWorld subset, beats the stronger of the text-adapted ShowUI-Aloha and Demo2Tutorial baselines by 3.95 to 7.79 and 1.65 to 6.96 points, and lifts GPT-5.5 MyPCBench success rate to 48.44% versus 37.50% for both baselines.
Skills reused without re-induction on 64 same-application MyPCBench Jitter variants still gain 3.93 to 8.64 points over No Skill, GPT-5.5-induced skills transfer to a Qwen3.5-9B executor with positive but much smaller gains, and ablations show that dropping indexed images, UI annotations, or key logs costs 5.02 to 10.63 score points and 14.06 to 21.88 success-rate points.
Skill induction is expensive, roughly double the baselines' token usage for GPT-5.5 on MyPCBench (334.5k versus 170k to 177k), GPT-5.5 takes more execution steps with skills than without, the main evaluation runs on the very tasks that were demonstrated, and the OSWorld numbers cover only a curated feasible subset rather than the full benchmark.

arXiv:2610.12984 PDF

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

Highlight HF pick · 39▲Agents Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang et al. Continual learning research has largely stayed in non-embodied settings. EmbodiedCLBench is a benchmark in embodied simulation for evaluating whether self-evolving agents can learn basic tasks and then recombine the acquired skills to solve unseen, more complex advanced tasks, with compositional generalization as the core evaluation principle. Experiments across multiple agent harnesses and learning methods show that advanced tasks are nearly unsolvable zero-shot, that experience from basic tasks yields consistent improvement that holds up as task complexity grows, but that the improvement does not scale with the number of learning samples, so current agents cannot extract more from additional experience. Code and data are released.

Continual-learning benchmarks for agents live in digital or symbolic environments, while embodied benchmarks never test whether an agent can reuse past experience. EmbodiedCLBench closes that gap by treating continual learning as compositional generalization: agents first complete basic household tasks in the TongSim simulator, then must recombine those skills to solve unseen advanced tasks, with no benchmark-specific training.

The benchmark holds 268 tasks in TongSim household scenes: 168 basic tasks spanning six categories with a 6-step shortest path, and 100 advanced tasks composing four atomic actions over 12 steps, all evaluated offline on pre-built state-transition graphs where the agent sees a first-person image, picks one action per step from candidates labeled optimal, suboptimal, or invalid, and gets three extra steps beyond the shortest path.
Three settings are compared across the OpenClaw, Claude Code, Codex, and Hermes harnesses with Qwen3.8-Flash and GPT5.6-Terra backbones: zero-shot, In-Context Learning which keeps two related basic-task trajectories in context, and Skill Learning which keeps only a Markdown skill summary the agent wrote from those trajectories.
Zero-shot success on advanced tasks is only 3–12%, and experience lifts it to 16–52% under In-Context Learning and 22–38% under Skill Learning with no consistent winner between the two methods, a gain the authors link to agents choosing 13–20 points fewer invalid actions in the Qwen-backed runs.
In the Qwen-backed scaling runs a single basic task supplies most of the benefit, raising success from 8–10% to 28–40%, yet adding up to six learning samples brings no consistent further gain, and the 20–30 point edge over zero-shot persists even at six atomic actions where zero-shot falls to 4%.
Limitations include evaluation as a multiple-choice walk over a frozen graph with LLM-generated, human-verified candidates rather than a live simulator, a closed-source TongSim build, learning that is prompt or summary reuse with no weight updates, and only 100 advanced tasks behind the headline numbers.

arXiv:2610.13021 PDF

Continually Post-training Small Models via Synthetic Self-distillation

Highlight HF pick · 32▲Large Language Models Woosung Koh, Hyunsoo Lee, Kyungjae Lee, Haeju Park, Dahyun Lee, Dasol Hwang et al. Continual post-training must add capabilities without erasing old ones, and Self-Distillation Fine-Tuning (SDFT) handles this by having the model rewrite its own training data, which fails on the small models that matter for low-latency on-device use because they lack the required in-context learning and instruction following. The authors show poor acquisition and forgetting both correlate linearly with the post-trained model's distribution shift, pointing to in-distribution supervision as the real requirement. SynthSelf therefore outsources the rewrite to a minimally larger model from the same family, combined with off-policy supervision and rejection sampling, keeping supervision quality high without pushing the small model off its own distribution. Across four acquisition datasets, 14 retention benchmarks, and three small model scales it beats six methods including off-policy and on-policy distillation.

Continual post-training must teach a model new capabilities without erasing old ones, and SDFT (self-distillation fine-tuning) does this by having the model rewrite demonstrations in its own distribution, but it fails on small models whose in-context learning and instruction following are too weak to produce usable self-rewrites. SynthSelf keeps the in-distribution rewrite but outsources it to a minimally larger same-family aligner, filters the rewrites by answer match, and then runs plain supervised fine-tuning on the survivors.

The diagnosis is that distribution shift is the culprit: across 23 math datasets fine-tuned into Qwen2.5-3B-Instruct, forgetting on five non-math benchmarks tracks the post-training KL shift at Spearman ρ = -0.85, and at small scale SDFT self-rewrites fall to roughly 60% accuracy and drop the chain-of-thought 8.2% of the time.
The aligner is chosen to minimize the student's negative log-likelihood on its greedy rewrites, subject to that NLL beating gold-SFT NLL and rewrite accuracy beating the base model, which in practice means Qwen2.5-7B or Qwen2.5-14B for 0.5B, 1.5B, and 3B Qwen2.5-Instruct students, with no new hyperparameters introduced.
Averaged over SciKnowEval chemistry, a four-subject science mixture, ToolAlpaca, and MedMCQA across the three student sizes, SynthSelf reaches 56.4 acquisition against 52.2 for SFT and 44.1 for SDFT, while matching the best baseline on 14 retention tasks at 41.7 to 41.8 (still about a point below the 42.8 base), and it beats off- and on-policy distillation using the identical aligner by +5.7 to +16.9 percentage points in acquisition.
Distribution match matters more than aligner strength: Qwen3.5-27B and Gemma-4-31B produce more accurate rewrites yet sit further from the student in NLL and yield equal or worse acquisition, and same-family aligners cut data-generation FLOPs by 62.9% to 76.9%.
The method requires a larger same-family model to exist, its advantage over same-aligner distillation disappears on MedMCQA where the aligner already knows the content, retention is matched rather than improved (0.5B chemistry retention drops to 27.8 from 31.8), and only the Qwen2.5-Instruct family is evaluated.

arXiv:2610.13113 PDF

Rethinking Relative Policy Differences Transfer of On-Policy Distillation

Highlight HF pick · 10▲Large Language Models Rong Wu, Daocheng Fu, Yu Yang, Xuemeng Yang, Siqi Li, Shu Zou et al. On-policy distillation (OPD) usually needs a teacher far stronger than the student, and recent work instead transfers the relative difference between an expert and a reference policy, but which differences carry useful signal and which targets can absorb them has been unclear. The authors frame the log-probability ratio between a proxy expert and a proxy reference as an implicit reward for KL-regularized policy optimization, which induces an on-policy update for the target, recovers standard OPD as a special case, and allows any-to-any transfer across models that differ in scale, training stage, or generation. Experiments show that differences induced by task-specific training such as reinforcement learning or supervised fine-tuning transfer consistently, whereas differences arising from model scale alone transfer weakly or not at all, that the reference filters out shared behavior to isolate the relevant residual, and that raw policy distance does not predict transferability while a well-chosen reference can rescue otherwise failed transfers.

Conventional on-policy distillation (OPD) needs a teacher far stronger than the student, since its reward is the teacher's own log-probability on student rollouts. The paper reframes the newer alternative of transferring a teacher's relative policy change as an implicit reward for KL-regularized optimization, yielding an any-to-any objective (A²-OPD) that recovers standard OPD as a special case and lets the proxy expert/reference pair and the target be chosen independently.

The log-ratio between a proxy expert and a proxy reference acts as a fixed token-level reward on the target's own rollouts, with a KL term anchoring the target to its own base, a centered advantage dropping the intractable normalizer, and experiments across Qwen3 and Qwen2.5 models trained on DeepMath-103K and Eurus-RL-Code using top-16 token distributions.
Differences created by task-directed training transfer reliably while scale-only gaps do not: RL and SFT proxy pairs lift the fixed Qwen3-8B target by 9.78 to 32.23 points on math, a 4B RL expert alone raises it from 18.43 to 45.27, yet the Qwen3-8B/Qwen3-4B scale pair yields −1.03, and two updates of similar KL magnitude (0.35 vs 0.42) give gains of 25.85 vs 3.42, so direction matters more than size.
The expert and reference play complementary roles: swapping the reference from Qwen3-4B base to an early-RL checkpoint collapses the target gain from 26.84 to 9.62 because the shared capability is filtered out, while a Qwen2.5-3B reference gives the largest gain of 29.09 despite the largest base KL distance of 1.23.
A useful difference is not universally transferable: Qwen3 targets gain 9.78 to 29.09, but Qwen2.5 targets under a Qwen3 reference land between −2.46 and 2.72, raw policy distance fails to predict this, and merely switching to a same-family Qwen2.5-3B reference rescues them with gains of 6.74, 2.25 and 5.76.
Limitations include testing only Qwen-family models with a fixed transfer strength of α/β = 1, much smaller gains on code than math (+2.3 vs +26.8 for the same proxy pair), lossy text-based token alignment across tokenizers, and "match the model family" being offered as an empirical fix rather than a general compatibility criterion.

arXiv:2610.13205 PDF

ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents

Highlight HF pick · 41▲Agents Chengyi Yang, Renxu Bao, Jue Hou, Zongpei Teng, Ziheng Wang, Yujie Lin et al. Coding agents are increasingly asked to build programs from scratch rather than fix issues in an existing repository, but verifiable training environments that force an agent to discover requirements and make design decisions are scarce, and current workflows cannot sustain the long-horizon loop of exploring a reference program before implementing it. ProgramGym pairs Program-Env, which turns repositories into verifiable environments through a feedback-driven loop that repairs sandbox builds and the quality of black-box tests and yields 1,648 environments across C, C++, Go, and Rust, with Program-Agent, a role-isolated workflow of analysis, implementation, and review plus revision management for compilation checks and failure recovery. On ProgramBench, fine-tuning Qwen3.8-27B on trajectories from these environments lifts its score from 34.87% to 42.84%, and running the base and fine-tuned models inside Program-Agent raises them to 55.71% and 56.25%, exceeding some flagship models.

ProgramBench-style program reproduction asks an agent to rebuild a program's observable behavior from its documentation and a reference executable, with no source or network, but almost no verifiable training environments exist and single-session harnesses forget early exploration evidence over long horizons. ProgramGym addresses both halves with Program-Env, an automated pipeline that turns GitHub repositories into sandboxed environments with a reference binary and validated black-box tests, and Program-Agent, a role-isolated analyst/implementer workflow with compile-gated revisions.

Program-Env wraps a Claude Code harness in feedback loops that split online dependency fetching from offline compilation, generate tests per functional module in parallel, and discard tests that fail on the reference, flake across runs, accept a dummy implementation, or carry weak assertions, yielding 1,648 environments in C, C++, Go, and Rust from 30,459 candidate repositories with 665,512 validated tests and mean coverage of 72.99%.
Program-Agent runs one Analysis session that records behavioral evidence from the reference program, then alternates Implement and Review sessions that pass only markdown handoff artifacts, with the Implement stage never seeing the reference, each round compiled from a clean snapshot, and failures rolled back to the last validated revision.
Fine-tuning on Program-Env trajectories raises Qwen3.8-27B from 34.87% to 42.84% Tests Passed under Claude Code, and Program-Agent with K=5 rounds lifts the base and SFT models to 55.71% and 56.25%, above Claude Opus 4.7 at 50.90% under mini-SWE-agent, with most of the iteration gain already present at K=3 (54.13%).
The scaffold does nearly all the work, since SFT adds only +0.54 on top of Program-Agent and Program-Agent adds just +1.29 for the stronger DeepSeek-V4-Flash-0731, and the flagship comparison is cross-scaffold rather than apples to apples.
Whole-program reproduction remains unsolved, with Resolved at most 0.50% and Almost Resolved at most 5.5% in every configuration, SFT actually dropping Almost Resolved under Claude Code from 3.50% to 1.50%, an environment set skewed to 1,492 easy versus 5 hard tasks, and construction costing roughly 83M tokens per environment.

arXiv:2610.13395 PDF

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

Highlight HF pick · 22▲Agents Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao et al. Recursive self-improvement (RSI) seeks to turn interaction experience into lasting gains, but external feedback is expensive and a world model's predictions can drift out of date as the policy changes. RSI-Anything, built on JEPA-Anything, uses predictive world models to select both actions and training data, lets observed feedback drive policy or workflow updates and conditional world-model adaptation, and gates retention of each updated policy-model pair behind component and compatibility checks. In code repair with a fixed coder, success rises from 26.3% with a static verifier to 29.2% with predictive selection and 31.2% after world-model adaptation at comparable test usage; graphical user interface (GUI) tasks show higher completion with 39.4% fewer actions, laboratory-planning scores more than double from 21.80% to 46.01% over five policy versions, and reaction optimization reaches 96.15% mean best observed yield versus 82.16% for Bayesian optimization on a 40-experiment budget. An end-to-end immune-science application links workflow evolution and predictive modeling to therapeutic hypothesis generation.

Recursive self-improvement needs external feedback that is costly to collect, and the world models that could stand in for it become unreliable as the policy they predict for changes. RSI-Anything builds on JEPA-Anything predictive world models to decide which candidate actions receive real feedback, then retains an updated policy and world-model pair only when it passes both component-level and compatibility checks.

Each round the world model ranks candidates and only the executed one yields a paired prediction and observation, which drives policy or workflow revision while the predictor adapts only when a reliability check flags it unreliable with sufficient data and budget, and a final pair check decides which policy and model versions carry into the next round.
In code repair with ReasonFlux-Coder-7B held fixed on 200 CodeContests tasks across three seeds, hidden-test success rose from 26.3% with a static verifier to 29.2% with fixed world-model selection and 31.2% after world-model adaptation, at roughly 3.7 online test calls per task throughout, though the first step's 95% interval of +0.17 to +5.50 points barely clears zero.
Five successive retained versions of a 4B policy on LabHorizon F200 Level-2 climbed from 21.80% to 46.01%, edging the Actor–Simulator–Selector reference at 45.32%, and in a flow Suzuki–Miyaura screen the mean best yield reached 96.15% versus 82.16% for Bayesian optimization at a matched 40-experiment budget.
On MiniWoB++ click-checkboxes success went from 51% to 99% with 39.4% fewer actions per episode and repeated clicks dropping from 237 to 0, but a World-based control reached 98% on the same task, so the GUI headline is barely distinguishable from the control.
No single experiment runs the full loop, since code repair fixes the coder and lab planning only tracks policy versions, so sustained joint policy and world-model improvement remains unproven, lab scores measure protocol similarity rather than physical execution, and the immune-science therapeutic hypothesis has not been tested experimentally.

arXiv:2610.13912 PDF

Applications 99

TTS-Q: Perception-Guided Post-Training Quantization of Text-to-Speech Diffusion Transformers

Xiaoqin Feng, Yuqi Li, Weilun Feng, Yi-Cheng Lin, Kai Li, Xianglong Wang et al. cross-listed Zero-shot text-to-speech (TTS) built on billion-parameter diffusion transformers is too large and slow for the on-device, assistive, and real-time uses that need it most, and post-training quantization (PTQ) methods were designed for language and vision rather than speech. The first 4-bit weight-and-activation (W4A4) PTQ study of a TTS diffusion transformer at 1B and 3.5B parameters finds that quantization leaves intelligibility near full precision but consistently erodes speaker timbre, with damage from both weights and activations though only the activation error has a timestep structure a precision schedule can exploit. TTS-Q selects hard-to-quantize items from task-matched calibration candidates and applies a budget-balanced mixed-precision schedule that gives late-step feed-forward activations more bits, recovering timbre best at both scales with no intelligibility cost, while the calibrated W4A4 reference also cuts peak GPU memory.

arXiv:2610.12474 PDF

GeopoliticalQA: Benchmarking Retrieval and Generation over Intelligence Reports

Veronika Solopova, Ivan Havlytskyi, Viktoria Skorik, Ostap Vykhopen Geopolitical and strategic intelligence reports are a rich but heterogeneous source for retrieval-augmented question answering, and no benchmark has targeted them. The authors build GeopoliticalQA, 5,525 document-grounded question-answer pairs over 415 open-source reports spanning factual, inferential, comparative, and risk-oriented questions, generated by a pipeline that profiles each report before writing questions and validated with expert annotation and an LLM-as-judge protocol. Evaluating 15 retrieval systems and six open-weight LLMs, they find document-level retrieval is similar across systems, while at the chunk level the mid-size open encoder bge-m3 matches embedding models an order of magnitude larger and beats commercial embedding APIs. Answer generation remains hard even with gold evidence, especially for inferential and risk-oriented questions.

arXiv:2610.12617 PDF

Critical Mass: Structure-Disjoint Splits Drift With the Libraries They Are Built On

Tomo Oga, Devesh Shah, Cailum M. K. Stienstra, Gabriel Asher, Antonio H. de O. Fonseca, Niall O'Connor et al. Molecular machine learning measures out-of-distribution performance with structure-disjoint splits, and for predicting structures from mass spectra the field settled on single-linkage clustering under the Maximum Common Edge Subgraph (MCES) distance at a threshold of 10, a rule fixed when libraries held 29,000 molecules that has never been checked at today's 227,690. Because MCES is NP-hard per pair, the authors introduce admissible bounds and a union-find skip that reduce about 26 billion molecular pairs to 11.5 million integer programs, a 2,259-fold saving, while provably producing the same lossless partition. At this scale the similarity graph percolates: a single component holds 95.5% of the library, capping any test set at 4.5%. What remains is systematically different, since a held-out compound outweighs a training compound 88.1% of the time, so large-library benchmarks using this rule conflate structural novelty with extrapolation to larger, more complex chemistry.

arXiv:2610.12694 PDF

PVBRF-ID: Private, Verifiable and Byzantine-Robust Federated Intrusion Detection for Oilfield Enterprise Networks

Mustafa S. Aljumaily, Ali Ataeemh Allami, Mustafa N. Mnati, Nawar S. Alseelawi, Shahab Abdulla cross-listed Federated intrusion detection across distributed oilfield sites faces a tension: Byzantine-robust aggregation rules need to inspect each client update in the clear, while secure aggregation hides those updates from the poisoning checks. PVBRF-ID is a two-server protocol in which sites secret-share their updates and the servers jointly test each update's norm and cosine similarity against a reference computed on a small clean root set, revealing only one accept bit per site, with commitments, Merkle roots, and a homomorphic audit making inclusion publicly verifiable. On a 15-class Edge-IIoTset task with ten non-identically distributed sites, it matched FedAvg on clean data and held macro-F1 at 0.696 under sign-flip attacks where FedAvg collapsed to 0.091, while an adaptive attacker controlling 30% of sites reduced it to 0.641 versus 0.046 for FedAvg, still 0.052 below plaintext multi-Krum. An update-inference attacker could recover each site's class mix from plaintext updates but not from a single server's view, though backdoors passed the filter and noise-based mitigation offered only heuristic protection.

arXiv:2610.12704 PDF

Faithful and Complete LLM-Assisted Automated Construction of Hierarchical Hardware Knowledge Graphs

Haotian Qiao, Robert P. Dick Hardware datasheets scatter dense specifications across tables, figures, and text, and even when an entire datasheet fits in a large language model's context window, one-shot extraction omits relevant properties. The proposed pipeline turns datasheets into hierarchical hardware knowledge graphs through chunk-based property extraction, hierarchical routing, and iterative refinement, with the LLM handling extraction and semantic reasoning while graph updates are applied through validated instructions and deterministic local execution that preserves source provenance. On 20 datasheets of varied component types and lengths, the pipeline reaches 98.8% property faithfulness versus 76.1% for a one-shot baseline, fully covers 83.6% of the baseline's faithful properties, and is covered only 10.5% in the reverse direction under a new inter-graph completeness measure that avoids exhaustive ground truth.

arXiv:2610.12762 PDF

DeFiRank: Federated Pairwise Ranking for Governance in Decentralized Financial Networks

Zahra Batool, Michael Duchesne, Kaiwen Zhang, Ulrich Aivodji Governance in decentralized financial networks typically depends on manual review by human committees, which is subjective, does not scale with transaction volume, and misses evolving behavioral patterns. DeFiRank is a federated pairwise ranking framework in which several institutions jointly train a model over their private transaction data without sharing it, scoring providers on signals of genuine network participation such as counterparty diversity, use of multiple applications, and transaction frequency, then promoting authentic contributors and demoting suspicious ones. On a real blockchain transaction dataset the federated model reaches NDCG@15 of about 0.87 against roughly 0.92 for a centralized baseline while matching its ranking stability at a Kendall tau of about 0.70, and it yields concrete decisions promoting 648 and demoting 39 unique providers.

arXiv:2610.12916 PDF

Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan, Sean Suchter cross-listed Centralized GPU clusters for machine learning inference are expensive, inelastic, and prone to long queues, while the GPUs in Apple Silicon Macs sit mostly idle during everyday use. Minions is a dynamic, distributed virtual GPU cloud that pools idle Apple Silicon GPUs across macOS devices to run large-scale inference workloads, providing flexible task execution, fault tolerance, and efficient resource utilization so teams can use hardware they already own instead of buying more. The authors report that idle macOS devices can be turned into a practical distributed inference platform in environments where centralized GPU resources are constrained.

arXiv:2610.12970 PDF

Scaling RNA Foundation Models through Iterative Refinement

Pengan Chen, Liang Hong, Xinyuan Liu, Han Huang, Ziqian Lin, Jiayang Chen et al. RNA foundation models normally couple parameter capacity with inference depth, prescribing one fixed computation graph per input. Eco-RNA is an 80-million-parameter model whose shared, input-conditioned Transformer core is applied repeatedly, so effective depth is chosen after training by setting an iteration count, trained with masked reconstruction, state regularization, and a one-sided loss anchoring deeper reconstructions to a shallow reference. On the 70-assay RNAGym fitness benchmark, going from one to four refinement passes raises aggregate absolute Spearman correlation from 0.242 to 0.286, and at four passes the model surpasses Evo 2 40B on all three aggregate fitness metrics with roughly 500 times fewer parameters. It also leads the evaluated baselines on non-coding RNA family classification, secondary-structure prediction, and chemical probing profiles.

arXiv:2610.13106 PDF

Hidden Data Challenges in Clinical Machine Learning

Carolin Ciss\'ee, Dennis Fast, Raneen Younis, Alexander L\"oser, Wolfgang Nejdl, Zahra Ahmadi Widely reused preprocessing pipelines for electronic health records encode assumptions about when information became available, which clinical episode a label refers to, and what missing values mean, and these assumptions are rarely audited yet propagate as pipelines are copied across studies. Using MIMIC-IV as a case study, the authors identify and quantify violations of three forms of alignment, temporal, semantic, and population, which surface as temporal leakage, feature-task misalignment, and biased imputation. Choices about timestamp semantics, episode definitions, endpoint construction, modality selection, and missingness handling can remain invisible in aggregate predictive performance while changing the signals a model actually relies on. The work closes with practical recommendations for temporally valid, clinically aligned, and reproducible preprocessing of MIMIC-IV and similar datasets, with code and a proposed preprocessing pipeline released.

arXiv:2610.13350 PDF

PPG-Bench: A Unified Benchmark for PPG Foundation Models

Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Sandeep Singh, Siddharth Panwar Photoplethysmography (PPG) foundation models (FMs) are hard to compare because published evaluations differ in datasets, task definitions, preprocessing, downstream heads, tuning, and baselines. PPG-Bench evaluates five public PPG FMs on nine open-access datasets and 26 tasks covering cardiovascular state estimation, activity, sleep, screening, and identification, under four downstream regimes from linear probing to end-to-end finetuning, with per-model hyperparameter optimisation and subject-disjoint cross-validation, and against three controls: a random/constant baseline, a deterministic pulse-rate-variability (PRV) feature set, and the generic time-series FM MOMENT. Across the 68 task-regime combinations, 35.3% show no meaningful advantage over the random/constant baseline, and in 83.8% no PPG-specific FM beats both the PRV features and MOMENT. Extra downstream head capacity generally helps but can sharply degrade models already at chance level, leading the authors to conclude that current PPG FMs cannot yet claim generalised, transferable representations and that future evaluations should include simple non-learned controls.

arXiv:2610.13435 PDF

O\'ido: Open-Vocabulary Speech Recognition on a $5 Microcontroller without a Neural Accelerator

Daniel Varela S\'anchez, Guillermo Mart\'in-Coello Ju\'arez cross-listed Speech recognition on cheap microcontrollers has in practice meant keyword spotting or a fixed command list. Oído is an inference engine that runs open-vocabulary English recognition on an ESP32-S3 (dual-core 240 MHz, 512 KB SRAM, 8 MB PSRAM, no neural accelerator) by executing a 13M-parameter Conformer-CTC model with int8 weights, activations, and attention on the chip's SIMD instructions, with a schedule built around flash and PSRAM bandwidth. On LibriSpeech test-clean and test-other it reaches 3.70%/8.23% word error rate (WER) on device, within 0.12 points of full precision, improving to 3.33%/7.15% with an on-chip 1.3M-parameter language model and 2.99%/6.67% with a transducer head, which is 2.3-2.8x fewer errors than the vendor's recognizer and fewer than Whisper tiny.en and Moonshine tiny at full precision on a laptop. On a physical board the firmware takes 1.97 times the audio duration on one core with transcripts identical to an instruction-exact emulator, so it is accurate but not yet real time, a streaming variant shows partial text while the user speaks, a Spanish model cuts errors 34-60% versus multilingual Whisper tiny, and code and models are released.

arXiv:2610.13447 PDF

A Local Agentic AI Framework to Generate Edge Case Scenarios for Safety Assurance

Victoria J. Hodge, Colin Paterson, Rob Alexander Generating edge-case scenarios is essential for testing and safety-assuring autonomous systems, but many recent language-model-based approaches rely on human-in-the-loop or rule-based strategies that limit how far into the long tail of scenarios they reach. The proposed framework runs entirely on local language models on a laptop, using think, act, observe, update, and continue loops with lightweight JSON encoding: one model generates scenarios grounded in the system requirements, system description, and environment description, a second independent model critiques each scenario against domain-specific criteria, and the refined scenarios run in simulation to collect metric scores that steer the next round either toward similar edge cases (exploitation) or toward novel regions of the scenario space (exploration). Post-hoc clustering on the metric scores plus thematic analysis of the generated scenario descriptions pinpoints functional gaps and their triggering conditions, producing traceable safety-assurance evidence without cloud models, data leaving the machine, or hand-written rules.

arXiv:2610.13479 PDF

Shortcut Learning in Geospatial Foundation Model Embeddings for Cross-Disaster Damage Assessment

Ali Shibli, Andrea Nascetti, Yifang Ban cross-listed Geospatial foundation models can support disaster damage assessment, but high accuracy on familiar disasters may reflect shortcut learning from event or scene characteristics rather than transferable damage evidence. The authors evaluate embeddings from TerraMind, Prithvi-EO-2.0, DOFA, and DINOv3 on the multi-disaster datasets xView2 and BRIGHT, and find that holding out an entire disaster instead of randomly splitting tiles drops mean balanced accuracy from 0.807 to 0.638 on xView2 and from 0.818 to 0.568 on BRIGHT. Further tests show pre-event images alone remain predictive, a classifier given only the annotated building count reaches balanced accuracy above 0.62, and shuffling image pairs within a disaster causes only modest losses. Feature transformations learned solely from training disasters, supervised contrastive learning on xView2 and PCA on BRIGHT, improve generalization to unseen disasters.

arXiv:2610.13766 PDF

FORTE: A Production Framework for Feature Evaluation and Deprecation in Industrial Recommender Systems

Vishnu Arun, Adam Obeng, Tianyou Wang, Zilin Ren, Mukuntha Narayanan Sundararaman, Han Wang et al. Production recommender systems can add features freely but struggle to retire them, since a removal must regress no metric while features interact, drift in importance daily, and are shared across models. FORTE is a model-agnostic framework that scores features daily with an efficient perturbation implementation, groups low-importance features into candidate sets evaluated by grouped ablation, and traces lineage to find every downstream consumer model. Across two large-scale systems its offline results predicted online A/B outcomes well enough to ship deprecations: a search-relevance model had 43.3% of its features removed, and a large ranking model shed 29.9% of its trainable parameters with serving throughput up 19.8% to 25.5% and no regression.

arXiv:2610.13806 PDF

Designing Protein Structures from Natural Language

Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith et al. De novo protein design models usually need a dedicated input representation for each property they condition on, which limits the design objectives they can support. Prompteina is an all-atom structure generation model conditioned on free-form natural language, trained on Prompteina-Data, a paired corpus of 1.8 million structures and over 50 million captions describing global and residue-level structural and functional properties, by adapting the pretrained Proteina-Complexa model through text-conditioned pretraining and finetuning that aligns residues with their descriptions. It outperforms baselines on 15 of 18 structural properties, leads on at least 13 of 20 binder-design targets with 14% higher interface confidence than the strongest baseline, and gains 57% more from iterative in silico feedback than the next-best method on the hardest targets. In wet-lab validation against GM2A, a target with no previously reported de novo binders, nine binders were identified, the strongest with an apparent dissociation constant of 107 nM.

arXiv:2610.13917 PDF

84 more specialized papers

Neuromorphic Satellite Intelligence: A Formal Framework for Event-Driven Architectures, Spike-Based Algorithms, and Quantitative Energy-Accuracy Trade-offs Suresh Kumar TP, Vikas Agnihotri
Acoustic Biomarkers for Alzheimer's Disease Detection from Spontaneous Speech Anh L Tran, Arjee Javellana Restar, Bereket Kindo
EarLog: Opportunistic Ear Physiology Sensing toward Personalized Hearing Risk Awareness Yongzhi Huang, Jiayi Zhao, Kaishun Wu
Energy-Efficient Spiking Neural Networks for Sensing-Aided Beam Prediction Nhan Thanh Nguyen, Phuong Nam Tran, Mengyuan Ma et al.
Behind-the-Meter Photovoltaic Visibility from Net Smart Meters: Deployment Regimes and Operational Insights Jack Zheng, Hao Wang
GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer Adinath Madhavrao Dukre, Abdullah Mohammad, Yifan Lu et al.
Structural and Representational Agency in LLM-Driven Timbre Generation and Interpolation Joseph Cameron, Alan Blackwell
Market Impact Estimation via Causal Foundation Models Dennis Thumm, Ying Chen
Enriched Holomorphic Neural Networks for 2D and 3D Laplace Problems Kalp Shah, Andrei Vlad Nicula, Mihnea Andrei B\^arsan et al.
Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective Dmitry Osin, Egor Surkov, Petr Mokrov et al.
Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data Georg Maierhofer, Miles Cranmer, Payel Mukhopadhyay
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Amey Vasulkar, Samuel Kernan Freire, Arie-Willem de Leeuw et al.
Data-driven Optimal Sensor Placement for Aerosol Monitoring in the UK Talha Ansar (University of Engineering and Technology, Lahore, Pakistan) et al.
Incremental Retrieval-Augmented Generation for Human-in-the-Loop Decision Support Federico Mazzoni, Andrea Failla
GeoFitBench-LLM: A Benchmark for Assessing Large Language Models' Capabilities in Geospatial Data Requirements Elicitation for Geo-Analytic Tasks Amirhossein Vahdat, Thierry Badard, Luc Lamontagne
Expert-level detection of low-voltage intervals in EEG as a robust biomarker of neonatal encephalopathy John M. O'Toole, Sean R. Mathieson, Aurel Luca et al.
Spatial Context or Memory? Sizing the Inputs of Neural Surrogates for Partially Observed Fields Ruige Kong
KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting Jonas Petersen, Pratik Jawahar, Riccardo Maggioni et al.
U4RASD at KnowledgeGraphEval-2026: Ontology-Constrained Encoder and LLM Systems for Arabic Relation Extraction Ahmad Abbas, Hussein Mazeh, Chadi Abou Chakra et al.
Trajectory Prediction for Autonomous Driving in Road Construction Zones Using Human-in-the-Loop Driving Data Shuo Zhang, Aman Sunesh, Semiha Ergan et al.
NV-Tesseract-AD: Curriculum-Enhanced Conditional Diffusion for Multivariate Time-Series Anomaly Detection Aditi Gautam, Manikandan Ravikiran, Saira Qureshi et al.
Provenance-grouped protocol for evaluating machine learning in perovskite photovoltaics Qiwei Han, Chi Zhou
Credit Attribution in Timestamped Soccer Networks Using Multivariate Hawkes Processes Phillip W. Kornberg, Kevin S. Xu
Polyconvex neural-network surrogates with learned paired directional features for anisotropic hyperelastic homogenization S. Ares de Parga, A. Cornejo, J. A. Hern\'andez et al.
Inverse Synthesis of Shock Time Series from Shock Response Spectra Using Ensembles of Conditional Variational Autoencoders Adam Watts, Charles Young, Ryan Bowering et al.
Unreliable Narrator Framing as a Cost-effective Mechanism for Satire Explanation of Australian and American Satirical News Kanay Gupta, David Y. Liu, Aditya Joshi
Localized Generative Design of Ship Hulls with Diffusion Models and High-Fidelity Assessment Hadi Keramati, Mohammed A. Hannan, Patrick Kirchen et al.
CoGameReg: Cooperative-Game-Based Collaborative Semi-Supervised Regression for Seismic Reservoir Property Prediction Zhiyu Guo, Jiwei Jia, Lei Li
AI-Mediation in Retail Markets: Revenue Effects and Assortment Optimization Omar El Housni, Marouane Ibn Brahim
Teaching neural decoders with language Xuan Wen, Yonghao Song, Yuhong Zhang et al.
Diffusion Precipitation Generation with Medium-range Weather Forecasting Guowen Li, Yang Liu, Mengxuan Chen et al.
Beyond Semantic IDs: Finite-Beam Reachability and Continuous-Query Generation in Generative Recommendation Yu Sun, Renkai Xiang, Zuodong Xiang et al.
NucleoNet: A Nuclear-Aware Determinant Ansatz for Multi-Geometry Neural Variational Monte Carlo Qiao Gao, Yang Huang, Enze Hou et al.
Benchmarking Embedding Geometries and Clustering Algorithms for Unsupervised Patent Claim Classification Sujith Subramanian, P N Bhargav Teja, Susen Sarkar et al.
Low-Rank Kolmogorov-Arnold Generator Flows for Interpretable Operator Learning Minjong Cheon
Lightweight Time-Frequency Vibration Sensing for Consumer Device Condition Monitoring Po-Heng Chou, Jen-Yu Chiu, Chun-Yu Yeh et al.
CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma et al.
Evolving Creatives with Clicks: CTR-Driven Advertising image Editing via Attribute Disentanglement and Preference Alignment Xu Zhang, Haohan Wang, Zheng Zhang et al.
BenCoDiC: A Large-Scale Contextual Cyberbullying Corpus for Bengali Political News Discourse FULL VERSION Md. Nesarul Hoque, Rudra Pratap Deb Nath, Umme Salma et al.
Looking Ahead and Back: Anticipatory Generative Modeling for Road-Network Trajectories Zhenhao Huang, Li Sun, Philip S. Yu
VIGOR: Verifiable Industrial Geometry Optimization and Repair in Scene Programs Fuqiang Wang, Zheng Guo, Xiaoming Wu et al.
SharP: Soft and Hard Prompt-Guided Augmentation with LLM for Low Resource Fake News Detection Ziyu Luo, Wei Li, Haoping Huang et al.
Tensor-based Fusion of Multi-modal Data for Predicting Cancer Risk Levels Ali Anaissi, Yuanzhe Jia, Qishun Peng et al.
Learning Schwarzschild Geodesic Motion with Lagrangian Neural Networks Sukru Caglar, Sena Kalabalik, Abdullah Umut Hamzaogullari et al.
Neural Surrogate Modeling of Ultrasonic Wave Propagation in Heterogeneous Media Burla Nur Korkmaz, Petro Moilanen, Ari Salmi et al.
Recovering Irreducible Three-Body Interactions with Lagrangian Neural Networks Sena Kalabalik, Sukru Caglar, Abdullah Umut Hamzaogullari et al.
MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design Jixin Tang, Ji Guo, Jun Zhao
Sustainable Cloud-Fog-Edge Systems for Smart Buildings: Algorithmic Adaptation under Memory Constraints and Dynamic Orchestration Strategies Mamadou Sow, Christophe Cerin
JEPA-kMMVAE: CSI Prediction for Time-Varying Wireless Channels Efstathios Chatziloizos, Konstantinos Vandikas, Aneta Vulgarakis Feljan et al.
SBNO : Schr\"odinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information Jaehyeon Park, Inhyeok Jeong, Noseong Park
PDE-PFN: Prior-Data Fitted Neural PDE Solver Jaehyeon Park, Mingu Kang, Dongseok Lee et al.
MoPrO-SD: Modular Prompt Optimization for Stance Detection Meysam Sabbaghan, Nikesh Gyawali, Arman Zareian Jahromi et al.
PlacePulse: Learning Transferable Place Representations from Structural Geographic Data Luc\'ia Garc\'ia-Duarte, Miguel \'Alvarez, Diego Vicente et al.
RegulaTutor: A Theory-Informed Framework for Long-Term Personalized Learning Yonggang Li, Boqian Han, Rui Sheng et al.
Desertification Forecasting in Semi-Arid Regions: Using Deep Learning to Probe the Observability Limits of Land Degradation Dynamics Filipe Moura Da Silva, Cl\'aubio Landney Lima Bandeira
In-Context Learning for System Identification: Probabilistic Formulation, Long-Sequence Handling, and Benchmark Validation Matteo Rufolo, Dario Piga, Filippo Pura et al.
Does Financial Speech Need Continual Learning Regularization? A Study of EWC and mini-KFAC on Earnings Call ASR Thai Quoc On, Homayoon Beigi
CNN-Transformer with Sigmoid Label for High-Resolution Sub-Degree DOA Estimation Lifeng Wang, Xintong Li, Aifei Liu
Local Diagnostics of Learned Event Representations for Anomaly Detection at the HL-LHC Donatella Genovese, Philip Ploner, Simone Scardapane et al.
Collective Autonomous Vehicle Routing Needs Social Awareness for Coexistence with Human Drivers Anastasia Psarou, Mikolaj Rams, Blazej Torbus et al.
Revisiting Accent Classification with DistilHuBERT: Speaker-Disjoint Evaluation and Cross-Corpus Generalization Xu He, Homayoon Beigi
A Scalable Learned Iterative Solver for Nonlinear MPC via Temporal Decomposition Lukas L\"uken, Sergio Lucia
Retrieval-Augmented Multimodal Language Model with Heterogeneous Evidence Integration for Clinical Prognosis Hamza Shafiq, Jialu Tang, Ting Dang et al.
Probabilistic Ship Trajectory Forecasting using Transformer Models on AIS Data Frederik E. T. Sch\"oller
$L_0$ Feature Selection via Monte Carlo Tree Search in High-Dimensional Financial Regression Cameron Mouangue, Tristan Cazenave
Probabilistic Forecasting of Electric Vehicle Charging in a Time-varying Hierarchical Infrastructure Guillaume Principato, Yvenn Amara-Ouali, Yannig Goude et al.
EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction Michelle Espranita Liman, Ioana Ionac, Alexander M\"uller et al.
CorrFlow: Exact and Permutation-Equivariant Flow Matching on the Correlation Elliptope Gautier Marti
Vocal Intensity Conversion with Continuous SPL Control via Adversarial Training Quentin Le Tellier, Albert Rilliard, Olivier Perrotin et al.
AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis Natapong Nitarach, Phume Ngampornsukswadi, Pittawat Taveekitworachai et al.
Private and Distributionally Robust Federated Multiclass Classification Andrea P. Hernandez, Michael Ibrahim, Nagi Gebraeel
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data Jayanie Bogahawatte, Sachith Seneviratne, Simon Maksour et al.
Electron-Density-Informed Molecular Property Prediction Using Surrogate Functionals Manuel V. Klockow, Simon Wagner, Roman Remme et al.
Verifiable Retrieval for Legal Reasoning through Structure-driven Knowledge Graph Ibrahim Siddig, Sviatoslav Tugeev, Munir Georges
Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training Ibrahim Siddig, Munir Georges
SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark Taeyoung Yu, Seonbin Jo, Jiwon Kim et al.
REFLECT: Role-Aware Detection of Sentiment, Emotion, Politeness, and Formality Sangpil Youm, Annabel Stocks Natalias, Hayden Heninger et al.
Error Rate is Not Enough: Costs of Augmentation in Low-Resource ASR Katsumi Ibaraki, David Chiang
Shared Canonical Component Removal Beyond Expert Views Jose Roberto Tello Ayala, Akl C. Fahed, Roy Perlis et al.
OpenPA: Forced-alignment-free Multi-granular Pronunciation Assessment with Mixed-transcript Adaptation Fu-An Chao, Bi-Cheng Yan, Berlin Chen
Studying Comics at Scale with Artificial Intelligence: The PAGE Framework and Three Studies of American Comics and Japanese Manga Emilio Ferrara
Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design Mason Minot, Simon Friedensohn, C\'edric R. Weber
DyssomniaNN: Multi-Scale Spiking Neural Modeling of Human Sleep Disorders and Perceived Restedness Daniel Szelogowski
Goal-Oriented Bundle Management and Flow Control for Deep-Space Communications via Reinforcement Learning Sajjad Baghaeea, Aimin Lib, Shanmugarajan Balasubramanian et al.

Large Language Models 67

On the estimation and validity of AI time horizons---a statistical look at the METR plot

Drew T. Nguyen, William Fithian cross-listed METR's 50% time horizon expresses an AI system's capability as the human completion time of software tasks it solves half the time, but the standard estimate assumes a task's difficulty for AI is linear in the log of human time. Using 228 tasks and 26 models, the authors refit the horizons with splines and item-response theory, producing a learned function that converts human time into AI difficulty. That function is nearly flat between 2 and 30 minutes and roughly linear elsewhere, so a horizon jump from 3 to 30 minutes is far easier than one from 30 minutes to 5 hours despite both being a 10x multiplier. The spline estimates score better under a cross-validated suite of proper scoring rules, and the paper adds diagnostic plots for assessing construct validity as time-horizon benchmarks expand to longer tasks.

arXiv:2610.12466 PDF

Algebraic Integer Quantization for Machine Learning Applications

Anthony Kostal-Vazquez, Vassil Dimitrov, Branden Butler Algebraic integer number formats (AINF) are proposed as a configurable framework for compressed computation whose key property is a straightforward reduction to pure integer arithmetic, enabling error-free multiply-accumulate operations. The authors build a matrix multiplication engine around this property that runs within a predictable factor of the most efficient numerical kernels on modern accelerators, and note that the algebraic integers can be chosen to match a workload's numerical characteristics in ways floating-point and fixed-point formats cannot. From this they derive algebraic integer quantization (AIQ), a lossy format for large language model (LLM) inference, and report that AIQ matches the accuracy of the industry-standard MX format at equal bits per parameter across a range of models and benchmarks, using degree-two (quadratic) algebraic integers with higher-degree and multidimensional generalizations left for future work.

arXiv:2610.12494 PDF

AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits

Shu Ishida, Aliasghar Khani, Tianyu Zhang, The Cong Luong, James Seale Smith, Adam Gaier Autoregressive decoding cannot revise earlier output, which suits it poorly to data with cross-references or future dependencies, and diffusion language models still cannot insert or delete arbitrary numbers of tokens in already-generated context. AutoIndexer trains a causal model on a Chain of Edits: it predicts marker tokens that open and close edits while moving a cursor through the existing token sequence, like a text editor, so it can insert, substitute, and delete tokens in hindsight. Training sequences are built by perturbing ground-truth labels with inverse chained edit operations, and the model learns to undo them, which also makes it robust to erroneous context and applicable to tasks such as code correction. The model shows post-generation context-editing abilities that most autoregressive and diffusion models lack without compromising performance on standard LLM benchmarks, and the code is released.

arXiv:2610.12545 PDF

Sample-Adaptive Zeroth-Order Post-Training

Haiyuan Wan, Yanyi Li, Weicheng Lin, Cong Fang Zeroth-order (ZO) optimization fine-tunes large language models without the memory cost of backpropagation, but existing methods such as MeZO sample training examples uniformly, ignoring their changing importance and leaving avoidable estimation variance. Sample-Adaptive Zeroth-Order Optimization (SAZO) extracts gradient-norm information from the directional differences the method already computes, tracks it with per-example exponential moving averages, and allocates queries adaptively, adding no extra model evaluations and only linear scalar state in the number of examples. The authors prove SAZO asymptotically matches the leading convergence term of a minimum-variance sampling oracle and converges strictly faster than standard methods. Across four model families it improves average downstream performance and reaches the same training loss with up to 6 times fewer queries.

arXiv:2610.12566 PDF

Contextual Correction of Parallel Draftsfor Speculative Decoding

Roy Eisenstadt, Harel Lidar, Tal Gonen, Omer Luxembourg, Itamar Zimerman, Lior Wolf et al. Speculative decoding speeds up large language model inference by having a lightweight drafter propose several tokens that the target model verifies in parallel, and diffusion-based drafters such as DFlash draft many positions at once, but tokens within a block are predicted without seeing the realized identities of preceding draft tokens. DCOCO (DFlash Contextual Corrector) keeps multiple candidate tokens per position and scores their contextual compatibility across adjacent draft positions in parallel, restoring intra-block dependencies without giving up parallel drafting. Across four verifier backbones and ten benchmarks it achieves the highest average throughput in every configuration, covering greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding. Under concurrent serving with 4 to 32 requests it beats the strongest competing method by 6 to 23% while other sequential correctors fall below uncorrected DFlash.

arXiv:2610.12570 PDF

One Model to Loop Them All: Elastic Latent Recurrence for Language Models

Faisal Ladhak, Shantanu Acharya, Santiago Akle Serano, Simeng Sun, Boris Ginsburg, Krishna C Puvvada Latent recurrent language models reuse a shared block of layers so one checkpoint can trade inference compute for accuracy by changing the loop count, but at one loop they typically fall well short of a non-recurrent model of the same size, a tax that makes a separate model the better choice whenever latency matters. Elastic Latent Recurrence (ELR) removes this tax through a gated state update with a learned step size, loop counts mixed within each optimizer update, and conditioning each pass on its loop index and budget; mixed depths can be introduced late in pretraining to limit added cost, and keeping only the last loop's attention KV cache holds memory at the one-loop level while retaining most of the gain. On 4B dense and 30B-A3B mixture-of-experts hybrid models, one-loop ELR lands within about a point of its matched baseline on a ten-task average and four loops add 7.25 and 4.75 points respectively. Carried through post-training, the recipe yields Koi-4B, a single checkpoint that matches Qwen3.5-4B with an estimated 18 on the Artificial Analysis Intelligence Index at one loop and rises to 29 at four loops.

arXiv:2610.12582 PDF

STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules

Ranajoy Sadhukhan, Euijin Hong, Sheng Cao, Zechun Liu, Shuowei Jin, Sha Li et al. Sparse computation can grow a language model's parameter count without proportionally increasing active compute, but sparse pretraining is expensive and existing upcycling methods target Mixture-of-Experts (MoE) models, where expert homogeneity, sample inefficiency, and system overheads limit scalability. STAPLE builds on STEM, which replaces feed-forward up-projections with token-indexed embeddings, but instead keeps the up-projections and adds the static sparse embeddings as a residual pathway, preserving pretrained computation while learning complementary token-specific information. On OLMo-2 1B it improves over continued dense training by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP, the gains grow with static capacity and persist on OLMo-2 7B, and it beats STEM, an Engram-style model, and an MoE upcycling baseline on average while decoding 2x faster than the MoE baseline with less than half its GPU memory and energy per token. A host-offloaded implementation that caches only the roughly 5% most frequent embeddings on GPU keeps generation throughput within 1.5% and training throughput within 3% of the dense baseline despite adding billions of parameters late in training.

arXiv:2610.12599 PDF

Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization

Bahareh Nikpour, Hossein Rajabzadeh, Maryam Dialameh, Vahid Partovi Nia, Parsa Omidi The key-value (KV) cache grows linearly with sequence length, batch size, and depth, and existing remedies such as Multi-Query Attention (MQA), Grouped-Query Attention (GQA), Multi-Head Latent Attention (MLA), and Tensor Product Attention (TPA) trade away expressivity, add positional-encoding machinery, or impose restrictive low-rank structure. Kronecker Product Attention (KPA) represents keys and values as sums of Kronecker products of compact, token-dependent matrix factors, strictly generalizing TPA since its rank-one outer products arise as a degenerate Kronecker split while non-degenerate factors capture richer block-structured patterns. A structured rotary embedding operator matched to the Kronecker form preserves the relative-position property of RoPE while rotating only the compact factors, removing the decoupled RoPE branch MLA needs. KPA matches or exceeds TPA, MHA, MQA, GQA, and MLA across four model scales and nine benchmarks under identical training while achieving up to 10x KV cache compression, with the Kronecker head split acting as a tuneable knob on the compression-accuracy trade-off.

arXiv:2610.12657 PDF

Semantic Views: Certified Reuse of LLM Labels Across Semantic Queries

Oktie Hassanzadeh, Dharmashankar Subramanian cross-listed Semantic query engines evaluate natural-language filters by calling a large language model oracle on each tuple, and existing optimizers either miss precision and recall targets, save few calls when they offer guarantees, or discard the oracle's labels once a query ends. SemViews keeps those labels as reusable semantic views, orders tuples by signals that cost no oracle calls, namely earlier view labels, a local embedding, and a helper model's score, refines that order with labels it must buy anyway, and certifies the answer with fresh stratified oracle samples so that precision and recall targets hold with probability at least 1 minus delta for any workload or catalog. Replayed against 725,000 recorded oracle labels on four workloads, it meets both targets on 99.63% of queries while saving 28% to 45% of oracle calls, where the only prior method with the same guarantee needs 1.32 to 1.71 times as many calls, and rephrased repeat questions cut calls by a further 27% to 36%.

arXiv:2610.12768 PDF

COFFEQA: Counterfactual-Guided Few-Shot Distillation for LLM Question Answering

Kavinda Kehelella, Sanghamitra Dutta Knowledge distillation compresses a large teacher into a small student but is usually data-hungry, and few-shot distillation turns out to be very sensitive to which training examples are chosen, with uncertain samples near the teacher's decision boundary actually hurting the student. COFFEQA addresses this in open-book question answering by pairing each factual example with a counterfactual, a minimally edited context that flips the teacher's answer, and filtering these pairs with a new Context Following Difficulty (CFD) score that checks the teacher actually follows the modified context to the intended alternative answer. On five in-domain and three out-of-domain QA benchmarks with T5 and Qwen2.5 students, counterfactual-augmented distillation beats factual-only baselines at the same few-shot budget, gaining 8.77 exact-match and 9.98 F1 points on TriviaQA with 32 pairs in the T5 setting.

arXiv:2610.12837 PDF

Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization

Jiaqi Tian, Ting Long On-policy self-distillation (OPSD) lets a large language model (LLM) improve by acting as both teacher and student under different conditioning contexts, but the authors find that queries answered correctly early in OPSD training can become incorrect later even as aggregate accuracy keeps rising. This forgetting coincides with degraded hidden-representation geometry, specifically looser intra-cluster compactness and weaker inter-cluster separation. Representation Geometry Regularization (RGR) adds a term that encourages within-cluster compactness and between-cluster separation during post-training and plugs into existing OPSD methods without changing their core optimization, and across multiple OPSD methods and benchmarks it generally reduces the capability degradation while preserving the gains.

arXiv:2610.12855 PDF

One-Step Text Generation by Seq-Drifting

Jiayu Liu, Ruizhi Zhao, Tianyun Ji, Yee Whye Teh, Zhenya Huang, Ning Miao Diffusion language models generate tokens in parallel but still need many sampling steps, so this work asks whether a model can emit an entire sequence in a single forward pass. Seq-Drifting applies drifting to sequences of continuous token representations, learning a direct map from Gaussian noise to text through a drift field whose per-position targets adapt to the current generated context, with attraction pulling each output toward contextually plausible token representations and repulsion discouraging collapse across samples and repetition within a sequence. The one-step generator reaches a generative perplexity of 36.01 on LM1B and is competitive on unconditional and conditional generation, with further experiments on translation, summarization, and reasoning. Code is released.

arXiv:2610.12879 PDF

Rethinking the Roles of Diffusion and Decoding: Predictive-State Flows for Parallel Language Generation

Jinjin He, Xinsong Feng, Lucheng Fu, Sinan Wang, Bo Zhu, Jiayun Wang In continuous diffusion language models (DLMs) the diffusion prior must encode fine-grained discrete token choices in continuous space because latents map almost deterministically to tokens, which makes the joint latent distribution hard to model. Predictive-State Flows (PSF) splits generative roles: a sequence is partitioned into blocks, each associated with a predictive state computed only from preceding context, a rectified-flow prior generates the sequence of predictive states, and a block-causal decoder then generates all blocks in parallel while remaining autoregressive within each block, reducing diffusion from L token-level states to K much smaller than L and sequential depth to L/K. On unconditional OpenWebText generation, a 197M-parameter model with K=8 reaches generative perplexity 22.8 at entropy 5.16 using 0.31 TFLOPs per 1024-token document, versus 24.0 and 7.2 TFLOPs for ELF-B with 32 steps, with MAUVE@512 of 0.93 versus 0.73. Controlled comparisons show content states that also encode their own blocks give 8-16% worse perplexity than predictive states, and replacing sampled predictive-state trajectories with a shared mean lowers MAUVE@512 by 0.11.

arXiv:2610.12898 PDF

Surfing the Activation Wave: Analyzing the Information and Geometry of LLM Hidden States for Speculative Decoding

Raghavv Goel, Sudhanshu Agrawal, Mukul Gagrani, Risheek Garrepalli, Mingu Lee Speculative decoding drafters such as EAGLE3 draw on a few target-model hidden layers chosen by fixed position heuristics, even though picking the best subset is a combinatorial problem that cannot be searched exhaustively. The authors score candidate layer sets on two criteria, informativeness measured by matrix entropy and diversity measured by centered kernel alignment (CKA), and study layer-wise entropy, entropy under concatenation, and pairwise CKA dynamics across depth, revealing distinct informational compartments in decoder representations; a token-entropy filter drops hidden states for low-predictive-entropy tokens to clean up the selection statistics. In EAGLE3-style speculative decoding, the entropy- and CKA-selected layer sets outperform fixed defaults, including the standard three-layer rule, while also explaining why some defaults remain competitive.

arXiv:2610.12939 PDF

What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training

Yuxuan Gu, Kequan Niu, Xiaocheng Feng, Yun Li, Yiqiao Yang, Bing Qin Supervised fine-tuning (SFT) on fixed demonstrations can override behaviors a model already acquired in pretraining, and recent token-reweighting methods use model compatibility as the signal, which conflates undesirable mismatches with exactly the task-specific decisions a demonstration is meant to teach. CIG-SFT instead scores each demonstrated token by conditional information gain: under teacher forcing, a frozen reference model's likelihood of the token with and without the task input, holding the gold solution prefix fixed, measures how much the task itself reduces uncertainty about that decision. This gain defines a KL-regularized projection of the current policy, which the authors show is equivalent to a scalar-weighted SFT objective whose logit gradient matches the projected target. On mathematical reasoning and code-generation benchmarks it consistently outperforms standard SFT and recent adaptive fine-tuning baselines, with the largest gains on harder out-of-distribution evaluations, without extra training data or online rollouts.

arXiv:2610.12995 PDF

Evaluating LLM Creativity through Long-Tail Performance

Yichen Wang, Tenghao Huang, Vishakh Padmakumar, Mina Lee Creativity benchmarks for large language models (LLMs) differ widely in task and metric design, raising the question of whether they measure a shared notion of creativity. The authors propose that creativity can be formalized as task-specific performance in the long tail of a reference input-output distribution, grounded in the novelty-quality product, and test this across 15 benchmarks and 26 open-weight models using n-gram absence from pretraining data as the novelty proxy. Creativity benchmarks sit further in the tail than standard benchmarks, and model rankings from the higher-tail subsets of a benchmark align more closely with human creativity judgments in LM-Arena, on both creativity and standard benchmarks. They introduce q@Tail(α), a cheaper diagnostic that measures mean quality on the tail fraction of an existing benchmark, and find that current creativity-promoting methods leave substantial tail performance untapped.

arXiv:2610.12998 PDF

KV-Repurpose: A Testbed for Evaluating Non-Prefix KV-Cache Reuse

Taiga Ikeda, Daisuke Miyashita, Jun Deguchi Key-value cache reuse in language model serving normally demands an exact prompt prefix match, so a retrieved document cached behind one system prompt cannot be reused behind another. KV-Repurpose is an open-source, vLLM-based testbed for segment-wise reuse, bundling cache-aware segment scheduling, prompt construction, capacity management, workload generation, and runtime monitoring. Naive segment reuse largely preserves output quality on single-document question answering and summarization but clearly degrades cross-document reasoning. On a Zipfian retrieval-augmented workload it reaches a 63.9% token-level hit rate against 3.3% for prefix caching at the same 30 GB budget, lowering time to first token, while prefix caching tops out at 17.9% even with 1 TB.

arXiv:2610.13060 PDF

Double-Take: Improving Capacity, Binding, and Selectivity In Long-Context Within Bidirectional Memory

Vakul Raghavan, Utkarsh Grover, Mingyang Mao, Chengwen Wang, Xiaomin Lin Extending context length does not guarantee distant information stays usable, since a fixed-size memory must hold many associations, link them across positions, and avoid being swamped by everything it sees. Studying large-chunk test-time training, the authors trace two failures: single-step writes accumulate interference and cap capacity, while self-associative writes cannot form cross-position bindings. Double-Take adds multi-step unpinned updates, cross-position mixing, and bidirectional change-sensitive writes for selective retention. Capacity alone leaves recall at the 0.10 chance floor, adding binding lifts it to 0.92, and a model trained only on 32 kb genomic contexts reaches 0.927 AUROC at 131 kb and 0.856 near 1 Mb while transferring zero-shot from human to mouse.

arXiv:2610.13073 PDF

Continually Post-training Small Models via Synthetic Self-distillation

Woosung Koh, Hyunsoo Lee, Kyungjae Lee, Haeju Park, Dahyun Lee, Dasol Hwang et al. Continual post-training must add capabilities without erasing old ones, and Self-Distillation Fine-Tuning (SDFT) handles this by having the model rewrite its own training data, which fails on the small models that matter for low-latency on-device use because they lack the required in-context learning and instruction following. The authors show poor acquisition and forgetting both correlate linearly with the post-trained model's distribution shift, pointing to in-distribution supervision as the real requirement. SynthSelf therefore outsources the rewrite to a minimally larger model from the same family, combined with off-policy supervision and rejection sampling, keeping supervision quality high without pushing the small model off its own distribution. Across four acquisition datasets, 14 retention benchmarks, and three small model scales it beats six methods including off-policy and on-policy distillation.

arXiv:2610.13113 PDF

DP-MIND: Differentially private memory-guided Long context Inference

Zhihao Liu, Yitian Cheng, Shuang Liu, Rui Wen, Yuke Hu, Zhan Qin Sending sensitive contexts to hosted large language models raises privacy concerns, and existing differential privacy (DP) inference methods are designed for short or moderate contexts and degrade badly on very long inputs. DP-MIND perturbs the input locally, processes it chunk by chunk, and carries a compact memory that preserves task-relevant information across the sequence. The memory construction policy is trained with reinforcement learning directly on perturbed inputs, rewarded for final answer correctness and memory consistency, and a local denoising module recovers perturbation-corrupted answer spans without exposing the original context to the external model. The authors give a theoretical privacy analysis and report that the framework outperforms existing DP and long-context inference baselines at context lengths up to the million-token scale.

arXiv:2610.13132 PDF

Neurograf v1: An End-to-End Romanian-First Language Model Family

Ionut Adrian Rotaru Training a language model family from scratch for a lower-resource European language is tested end to end for Romanian. The pipeline comprises RoSfera, a 635 GB corpus of 175.7 million documents, an 80k byte-level BPE tokenizer, five decoder-only models from 160M to 6B parameters trained on 131 billion tokens, three instruction-tuned assistants, and two contrastively trained embedding models. Validation loss and Romanian factual and cultural knowledge scale consistently with model size: on the 1,000-question WWTBM benchmark, Neurograf-6B reaches 69.89% accuracy on Romanian-context questions and Neurograf-3.2B reaches 67.38%, while a compute-optimal analysis gives a parameter-scaling exponent of about 0.423. The authors frame the release as an experimental demonstrator and document practical challenges in diacritic preservation, corpus quality, data composition, and tokenization efficiency.

arXiv:2610.13134 PDF

PreSTO: Predictive Subtree Prefetching for Fast LLM Power Sampling

Nan Jiang, Panagiotis Theodoropoulos, Wei Deng Power sampling sharpens a large language model's output distribution at test time to improve reasoning, but its PowerMH sampler is slow because each Metropolis-Hastings (MH) accept-or-reject decision sets the state from which the next proposal is generated. Predictive SubTree Prefetching (PreSTO) views future MH outcomes as a tree of accept/reject decisions, generates the proposals for a prefetched subtree in one batched model call, then follows the realized path with the original acceptance rule, which provably leaves the MH transition kernel unchanged. Because MH reaches only a sparse, nearly unary part of each subtree, a predictive traversal rule chooses the subtree expected to complete the most transitions. Across seven datasets and five base LLMs the method gives median speedups of 1.71x, 1.91x, and 2.33x on PowerMH, EntropyCut, and MultiTryMH, making about 3.9x fewer model calls that are each only about 2.1x longer.

arXiv:2610.13171 PDF

Rethinking Relative Policy Differences Transfer of On-Policy Distillation

Rong Wu, Daocheng Fu, Yu Yang, Xuemeng Yang, Siqi Li, Shu Zou et al. On-policy distillation (OPD) usually needs a teacher far stronger than the student, and recent work instead transfers the relative difference between an expert and a reference policy, but which differences carry useful signal and which targets can absorb them has been unclear. The authors frame the log-probability ratio between a proxy expert and a proxy reference as an implicit reward for KL-regularized policy optimization, which induces an on-policy update for the target, recovers standard OPD as a special case, and allows any-to-any transfer across models that differ in scale, training stage, or generation. Experiments show that differences induced by task-specific training such as reinforcement learning or supervised fine-tuning transfer consistently, whereas differences arising from model scale alone transfer weakly or not at all, that the reference filters out shared behavior to isolate the relevant residual, and that raw policy distance does not predict transferability while a well-chosen reference can rescue otherwise failed transfers.

arXiv:2610.13205 PDF

CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation

Jingru Yang, Jinhang Wu, Minjing Shi, Shengfeng He, Xuyan Li, Man Tan et al. Cross-tokenizer on-policy distillation (OPD) must align teacher and student sequences before comparing scores, and decoded-text alignment gives chunk-level correspondence but leaves open how a chunk's signal should be spread across student tokens, while tokens with no teacher counterpart are left unconstrained. CAST (Chunk-aware credit Assignment with Selective Token Anchoring) redistributes each aligned chunk's advantage across its student tokens according to predictive uncertainty and local smoothness of the teacher signal, preserving the chunk total through weights that exactly solve a KL-regularized credit-allocation problem, and anchors unaligned positions to the student's initial policy with a KL penalty. Because credit assignment is decoupled from chunk construction, the method also works with punctuation-defined or fixed-length chunks under a shared tokenizer, and across six math, science, and code benchmarks it beats the strongest evaluated baseline by 2.45 and 1.32 average points when distilling Qwen3-8B into Mistral-7B and Llama-3.1-8B, and by 2.10 points for shared-tokenizer transfer to Qwen3-1.7B.

arXiv:2610.13257 PDF

CopRoute: Dependence-Aware Conformal Routing for Language Models

Siva Rajesh Kasa, Santhosh Kumar Kasa, Karan Gupta Set-based routing calls several language models per query and tries to ensure at least one answers correctly while limiting inference calls, but existing risk-controlled policies score models independently and so ignore that models sharing data or architecture tend to fail on the same queries. CopRoute orders models greedily by how many additional queries each one covers beyond those already selected, decides per query how deep to go in that order, and calibrates a conformal cutoff that certifies the all-miss risk of the deployed sets. The authors show the ordering achieves the best guarantee available to any efficient fixed-ordering method for expected calls to the first correct model, and on RouterBench and RouterEval it attains the fewest certified calls on nine of ten datasets at a 0.05 target risk, making up to 79% fewer calls (median 24%) than the state-of-the-art risk-controlled policy at the same certified risk.

arXiv:2610.13276 PDF

Beyond the Final Embedding: Retrieving Memories from Token-Level Evidence

Han Luo, Zhan Zhuang, Xu He, Jujia Zhao, Yu Li, Suqin Yuan et al. cross-listed Memory-augmented assistants retrieve past conversations and documents with single-vector embeddings, even though relevance often hinges on a small passage inside a long memory. The proposed method reuses the contextual token representations that the embedding model already computes: each query token retrieves a small set of matching tokens across candidate memories, the match scores are aggregated per memory, and the result is combined with whole-memory similarity for ranking, with no extra training required; an optional low-rank query adapter can be trained while the encoder and stored memory vectors stay frozen. The training-free variant improves retrieval across eight benchmarks and multiple model families, and with query adaptation the gains reach 13.3 and 14.3 percentage points in average Recall@10 and NDCG@10 on Qwen3-Embedding-4B, translating into higher question-answering accuracy on long-term memory benchmarks.

arXiv:2610.13277 PDF

Do Prompt Optimizers Break Requirements They Are Not Scored On? A Preregistered Study

Sangdo Han, Moonsoo Kim Prompt optimizers maximize one scored metric, but prompts usually also carry unscored requirements such as output format, citations, or abstention, and this preregistered study asks whether optimization silently violates them. The authors run GEPA, MIPROv2, and TextGrad on HotpotQA, SQuAD 2.0, and IFEval and count per-item violations with deterministic checks; by the preregistered criterion all three optimizers regress, most sharply on HotpotQA, where citation validity drops 10.7 to 41.6 points while exact match rises 21 to 34 points. A post-hoc decomposition attributes most of the damage to truncation: optimized prompts elicit reasoning before the required JSON object, so under the 200-token cap 13.9% to 47.0% of responses end before any answer appears, while among parseable outputs net regression is within 3 points of zero and a 1,024-token cap eliminates the extraction failures. The side effect is narrower than hypothesized: optimized prompts are more accurate but break an explicit output contract that a lenient target metric fails to expose.

arXiv:2610.13287 PDF

PRIME: Pre-Retrieval Index Modeling Enhancement for Structured Document Retrieval-Real

Cheuk Hei Chung, Li Liu cross-listed Retrieval research for LLM agents has concentrated on query-side retrievers and generators while reducing the corpus side to chunking, leaving index quality largely unexplored. PRIME (Pre-Retrieval Index Modeling and Enhancement) optimizes the index with two tracks: an LLM-driven Semantic Restructuring track that rewrites long documents into structured, information-dense sections, and a deterministic Metadata Synthesis plus Hierarchical Retrieval track that extracts semantic anchors and heading structure for coarse-to-fine section retrieval. Deterministic metadata matches structure-aware chunking's retrieval quality while cutting embedding computation 45.8 times and index vector count 22.4 times, Semantic Restructuring adds a further 13.7% Answer-F1 in full deployment, and across QASPER, BRIGHT, and NFCorpus the authors identify three regimes governing whether index-side work helps or hurts, noting that the effects vanish under a strong reranker.

arXiv:2610.13296 PDF

Agreement Without Alignment: Label-Free Rubric Clarification Increases LLM-Judge Consensus but Not Agreement with Humans

Sangdo Han, Moonsoo Kim When panels of LLM judges disagree, the rubric is often blamed, so the authors test a label-free procedure that decomposes each rubric into sub-criteria, ranks them by cross-judge disagreement on a development split, and has an editor model rewrite only the most contested clauses, with human labels used solely for evaluation. In a preregistered protocol with five judges from three providers on SummEval, the HealthBench meta-evaluation, and USR Topical-Chat, totaling 129,770 judge requests, the procedure raised inter-judge Krippendorff's alpha on SummEval by 0.069 over the original rubric and by 0.074 over a control that edited an equal number of random clauses. Agreement with expert human ratings showed no detectable change (Kendall tau difference of -0.007), and on the other two datasets neither consensus nor human agreement met the preregistered criteria, so disagreement-targeted edits make judges agree with each other more without bringing them closer to human annotators.

arXiv:2610.13303 PDF

Host-Aware Web Crawling for LLM Pretraining Data

Malte Ostendorff Web crawls supply much of the pretraining data for large language models (LLMs), yet quality filters discard a large share of what is fetched, raising the question of how a crawler should allocate its budget across and within hosts. Combining document-level quality classifier scores, web-graph centrality metrics, and Common Crawl's per-host coverage, the authors find that about half of the variation in quality scores lies between hosts, that a few sampled pages predict a host's remaining quality well, and that the harmonic centrality Common Crawl uses to guide allocation offers limited gains over random host ordering and sometimes does worse. Sampling a few pages per host and then crawling deeper into hosts whose samples score highly collects several times more high-quality pages per fetch than uniform depth or centrality-based allocation, and the authors estimate that tens of millions of high-quality pages remain uncrawled in each crawl examined, concentrated on a small share of hosts.

arXiv:2610.13321 PDF

Taking Tokenization off the Host for Agentic LLM Serving

Mohammad Siavashi, Pascal Schwinte, Ali Banaei, Gerald Q. Maguire Jr., Marco Chiesa, Dejan Kosti\'c cross-listed In agentic serving, prefix caching skips most of the prefill, but every request still has to be tokenized in full on the CPU before the cache can be consulted, so tokenization becomes the dominant share of time-to-first-token (TTFT) and a single frontend can only feed a limited number of GPUs. GraphTok moves the whole encoding pipeline, from pre-tokenization through byte-pair encoding (BPE) merging to token-ID assembly, onto the GPU as one replayable CUDA graph, hands short inputs to a CPU engine where the GPU cannot win, and plugs into unmodified vLLM and NVIDIA Dynamo. It reproduces the reference token IDs bit-for-bit for GPT-2, Llama 3, Qwen 2.5, DeepSeek-V3, and Gemma 3, runs 4.7-6.0x faster than the gigatoken CPU tokenizer on Llama 3, and lowers mean TTFT by 2.1-2.4x in vLLM with a cached prefix. Within a 2 s P99 target, one Dynamo frontend sustains 518 requests/s of 121k-token traffic, against 296 with gigatoken and 18.5 with the default tokenizer.

arXiv:2610.13325 PDF

DuoRate: Two-Rate Per-Expert Precision Allocation for Mixture-of-Experts Serving

Mengqi Guo, Zhe Wei, Yuan Yuan, Jiunn Bin Lim, Boyi Pan, Michael Bi Mi Serving a Mixture-of-Experts (MoE) model pays for expert precision twice: the resident footprint counts every expert the router can reach, while each batch-one decode step reads only the experts it selects, and published per-expert quantization allocators constrain only the former. The authors call this the store-read gap and note that the released BitsMoE checkpoint stores 2.54 bits per weight (bpw) yet reads 7.50 bpw during decode. DuoRate stores a single uniform tier of a prefix-readable weight stream plus a precomputed distortion tensor, then at deployment profiles a workload's own prefill without labels and solves a Lagrangian rate-distortion allocation under the decode cost, which needs one dot product per expert and a bisection and from which routing frequency provably cancels. On Qwen3-30B-A3B, reading 2.3 bpw beats every 2-bit-class baseline by 6.3 to 47.0 points of average accuracy, and at 2.9 bpw it matches 4-bit formats that read 1.4-1.6 bpw more, trailing BF16 by only 0.25 points on GSM8K.

arXiv:2610.13328 PDF

Understanding Mamba-3 Through Oscillatory Dynamics

Shutian Liang, Yue Song Mamba-3 adds input-dependent rotations to its recurrent state updates, but what those rotations contribute beyond retaining state has been unclear. The authors interpret the rotations as underdamped second-order dynamics, giving a framework for studying the role of oscillation in sequence modeling. In controlled comparisons, underdamped dynamics outperform non-oscillatory damping regimes on state-tracking tasks, and language-model pretraining shows gains in long-context prediction and length generalization. Interventions on pretrained models establish that the learned phase matters, and state analysis suggests the rotation regulates how retained information is combined with incoming updates.

arXiv:2610.13341 PDF

Overcoming Data Imbalances with Expert Selection Biases

Ayush K. Tarun, Clara Meister, Antoine Bosselut Mixture-of-Experts (MoE) models can in principle let experts specialize by domain, language, or modality during pretraining, but when training data is imbalanced across such categories, minority categories end up squeezed into a small slice of the expert pool or spread thinly across it, limiting their effective capacity. The authors add category-conditioned expert-selection biases that steer tokens from each category toward designated expert subsets, inducing soft category-aligned specialization while still allowing cross-category sharing, and in particular letting them over-allocate expert capacity to data-scarce categories. Gains appear across modalities, low-resource languages, and code in models up to 10.6B parameters, equivalent to 18.5% more baseline training tokens at the smallest scale and 2.7% at the largest. Because the intervention only changes routing biases, inference proceeds as in a standard MoE.

arXiv:2610.13370 PDF

From Structure to Validity: Mechanism-Guided Generation and Refinement of LLM Rubrics

Feifei Qian, Xuan Chen, Lixin Cui, Chaoyang Mei, Fanlin Meng, Shuyao Xiao et al. Rubrics are widely used to evaluate and optimize language models on open-ended tasks, but most work focuses on generating better rubrics without understanding what makes a rubric design effective. The authors run a systematic mechanistic analysis across criterion organization, criterion operationalization, and score aggregation, finding that structured organization yields more comprehensive criteria, semantic specification alone does not guarantee sound criterion-level judgments, and sophisticated score aggregation adds little. Guided by this, they introduce Adaptive Rubric Tree Search with Hybrid refinement (ARTS-Hybrid), which pairs adaptive tree construction with semantic-first, response-calibrated refinement. Across benchmarks it improves both end-to-end evaluation performance and intrinsic rubric quality, and the rubrics it generates also serve as effective signals for model optimization.

arXiv:2610.13375 PDF

Branching On-Policy Distillation

Dongyang Chen, Shudong Liu, Shubo Li, Duoxun Tang, Jinwei Liang, Zehuan Wu et al. On-policy distillation (OPD) supervises a student on its own generated trajectories, but each trajectory realizes only one continuation at every decision point, so a finite sampling budget covers few alternatives. Branching On-Policy Distillation (BranchOPD) selects informative decision points using student uncertainty and teacher-student disagreement, resamples student continuations from shared prefixes, and applies token-level teacher supervision to both main trajectories and branches, operating at token-level decisions in mathematics and within-turn decisions in agent tasks; a protocol-boundary repair (PBR) component adds targeted supervision under protocol constraints to curb repetition and wasted generation. Shared-prefix reuse cuts student-generated token costs by about 19% on ALFWorld and 31% on WebShop relative to seven independent trajectories, and BranchOPD beats both vanilla OPD and seven-rollout multi-rollout OPD on every benchmark, with gains over vanilla OPD of 18.55 points in mean mathematics accuracy over 16 sampled answers and 13.8 points in ALFWorld in- and out-of-distribution success; ablations show branching at uncertain decisions expands subsequent behavior coverage and outperforms random or turn-initial branching.

arXiv:2610.13481 PDF

Not All Correct Fixes Run Alike: An Empirical Study of Patch Performance in Repository-Level Issue Resolution

Han Peng, Xing Hu, Feifei Niu, Huan Li, Xin Xia cross-listed Benchmarks such as SWE-bench grade a patch only on whether it passes correctness tests, so two passing patches with very different runtime behavior score identically. The authors repurpose the built-in FAIL_TO_PASS (F2P) tests as performance workloads, running Claude Opus 5, GPT 5.6, DeepSeek V4 Pro, and Qwen 3.8 Max through SWE-agent on SWE-bench and SWE-bench-Live to collect 262 instances with at least two distinct correct solutions, then compare instruction counts and wall-clock time with Welch ANOVA. The method detects performance differences in almost half the instances, and patch runtime performance is independent of issue resolution rate, with model rankings nearly reversed; a case study shows performance analysis exposing architectural regressions that correctness tests missed.

arXiv:2610.13563 PDF

Measuring Artifact Reliance in Machine-Generated Text Detectors

Anastasia Voznyuk, Andrey Grabovoy Machine-generated text detectors score near-perfectly on scientific-literature benchmarks, which the authors trace to how the documents were assembled rather than to real detection ability. Across ten scientific subsets from six popular resources, a logistic regression on 27 superficial features reaches 0.833 to 0.997 macro-F1 and beats every evaluated detector on the three arXiv subsets, with the dominant cue being a PDF extraction artifact: human texts keep their line breaks while generated texts arrive as a single block. The proposed format counterfactuals show that re-wrapping a machine-written abstract without changing a word flips 20 to 27 percent of a public detector's correct decisions, and normalizing formatting before training shrinks the generalization gap and roughly doubles out-of-domain macro-F1 on format-clean benchmarks.

arXiv:2610.13583 PDF

zip2zip++: Towards Inference-Time Token Compression with Minimal Quality Loss

Saibo Geng, Andrea Gentilini, Xinxian Ma, Nathan Ranchin, Yunzhen Yao, Robert West Dynamic vocabularies built on the fly with Lempel-Ziv-Welch (LZW) compression let a language model cover the same text with fewer tokens, but earlier work found the saving cost reasoning accuracy and perplexity. The authors attribute the gap to four fixable issues: a shared embedding cannot serve both input and output roles for merged tokens, naive RoPE ignores that a merged token spans several positions, merging digits is disproportionately harmful, and dynamic vocabularies create segmentation ambiguity that confounds evaluation. Decoupled input and output dynamic embeddings, size-aware positional encoding, and a modified codec with selective merges and multi-view segmentation improve perplexity on every corpus tested across four models and raise math accuracy, closing roughly half the quality gap between zip2zip and the uncompressed baseline at a similar compression ratio, all without modifying the pretrained model.

arXiv:2610.13605 PDF

Multi-Stream Attention Residuals

Xiaohan Qin, Xiangdong Zhang, Tuo Dai, Sunan Zou, Huaijin Wu, Yebin Yang et al. Attention Residuals (AttnRes) let a layer retrieve from earlier layers through content-dependent softmax attention along the depth axis, but only as a single stream. MSAR generalizes this to multi-head depth attention, deriving stream diversity from multi-scale causal convolutions over the sequence so that different heads retrieve historical representations through distinct contextual views, then recombining them with lightweight adaptive mixing; because single-stream sources are expanded on demand, no persistent multi-stream storage or communication is needed. At 18B and 28B mixture-of-experts scales it outperforms standard residuals, manifold-constrained Hyper-Connections (mHC), and AttnRes on all 12 downstream benchmarks, and scaling-law experiments show an iso-loss compute speedup of 1.58x over the standard residual baseline, versus 1.35x for AttnRes and 1.30x for mHC.

arXiv:2610.13630 PDF

Adaptive DHiSparse: Speculative Decoding over a Load-Adaptive KV Hierarchy

Naji Alazhar, Wassel Alazhar Long-context LLM serving exhausts KV cache memory well before the GPU reaches its useful throughput. HiSparse already keeps history in host memory and fetches only selected entries, but it uses a single cache budget regardless of load and does not support draft-model speculative decoding such as DSpark and DFlash2. Adaptive DHiSparse runs speculative decoding over a hierarchical sparse KV cache and adapts residency to load, protecting the required working set, keeping extra history while GPU memory is spare, and recovering the rest through exact fused retrieval, with the drafters' proposal and acceptance rules left untouched. At 100K context and concurrency 128, GLM-5.3 with DFlash2 reaches 793 output tokens per second per decode GPU, 1.45x HiSparse, sustains up to 4x the concurrency under the same latency bounds, and on DeepSeek-V4.1-Flash at 500K context and concurrency 512 delivers 1.85x the output of resident serving with DSpark.

arXiv:2610.13708 PDF

Orthros: Phase-Aware Heterogeneous Attention for Efficient Transformers

Yuhong Chou, Zehao Liu, Yuqi Pan, Qian Liu, Xianwei Chen, Zhongzhu Zhou et al. Linear and sliding-window attention reduce Transformer cost from quadratic to linear but weaken in-context recall, so practical models interleave full attention layers and the quadratic bottleneck persists for long contexts. Orthros attention instead decouples by phase: a single module switches its computation pattern between linear-complexity prefilling and full-attention decoding, which gives lower asymptotic cost than a vanilla Transformer in the common regime of heavy prefill and light decode. Experiments show competitive language modeling across model sizes and architectural comparisons, plus a data-efficient Transformer-to-Orthros conversion that even surpasses the original Transformer. It achieves a 21.9x prefilling speedup over Transformers with end-to-end inference gains, near-perfect 128K needle-in-a-haystack results, and Transformer-level performance on real-world in-context recall tasks.

arXiv:2610.13735 PDF

From Pixels to Privacy-Sensitive Features: An Information-Theoretic Framework for Reconstruction Bounds

Henry Bourne, Robert F. Allison cross-listed Accelerators with native 4-bit support only deliver their full throughput when both weights and activations are quantized, but activation quantization is harder than weight quantization and typically requires calibration data. DualQuant is a calibration-free post-training method that directly minimizes weight-reconstruction mean squared error (MSE) and obtains per-channel activation rescalings as a by-product, based on the observation that scaling factors from weight reconstruction also work well for activations. The non-differentiable problem is solved by variable splitting and block coordinate descent with closed-form updates, and the method supports the full MX family and NVFP4 formats. Across Llama and Qwen models on WikiText-2, C4, and zero-shot tasks it matches or beats calibration-based SmoothQuant and substantially outperforms calibration-free baselines, and a real packed W4A4 path for Llama-3.1-8B on an RTX 4090 delivers up to 2.61x memory reduction and 2.46x throughput speedup over BF16.

arXiv:2610.13768 PDF

ReplaySSM: Cache SSM Inputs, Not State

Ze-Wei Liou, Harris Nover, Amey Naik, Andrii Skliar, Igor Shovkun, Ido Shahaf et al. State-space models (SSMs) keep a fixed-size recurrent state, but reading and writing it on every decoding step creates heavy state I/O with low arithmetic intensity, leaves speculative decoding without a cheap rollback, and feeds quantization error back into later updates. ReplaySSM unrolls the recurrence so the state is represented as a periodic checkpoint plus a buffer of recent SSM inputs, which avoids state writes on most steps, makes speculative rollback a buffer-pointer operation, and lets a low-precision checkpoint serve output computation while a higher-precision one serves recurrent updates. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it speeds up standard decoding by 1.20x to 1.47x over a tuned vLLM baseline at batch size 256 across four hybrid models and gives about 1.8x the throughput of vLLM's speculative decoding on two 120B-class models. Separate state precisions also reduce output KL divergence and accuracy losses compared with a shared low-precision checkpoint.

arXiv:2610.13778 PDF

Tracing and Taming Hallucinated Terms via Subsequence-Association Filtering and Breadth-Aware Selection for LLM Query Expansion

Amit Kumar Jaiswal, Sagar Uprety, Prayag Tiwari cross-listed Large language models are now a standard source of query expansion terms, but they inject non-existent entities and off-topic terms that add ranking noise, and existing remedies filter them with agreement heuristics rather than any account of why a term appeared. SAFE-QE is training-free and retriever-agnostic: it scores each candidate term by corpus-conditioned cross-context reproducibility, down-weighting terms that recur under randomized query contexts yet lack support in the pseudo-relevant set, then casts final selection as constrained coverage maximization to keep a set that is broad but consistent under resampling. A companion metric, the Hallucinated Expansion Term Rate, counts injected terms no relevant document supports, and on TREC and BEIR collections the method matches generative and corpus-steered expansion on nDCG@10 while injecting 4.3 unsupported terms per query against 27.8 for Query2doc and 57.5 for HyDE.

arXiv:2610.13810 PDF

QTestBench: A Test Framework for Evaluating LLM-Generated Quantum Code

Meriem Terki, Nouhaila Innan, Minghao Shao, Muhammad Kashif, Alberto Marchisio, Muhammad Shafique Whether language models can write trustworthy unit tests for quantum software is unclear, because quantum programs are probabilistic, depend on numerical tolerances, and are judged by domain-specific checkers, so a well-formed test can still disagree with the benchmark's verdict. QTestBench prompts models for syntactic, semantic, and behavioral tests, injects candidate solutions at runtime, and compares each test's verdict against an independent benchmark checker under three validation cohorts, never treating the tests as ground truth. Across five models on eight PennyLane QHack challenges and ten Qiskit HumanEval tasks, generated suites almost always execute but discriminate poorly: the best generator reaches a test quality of 0.71 on QHack and only 0.34 on Qiskit HumanEval. Oracle-verified synthetic mutants stabilize bug-detection estimates on the tasks where no naturally faulty solutions exist.

arXiv:2610.13811 PDF

Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection

Juan Terven, Diana-Margarita C\'ordova-Esparza, Ruben Gomez, Carlos Miranda Rocha, Julio-Alejandro Romero-Gonz\'alez, Edgar Chavez-Urbiola A model that answers only when it should can be built by training the model itself or by training a separate component to vet its answers, and both routes are compared on Qwen3-14B for factual questions under a score that rewards correct answers, penalizes mistakes, and allows abstention. Across three training seeds and two held-out confidence thresholds, a learned answer selector beats the model's own calibrated confidence filter on threshold-averaged utility, and on TriviaQA coverage rises from 20.2% to 30.9% while the pooled error rate among scored answers falls from 23.0% to 14.5% under exact-match grading, a result that survives a partial human-reviewed audit. The reinforcement-learning variants tested show no consistent gains, and on NQ-Open the selector beats the filter yet still scores worse than always abstaining, so dependable abstention across datasets stays unresolved.

arXiv:2610.13821 PDF

Same Query, Different Answer: Run-to-Run Variance in Semantic Query Processing

Oktie Hassanzadeh, Dharmashankar Subramanian cross-listed Semantic query engines ask a language model to judge individual rows, and benchmarks rank those engines from a handful of executions at temperature 0. Re-running the text and tabular workloads of SemBench, LRO-Bench, and UDA-Bench through their own harnesses while logging all 1.2 million model calls shows that at temperature 0 between 0.07% and 10.2% of rows flip their operator output across five executions, and up to 24.0% on one task. Query plans amplify this exponentially in how many rows an answer exposes: on UDA-Bench a few percent of extracted values change but roughly two thirds of the answers change, and 21% of SemBench's published text results shift in quality between runs, including for engines whose plans are deterministic. Sixteen of 28 tests of LRO-Bench's implementation recommendations are not settled by the executions the benchmark runs, voting damps the variance without eliminating it, and the authors derive reporting rules for benchmarks and design rules for engines.

arXiv:2610.13823 PDF

On the Design Space of Continuous Diffusion Language Models

Oscar Davis, Daan Roos, Alice Bizeul, Amitis Shidani, Victor Turrisi, Louis B\'ethune et al. Continuous diffusion language models now rival discrete diffusion, yet which training choices actually matter remains poorly mapped compared with images or video. Two levers are dissected: self-conditioning turns out to fix a train/inference mismatch by exposing the model to inference-like states that training rarely visits, and the same benefit is recovered without the extra forward pass by a two-level interpolant called DoubleCat, which more than triples TinyGSM accuracy. Adding discrete reveals gives SemiCat, which hits 49.94% on TinyGSM with standard hyperparameters and a tuned sampler, matching discrete diffusion, and at 1.7B parameters beats discrete diffusion on four of six likelihood-based evaluations and one-hot flows on unconditional generation. Cross-entropy training is also shown to push embeddings toward a geometry whose decoding error rate closely matches the one-hot case.

arXiv:2610.13834 PDF

Specialists Learn Faster: Knowledge Organization in Pre-training Shapes Knowledge Update Efficiency

Sheng Zhao, Yuhao Zhu Work on keeping a language model's knowledge current mostly optimizes the update procedure; the question here is how knowledge organization during pre-training affects later update cost. Training on chronologically ordered facts yields specialists that excel on the most recently seen period, while the same data shuffled yields generalists that perform evenly across periods, and specialists need far fewer update steps than generalists to acquire new knowledge. A law derived from first principles relates update cost to optimization dynamics and closely predicts the measured results, motivating a data-ordering strategy that preserves stable historical knowledge while keeping current knowledge cheap to refresh, with the effect also appearing in real pre-trained models, existing knowledge-update tools, and other orderings.

arXiv:2610.13835 PDF

Can Language Models Play the Wikipedia Game?

Alexander Hickey, Joanna Knight, James Bishop Language models are increasingly used to retrieve information from document collections too large to index, which is hard when specialist linguistic or scientific knowledge is needed. The Wikipedia Game, in which a player must reach a target article by following only article links, serves as an analogous navigation task for comparing sentence embedding models and large language models across five Wikipedia languages and five academic-literature datasets, with very large local graphs of Wikipedia pages and OpenAlex paper citations built so the study needs no third-party APIs. Against human and TF-IDF baselines, no model significantly outperformed humans, though the best models (GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B) reached comparable success rates. Performance and model rankings varied across languages and topics, which matters for practitioners choosing among these techniques.

arXiv:2610.13948 PDF

16 more specialized papers

An Input-Dependent Readout Phase for Resonant Sequence Models Shivesh Rane
MolP2SBench: Do LLMs Really Reason from Molecular Properties to Structure? Yuhang Xie, Zhiyi Shi, Zeyi Zhang et al.
Self-Preservation in Language Models Dongkyu Cho, Rumi Chunara, Ali Behrouz
perturber: A Stateless Service for Semantics-Preserving Text Perturbations Khoi-Nguyen Tran
Byzantine-Resilient Federated Large Language Model via Hypergraph Message Passing Kai Li, Hanlin Cai, Zihao Liu et al.
Static Gradient Attribution Underperforms a Density-Matched Random Mask on Loss-Based Forgetting Within LoRA B-Matrix Yu Sun, Junwei Zhou, Zuodong Xiang et al.
Time Delay System for Sequence Modeling Youshan Zhang, Jialu Li
Personalized Federated Direct Preference Optimization via LoRA Adapters Zhanhong Jiang, Soumik Sarkar
Language Models Have Interpretable Narrative Mechanisms David Yiyang Liu, Aditya Joshi, Paul Dawson et al.
RAMP: Recurrent Adaptive Mixed Precision for Efficient Prefill in Recurrent-Depth Language Models Ke Wan, Chen Chen
Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation Shi-Yu Tian, Yong Liu, Ximan Liu et al.
BLiMPPI: Benchmark of Linguistic Minimal Pairs for Polarity Items Arzu Burcu G\"uven, Anna Rogers, Leonie Weissweiler et al.
Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels Weiqi Sun, Qianyue Wu, Wenjia Wang et al.
Not Just an Evaluation: A Case Study of Discourse Particle Sensitivity in Language Models William Sheffield, Jiahui Chen, Amy Zhang et al.
Task Context Shapes Component Roles in Language Models Sovesh Mohapatra, Dani S Bassett
Structural Automata for Neural Semantic Parsing Xiao Zhang

Other 60

Equal-Budget Monte Carlo Tree Search versus Randomized-Restart List Scheduling for DAG Workflows: A Controlled Equal-Budget Computational Study

Yuyu Zhang Monte Carlo Tree Search (MCTS) is frequently proposed for scheduling directed-acyclic-graph (DAG) workflows, but studies rarely give a simpler method the same search effort. Under a preregistered protocol with a locked analysis plan and an independent validity checker, the authors compare an MCTS scheduler against a randomized-restart list scheduler with no search tree, plus a textbook-UCT MCTS arm, all with an identical budget of built and evaluated schedules. At the main budget of 2048 schedules, MCTS produced makespans 2.88% worse than the restarts, a deficit attributed to a selection rule that locks onto the first-expanded child while restarts draw gains from the tails, though the authors note the causal link to makespan remains an inference; swapping in textbook UCT removes most of the gap without reversing its sign. Restarts win on sparse irregular workflows, no difference is detectable on regular layered ones, and the claims are confined to a synthetic domain of 8 to 32 tasks, 2 to 8 processors, and budgets up to 2048.

arXiv:2610.12542 PDF

Deploying RDMA for Large-Scale Inference and Training Workloads

Yibo Huang, Yi Cui, Yukai Huang, Huaping Zhou, Jiancong Tong, Zhaoyu Wu et al. Remote direct memory access (RDMA) can cut communication overhead in distributed machine learning, but variable-size messages, concurrent connections, bursty incasts, and service failures mean that replacing TCP with hardware offload alone is not enough in production. ByteExpress is a software communication layer over hardware RDMA deployed in ByteDance's parameter-server-based inference and training systems, combining an RDMA-aware allocator, zero-copy transfers, QP multiplexing, receive-credit flow control, a software send window, and RPC-assisted recovery, and the authors distill three lessons: buffer management must span the application and transport layers, scalability needs coordinated management of connections, receive capacity, and outstanding data, and availability requires recovery above the transport layer. In production inference it reduces mean and P99 model parameter fetch latency by 35.7% and 43.7% and total CPU usage by 10.69% versus kernel TCP, and in training it cuts P99 parameter push latency by up to 74.0% with no PFC pauses observed at runtime.

arXiv:2610.12815 PDF

Fear Is Not Always Fear: Variation in Emotion Label Equivalence Across Datasets

Emily \"Ohman, Anna Koufakou Emotion classifiers are routinely reused across corpora on the assumption that labels such as fear or surprise mean the same thing everywhere. The authors test this for five labels shared by GoEmotions, XED, and BRIGHTER, comparing classifiers on identical inputs and under cross-corpus transfer. Scale disagreements are large but mostly fixable, with one fear classifier flagging 71 percent of texts versus 8 percent for another while agreeing on ranking, and a fitted intercept removes most of the mismatch, yet transferred models lose 55 percent of their native discrimination in aggregate and for surprise the resources disagree about which texts express the emotion at all. On two external tweet collections, classifiers trained on different resources give opposite answers about which collection shows more surprise, even after rank normalization.

arXiv:2610.12841 PDF

Neural networks are more modular than single neurons suggest

Shreya Saha, Benjamin Bergen, Meenakshi Khosla Because individual neurons in trained networks usually participate in several functions, the networks are often described as non-modular. The authors define a nested hierarchy of modularity, from unit-level (disjoint neurons), to orthogonal (a rotation reveals independent subspaces), to linear (a bounded linear map separates computations), and build a unified framework that learns decompositions at each level and tests them causally through selective ablation. Across vision models, language models, and multi-task recurrent neural networks (RNNs), networks lacking unit-level modularity nonetheless show clear subspace-level modularity, with content versus style and syntax versus semantics dissociating under ablation despite sharing neurons, and the recovered subspaces align differentially with higher versus lower visual cortex and with language cortex. In RNNs previously reported as non-modular, the decomposition recovers sharp double dissociations between cognitive tasks.

arXiv:2610.12842 PDF

When Generative Models Act as Simulators

Sevan Brodjian, Michael Hobley, Pietro Perona A simulator exposes an editable state, whereas a generative model trained only on observations has no built-in way to edit the world it has implicitly learned. The authors study editability, asking whether changing a model's latent representation yields the next-step outputs that editing a simulator's state would produce, and compare two post-hoc mappings learned from observations paired with simulator states: steering along linear probes of the state, and a directly learned inverse map from state to latent. In the board game Othello and in Rayworld, a continuous world of moving discs seen through partial renderings, the state is decodable in every variant, yet which mapping succeeds depends strongly on the environment. In Othello editability degrades smoothly as the rules are simplified and small rule changes have large effects, while in Rayworld probes fail to steer along the continuous state but an inverse map can still set the model from it, so a model can be driven by a simulator it may never have learned.

arXiv:2610.12847 PDF

When Solvers Get Moody: Quantifying and Predicting Performance Variability in Mixed Integer Programming

Mourchid Adegbindin, Brenda Dietrich, Jamol Pender, Andrea Lodi cross-listed Mixed-integer programming (MIP) solvers can produce very different runtimes under seemingly neutral perturbations such as random seeds, which complicates benchmarking and the interpretation of computational studies. The Weighted Variability Score (WVS) aggregates dispersion in normalized runtime, final optimality gap, LP iterations, and branch-and-bound nodes across repeated runs, and is measured on a large subset of MIPLIB 2017 with 100 randomized runs per instance under a four-hour limit. The resulting distribution is right-skewed with a heavy upper tail, and a tree-based predictor using static descriptors and root-level features predicts WVS with a mean absolute error of 0.058 and Spearman rank correlation of 0.65 on a sealed test split, with root degeneracy, matrix structure, coefficient ranges, and presolve reductions as the main drivers. As a side result, two SCIP runs proved optimality for the previously open instance 'neos-5045105-creuse'.

arXiv:2610.12876 PDF

HART: Hybrid Attention-RWKV Time series Foundation Model

Qingwen Lin, Boyan Xu, Shifeng Xie, Zhifeng Hao, Keli Zhang, Ruichu Cai Non-autoregressive multi-step forecasting (NAF) lets time series foundation models predict a whole horizon in one forward pass, but RWKV-7 degrades at long horizons in this setup because future patches are represented as missing inputs and the recurrent state must evolve without new observations. An analysis of the generalized delta rule identifies Placeholder Idling Decay (PID), where repeated state transitions attenuate historical information when placeholder-driven writes do not compensate, supported by layer-wise measurements of reduced state-injection norms. HART alternates attention and RWKV blocks so the model retains direct access to historical tokens alongside recurrent modeling, improving over pure RWKV at medium and long horizons on GIFT-Eval and TIME-Benchmark and beating a comparably sized pure Transformer in aggregate. With 36M parameters it matches the long-horizon MASE of Toto-2.0-1B on GIFT-Eval (0.752 vs 0.754) using roughly 1/28 of the parameters.

arXiv:2610.12883 PDF

Towards Easier-to-Tune Training Algorithms: New Results from AlgoPerf

Priya Kasimbeg (Alice), Ahmed Khaled (Alice), Aahlad Chandrabhatta (Alice), Yifan Wang (Alice), Andy Liu (Alice), Niccol\`o Ajroldi (Alice) et al. Because hyperparameter tuning and optimizer-model co-design confound optimizer comparisons, the updated AlgoPerf V1 benchmark scores training algorithms on wall-clock time-to-target across nine workloads on fixed hardware, and this report focuses on the self-tuning ruleset in which a submission must use a single configuration for every workload. Of 13 PyTorch and JAX submissions, Schedule-Free AdamW ranked first. An analysis of step efficiency versus wall-clock cost shows that matrix-preconditioned and two-loop methods are sample-efficient but need more aggressive computational optimization to compete on runtime. Hyperparameter saliency and retention analyses find that learning rate and momentum remain highly sensitive across domains and algorithms, pointing to the need for methods that are more robust to tuning and generalize to new tasks.

arXiv:2610.13036 PDF

How Much Computation Does Every Forecast Need? Structured Block Skipping in Two Time-Series Foundation Models

Max Bourgeat Time-series foundation models apply the same depth to every input even though forecasting windows differ in structure and difficulty, so the authors test whether whole Transformer blocks can be bypassed at inference in the pretrained Chronos-Bolt and TimesFM 2.5, using bypasses that reproduce the dense model exactly when all blocks are enabled. With masks selected only on calibration windows and evaluated on 1,088 held-out windows from ETTh1, Traffic, Weather, and Exchange, keeping 7 of 11 candidate blocks in Chronos-Bolt and 16 of 20 in TimesFM leaves mean relative forecast quality little changed while cutting median A100 batch-1 latency by 11.2% and 13.4%, whereas more aggressive skipping in TimesFM substantially harms Traffic and Weather. Different equal-budget masks win on different windows and a retrospective oracle beats the best static mask on every dataset, but dataset-specific masks transfer inconsistently and Chronos representations do not support reliable routing for unseen domains, so the work establishes cross-architecture structured redundancy and input-dependent opportunity rather than a deployable dynamic inference method.

arXiv:2610.13208 PDF

ReSFlow: Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation

Xiaohuan Jia, Renzhe Xu, Xiao Wang, Jiayun Wu, Shaohua Fan Flow-based generative models for discrete data learn continuous vector fields on statistical manifolds, but because the learned flows are not constrained to follow geodesics they need many numerical integration steps at inference. Rectified Statistical Flow (ReSFlow) adapts the straight-and-fast idea of Rectified Flow from Euclidean space to the statistical manifold, explicitly learning trajectories that respect its Riemannian geometry, with a theoretical guarantee that rectified paths between prior and target approximate geodesics and thereby reduce transport cost. The authors claim one-step generation from pure noise to discrete data with superior quality across image, text, and biological tasks, and release code.

arXiv:2610.13225 PDF

Tabular Foundation Models Are Effectively Shallow

Irene Cannistraci, Julia E. Vogt In-context-learning tabular foundation models (TFMs) solve new tasks without training but stack many transformer blocks, while gradient-boosted decision trees (GBDTs) stay competitive on CPUs, raising the question of whether tabular tasks need such depth at all. The authors introduce a closed-form, label-free affine translator that replaces a contiguous window of up to 96% of a model's blocks while leaving the rest untouched, and evaluate eight frontier TFMs on 171 classification datasets. Replacing half of a model's blocks costs at most 0.002 median AUROC across all architectures, and one half-depth model still outperforms every other full-depth TFM and tuned GBDTs. Families are overparameterized in different ways, with TabPFN-3 relying on its encoders and keeping its performance whichever single block is retained, other single-axis models needing only their first block, and dual-axis models concentrating computation in one critical block, and removing depth saves up to 1.5B parameters and speeds GPU inference by up to 6 times at a median cost of 0.03 AUROC, while a single-block TabPFN-3 serves cached queries 2.4 times faster with 6 times less memory at negligible cost.

arXiv:2610.13253 PDF

One Thing Leads to Another: The Essentials of Time Series Causal Discovery

Kurt Butler, Damian Machlanski, Michael P. J. Camilleri, Zhihua Liu, Panagiotis Dimitrakopoulos, Feng Chen et al. cross-listed Time series causal discovery, the detection of causal relations between time series, grew out of signal processing research half a century ago and has since fragmented across statistics, dynamical systems, and machine learning. The tutorial provides a unified overview across those communities, introducing the key principles behind modern discovery methods together with their assumptions and limitations, and explaining how such methods are evaluated, benchmarked, and tuned. It also surveys why causal discovery for time series is becoming central to downstream work in agriculture and Earth science, healthcare, and artificial intelligence systems, and it is pitched partly as an invitation to signal processing researchers to take up causal problems. A companion repository collects code packages and related resources.

arXiv:2610.13430 PDF

A small cognitive core explains human behavior at scale

Marcel Binz, Eric Schulz Psychology has accumulated many separately developed principles of cognition without testing how much they explain in combination. COntextual Retrieval and Encoding (CORE) assembles a small set of them into a single model: a hierarchy of episodic memories into which prediction errors are written, with context-dependent encoding, retrieval, and forgetting. Fitted jointly to 7,036,442 choices from 137 experiments, it predicts human choices better than black-box models, reproduces 75 percent of reported behavioral effects in simulation, and exhibits latent signatures of Bayesian learning and event segmentation.

arXiv:2610.13603 PDF

SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data

Bastian Sch\"afer, Dominika Matus, Lennart Purucker, Ricardo Knauer, Erik Rodner, Frank Hutter et al. Practitioners choosing a feature selection method face benchmarks built on outdated models, ad hoc protocols, or synthetic data that does not reflect real problems. SelectArena compares 15 widely used feature selection methods on 71 manually curated real-world tabular datasets spanning classification and regression, with standardized protocols and state-of-the-art tabular foundation models as downstream learners, measuring predictive performance, runtime, stability, and validity. Tree-based CART and RFImportance together with F-Test are Pareto-optimal on performance versus runtime, with F-Test also leading on stability and validity and emerging as the most balanced choice. Search-based methods are slow, unstable, and low-performing, and the benchmark is released as an extensible platform.

arXiv:2610.13644 PDF

TabReliBench: Evaluating Tabular Foundation Models Beyond Predictive Accuracy

Tianqi Zhao, Jia Guo, Yan Shuo Tan, Qiong Zhang Tabular foundation models (TFMs) are prized for predictive accuracy, but the reliability of the posterior predictive distributions (PPDs) many of them provide is understudied. TabReliBench is a large classification and regression benchmark that assesses point prediction, distributional quality, marginal and subgroup calibration, and interval efficiency, with dataset-level analyses. TFM rankings shift between TabArena and this benchmark and across reliability metrics; regression TFMs usually reach near-nominal marginal coverage with narrow intervals, but this does not imply valid conditional coverage, and classification difficulty tends to accompany poor calibration while regression accuracy does not predict calibration. Ensembling improves predictive performance without consistently improving PPD calibration, and conformal post-processing helps little and does not reliably close subgroup gaps, so selecting a TFM requires multidimensional evaluation.

arXiv:2610.13654 PDF

On the Encoding of Calendar Features in Covariate-Informed Time Series Foundation Models

Alexander Kreusel, Maximilian Beichter, Alexandra Nikoltchovska, Benjamin Sch\"afer, Ralf Mikut, Frederik Zahn et al. Time series foundation models (TSFMs) increasingly accept covariates, and calendar features are the most common categorical covariate since they exist for every timestamp, yet how they are encoded is usually treated as an implementation detail. The authors systematically compare seven calendar encodings, three target-agnostic ones (ordinal, one-hot, cyclic) and four target encodings, across four state-of-the-art TSFMs on nine datasets shaped by weekends and holidays, in a zero-shot rolling evaluation. Encoding choice affects accuracy for every model, with MASE differences of up to 9.9% on non-holidays and 83% on holidays within a single model, and the gains cluster around holidays where the target history carries no signal. A SHAP-based attribution analysis explains why some model-encoding pairs degrade performance, and the authors argue that benchmarks that impose a fixed encoding can distort comparisons between covariate-informed TSFMs.

arXiv:2610.13676 PDF

Calibration in One Go: Jointly Learning Accurate and Calibrated Neural Networks

Nektarios Kalampalikis, Pablo M. Olmos, Isabel Valera Neural networks are often poorly calibrated, and post-hoc calibrators fix this with a separate calibration set, extra parameters, and a second optimization and model-selection pass. The authors show that standard post-hoc calibrators can instead be trained end-to-end with the backbone through a bilevel Stackelberg formulation that optimizes accuracy and calibration jointly with cross-entropy alone, adding only one hyperparameter, the calibrator learning rate, and needing no separate calibration set or loss. The framework supports temperature scaling, matrix scaling, and neural calibrators, but using the final classification layer itself as the calibrator yields accurate, well-calibrated models with no extra parameters. The approach outperforms state-of-the-art calibrators with little added complexity and avoids the multiplicative cost of separate calibration components in settings like multi-task learning.

arXiv:2610.13693 PDF

Spanning Diffusion and Flow-based Posterior Sampling via Exact Gaussian Conditioning

Teo Deveney, Lea Bogensperger, Hok Shing Wong Flow-matching and diffusion priors for linear inverse problems are usually built for one trajectory type, and plug-and-play variants implicitly decorrelate the guided trajectory, producing broad priors that struggle when measurements are sparse. beta-Flow introduces a correlation parameter that interpolates across flow, diffusion, and decorrelated priors plus a confidence parameter setting how strictly the path distribution is enforced, with each step needing one denoiser evaluation followed by an exact Kalman update that uses no network gradients. Many existing guided samplers can be read as instances or approximations of this scheme, each silently assuming a different trajectory prior and a different stand-in for the Kalman gain. Performance is at state-of-the-art level on seven inverse problems across four datasets by PSNR and SSIM, and strict high-correlation priors help most when observations are uninformative while low correlation wins when they are informative.

arXiv:2610.13812 PDF

Revisiting Tabular Deep Regression: Objectives and the Limits of Global Early Stopping

Danil Rudenko, Georgii Bulgakov, Nikolay Kartashev, Ivan Rubachev, Daniil Krasylnikov, Artem Babenko The choice of loss function for deep tabular regression is rarely studied systematically, and proposed alternatives to Mean Squared Error (MSE) are hard to compare because they were evaluated on different datasets, backbones, and protocols. A re-evaluation across a broad collection of established tabular datasets and several neural backbones finds that most proposed objectives give limited or inconsistent gains over a well-tuned MSE baseline, while quantile loss delivers consistent aggregate improvements over MSE across the tested backbones. The authors then show that regions of the input space with different target-noise levels reach their best performance at different epochs, so a single global early-stopping checkpoint forces a compromise even when the objective can model heteroskedastic targets. A heteroskedasticity-aware extension of quantile loss using input-dependent label smoothing partially remedies this and improves on plain quantile loss on several backbones.

arXiv:2610.13909 PDF

41 more specialized papers

Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance Marco Zullich, Emily Schiller, Ivan Gentile et al.
Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics Simon Olsson, Frank No\'e, Grant M. Rotskoff et al.
CircuitTree: Composing Circuits into A Syllable Transcriber Jiankun Wei, Ewan Dunbar, Gerald Penn
An all-transistor physical-learning transformer can be trained to perform in-context learning Michael C. Abbott, Benjamin B. Machta
CLUE: A Contrastive Learning Framework for Unsupervised Event Type Induction and Annotation Pratik Maitra, Lauren Kwon, Qi Li
Activating the Annotator: Operationalising Lived Experience in Data Annotation Sheza Munir, Ramaravind Kommiya Mothilal, Syed Ishtiaque Ahmed
Kernelized Parameter Sensitivity Atlases for Amortized Conditional Interpretability Davide Wiest, Ahanaf Hasan Ariq, Abdul Monaf Chowdhury
Coordinate-Free Automatic Differentiation Alex Guerra, Joshua Aduol, Yaqian Tang et al.
Large Language Models, Machine-Assisted Cultural Analytics, and the Agentic Turn Andres Karjus
Trainable Quantum-Native Language Models: Dissipative Memory, Capacity Limits, and Hardware Validation Deniz Yunus G\"o\u{g}\"u\c{s}
Scalable Learning of Determinantal Point Processes with Kernel-Structured Interactions Takahiro Kawashima, Hideitsu Hino
Towards Defining, Measuring, and Leveraging Creativity in AI Applications Jingyi Yang, Alexander Tuzhilin
GIST: Generator-state-Informed Supervision for Tabular Foundation Models Zhangyang Peng, Jingze Shi, Tongyi Zhang et al.
Scan-Parallel Variational Inference for event-driven Latent Jump Models Vansh Bansal, Min Chen, James G. Scott
Target-Constrained Pooling for Trajectory Extrapolation Ning Wang, Claire Cardie, Sainyam Galhotra
Human and LLM Facilitation in Collaborative Design Meetings: A Computational Behavioral Comparison Mohammed Azeez Khan, Aaron Dsouza, Ashutosh Mishra et al.
From 2.1M Hours of Raw Broadcast Audio to Finnish Speech SSL: Confidence Filtering and Architectural Analysis at Scale Yaroslav Getman, Tam\'as Gr\'osz, Tommi Lehtonen et al.
Fast Gradient-Based Approximation of Continuous Dynamic Time Warping Pierre Houedry, Romain Tavenard, Nicolas Courty et al.
Les Mis\'erables Dialogue: A film-based network representation of social communication Theodora Moldovan, Georgios Panayiotou, Davide Vega
Verifiable Sampling for Delegated Generative Quantum Machine Learning Matteo Inajetovic, Anna Pappa, Petros Wallden
How Reliable are AMR Parser Outputs? Uncertainty Estimation for AMR Parsing Anuj Kumar, Shintaro Ozaki, Yamuna Prasad et al.
Data-Adaptive Mahalanobis Metric Learning for Cross-Head Attention in Transformers Zhenke Duan, Xin Li, Jiqun Pan et al.
BLOKS: Scaling Kernel Solvers with Sketched Block Preconditioning Giordano Vitale, Giacomo Meanti, Marco Rando et al.
Towards the Roof of the World: NLP Progress in Tibetan Languages Xi Cao, Yijie Li, Yuan Sun et al.
Towards Enriching Design-Time Models with Evidence from Simulation and Operational Activities Barbara da Silva Oliveira (UniCA, Laboratoire I3S - COMRED, KAIROS) et al.
A Better Metric Will Not Save You: What You Compare Beats How You Compare It Noureddine Khaous (LLF - UMR7110, UPCit\'e), Guillaume Wisniewski (LLF - UMR7110 et al.
FedPCD: Priority-Constrained Federated Aggregation Yousef Irshaid, Dara Varam, Mohamed I. AlHajri
PANDA: Prior-guided Attentional Dual-path Architecture Rishabh Jain, Federico Belotti, Stefano Coniglio et al.
Learning Size Independent Constraint Models from Input Data and Positive and Negative Example Solutions Helmut Simonis
Scalable Sparse Autoregressive Variational Inference for Non-Gaussian Spatial Fields Paul F. V. Wiemann, Matthias Katzfuss
Retractions Revisited: A Replication, Extension, and Mechanistic Evaluation of Linguistic Obfuscation Effects in Fraudulent Science David M. Markowitz
VQIL: Visible Quantum Interference Learning for Perturbation Response Ranking on Graphs Seokjun Lee, Sergio A. Ortega, Daniel K. Park et al.
Transformer-based In-Context Score Operator Learning for Diffusion Models Wuzhe Xu, Zhaiming Shen, Yulong Lu et al.
Machine Unlearning for Federated Socially Fair k-Means Clustering Aravind Reddy, Deeraj S K, Mohd Wasif Raza et al.
Amortized Counterfactual Prediction with Graph Neural Networks Hendrik Stumpf, Francesco Quinzan
How Well Do Prior-Data Fitted Networks Amortize Bayesian Inference? Manuel H\"ulskamp, Julius Kobialka, Emanuel Sommer et al.
When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors Adarsh Jamadandi, Nicolas Keriven, Aline Roumy
Your Discrete Graph Diffusion Model is Secretly Memorizing Adarsh Jamadandi, Nicolas Keriven
Constrained Probabilistic Numerics: Gaussian Process Solvers for Linear PDEs under Linear Inequality Constraints Boyuan Deng, Kshitiz Upadhyay, Michael Shields
Diversity-Aware Ensemble of Quantum-Classical Models via Combinatorial Fusion Jingyan Xu, Yuanhong Wu, Samuel Yen-Chi Chen et al.
TabMF: Learning Dataset Representations with Tabular Foundation Models Rafiq Kamel, Nicolas Atienza, Aziz Fouch\'e et al.

Agents 58

SWE-Duel: Towards a Self-Scaling Arena for Coding Agents

Wei Minn, Zhensu Sun, Lwin Khin Shar, David Lo Static coding benchmarks leak into training data and saturate as agents improve, so the authors propose a self-scaling arena whose tasks and oracles are created at evaluation time by the participating agents themselves. In SWE-Duel, a Red agent implements a new feature in a real repository, embeds a subtle bug with sealed triggering tests, and the pull request is admitted only if it passes the feature and existing tests, fails exactly the bug tests, and a fresh session of the same agent can repair the bug blind; a Blue agent then reviews the pull request without being told a bug exists, wins only if its revision passes all three suites, and the roles swap. In a tournament of 11 model-harness participants on 15 repositories in five languages over 1,500 turns, GPT-5.5 with Codex CLI ranked first as the strongest solver, repairing 98.3% of the defects it reviewed, while Claude Opus 4.8 with mini-swe-agent ranked second as the strongest task author with a 27.3% Red win rate. Within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% versus 4.0%), preliminary evidence that task difficulty scales with authoring capability.

arXiv:2610.12569 PDF

Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning

Wei Yang, Qing Ping, Youzhi Luo, Yawei Wang, Abhishek Kumar, Panpan Xu et al. Improving interactive language model agents increasingly depends on synthetic experience, but existing data synthesis methods face a feedback-to-decision gap: observed learner weaknesses do not directly say how future synthesis should change or whether the current synthesis space can even express them. EvoSyn is a learning-feedback-driven multi-agent framework organized around a revisable state that specifies which interaction factors can be deliberately generated and where synthesis effort should go, with role-specialized agents translating learner feedback into revisions, a Designer allocating synthesis directions, and a Planner grounding them into executable experience. This supports both adaptation toward weak or under-covered regions within the current space and expansion of the space by promoting newly discovered factors into explicit generation controls. On τ²-Bench and CAR-Bench, the framework improves average task success by 13.3 points over the strongest synthesis baseline while producing higher-quality, structurally broader synthetic experience and discovering controllable factors that go unnoticed without learner feedback.

arXiv:2610.12593 PDF

SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire

Scott C. Lowe, Sander W. Keemink Most agentic benchmarks are short and episodic, so strategic reasoning over long horizons remains under-studied and hard to separate into stored knowledge, reasoning, and luck. SpireBench has models play the complete commercial deck-builder Slay the Spire at its highest difficulty, where early decisions determine long-term success, and the harness is the first to give a model the full original game. Across 20 large language models (LLMs), frontier models beat heuristic baselines and match the average human player but remain far behind experts, winning the first act as often as experts and then falling behind; what separates strong from weak models is per-fight play, while all of them misjudge risk over a run by preserving health rather than spending it to grow stronger. A self-consistent renaming of every game term shows part of their performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs, which are often wrong and overly cautious.

arXiv:2610.12596 PDF

Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

Elias Lumer, Sahil Sen, Vamse Kumar Subbiah, Kevin Paul Prompt optimization is a cheap, human-readable alternative to reinforcement learning for adapting language-model agents, but on factual tasks an optimizer that reads training examples can simply write the answers into the prompt, so in-distribution gains may be memorization. The authors apply reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluate under a topic-disjoint split, so no topic is shared across train, validation, and test. Held-out citation reward rises from 0.702 to 0.781, replicating across three seeds, and although nearly half of the winning prompt is a memorized answer key, an ablation shows that content earns no reward even on the exact training questions it encodes. Verifier ablations show decomposed per-dimension feedback is the causal ingredient, since collapsing it to a single scalar reward makes held-out transfer negative, and the gain survives an out-of-family judge and is largely reproduced by a hand-written version of the discovered recipe.

arXiv:2610.12601 PDF

SynAct: Synergistic Active Feedback Learning for LLM Agents

Ning Liu, Ian Li, Kristina Klinkner LLM agents running multi-step workflows fail either through recoverable slips, which retries can fix, or knowledge gaps, which they cannot, and scarce expert feedback is only worth spending where it changes the outcome. SynAct is an active learning framework that ranks individual decisions by the mutual information between the chosen action and the terminal outcome across repeated rollouts, queries an expert at the highest-value decisions, and distills corrections into reusable rules the agent applies without expert access at deployment. The authors prove that with k rollouts and at least m observations per action, outcome-constant decisions have zero mutual information while outcome-separating ones have at least H2(m/k) bits, giving exact separation, and that success-based acquisition goes blind when all rollouts fail. Across four agent benchmarks and four models from three families, held-out accuracy improves in all 16 settings by 11.7 percentage points on average, the same signal beats budget-matched cascade, uncertainty, and expert-judge routers in 15 of 16 inference-time routing settings, and one targeted correction outperforms seven unaimed retries.

arXiv:2610.12651 PDF

CLEAT: Learning Value-Guided Clarification for Interactive Agents

Jiatan Huang, Zheyuan Zhang, Ziming Li, Tianyi Ma, Mingchen Li, Yanfang Ye et al. LLM agents must align their actions with goals and constraints users leave unstated, and asking a clarifying question only helps if the agent can turn the answer into better decisions, a relationship that shifts as the agent trains. CLEAT (CLarification through valuE-guided Agent co-Training) co-trains a clarification-value controller and the task policy from shared interaction outcomes, defining residual query value as the task utility gained by asking rather than proceeding, estimated through paired rollouts from the same history. The controller guides subsequent rollout collection while outcomes from the updated agent supply fresh value labels, so clarification guidance adapts as the policy learns. With Qwen3-4B, CLEAT achieves the highest mean score in all ten evaluation settings across four benchmarks, with gains over the strongest baseline reaching 13.2 points on UserGym Travel and 6.5 points on IN3 underspecification judgment, and ablations show co-training helps the policy even when the controller is removed at inference.

arXiv:2610.12655 PDF

Recurrent Harness for Agentic Continual Learning

Yuxin Xiao, Zhiyun Lu, Jianyu Wang, Yangtian Zhang, Yang Song, Qifan Wang et al. Language-model agents operating over long interaction streams need to adapt online to distribution shift, but existing harnesses either update their state only through forward information flow or require extra replay or validation data. ReHarness is a recurrent harness that adds backward revision to forward read and write under a strictly online streaming protocol: after each instance, the resulting trajectory serves as feedback to revise the reflections that were retrieved for that instance, propagating the signal backward through the textual harness state. Across three continual learning benchmarks and three language-model backbones, ReHarness consistently outperforms raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches. Analysis shows the revisions are source-grounded, correct overgeneralized or outdated experience under shift, and improve both stability and plasticity.

arXiv:2610.12668 PDF

MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents

Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni, Mudit Arora, Hugo Lin, Justin Chiu et al. Task-oriented conversational agents break down when users are uncooperative, and existing function-calling benchmarks contain few adversarial trajectories to train on. MADS (Multi-Agent Dialog Simulation) is a synthetic data pipeline that generates structured, diverse dialogues and injects adversarial conversation to produce validated unhappy-path trajectories. From it the authors build MADS-Bench, which spans six domains, 55 personas, and 1095 conversations, about 66% of which contain failure-prone scenarios, and they assess robustness by training models with supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR).

arXiv:2610.12757 PDF

EvoBelief: Falsifiable World Models for Self-Improving LLM Agents

Janet Wang, Jiang Guo, Yi Xiang, Jiaqing Yuan, Mingwen Dong, Hideo Kobayashi et al. Self-evolving LLM agents usually improve how they act through accumulated experience, reflection, or skills, yet even procedurally capable agents fail when their beliefs about environment dynamics are incomplete, wrong, or stale. EvoBelief is a self-evolving world model that stores environment knowledge as falsifiable hypotheses: unexpected outcomes and decision-relevant unknowns trigger hypothesis formation, later observations and controlled experiments validate a rule before it can guide predictions or plans within its supported scope, and counterexamples narrow or retire rules as the environment changes, all with the language model frozen and each rule linked to its evidence and the decisions it changed. Across four benchmarks it improves task performance and reuses rules across tasks, raising ALFWorld success over a matched frozen world model by 23.9% with Claude Haiku 4.5 and 31.3% with GPT-5.4, recovering from unannounced dynamics changes on MiniGrid, and in the Reactor Lab setting of DiscoveryWorld completing 83.3% of Normal-tier and 58.3% of Challenge-tier instances versus 0% for every baseline.

arXiv:2610.12789 PDF

Meta-Reflective Evolution Effectively Improves Production AI Systems

Seung Won Wilson Yoo, Chang Zeng, Aditi Somayajula, Sahithi Ankireddy, Rohan Kosalge, Avishree Khare et al. LLM-guided evolutionary optimizers typically select a candidate from the search frontier and mutate it in one LLM pass based only on that candidate's failures, which stalls on production AI systems where isolated failures reveal little about hidden interactions. CoCoEvolve (Cortex Code Evolve) is built around meta-reflection, in which an agentic mutator selectively retrieves evidence from the entire evolutionary history and probes the live system before verifying and proposing mutations. Applied to data agents, LLM-powered data pipelines, and data engineering workflows, it reaches 92.8% held-out hard accuracy on DABStep, topping the public leaderboard against 78.4% for Meta-Harness and 30.4% for GEPA under matched conditions, lifts the data-eng-bench test solve rate from 59.1% to 77.4%, beats every baseline on six pipeline workloads, and scores 11 points above GEPA on HotpotQA, with matched ablations attributing gains to each component.

arXiv:2610.12794 PDF

When Charts Deceive Web Agents: A Paired Benchmark for Misleading Visualization-Grounded Tasks

Yisheng Li, Dongming Han, Hongkun Pan, Bingcheng Mao, Zijian Wang, Yunlong Mao et al. Web agents consume charts and dashboards as intermediate evidence, and a misleading visualization can steer their downstream actions without any explicit malicious instruction. MisVisAgentBench is a paired executable benchmark of 280 task instances, each available in a clean and a misleading version that differ only in the chart evidence, together with a diagnostic framework that separates misleading-induced failures from baseline task difficulty and records execution effects such as extra browsing steps and hesitation. Across thirteen vision-language web agents, swapping clean charts for misleading ones lowers average task success by more than 34 percentage points, and the failures mostly follow the specific action branch the deceptive chart points toward, indicating systematic redirection rather than random disruption.

arXiv:2610.12823 PDF

TreeHCA: Hindsight-Guided Exploration and Credit Assignment for Multi-Turn LLM Agents

Debajoy Mukherjee, Jeremy Carleton, Sushil Vemuri, Srinivas Shakkottai, Dileep Kalathil Outcome-supervised group-relative policy optimization (GRPO) gives every token in an agent trajectory the same advantage, so helpful and unhelpful decisions receive identical feedback, and hindsight credit assignment (HCA) can only sharpen that signal when the sampled trajectories contain informative alternatives at the decisive points. TreeHCA uses changes in the likelihood of the reference answer to decide which intermediate histories to revisit, samples alternative continuations there to form a tree of reasoning and tool calls, applies a derived inverse-hindsight importance correction for the shifted sampling distribution, and backs terminal outcomes up the tree into turn-level advantages. Across seven question-answering benchmarks and three Qwen backbones, it beats all baselines in aggregate exact match.

arXiv:2610.12828 PDF

Co-Modifying Application Code and Deployment Infrastructure: A Causal Study of AI Agent Pull Requests

Mohammad Wathiq Soualhi, Basem Suleiman cross-listed Software engineering agents are mostly evaluated in sandboxes with local tests, yet shipping real features often means editing continuous integration (CI/CD) pipelines, containers, and provisioning manifests alongside application code. Using a deterministic four-tier file taxonomy and exact strata matching over 26,999 pull requests from AIDev v5 and MOSAIC-3M, covering six agent architectures and a matched human baseline, the authors isolate the integration overhead of dual-scope pull requests that touch both application code and deployment infrastructure. Such PRs make up only about 4.5 percent of agent contributions but split into two modes: when explicitly reviewed they take a median 13.13 hours longer and 3.24 times as long to merge, while agent identity reduces the odds of human review by 63.7 percent compared with human peers, even for external contributors, so roughly half of dual-scope agent PRs merge uninspected in a median of under 31 minutes. The authors argue for deployment-aware benchmarks and path-based access controls.

arXiv:2610.12849 PDF

Collaborative Theory of Mind: Learning to Self-Revise by Anticipating User Responses

Fanzhi Zeng, Anubrata Das, Maytal Saar-Tsechansky, Liu Leqi Large language models often attempt solutions before understanding what a user needs and fail to revise mistaken assumptions after feedback in multi-turn collaboration. CToM (Collaborative Theory of Mind) trains an assistant to draft a response, predict the user's reply, and use that prediction to revise internally before answering, with the user's actual reply serving as the training target under a negative log-likelihood objective combined with group-relative policy optimization (GRPO). Across collaborative mathematical reasoning, coding, and writing tasks the method outperforms all baselines, raising final accuracy on collaborative math from 74.1% to 79.6% relative to naive GRPO, and ablations show the prediction objective cuts the number of user replies by 8% on coding and 17% on writing at the same reward. Trajectory analyses find that user-response predictions become more accurate after training and that more accurate predictions correlate with larger gains from self-revision.

arXiv:2610.12872 PDF

Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory

Junyou Guo, Chongyang Tao, Yan He, Yong Wang, Shuai Ma Long-term multi-party conversations accumulate decisions that are repeatedly updated across participants and stages, and existing memory frameworks for LLM agents mostly store time as metadata rather than using temporal structure to recover evidence and reconstruct how events progressed. T-Mem organizes query-relevant evidence through two temporal mechanisms: Temporal Evidence Discovery exploits the fact that related messages cluster within short intervals to recover overlooked evidence, while Temporal Memory Structuring models state transitions and supporting relations over time and compresses low-information spans into a forest that is then linearized for answer generation. On two long-term multi-party memory benchmarks the framework consistently outperforms competitive memory systems across LLMs, parameter settings, domains, and query types.

arXiv:2610.12874 PDF

Collaborative Learning and Reasoning over Model Slices

Yuheng Wang, Runde Yang, Zhuohan Bao, Qiran Zhang, Huatao Li, Yiyang Zheng et al. Single-model chain-of-thought reasoning is prone to error propagation because every step builds on the model's own earlier outputs, leaving blind spots that are hard to reassess independently. The authors propose collaborative learning: model slicing replaces one large model with several smaller, heterogeneous models under a comparable parameter budget so that earlier reasoning can be re-examined by differently parameterized models, and pyramid supervision combines protocol distillation with step-level supervision to train both the reasoning and the interaction decisions. On multi-hop question answering, mathematical reasoning, code generation, and constrained open-ended generation, the trained slices match or beat the original large model and established multi-agent baselines, including reinforcement-learning-based ones, while running more efficiently than those multi-agent baselines and showing task-dependent collaboration patterns.

arXiv:2610.12927 PDF

SEEDS: Preview-Guided Evidence Admission for Search Agents

Jiajin Liu, Chenxi Xu, Dongzhe Fan, Yuanfu Sun, Qiaoyu Tan Search agents usually feed retrieved documents into their context by fixed or rank-based rules, even though a document's usefulness depends on where the agent currently is in its reasoning. The authors frame this as an evidence admission problem and introduce SEEDS (Selective Evidence Entry During Search), which shows the agent lightweight previews of retrieved candidates and lets it decide which ones, and how many, to inspect in full given its current reasoning state; the method works with an existing reasoning agent as-is or can be further optimized by training. Across nine knowledge-intensive question-answering benchmarks with both local corpora and live web search, SEEDS improves search effectiveness while substantially cutting full-evidence consumption, and analyses show it picks useful candidates that fall outside the top-ranked retrieval prefix.

arXiv:2610.12932 PDF

SmartSalesBench: A Resistance-Aware Multi-Turn Benchmark for Evaluating LLM Sales Agents

P N Bhargav Teja, Velpula Lokesh Reddy, Madhu Basumata, Susen Sarkar, Rajasekhara Reddy Duvvuru Muni Task-oriented dialogue benchmarks assume cooperative users and measure task completion, which says little about how large language model (LLM) sales agents cope with buyers who push back. SmartSalesBench pits agents against twelve psychologically grounded resistant buyer personas in conversations of up to twenty turns, spanning five agent profiles, fifteen smartphone products across six price tiers, and four controlled experiments totaling 13,680 sessions, with both agent and buyer generations grounded through retrieval-augmented generation (RAG). It defines nine metrics in three tiers, primary outcome metrics CR, BRS, and RIR, secondary quality metrics, and exploratory metrics reserved for future reporting, scored by an LLM judge; a human assessment with n = 50 reached kappa 0.71 and a Pearson correlation of 0.81 with BRS.

arXiv:2610.12975 PDF

Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction

Yuqi Huang, Hui Li, Yifei Cao, Zikun Cai, Tianci Xie, Yangfan Gao et al. Computer-use agents still struggle with unfamiliar software and workflows that require human domain knowledge, and although human demonstrations convey that knowledge through the same GUI, reusing them means interpreting actions in context and separating general workflow rules from task-specific details. Work2Skill provides WorkRecorder, which aligns screen recordings, mouse and keyboard events, and accessibility-tree data, and uses it to build WorkDemo-350, a synchronized multimodal dataset of 350 screened human desktop demonstrations; an agentic skill-induction method then inspects this evidence on demand to write reusable skills capturing workflow rules, step dependencies, and completion checks. With four execution models equipped with induced skills under shared settings, Work2Skill scores highest, raising mean scores over a no-skill baseline by 4.67 to 10.54 points on MyPCBench Bounded Action and 3.11 to 7.50 points on a 286-task subset of OSWorld-Verified.

arXiv:2610.12984 PDF

Hindsight Structured Advice Distillation: Steering Executor Agents with Advisors Trained on Their Failures and Task Characteristics

O. Ozan Koyluoglu Language-model agents deployed in specialized domains such as finance are often capable enough for a task but miss domain-specific details, and fixing this through retraining is expensive or impossible when the executor is closed, while writing the know-how into a static prompt makes it long and identical for every input. Hindsight structured advice distillation (HSAD) trains a small advisor model that reads each input and writes structured advice appended to a frozen executor's prompt. The advisor is trained with privileged information: on a graded training pool, the executor's failures are analysed, several strategies are run to see which worked on each input, and a teacher that sees grades and traces writes the advice that would have helped, which the advisor then learns to produce from the input alone via supervised fine-tuning. On FrontierFinance, an 8B advisor raises the qualification rate of Samaya's leaderboard-leading research agent from 52.9% to 56.6%, beating the agent's own high-effort configuration at 37% lower cost per query, with further gains on WebShop, BIRD text-to-SQL, and ExpertLongBench.

arXiv:2610.13000 PDF

EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang et al. Continual learning research has largely stayed in non-embodied settings. EmbodiedCLBench is a benchmark in embodied simulation for evaluating whether self-evolving agents can learn basic tasks and then recombine the acquired skills to solve unseen, more complex advanced tasks, with compositional generalization as the core evaluation principle. Experiments across multiple agent harnesses and learning methods show that advanced tasks are nearly unsolvable zero-shot, that experience from basic tasks yields consistent improvement that holds up as task complexity grows, but that the improvement does not scale with the number of learning samples, so current agents cannot extract more from additional experience. Code and data are released.

arXiv:2610.13021 PDF

Don't Guess, Just Ask: Evaluating Interactive Requirement Clarification for Repository-Level Software Development

Mingyu Chen, Zhiqing Zhong, Manyi Wang, Linxi Liang, Rouyu Chen, Pinjia He cross-listed Software users usually begin with vague intent, and in repository-level development a misunderstanding at the start propagates into costly rework, yet existing benchmarks evaluate code generation from complete or underspecified specs and skip clarification, while prior clarification studies stop at function-level tasks. ReqClarifyBench contains 341 repository-level tasks across game, miniapp, and web development, each with a vague user intent, a human-validated requirement document, and code evaluation artifacts, and uses LLM-based user simulators with personas derived from real-world conversations. Evaluating six LLMs on clarification effectiveness and requirement-document quality shows that stronger clarification does not necessarily yield better requirement documents, that performance varies substantially across user personas, and that coding-agent harnesses do not improve document quality over their base models.

arXiv:2610.13022 PDF

Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution

Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li et al. Self-evolving agent systems either update model weights from self-generated experience or rewrite the harness that governs prompts, tools, context, and execution feedback, but existing methods pursue the two separately. Evo² co-evolves them: a single policy plays two roles, a Proposer that writes harness edits and a Reflector that diagnoses their effects and guides revision, and the harness edits plus their execution outcomes in turn supply training signal for both roles. The EvoTIDE pipeline recovers usable trajectories through edit extraction and prefix continuation, then combines supervised fine-tuning with interleaved GRPO across the two roles. On coding benchmarks it reaches 82.1% on Polyglot and 48.3% on SWE-bench Verified versus 62.6% and 41.0% for the strongest baseline, the evolved harnesses transfer to other tasks without further evolution, and the dual-role harness loop alone lifts a GPT-5-mini agent from 24.4% to 48.3%.

arXiv:2610.13025 PDF

Travel Back To Move Forward: Retrospective Exploration and Evolving for Long Horizon Agents

Ziang Ye, Wentao Shi, Lei Chen, Zhengzhou Cai, Yu Wang, Zhiyuan Yao et al. Long-horizon agents that err early often cascade into dead-end trajectories they cannot escape. ReTRAIL lets an agent rewind the environment to an earlier state while keeping the knowledge gained from the abandoned branch, training one policy with State-Anchored Group Relative Policy Optimization (GRPO) over alternative continuations plus a branch-advancement reward that shapes reflection. Across MineSweeper, Sokoban, ALFWorld, and ScienceWorld, adding a rewind action on its own gave little benefit and sometimes degraded performance, with the gains coming from pairing rewinding with reuse of what the failed exploration revealed.

arXiv:2610.13058 PDF

CB-Orchestrator: Adaptive Workflow Optimization for LLM Agents via Contextual Bandits

Jiahang Sun, Zhiwei Shang, Zhipiao Liu, Hongwei Yang, Guoqing Xie, Shuang Qiu et al. Automating the design of agentic workflows usually burns a large API budget collecting training-phase feedback and fails to reuse workflows that already work well. CB-Orchestrator splits generation from selection, first building a candidate pool by evolutionary search, then treating per-query workflow choice as a contextual bandit so exploration and exploitation balance with far fewer feedback samples. Across five benchmarks it beats all baselines, and on MATH it outperforms recent methods while cutting training-phase API token overhead by at least 40.55% and total end-to-end training tokens by at least 29.73%.

arXiv:2610.13064 PDF

LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence Management

Songtao Huang, Zhijie Zhong, Xiaohan He, Shuaiyu Zhang, Zhiyin Yu, Li Kang et al. Agents reasoning over document collections far larger than their context window tend to tangle acquired evidence with evolving conclusions inside a growing interaction history, producing unbounded exploration and conclusions that drift from the evidence. LongHarness makes that interaction explicit through Evidence-Conclusion Coevolution, where unresolved conclusions steer retrieval toward missing information and newly acquired evidence revises the conclusions that directed the search, paired with lifecycle management that keeps stored evidence current and traceable. On OfficeQA Pro it lifts GLM-5.3-Flash exact match from 15.8% to 60.2% on V1 and from 4.4% to 61.1% on V2 against direct reasoning with the same backbone. It also beats existing agent harnesses, by 12.0 points over Claude Code on V1 and 18.9 points over Codex on V2.

arXiv:2610.13075 PDF

When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows

Yuchen Ling, Shengcheng Yu, Zhenyu Chen, Chunrong Fang Agents make many model calls per task, so a faster and cheaper decision model such as Jev promises large savings, but whether per-decision savings become efficient task completion is untested. Across five agents in two settings, embodied tasks in ALFWorld and mobile app use in AndroidWorld, spanning 4,500 task executions, the authors jointly measure task effectiveness, decision- and task-level resource use, and execution behavior. Decision cost and latency fall for every agent, yet task success declines substantially in two mobile agents and mean cost per successful outcome rises in one agent even as spending per attempt falls. Auxiliary reasoning and extra environment interactions absorb the savings, while failed recovery and premature completion eat into delivered results.

arXiv:2610.13092 PDF

VETR: Execution-Grounded Self-Improving Agent for Symbolic Graphics Reasoning

Yingbo Ma, Kuangshi Ai, Gordon Dai, Chenglong Yu, Chaoli Wang, Hua Shen Visual agents that treat scenes as executable programs improve slowly because the final render is ambiguous supervision, and existing refinement loops ignore which proposed symbolic edits survived, vanished, or interfered during execution. VETR keeps aligned program-and-render records with bounded runtime observations and adds an execution refiner that compiles observed writes and state changes into constrained alternative programs, validating their endpoints and preservation conditions against captured runtime state before an ordinary visual verifier judges goal relevance. Repair construction requires no additional language-model calls, only extra executions. On BlenderGym (280 tasks) and BlenderBench (27 tasks) with GPT-5.6-luna and Qwen3.8-27B, it beats refinement baselines including BlenderAlchemy and VIGA.

arXiv:2610.13098 PDF

Can Failed Trajectories Reveal Where Agents Go Wrong? Agentic Failure Attribution Without Step-Level Labels

Do Nguyen-Thanh, Junlin Fang, Bing Li, Zhen Fang, Sean Du Pinpointing the decisive step that doomed a failed agent trajectory normally needs expensive language-model prompting, step-level error annotations, or a reference set of verified successes. SOAP works from trajectories known only to have failed, building a spectral reference space from model representations and scoring each step by its projection behavior. Because later steps inherit earlier mistakes, independent scorers tend to blame downstream consequences, so SOAP derives soft step dependencies from attention patterns and propagates downstream error evidence back toward its likely upstream source, which improves attribution over competitive baselines.

arXiv:2610.13105 PDF

Locate, Judge, Read: Role-Specialized Tool Agents over a Knowledge-Point Index for Open-Domain Multimodal Document QA

Wenlin Lin, Fanzhe Wei Retrieval-augmented agents for question answering over large heterogeneous document collections usually hand one model both search and reading tools and let it schedule itself. KnowForge separates the jobs instead: a hybrid-retrieval knowledge-point index, a Locator agent that can only search, a code-scheduled bridge judge that forces the missing second hop, a typed and size-bounded evidence pool that merges every search's ranking, a Deep Reader agent that can only read text, images, and tables, and a synthesizer with a format checker. With a self-hosted Qwen3.6-35B-A3B the pipeline reaches F1 47.5 on the full open-domain M3DocVQA, 11.0 points above the published M3DocRAG system, 73.5 accuracy on MultiHop-RAG above the strongest published RAG and GraphRAG results, and 47.0 on MMLongBench-Doc, level with GPT-4o reading page images. Ablations show the Deep Reader's value tracks how visual each corpus is, and merging both agents into one with all tools and twice the rounds loses 1.9 F1 because it keeps searching instead of reading.

arXiv:2610.13128 PDF

Competence-Aware Orchestration for LLM-Based Multi-Agent Systems

Songmiao Wang, Yuchen Wang, Chenglin Yu, Hongxia Yang, Ming Li Multi-agent systems (MAS) built on large language models improve complex reasoning through role specialization, but more collaboration is not always better, since richer coordination adds token cost and redundant interaction that only pays off on hard queries. The proposed two-stage framework first models the backbone model's task-conditioned competence on the current query, then uses that signal to decide which roles to activate and how to shape the communication graph. Across six benchmarks the method averages 93.81%, and with six backbones of varying strength it adapts as intended: for higher-competence backbones it raises average accuracy from 87.22 to 89.97 while cutting token cost by 36.26%, and for lower-competence backbones it adds 4.58 points over existing methods. The same trend holds on backbones not seen during development.

arXiv:2610.13144 PDF

Quality, Not Just Outcome: A Structural Scoring Framework for Selecting Coding-Agent Trajectories for Supervised Fine-Tuning

Sida Li, Yuanxin Liu, Shicheng Li, Linli Yao, Xu Sun, Haoli Bai et al. Open coding agents are usually fine-tuned on interaction trajectories selected only because the final patch passed the benchmark test suite, which lets inefficient behavior such as repetitive file inspection, unstable tool usage, or delayed modification become part of the imitation target. TQS (Trajectory Quality Score) scores trajectories from their structure alone, with no LLM judge and in linear time, using signals for information focus, intent-action alignment, and policy dynamics organized around structured convergence. Ranking 155,144 successful trajectories from a public open-license corpus and fine-tuning coding agents on filtered subsets, the authors find that models trained on top-ranked trajectories consistently beat those trained on random or bottom-ranked subsets on SWE-Bench-Verified. They conclude that process-level trajectory quality matters for coding-agent supervision and that structural filtering is a scalable alternative to LLM-based data selection.

arXiv:2610.13152 PDF

SOAR: Source-Aware Reinforcement Learning with Precise Credit Assignment for Search Agents

Shengze Huang, Zihao Tang, Jinluan Yang, Xin Yu, Zhenghao Lin, Xiao Liu et al. Search agents trained with reinforcement learning (RL) often fabricate evidence, bypassing retrieved external information and answering from unreliable internal knowledge, which the authors attribute to the opacity of monolithic reasoning trajectories. SOAR (Source-Aware Reinforcement learning) splits external and internal evidence into parallel, observable streams through a Source-Aware Workflow and adds a Source-Consistent Reward, so that credit can be assigned precisely between the two knowledge streams. On 3B and 7B backbones the method outperforms baselines on knowledge-intensive benchmarks by 7.3% and 5.4% while reducing evidence fabrication by 74.8%.

arXiv:2610.13157 PDF

Scaling Agentic Data for Long-Horizon Terminal Intelligence

Shun Zou, Ziyu Ma, Yi Zou, Yong Wang, Lin Chen, Zehui Chen et al. Short agent trajectories show individual commands and tool calls but not how an agent accumulates evidence through exploration, adjusts strategy from feedback, and turns intermediate results into deliverables, and synthesized environments often pre-complete the very work the agent was meant to do. TerminalHorizon is an automated data synthesis engine that starts from real professional work, uses independent execution to find key work omitted during environment construction, uses that evidence to rebuild the environment, and then progressively extends tasks under the same goal with new requirements that build on prior results. The authors construct 1.5K task environments and fine-tune Qwen3.5-27B on the resulting data to obtain TerminalHorizon-27B, reporting substantial gains on all 12 benchmarks spanning coding-agent tasks, agentic tasks in real production settings, and general reasoning. Case studies with a 12-hour research budget each show the model continuing to experiment beyond an initial solution and retaining effective improvements in its final deliverables.

arXiv:2610.13166 PDF

RoleWorld: Knowledge Transfer in World Modeling with Functional Role Abstractions

Boyuan Xiao, Pu Wang, Ji Feng, Yao-Xiang Ding World models let large language model (LLM) agents anticipate the effects of actions, but recent symbolic or program-based world models organize their rules around environment-specific objects, which limits transfer to new environments with similar interaction structure and dynamics. RoleWorld represents how objects behave through functional role abstractions, namely reusable role prototypes combined with environment-specific parameter values. A teacher LLM extracts role prototypes, executable transition rules over role instances, and a decision-making skill from interaction trajectories, and in a target world the actor LLM alone assigns roles, estimates parameters online from action feedback, and actively tests a role hypothesis when its uncertainty affects the current decision, with no parameter updates and the rules held fixed. On the MARS benchmark the authors report robust generalization across environments with significantly altered object behavior.

arXiv:2610.13172 PDF

JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models

Mengdi Liu, Chenghao Jia, Hong Chang, Mingjie He, Shiguang Shan, Xilin Chen Long-horizon autonomous research systems that cycle through generating, executing, and evaluating hypotheses are throttled by slow, costly executions that delay feedback in a rapidly expanding search space. JANUS replaces reactive trial and error with a Predict-Verify-Evolve loop driven by a co-evolving world model: within an evolving graph, the research agent proposes candidates, the world model scores them by forecasting value improvement, information gain, and exploration utility, and branches are selected, refined, or pruned before any execution. After ground-truth feedback arrives, experience is distilled into hierarchical skills for the research agent, exploration-strategy skills for the world model, and coordination protocols between them, so candidate generation and predictive judgment improve together. On MLS-Bench-Lite and real research tasks the authors report that the framework outperforms state-of-the-art baselines in both evolving efficiency and ceiling.

arXiv:2610.13194 PDF

State-OPD: State-Grounded On-Policy Distillation for Multi-Turn Agents

Zhentao Zhang, Yanzhao Zheng, Yanlin Liu, Yuanqiang Yu, Tianze Xu, Chao Ma et al. On-policy distillation (OPD) transfers reasoning from frontier models to smaller ones, but extending it to multi-turn agentic environments is hard because the student's own reasoning both distorts teacher supervision and shapes which actions get sampled. Analyzing interaction states, which separate the objective task and action-observation record from the student's internal reasoning trace, the authors find that removing prior student reasoning improves the teacher's continuation success and changes its token-level targets, especially at later turns, and that resampling complete reasoning-action responses yields more action diversity than resampling actions under a fixed reasoning trace. State-Grounded On-Policy Distillation (State-OPD) grounds teacher supervision in interaction states and reallocates the sampling budget to generate multiple complete responses at high-entropy states while keeping full-history student generation. Across ALFWorld, ScienceWorld, and WebShop it achieves the highest macro-average success among evaluated methods for Qwen3-1.7B and Qwen3-4B students, and for Qwen3-4B it lifts macro-average success from 26.07% with the strongest baseline to 31.74% and surpasses its Qwen3-32B teacher on ALFWorld Seen, with fewer supervised turn responses and far less repetition of failed actions.

arXiv:2610.13238 PDF

Agent4Spec: An LLM-based Agentic System to Optimize API Specifications through Automated Validation and Market Analysis

Yongjian Tang, Costel Moraru, Doruk Tuncel, Hanya Elhashemy, Steffen Klepke, Thomas Runkler cross-listed Large language models (LLMs) can draft OpenAPI specifications from source code, but existing work treats generation as a one-off step and ignores ongoing validation, maintenance, and evolution against competing products. Agent4Spec is a multi-agent framework with two pipelines: one detects and self-corrects OpenAPI syntax errors, and the other performs competitive research against industry leaders, identifies functional gaps, and generates the missing endpoint specifications. Evaluated on two curated datasets covering general-purpose and industrial APIs with nine state-of-the-art LLMs, the best-performing model detected 100% of syntax errors and implemented over 95% of proposed corrections successfully, and more than half of its suggested endpoint additions were adopted.

arXiv:2610.13273 PDF

Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning

Mengdi Liu, Chenghao Jia, Hong Chang, Shiguang Shan Chemical reasoning agents that call external tools have so far been autoregressive, and the authors argue that left-to-right generation with causal attention hampers bidirectional molecular understanding, globally coordinated tool planning, and integration of evidence from multiple sources. ChemDiffAgent instead casts the task as iterative denoising over interleaved reasoning and tool-call trajectories with a masked diffusion language model, so tool-use decisions are made jointly under full-trajectory context, and it is post-trained in two stages: agentic supervised fine-tuning followed by variance-reduced preference optimization. On a new benchmark spanning six chemistry tasks in single-turn and multi-turn tool-use settings, it outperforms autoregressive agents by roughly 20% overall under matched budgets, and the 8B model is reported to match or in some cases exceed much larger frontier models such as GPT 5.5 and Claude Opus 4.7. Code is released.

arXiv:2610.13330 PDF

AgentTune: A Unified Training Lifecycle for Tool-Using LLM Agents

R E Zera Marveen Lyngkhoi, Abhivansh Gupta, Vidushee Vats, Ram Mohan Rao Kadiyala, Vinay Kumar Sankarapu, Pratinav Seth Building, training, evaluating, distilling, and maintaining tool-using agents is typically handled by separate tools for each stage, each with its own trajectory format. AgentTune is a source-available Python library built around a two-tier EventLog representation that preserves raw agent trajectories while projecting them consistently for training, evaluation, and auditing. On that substrate it layers a shared agent runtime, tool-calling reinforcement learning, agentic distillation, retrieval-augmented generation, decision workflows, evaluation utilities, and a planned path toward self-healing agents, all exposed through composable interfaces so strategies, training methods, and recovery mechanisms can be mixed without separate pipelines. The framework is demonstrated on agentic tasks including legal contract search and domain-specific applications, illustrating trajectory-based optimization, tool-use evaluation, and cross-task agent development.

arXiv:2610.13335 PDF

Test-Time Memory Construction

Zijun Wang, Yifan Ji, Haidong Xin, Zhenghao Liu, Yukun Yan, Zulong Chen et al. Large language models (LLMs) struggle to answer questions over long interaction histories when the relevant evidence is scattered across distant sessions, and existing memory systems build elaborate memory representations offline before any query arrives, which is costly and cannot adapt to what a specific question needs. T²Mem defers memory construction to inference: offline it extracts only lightweight episode-level metadata capturing key events and temporal links as retrieval cues, keeping pointers back to the raw interactions, and at test time the LLM adaptively retrieves and organizes metadata and raw turns, building query-specific memory as its reasoning unfolds. On LoCoMo and LongMemEval it consistently outperforms strong baselines across several backbone LLMs while substantially reducing memory-construction overhead, and further analysis shows the constructed memory adapts to the information needs of individual queries. Code is released.

arXiv:2610.13348 PDF

SWE-Critix: Post-Training Generalist Verifiers for Assessing Coding-Agent Trajectories

Jiajun Hu, Chun Yong Chong Coding agents increasingly tackle end-to-end software engineering tasks with no oracle evaluator defining success, so a verifier that can judge and explain an agent's behavior is needed, yet trajectories can run to hundreds of tool and repository interactions, and existing verifiers either use untrained general-purpose LLMs or train critics that emit only scalar rewards. SWE-Critix is a post-training pipeline that teaches a model to read a task and its execution trajectory, predict whether the task was completed, and write a natural-language analysis justifying the verdict, combining supervised fine-tuning (SFT) with reinforcement learning (RL) on trajectories from CodeForge-Preview. On a held-out test set, classification accuracy rises from 57% for the base Qwen3-30B-A3B reasoning model to 66% after SFT and 78% after RL. Models, datasets, and training code are released.

arXiv:2610.13356 PDF

AgentSWE: Can Coding Agents Build the Agent You Actually Want?

Jiahao Wang, Hongjin Qian, Yuyang Hu, Jiajun Zhang, Zheng Liu cross-listed As coding agents are commissioned to build other agents tailored to a user's workflow, there is an unstated assumption that the delivered agent faithfully serves its requirement rather than merely appearing to, and no benchmark had tested it. AgentSWE spans three stages of the agent development lifecycle, creating a complete agent from a natural-language requirement, editing production agent codebases, and optimizing toward a measurable target, across 25 tasks framed as commissions: the builder iterates against a few illustrative cases while acceptance runs the frozen submission on held-out cases graded by requirement-derived rubrics under hard pass/fail gates, or against the source benchmark's own objective for optimization. Across a range of coding agents, models, and harnesses, none comes close to delivering what the requirement asks for, and the gap between what a builder demonstrates and what it delivers is the norm rather than the corner case. Analysis of the failure trajectories identifies recurring modes including overfitting to the visible cases, hollow runtimes behind code that looks complete, and unsupported claims of completion.

arXiv:2610.13361 PDF

ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents

Chengyi Yang, Renxu Bao, Jue Hou, Zongpei Teng, Ziheng Wang, Yujie Lin et al. cross-listed Coding agents are increasingly asked to build programs from scratch rather than fix issues in an existing repository, but verifiable training environments that force an agent to discover requirements and make design decisions are scarce, and current workflows cannot sustain the long-horizon loop of exploring a reference program before implementing it. ProgramGym pairs Program-Env, which turns repositories into verifiable environments through a feedback-driven loop that repairs sandbox builds and the quality of black-box tests and yields 1,648 environments across C, C++, Go, and Rust, with Program-Agent, a role-isolated workflow of analysis, implementation, and review plus revision management for compilation checks and failure recovery. On ProgramBench, fine-tuning Qwen3.8-27B on trajectories from these environments lifts its score from 34.87% to 42.84%, and running the base and fine-tuned models inside Program-Agent raises them to 55.71% and 56.25%, exceeding some flagship models.

arXiv:2610.13395 PDF

Web$hell: A Unified Interface for Agentic Web Access

Sunghwan Kim, Deogyong Kim, Sangam Lee, Donguk Kwon, Dongha Lee Large language model (LLM) agents increasingly need to gather information broadly, perform browser actions, and revisit evidence within one long-horizon task, but existing access models, whether search-augmented, browser-based, or filesystem-based, mostly support either information gathering or browser interaction, and combining them brings cost or performance trade-offs. Web$hell unifies web access under a single model: using a common schema defined by WebContract, a WebCompiler turns live webpages into addressable information and action objects, which are kept in a WebCache and manipulated through one shared set of operations so access is consistent, operable, and persistent. Across web search, web navigation, and deep research benchmarks, the unified interface achieves strong task performance at lower cost than existing access models, and the results point to interface design as a lever for more capable web agents.

arXiv:2610.13416 PDF

Agentic Large Language Models

Max van Duijn, Michiel van der Meer, Aske Plaat, Niki van Stein Large language models (LLMs) already shape decisions through writing and question answering, and the book asks what changes when they can act in the world directly and interact with one another. Agentic LLMs combine language models with the ability to take actions, and the text covers how agentic LLMs work, how they act, how they interact with each other, and what this means for users. It is a book-length treatment rather than a single experimental result.

arXiv:2610.13455 PDF

Proactive Agentic Orchestration

Xueyang Feng, Chaozhuo Li, Xiaohe Bo, Jiakai Tang, Chenxi Li, Qiwei Ye et al. Agentic orchestration extends the execution horizon of large language model agents by delegating tasks to sub-agents with independent contexts, but current methods fix each sub-agent's goals and harness at creation time and learn the orchestration policy only from sparse, delayed feedback. POrchestra lets sub-agents proactively message the orchestrator at any logical point to request harness adjustments while continuing their local execution, and those revision requests drive On-Policy Orchestration Distillation (OPOD), which attributes each request back to the preceding orchestration decision to produce targeted process supervision alongside the final reward. Across three long-horizon benchmarks spanning deep research, dynamic digital assistance, and software engineering, POrchestra raises task success by an average of 9.1% over the strongest baseline, OPOD lifts small-orchestrator success by an average of 10.0% over existing training baselines, and the system reaches a more favorable cost-performance trade-off; the code is public.

arXiv:2610.13485 PDF

Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits

Changye Li, Zijian Zhang, Muheng Li, Yiduo Pang, Tingjun Wu, Menghan Li et al. Long-horizon machine learning engineering (MLE) agents usually rely on custom harnesses that manage tools, skills, and accumulated experience, but the authors find that reasoning intensity and progress both decay as a single session lengthens (long-horizon deceleration) and that the benefit of custom harnesses vanishes as the underlying model improves (harness staleness). They model alternating agent sessions on a shared workspace as a context-reset semi-Markov decision process (CR-SMDP), derive sufficient conditions and an experiment-budget bound under which alternation beats a single agent, and instantiate this as Humanize MLE Agents (HMA), which alternates two agents with at most k experiments per session and a context reset between sessions, with no task-specific harness or extra prior knowledge. On 75 Kaggle tasks in MLE-bench, HMA reaches a state-of-the-art any-medal rate of 78.2%, 8.0 points above the strongest open-source baseline with a quarter of its time budget, improves over its constituent agents by 6.4 points on average across five heterogeneous agent pairs, and a kernel-optimization case lands within 1.6% of the community leaderboard leader.

arXiv:2610.13494 PDF

MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures

Loan Bernat (LAAS-GEPETTO), Matthieu Grard (LAAS-RAP), Ariane Herbulot (LAAS-RAP), Florent Lamiraux (LAAS-GEPETTO) Long-horizon embodied agents must retain task-relevant information, coordinate robots with heterogeneous capabilities, and revise decisions as execution unfolds, but existing benchmarks rarely test dynamic task execution and multi-agent coordination together. MAGMA-Bench is an architecture-agnostic benchmark for persistent task reasoning in dynamic multi-robot environments that scales task horizons and injects controlled online perturbations such as execution failures, task interruptions, and mission updates, using a structured skill-level interface to separate task-level reasoning from perception and low-level control. Evaluating locally deployable open-weight planners, multi-agent state-tracking strategies, and frontier models, even the strongest models fail a substantial fraction of clean episodes and degrade further under online perturbations; supervised adaptation greatly improves a small model, but it remains well behind on longer and multi-robot missions. The release includes the benchmark, 261,646 task-level training examples, and an expert-trace generation pipeline.

arXiv:2610.13534 PDF

Training User Simulators with Latent Intrinsic Feedback

Tarek Naous, Craig Boutilier, Wei Xu, Ofer Meshi User simulators are usually built to generate realistic user utterances, but they have not been used as a direct source of feedback on an assistant's responses. The authors train LLM-based simulators on real user-assistant interaction logs and equip them with latent evaluation states: internal thoughts, reactions, and scalar ratings that judge the assistant from the user's first-person perspective at every turn. These simulators are more realistic and more critical than prompted baselines, and their turn-level evaluations correlate better with human judgments than those of prompted simulators of the same size. Used as a reward source for multi-turn reinforcement learning, they match setups that rely on external judge LLMs while cutting RL compute by more than half.

arXiv:2610.13617 PDF

Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify

Zhengyang Zhang, Jing Xiong, Hongzhou Lin, Juihui Chung, Zijian Zhang, Song Han et al. Single agent sessions remain brittle on hard olympiad-level problems: early errors propagate, missing domain knowledge stalls progress, and apparent success can hide a misread or weakened task. HOA is an open flow-level framework that coordinates multiple agent sessions through three feedback channels: an independent reviewer that critiques each candidate, a structured wiki that supplies task-specific knowledge the base model may lack, and a verifier that checks generated Lean proofs against trusted challenge statements. With GPT-5.6 Sol, it reports perfect scores on the IMO, IPhO, and IOI 2026, near-full marks on IChO 2026, solves all 672 problems on PutnamBench, and ranks first on Lean-Eval with 251 of 304 problems. With the open model Kimi-K3 it also reaches full scores on the IMO, IPhO, and IOI and 97.1% on IChO, suggesting systematic feedback can bridge large gaps in base-model capability.

arXiv:2610.13632 PDF

A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models

Ivan Moloshnikov, Dmitry Kunitsyn, Alexander Sboev, Maksim Skorokhodov, Roman Rybka, Yuanchao Liu Many machine learning methods are two-level: an expert picks a class of models, and then the choice of key components within that class is critical and usually manual. The proposed multi-agent large language model (LLM) framework uses three roles, analyst, coder, and recorder, with an experiment log and a sandboxed execution environment, and generates code only for specified components while leaving the overall network structure unchanged, rather than rewriting whole scripts or searching a discrete architecture space. It is validated on tuning a spiking neural network with correlation learning and on Physics-Informed Neural Networks (PINN) for high-order differential equations across five tasks. The agent beats baselines throughout, with classification accuracy 13 points higher on MNIST and 6 points higher on Digits, a roughly 2491-fold reduction in error against the analytical single-soliton solution of the generalized nonlinear Schrödinger equation (GNLSE), and a 14-fold drop in mass drift on the Kuramoto-Sivashinsky (KS) problem.

arXiv:2610.13663 PDF

A Baseline Assessment for the Agentic Web Technology Stack

Vagner Figueredo de Santana, Diogo Cortiz The rapid churn of protocols, frameworks, and libraries for LLM-based agents makes it hard for developers to know which combinations work together to build an Agentic Web solution, creating skilling, governance, and infrastructure hurdles before a team even gets started. The authors present MASTIF (Multi-Agent System TestIng Framework), a system for running Agentic Web evaluations, and use it for a baseline assessment of 3 LLMs, 3 protocols, and 4 agent implementation frameworks on 300 tasks from the Mind2Web benchmark. Choosing a task-specific technology stack can raise task completion rate, efficiency score, and social value stability while reducing verbosity and time to completion. The results are offered as baseline metrics for anyone designing, assessing, or building multi-agent systems.

arXiv:2610.13703 PDF

Matched-Context Group Policy Optimization for LLM Agent Training

Chenxu Zhang, Yaorui Shi, Wenlong Zhang, Jianchao Tan, Zihan Wang, Yiming Cheng Training long-horizon language model agents requires assigning credit to intermediate decisions from sparse terminal feedback, and comparisons between sibling rollouts are only useful when they share a comparable context. Matched-Context Group Policy Optimization (MCGPO) builds weighted action preferences from already collected trajectories: comparison contexts are defined by pre-action state summaries and rubric progress, ordered exact and coarse matching rules select pairs, fixed folds separate candidate decisions from the terminal outcomes used to orient them, and support-dependent evidence weights scale an auxiliary pairwise objective alongside the shared GraphGPO task loss, all without a critic. Across six environment-backbone settings with three seeds each, MCGPO improves mean success over matched GraphGPO by 0.47 to 1.46 points on ALFWorld and WebShop with Qwen2.5 1.5B and 7B models, and by 3.96 and 5.16 points on WebShop with Qwen3-4B and Qwen3-8B, with factorial controls supporting the opposite-fold evidence and weighting design.

arXiv:2610.13755 PDF

MC-EvoBench: Auditing Experience-Driven Self-Improvement in Minecraft Agents

Jingwei Song, Zhisheng Chen, Bingfan Zeng, Jing Tang, Yubo Hou, Jinhao Jing et al. cross-listed Claims that an embodied agent improves itself from experience are hard to verify, because a saved memory or a successful final action can reflect carried-over resources, unequal exposure to the source task, or early execution failures rather than a durable new skill. MC-EvoBench is a controlled Minecraft benchmark with 19 registered evolution chains, representing each task as a task-state graph with stage-level environment verifiers, assigning matched source and target roles, resetting the target world, and letting only one declared agent artifact cross the boundary, measured under no-prior, related, unrelated, and shuffled source conditions at matched step budgets over five seeds. Planner-backed agents consistently outperform controller-only variants on direct capability, but related experience produces no stable transfer, with every reported interval including zero. Stage audits also show that early wood-collection and crafting bottlenecks frequently leave later capabilities unobserved.

arXiv:2610.13829 PDF

Returning to Development States: When Full Sandbox Snapshots Help Coding Agents

Daniel Thi Graviet, Tanmay Shah Coding agents often make real progress before failing, and retrying from scratch throws away the machine state that progress created, which turns failure recovery into a search problem over intermediate environment states. The authors implement Go-Explore-style search over full Daytona sandbox snapshots so an agent can restore and branch from promising earlier states instead of restarting from the initial one. Higher and more efficient solve rates depend mainly on progress that lives outside the repository, such as databases, running services, and other runtime state, indicating that full-snapshot search pays off for long-horizon work spanning several services whose state cannot be rebuilt from a single diff.

arXiv:2610.13841 PDF

RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics

Weiliang Chen, Song Liao Tool retrieval for agents usually optimizes semantic relevance alone and ignores whether a tool will actually be available at execution time. RATED (Reliability-Aware Tool Retrieval via Execution Dynamics) pairs a frozen dual-encoder with contrastive fine-tuning for semantic matching with a lightweight contextual bandit that re-ranks candidates from real-time execution feedback, so runtime failures are absorbed without retraining the model. The authors also release EcoAgent, a benchmark of 5,172 real-world tools and over 210,000 query-tool pairs for offline retrieval and dynamic reliability evaluation. RATED improves mean reciprocal rank by 52.2% relative to the ToolLLM API retriever on real-world queries and, under outage scenarios with a 30% failure rate, cuts the fault selection rate from 86.2% to 2.9% while adding under 1 ms of latency over encode-only inference.

arXiv:2610.13889 PDF

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao et al. Recursive self-improvement (RSI) seeks to turn interaction experience into lasting gains, but external feedback is expensive and a world model's predictions can drift out of date as the policy changes. RSI-Anything, built on JEPA-Anything, uses predictive world models to select both actions and training data, lets observed feedback drive policy or workflow updates and conditional world-model adaptation, and gates retention of each updated policy-model pair behind component and compatibility checks. In code repair with a fixed coder, success rises from 26.3% with a static verifier to 29.2% with predictive selection and 31.2% after world-model adaptation at comparable test usage; graphical user interface (GUI) tasks show higher completion with 39.4% fewer actions, laboratory-planning scores more than double from 21.80% to 46.01% over five policy versions, and reaction optimization reaches 96.15% mean best observed yield versus 82.16% for Bayesian optimization on a 40-experiment budget. An end-to-end immune-science application links workflow evolution and predictive modeling to therapeutic hypothesis generation.

arXiv:2610.13912 PDF

Theory 49

Amortised Inference through One-Step Implicit Sampling

Vincent Pauline, Kirill Tamogashev, Arran Carter, Sanghyeok Choi, Stefan Bauer, Esmeralda S. Whitammer Drawing unbiased samples from high-dimensional Bayesian posteriors usually takes many steps, whether MCMC iterations or integration steps in a diffusion-based amortised sampler, and scalable one-step amortised inference has remained open. The one-step implicit sampler (OSIS) uses short-time MCMC dynamics to define a local target for the evolution of an implicitly modelled distribution, so that in the asymptotic limit the modelled distribution is a fixed point of the MCMC kernel and therefore samples from the target. The authors analyse convergence, draw connections to other generative modelling and sampling algorithms, and study design choices. OSIS performs well against multi-step amortised samplers and non-amortised MCMC baselines while being far more efficient at sampling time.

arXiv:2610.12679 PDF

A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges

Xiaoyu Li, Zhizhou Sha, Chiwun Yang, Dai Shi Hundreds of parameter-efficient fine-tuning methods exist but are typically compared only by parameter count and benchmark score. The authors instead treat each method as a smooth map from a small parameter space into weight space anchored at the pretrained checkpoint, forming a slice category in which the frozen model is initial and full fine-tuning terminal, and show that three invariants of the map, its first-order image, its fibres under the optimizer's metric, and base change, govern expressivity, training dynamics, and operations like quantisation and merging. Proved results include that zero-initialised LoRA starts at the apex of a cone, that no exact initialisation lets LoRA reach an update of rank above twice its own, that exactly orthogonal methods never alter a weight's singular values no matter how often they are merged, and that under idealised Adam LoRA+ from zero initialisation follows plain LoRA's trajectory with a rescaled scale. The atlas classifies 382 methods from 138 papers by seven coordinates, with 403 simulation arrows and 485 obstructions between them.

arXiv:2610.12728 PDF

Why Heavy-Tailed Weight Spectra Predict Model Quality

Joseph Wilson, Chris van der Heide, Liam Hodgkinson, Zhichao Wang, Fred Roosta, Michael W. Mahoney cross-listed Heavy-tailed eigenvalue spectra in trained weight matrices are known to predict deep neural network (DNN) quality without access to training data, but these metrics have lacked a rigorous statistical justification. The authors use a probabilistic framework to derive a marginal likelihood for a DNN evaluated using only its final-layer weights, and show across large convolutional networks and large language models that this quantity predicts model quality. Taking the limit for spectra that follow the heavy-tailed High-Temperature Marchenko-Pastur (HTMP) distribution, they show the log-marginal likelihood is convex in a heaviness parameter, which yields an optimal heaviness that empirically predicts over-fitting in large-scale models.

arXiv:2610.13123 PDF

Why Linear Models Can Still Work in a Complex and Changing World

Fangzhou Lin, Dengzhe Hou, Yinxing Li, Kazunori D Yamada Linear models often predict well in domains whose underlying mechanisms are plainly nonlinear, and the authors argue this happens because the world's complexity exceeds what can be learned reliably from finite, partially observed data. Flexible models can latch onto curvature, thresholds, and interactions that hold in one hospital, period, population, or measurement process but fail elsewhere, so when that instability sits mainly in the nonlinear component, dropping it lowers prediction error in a new environment. A squared-error proposition formalises the case where a nonlinear rule learned in training raises target-environment error because it transfers too weakly, and the argument is tied to prior work on bias and variance, regularisation, measurement error, distribution shift, invariance, Rashomon sets, and underspecification. The conclusion is that model complexity should be chosen by cross-environment stability as well as training accuracy, since partial observation alone is not enough, stable nonlinear structure still favours flexible models, and linear models remain exposed to linear spurious relationships.

arXiv:2610.13404 PDF

Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models

Ran Tian, Courtney Paquette Intermittent loss spikes during large language model (LLM) pre-training remain poorly understood both theoretically and empirically. The authors show that simple quadratic models fed heavy-tailed, power-law distributed inputs reproduce these spikes, and build a theoretical framework and simulation algorithm that models the per-layer training dynamics of deep neural networks (DNNs) with such layer-wise random feature models. In a case study of an Adam configuration with β2 set close to 1, which is known to induce spikes in Transformer language models, the simulation predicts the bumps that signal spiky behavior from a given checkpoint, and a β2 schedule derived from the observed metrics empirically improves performance.

arXiv:2610.13524 PDF

Rationally Linearized MLPs

Arijit Sehanobish, Byeongchan Kim, Avinava Dubey, Min-hwan Oh, Krzysztof Choromanski Linearizing multilayer perceptrons (MLPs) into kernel form makes them cheaper to evaluate, but polynomial approximations of activation functions need high degree to be accurate. RaLiMPs replace them with low-degree rational approximations, which the authors show reach comparable accuracy at substantially lower degree, and use them to build kernel representations with narrow bottlenecks that compose efficiently across layers. The approach comes with theoretical approximation guarantees and, in experiments spanning implicit 3D models to large language models, achieves competitive or improved accuracy while substantially reducing inference cost.

arXiv:2610.13568 PDF

43 more specialized papers

A Sampling Complexity-aware Framework for Discrete-time Fractional-Order Dynamical System Identification Xiaole Zhang, Vijay Gupta, Paul Bogdan
Efficient Learning of Survival Analysis Models Under Outcome-Dependent Truncation and Censoring Itai Zilberstein, Ioannis Anagnostides, Arman Kilic et al.
Transformer-Based Masked Diffusion Models Provably Learn Distributions with Latent Structure Jingda Wu, Changxiao Cai
Toward Maximizing Algebraic Connectivity via Density-Dependent Backbone Construction and Reinforcement Learning Completion Zeru Zhu, Marco Gamarra, Ji Liu
Primal-Dual Flow Matching for Sample-Wise Constrained Generation Zhengyan Huan, Peter Y. Lu, Shuchin Aeron
Product Monte Carlo: Expectation Under a Product of Densities Pietro Carlotti, Sifan Liu, Yuling Yao
Minimax Rates for Estimating Powered Sliced Wasserstein Functionals Binh Thuan Tran
Neural expectation-maximization and meta-models Zhuolin Fu
The Expressive Power of Random-Feature MPNNs Across Graph Sizes Robert R. Nerem, Mitchell Black, Antonis Vasileiou et al.
Two-way Guidance: Simultaneous Learning of Summary Statistics for Generative Bayes Percy S. Zhai, Max Pachman, Peverel D. Shipley et al.
Universal Bounds for Self-Normalized Kullback-Leibler Estimation Buu Phan, Lester Mackey, David Alvarez-Melis
Robust Langevin Sampling under State-Dependent Heavy-Tailed Noise with Applications to LLM Fine-Tuning M. Berk Sahin, Ege C. Kaya, Arda Fazla et al.
Generalized Determinantal Point Processes for Attraction and Repulsion with Distinct Interaction Kernels Hideaki Ishibashi, Takahiro Kawashima, Hideitsu Hino
Controlled-Minimax: Stabilizing Adversarial Optimization via Control Theory Tianjiao Luo, Jianfei Chen, Jun Zhu
Beyond Second-Order Information: Paid Logistic Regression with Noisy Features Sanghwa Kim, Kyoungseok Jang, Se-Young Yun
Sparse Dictionary Learning for Measure-Valued Data via Linearized Optimal Transport: Efficient Algorithms and Sample Complexity Brendan Mallery, Shuchin Aeron, James Murphy
From Imbalance to Evidence: Deep Nonparametric Testing for Complex Data under Severe Sample Imbalance Guanhao Zhou, Xiufan Yu
DataCupid: Stable Matching Repair for Training Data Acquisition with Delayed Feedback He Sun, Junyuan Mao, Wenhao Jiang et al.
Forecasting Corpus-Sampling Disagreement in Masked Text Reconstruction Hare Krishna
Gaussian-Width Bounds for Offline-to-Online Linear Bandits Kushagra Chandak, Volodymyr Tkachuk, Xiaoqi Tan et al.
DAG Structure Learning with Node-Varying Heavy-Tailed Disturbances Robin Francis, Sundeep Prabhakar Chepuri
Model-free Bayesian Clustering and Multimodal Posterior Summarisation Qing Li, Sara Wade
Accelerating Linear Bandits with Persistent Candidate Sets Bingkui Tong, Daria Korovaitceva, Marina Sheshukova et al.
Learning by Compression: Risk Estimates Attain the Law-of-Large-Numbers Rate Simone Garatti, Marco C. Campi
Harmonic analysis and circuit approximation of Sobolev functions on quantum input spaces Philipp Schmocker
Asymptotic Analysis of Discrete-Time Self-Attention Dynamics in Transformers Claudio Altafini
Causal Effect Identifiability in Causal Abstractions: A Strict Hierarchy of Criteria Cl\'ement Yvernes, Marianne Clausel, Emilie Devijver et al.
Simultaneous Competitive and Regret Guarantees for TCP Acknowledgment Tsubasa Harada, Yasushi Kawase, Hanna Sumita
Lower Bounds for Randomized Linear-Oracle Optimization over Strongly Convex Sets Sebastian Pokutta
Some remarks on the connection between generative diffusion and optimal transport Giulio Biroli, Marc M\'ezard
Kernel-based External Validity of Experimental Studies Federico Matteucci, Gerhard Neumann
Towards Understanding the Training of GFlowNets: A Theoretical Perspective Rui Hu, Yu Chen, Zhuoran Li et al.
Studying the Effects of Task Diversity Using Small Mathematical Worlds Emily Xia Meng, Xuanyu Yang, Junaid Hasan et al.
DualQuant: Dual-scale Calibration-free Weight-and-Activation Quantization for LLMs Mahsa Salmani, Sayeh Sharify, Hesham Mostafa
Geometry-Aware Multi-Objective Optimization: Norm-Constrained Oracles with Entropic Weight Updates El Mahdi Chayti, Alireza Mirrokni, Martin Jaggi
Optimal Algorithms for Nonsmooth Nonconvex Stochastic Optimization with Constraints Gon Buzaglo, Guy Kornowski
Randomized Momentum Transport with Cheap Predictors El Mahdi Chayti, Martin Jaggi
Adaptive Variance Reduction without Knowing Oracle Smoothness Wei Jiang, Yibo Wang, Zechao Li
Finite-Sample Guarantees for Fully Discrete Conditional Diffusion Models Qiuyun Zhu
Partially Informed Stochastic Optimization via Zeroth-Order Residual Tracking El Mahdi Chayti, Alireza Mirrokni, Martin Jaggi
Prediction Is Not a Sufficient Memory State Robert Jenkinson Alvarez
KL Before and After Mixing: A Round-Complexity Separation in Parallel Sampling Tianyi Wu
Constructive Bounds for the Topological Expressivity of Neural Networks Kosio Beshkov

Multimodal 27

Representation Alignment from a Few Paired Data via Distilled Optimal Transport

Julien Kordes, Reuben Dorent Aligning two independently pretrained unimodal encoders normally requires millions of paired examples, whereas this work targets the regime of only 100 to 500 pairs alongside abundant unpaired data on each side. Every representation is re-expressed by its similarities to the paired anchors, which places both modalities in a shared coordinate system. The two point clouds are then matched with a single entropic optimal transport problem over the whole pool, whose marginal constraints counteract the hubness that defeats nearest-neighbor matching, and the resulting plan is distilled into two lightweight encoders so that the correspondence becomes inductive. Only the anchors need pairing, since the unpaired collections can be drawn independently from each marginal, and the method outperforms existing approaches at the same anchor budget on Flickr30k, MS-COCO, and ChEBI-20 for image-text and molecule-text alignment.

arXiv:2610.12781 PDF

People Making Conversation: Evaluating How Naturally Speech Models Converse

Arya Bastani, Iman Modarressi A real-time conversational speech model must decide at every pause whether to speak, but existing evaluations measure boundary detection, policy adherence, or aggregate statistics rather than that decision. The authors place six systems at 427 pauses and 482 turn ends from 176 unscripted conversations in Seamless Interaction, give each the conversation so far, and score its output against what the person actually did: whether and how soon it speaks at a pause, its latency, length, and wording at a turn end, and whether it can answer a question about the preceding 30 to 120 seconds. The hosted APIs GPT Realtime, Gemini 3.1 Flash Live, and Grok Voice respond at almost every pause and reply longer than the person, with latency of about 1 to 4 seconds under silence-based endpointers and about 8 seconds under semantic turn detection deciding whether the words land inside the pause, whereas the full-duplex models PersonaPlex and Moshi are the only systems that regularly stay silent and the only ones that interrupt within a human reaction time. Given the history as text, hosted systems recall 41 to 60% of conversation facts under a Claude Opus 5.5 judge, while full-duplex models hearing it as audio recall 17 to 22%, and the labels, points, harness, and every output are released.

arXiv:2610.12802 PDF

MathVis: Can AI Reason through Drawing? A Benchmark for Mathematical Figure Generation

Jingjing Wei, Hao Zhao, Xinke Li, Linxi Yu, Xiaosong Gao, Shanshan Xiao MathVis tests whether large language models can draw figures for frontier mathematical problems and reason with them, using 512 tasks that each pair a statement with a reference diagram, executable TikZ or SVG code, a gold intermediate representation, atomic rubrics, and error codes. GPT-6-Astra, Claude Fable 5, and Gemini 3.1 Pro Preview reach pass@1 scores of 31.25%, 26.56%, and 22.85%. The models tend to answer in formal languages, symbols, and graphics programs rather than reasoning explicitly through diagrams, and even solutions that contain all the required elements are often visually confusing to humans.

arXiv:2610.12954 PDF

GeoGR: Spatial-Semantic Token Compression and Group Routing for Efficient 3D VLMs

Yongxin Zheng, Hao Wu, Fang Li, Hu Fu, Peiran Yin, Xudong Wang et al. cross-listed Multi-view inputs let pretrained image and video vision-language models (VLMs) reason about 3D scenes, but the resulting long visual token sequences make inference costly, with redundancy both from repeated views of the same regions and from processing every visual token at every LLM layer. GeoGR attacks both sources: at the projector, GeoSemZip consolidates cross-view observations into 3D voxel tokens, selects a compact set by combining semantic saliency with spatial coverage, and pools the unselected voxels into contextual anchors. Inside the LLM, GroupRoute exploits the finding that query-conditioned token relevance is locally stable across layers but drifts over longer intervals, letting low-relevance tokens skip whole layers while keeping their hidden states so they can rejoin at later anchor layers under a compute budget. With compression-aware post-training on a 67K-sample subset, GeoGR retains 95.2% and 96.8% of the dense models' performance on LLaVA-OneVision-7B and Video-3D LLM-7B while accelerating LLM prefill by 6.8x and 6.1x.

arXiv:2610.13035 PDF

Does Location Matter? A Black-Box Causal Audit Reveals Model-Dependent Audio Grounding in LALM Hallucination

Mohammad Khalooei, Mohammad Sabokrou Large audio-language models often affirm sounds that are not present, and the dominant fix, Audio-Aware Decoding (AAD), contrasts logits against an all-zero waveform that reviewers have flagged as out of distribution. This audit tests the claimed evidence directly instead: locate a claim's evidence region by CLAP similarity, remove it by spectral gating, and compare against a random-window control across a 2x2x2 grid on three models. Raw detection favors the blank baseline on Qwen2-Audio-7B (0.613 against 0.721 AUROC), but that reflects a duration confound, since blanking removes the whole clip against a 2-4 second window; holding duration fixed, localized removal beats random removal in every cell on two models. Applying the same evidence to decoding, a gated ensemble of blank and language-prior contrasts lifts Qwen2-Audio accuracy from 57.1% to 71.9%, and the grounding profile does not transfer between models.

arXiv:2610.13112 PDF

Monarch-MoE: Structured Experts for Multimodal Continual Instruction Tuning

Xi Zhang, Boyu Diao, Zesen Chen, Ruiqi Liu, Hangda Liu, Yongjun Xu Multimodal continual instruction tuning tends to forget earlier tasks, and even multi-expert LoRA does not fully isolate updates because a dense low-rank expert touches many output features. Monarch-MoE replaces the dense LoRA factors with block-diagonal factors joined by a fixed permutation, so each block-pair update affects only one output block while every input block still connects to every output block, and it pairs this with batched computation, fused forward operators, and an optimized backward pass. Across three continual-learning methods on MLLM-DCL and UCIT with LLaVA-1.5-7B, first-task forgetting drops to as low as 0.52 and 0.10 percentage points while adapter parameters shrink by roughly 75% and 50%, with method-dependent trade-offs in final performance, and the whole pipeline fine-tunes and runs on a single 24 GB RTX 4090.

arXiv:2610.13215 PDF

What Do Audio-Language Judges of Generated Music Listen For? Four Probes with Controls from Expert Ratings

Mateusz Modrzejewski, Tomasz Radzikowski cross-listed Open audio-language models are increasingly used to judge generated music, and are validated by correlating their ratings with expert ratings, yet correlation alone does not reveal whether a judge responds to the audio, the text prompt, or the phrasing of the question. The authors propose four probes, removing the audio, swapping in another model's clip for the same prompt, reversing the order of two clips in pairwise judging, and rewording the instruction, with three of them read against a control computed from the human ratings, and apply them to five open judges on all 2,748 clips of MusicEval plus the SongEval songs. Without audio, no judge beats what the prompt alone predicts from the human ratings, and with audio the strongest judge's quality score tracks the experts' prompt-alignment rating more than their quality rating. The four smaller judges' pairwise verdicts follow a clip's slot in the pair rather than its quality, and rewording the instruction shifts most judges' correlation about as much as switching judges, so the authors recommend reporting all four probes with any music judge.

arXiv:2610.13247 PDF

CV-OPD: Contrastive Vision On-policy Distillation

Haowei Wang, Hao Lin, Qi Zhu, Jiayi Ji, Deyi Ji, Liujuan Cao et al. cross-listed On-policy distillation (OPD) gives vision-language models (VLMs) dense token-level supervision from a teacher, but the authors find that distilled students often answer correctly even when the relevant image regions are masked or replaced with noise, and they argue theoretically that answer-conditioned distillation opens a shortcut because nothing constrains the student to use visual evidence. Their fix, CV-OPD, pairs the usual positive teacher with a negative teacher that sees an answer-irrelevant crop, applies a Contrastive Divergence Loss that pushes the student's distribution toward the positive teacher and away from the negative one, and uses a Discrepancy Aware Gate to restrict optimization to tokens where the two teachers disagree at the token and sample level. Across a range of fine-grained visual benchmarks, the method substantially reduces the masking shortcut while consistently improving visual reasoning over standard distillation.

arXiv:2610.13264 PDF

When MLLMs miss what humans read: diagnosing and learning to read squint text

Chen Peng, Shan Huang, Jianbo Huang, Linrui Qin, Qibo Sun, Mengwei Xu cross-listed Multimodal large language models (MLLMs) can fail to read text embedded in an image's coarse structure, the kind a human recovers by squinting, and instead report that no text is present, a failure the authors traced through a production content-moderation pipeline. They release SquintBench, 1076 production-derived moderation images, and SquintCtrl, a controlled synthetic suite for diagnosing the mechanism; frozen-weight analysis shows the hidden text leaves a coarse signal but glyph identity is weakly represented and becomes far more accessible once the correct spatial regions are supplied, implicating spatial selection as the main bottleneck. The proposed Squint OPSD distills on-policy from a frozen teacher that sees a downscaled view plus the ground-truth transcription, while the student reads only the original image at inference, reaching 88.61 average normalized Levenshtein similarity (ANLS) on SquintBench versus 26.50 for the unadapted backbone and 68.60 for the best off-the-shelf model given a downscaled view.

arXiv:2610.13465 PDF

S-Space: Exploring Spatial Workspace in Multimodal Language Models

Zhiyuan Shi, Xiang Li, Xingzhuo Guo, Shangchen Miao, Weiliang Chen, Qiyu Dai et al. cross-listed Multimodal language models (MLLMs) show promising spatial reasoning, but the internal mechanisms behind it are poorly understood. The authors identify S-Space, a linear subspace in which object-token representations encode continuous horizontal, vertical, and distance coordinates; although the axes are found by differentiating spatial-answer scores with respect to object-token activations, the subspace is intrinsic and stays readable under non-spatial questions. Perception writes object locations into it, internal reasoning transforms it for viewpoint changes, and verbal responses selectively read from it, as causal interventions alter spatial judgments while largely leaving non-spatial responses intact; it also assigns locations to unseen entities and to the observer and links physical directions to abstract linguistic meanings. Used as a diagnostic, it reveals systematic coupling between cardinal and image-relative reference frames, and a perception-computation decomposition shows that MLLMs form capable spatial representations but fall short in the spatial transformations needed to reason reliably across viewpoint changes.

arXiv:2610.13527 PDF

AudioTETRIS: Token Elimination via Transitional Importance Signals for Efficient Audio LLMs

Shreya Sajid, Anna Vettoruzzo, Joaquin Vanschoren cross-listed Large Audio Language Models (LALMs) inherit hundreds of audio tokens per short clip from their encoders, and quadratic self-attention makes that expensive, while existing training-free pruning treats audio tokens as generic features and erases the acoustic transitions that carry speech information. AudioTETRIS uses the audio encoder's representation geometry as a prior on token importance, fuses it with attention-based scores inside the decoder, and performs progressive top-k token selection across decoder layers with no extra training or auxiliary models. Evaluated on Qwen2-Audio-7B, Phi-4-Multimodal, and Audio Flamingo 3, it keeps near-uncompressed accuracy on several speech tasks with only 25 percent of audio tokens, matches or exceeds full-token accuracy at moderate compression, and beats existing training-free baselines.

arXiv:2610.13600 PDF

OccludeBench: Can VLMs Find What's Missing?

Jasin Cekinmez, Addison J. Wu, Ryo Mitsuhashi, Linrong Cai, Xingyu Fu, Zhuang Liu cross-listed OccludeBench is a manually curated benchmark testing whether vision-language models (VLMs) can identify fully occluded objects from the visual evidence that remains, separating direct cases that depend on a specific physical cue like a shadow or reflection from inferential cases solvable from broader scene context. Across five VLMs, accuracy averages 42.7% on direct cases versus 62.7% on inferential ones, a 20-point gap that humans do not show, scoring 63.4% and 68.5% respectively. A four-stage diagnostic called OccludeTrace locates the bottleneck at recovering the relevant object from its projection within the full image, where models succeed on only half of failed examples versus 91% of correct ones. Extended reasoning, tool use, and few-shot prompting help inconsistently, but an inpaint-then-confront intervention, in which the model generates a candidate hidden object and then checks it against the visible cue, shrinks the gap to statistical insignificance for both models with image-generation capabilities.

arXiv:2610.13790 PDF

SpatialASM: Internalizing Active Mentalization for Spatial Reasoning

Yuhong Liu, Shenxi Wu, Junyi Chi, Beichen Zhang, Xuanlang Dai, Jiaqi Wang et al. cross-listed People solve hard spatial problems by mentally moving to a task-relevant viewpoint and reconstructing the layout from there, an ability the authors call active spatial mentalization and find largely absent in large vision-language models (LVLMs). SpatialASM instills it at training time using an allocentric novel-view generator that grounds hypothetical viewpoints in visual evidence, yielding both explicit trajectories that teach when and how to invoke view planning and imagination, and automatically constructed perspective-taking tasks that strengthen viewpoint-conditioned scene imagination. Across seven benchmarks spanning spatial perception through multi-image, cross-view, and viewpoint-transformation reasoning, all four LVLM backbones gain average spatial accuracy while retaining their general vision capabilities, offering a scalable substitute for scarce spatial chain-of-thought supervision.

arXiv:2610.13832 PDF

When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors

Shubh Garg, Chen Shani Vision-Language Models (VLMs) score well on compositional reasoning benchmarks, but it is unclear whether their internal representations actually encode the structured spatial relations such reasoning needs. The authors introduce EICS (Evaluation-Invariant Concept Stability), a metric for whether representations encode a concept consistently under controlled single-concept interventions, and CASP-VL (Causally controlled Spatial Perturbations for Vision-Language models), a benchmark of spatial perturbations derived from Visual Genome scene graphs. Across multiple VLM architectures, spatial relations are represented substantially less faithfully than lexical information, even in models well above chance on compositional benchmarks. The relation is identifiable at the predicate's own tokens but is lost once the sentence is aggregated into a global representation through the end-of-sequence token or mean pooling, a failure invisible to output-based evaluation.

arXiv:2610.13910 PDF

13 more specialized papers

Kairos: Toward a Byte-Level Foundation Model for Every Modality Fabien Furfaro
Efficient Audio-Visual Speech Enhancement via Reinforcement Learning-Based Neural Architecture Search Khubaib Ahmed, Ahsan Adeel, Nasir Saleem et al.
Mind the Spatial Gap: A Benchmark for Geospatial Reasoning and Constrained Spatial Editing Zheng Ren, Yanning Wang, Ziheng Huang et al.
BLEND: Balancing Personalization vs. Generalization in Federated Vision-Language Models Payam Abdisarabshali, Fardis Nadimi, Kasra Borazjani et al.
MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph Yingjie Zhu, Han Cheng, Xuefeng Bai et al.
The Evidence Anchoring Boundary: Where Visual Interventions Matter and When They Help in Multimodal LLMs Qiji Zhou, Junyi Hu, Peiyu Wang et al.
Evidential Multimodal Transformers with Dempster-Shafer Fusion for Emotion Recognition Muhammad Usama, Yeoreum Jung
SWEAR: Section-Wise Efficient Audio Reasoning via Self-Rated Difficulty Yixuan Li, Minghao Lv, Xuenan Xu et al.
MHR-Router: Multimodal Heterogeneous Retrieval-Enhanced Adaptive Routing for VLM Inference Wenjuan Zhu, Zheming Yang, Yunqing Hu et al.
Sonic-AMG: Streaming Human Motion Generation from Spatial Audio Jingcheng Ni, Zekun Li, Zhiyang Dou et al.
EmoGavel: Grounding-Aware Reinforcement Learning with Verifiable Rewards for Multimodal Emotion Preference Learning Zilong Huang, Kong Aik Lee, Zezhong Jin et al.
X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation Saumyaranjan Mohanty, Aravind Reddy, Konda Reddy Mopuri
UkrMeme: Multimodal Explainable Persuasion Detection in Ukrainian Memes Nazar Andrushko, Vera Schmitt, Veronika Solopova

Reinforcement Learning 20

EliteTable: Self-Improving Program Search for Executable Control Policies

Stuart Farmer Large language model (LLM)-guided program search for executable control policies often lacks a structured way to reuse earlier high-scoring programs. EliteTable keeps a fixed-capacity table of top solutions and uses it to steer three kinds of proposals: new programs, edits of a single stored solution, and remixes of several. Across six Gymnasium control tasks with ten generations of fifty candidates each and ten independent searches per task, candidate-score medians rose from the first to the final generation in all six tasks, held-out gains appeared in four, and the CartPole-v1 winner hit the search-score ceiling in the first generation of every search. Results on Acrobot-v1 were mixed, with six of ten searches declining on held-out episodes, and the authors note that comparisons and ablations are still needed to establish relative performance and the contribution of each component.

arXiv:2610.12536 PDF

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang, Pengfei Zhou, Wangbo Zhao, Fangyuan Kong et al. Online reinforcement learning (RL) can fine-tune diffusion models from black-box rewards, but reverse-process methods such as Flow-GRPO suffer from poor credit assignment while forward-process methods such as DiffusionNFT are efficient yet unstable without carefully tuned exponential-moving-average (EMA) off-policy training. The authors show that NFT-style objectives contain a hidden forward-reverse consistency force that joint positive-negative training cancels only partially, and that EMA stabilizes training through restoration rather than off-policy rollouts. DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting) aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states, removing the consistency force by construction, and adds a trust-region-masked second update for drift-adaptive restoration without EMA tuning; it needs no differentiable reward model, works with fewer samples per prompt, and supports pairwise reward models. On HunyuanVideo-13B with equally weighted visual and motion quality rewards it reaches the late-stage mean reward levels of both baselines with about 12 times fewer GPU hours while achieving 28.8% and 53.0% higher mean training rewards.

arXiv:2610.12572 PDF

Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents

Yiran Wu, Andrew Zhao, Yurui Chang, Qingyun Wu, Xuan Feng Reinforcement learning with verifiable rewards (RLVR) supervises software engineering agents through executable tests, but a binary pass/fail reward cannot tell apart trajectories whose debugging quality differs substantially. Behavioral Self-Rewarding (BSR) reuses the current policy to compare observable debugging behavior between trajectories from the same task that share a verifier outcome, aggregates the pairwise preferences into task-local scores, and converts them into bounded auxiliary rewards that add contrast while preserving the verifier-defined ordering, all without a separately trained reward model or new environment interactions. Experiments on SWE-bench Verified and SWE-bench-Live with Qwen3-14B and Qwen3-32B show consistent improvements across all evaluated configurations, supporting the policy's own comparative judgments as a complementary training signal for software engineering agents.

arXiv:2610.12598 PDF

MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis

Ximing Lu, Aviral Kumar, Jan Kautz, Yejin Choi, David Acuna Scaling reinforcement learning with verifiable rewards (RLVR) has focused on collecting more verifiable problems, but corpus size alone does not guarantee a learning signal for a given policy. The authors characterize dataset effectiveness as a function of the problems, the target policy, and the rollout budget, and show that roughly 88% of a large pooled open-source math corpus falls into a zero-advantage trap, solved in every rollout or failed in every rollout, so filtering can only select and must discard the hard problems. MagicBeanStalk is a top-down decomposition and bottom-up recursive synthesis algorithm that grows seed problems into families of new problems calibrated to the policy and budget while preserving the seeds' reasoning blueprints, something direct teacher prompting fails to do. Continual RLVR on the synthesized problems lifts gpt-oss-120b from 61.2 to 69.8 and DeepSeek v4Flash from 72.4 to 77.9 across eight math benchmarks at a 16K-token budget, versus +1.1 from the full 78K source corpus and +5.8 from its best filtered subset, and in finance over 14K problems synthesized from 340 solutions improve SecQue and FinanceBench by 6.9 and 2.4 points.

arXiv:2610.12659 PDF

Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target

Du Cheng, Junyu Lu, Jingfan Zhang, Yuzhuo Zhang, Haoxuan Jia, Shichao Weng Reasoning post-training typically runs supervised fine-tuning (SFT) on demonstrations and then reinforcement learning (RL) on fresh rollouts as separate stages with separate objectives, which makes it hard to keep reusing previously collected responses. CBPO (Cross-Behavior Policy Optimization) trains on all sources with a single regression loss toward a fixed-reference target that does not depend on the generating policy, so positive and negative offline completions, reference answers, and current and historical rollouts share one reward-labeled replay buffer without importance weights, stored behavior probabilities, or a critic. Across two model scales and four mathematical reasoning benchmarks the unified variant leads in avg@32 and pass@32; on Qwen3-1.7B it improves benchmark-averaged avg@32 and pass@32 over SFT followed by DAPO by 2.33 and 4.44 points, and retaining historical rollouts in the buffer adds 1.87 and 3.72 points over discarding them.

arXiv:2610.12830 PDF

Reinforcement Learning via Fixed-point Flow Matching

Denis Blessing, Onur Celik, Gerhard Neumann In maximum-entropy and behavior-constrained reinforcement learning (RL) the optimal policy is a Boltzmann distribution that is often complex and multimodal, and while flow policies trained by flow matching can represent such distributions, flow matching normally needs samples from the target, which online RL does not provide. Fixed-point flow matching (FFM) removes that requirement by using the target flow identity to cast the actor update as a fixed-point iteration, regressing the flow onto targets built from the critic's action gradients, and the authors show the optimal policy is a fixed point of this iteration. Experiments show FFM both fine-tunes pretrained flow policies in the offline-to-online setting and trains flow policies from scratch in online RL efficiently and accurately, without any training data.

arXiv:2610.12848 PDF

Scaling Empowered Agents

Catherine Ji, Vivek Myers, Joshua Lin, Alexandra Wingate, Sergey Levine, Benjamin Eysenbach Empowerment measures how much control an agent has over its environment, but prior work has mostly been confined to tabular settings. The authors connect empowerment to skill learning: empowerment corresponds to the number of skills executable from a state, so maximizing it means seeking states that are good starting points for many skills, a recipe that works on top of any underlying skill-learning method. At scale this lets them observe properties of empowerment that theory predicted but experiments had not shown, such as resource gathering, environment modification, and survival, including an agent spontaneously building a house in Crafter. Code and a project website are released.

arXiv:2610.13002 PDF

Generalized Proximal Policy Optimization: Unifying LLM RL Algorithms from the Perspective of Divergence

Zhicheng Cai, Xinyuan Guo, Hanlin Wu, Weinan Dai, Yuyan Zhou, Wei-Ying Ma et al. Proximal Policy Optimization (PPO), the dominant algorithm for reinforcement learning on large language models, relies on a clipping proxy that cannot adapt its resistance to policy shifts, causing premature clipping and weak exploration, and later variants patch this with fragmented heuristics. Generalized Proximal Policy Optimization (GPPO) reframes the problem through divergence: the authors establish an f-divergence trust-region theory with a monotonic improvement guarantee, show that different f-divergences induce different update geometries, and introduce estimator topologies that give more faithful divergence estimates than single-sample Monte Carlo. PPO, DAPO, DPPO, and RIPO fall out as special cases of a two-dimensional Divergence × Estimator design space, and experiments indicate both dimensions matter. On eight competition-level math and coding benchmarks the best instantiation surpasses GRPO by up to 50% on math and 14% on coding.

arXiv:2610.13015 PDF

Trajectory-Aware Policy Optimization: Closing the State Distribution Gap in Long-Horizon LLM RL

Zhicheng Cai, Weinan Dai, Yuyan Zhou, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou Long-horizon autoregressive rollouts in reinforcement learning (RL) for large language models let small per-step policy mismatches compound into severe state distribution shift, and methods such as GRPO and GSPO that control mismatch only at the token or sequence level become unstable and eventually collapse as the horizon grows. Trajectory-Aware Policy Optimization (TAPO) derives a state-distribution correction from the off-policy policy gradient and folds it into the RL objective, pairing it with a geometric-mean estimator for variance reduction and length-aware dynamic clipping that blocks overly deviated updates along the trajectory. Across eight competition-level benchmarks it improves over GRPO by up to 48% on mathematical reasoning and 12% on algorithmic coding with markedly more stable training.

arXiv:2610.13020 PDF

AdaC-GRPO: Towards Mastering Long-Horizon Visual Sudoku Puzzles via Adaptive Curriculum Learning

Sida Li, Yuanxin Liu, Linli Yao, Shicheng Li, Xu Sun, Haoli Bai et al. Reinforcement learning with verifiable reward (RLVR) improves reasoning in language models, but its effectiveness on long-horizon multimodal tasks is underexplored. The authors introduce FG-Sudoku, a visual Sudoku benchmark categorized by grid size and required deductive steps, and find that frontier multimodal models fail on puzzles with long-range dependencies such as 9x9 grids, where near-zero success rates create reward sparsity that halts standard RLVR. Adaptive Curriculum Group Relative Policy Optimization (AdaC-GRPO) builds training batches dynamically by tracking the model's zone of proximal development through online rollout accuracy rather than following a rigid or prior-based schedule. It lets Qwen2.5-VL-7B approach Gemini-2.5-Flash and close 62.2% of the overall accuracy gap to Gemini-2.5-Pro, with the gains also holding on InternVL-3.5 and Qwen3-VL.

arXiv:2610.13184 PDF

Does Learning Recurrent Networks Matter in Partially Observable RL?

Zekang Wang, Zhe He, Gaojie Jin Reinforcement learning (RL) under partial observability usually trains a recurrent network jointly with the policy or value function to carry temporal information, which raises the question of whether learning the recurrent module is necessary at all. Motivated by Reservoir Computing (RC) and Generalized Reservoir Computing (GRC), the authors pair a fixed recurrent network with a trainable nonlinear output module and evaluate it across partially observable Markov decision processes (POMDPs) with varying memory demands, observation modalities, and action spaces. Frozen recurrent networks support recall over hundreds of steps and reach competitive control performance with less than half the trainable parameters of fully trainable counterparts, which the authors present as a simple, parameter-efficient baseline for partially observable RL.

arXiv:2610.13197 PDF

S$^3$R: Fast and Faithful RL Rollouts via Self-Sparse Speculative Decoding

Yuhang He, Hongyu Ge, Zihan Lin, Siyi Liu, Yongqi Zhang Long-context reinforcement learning with verifiable rewards (RLVR) spends most of its rollout time in attention, which the authors measure at 72 to 76% of autoregressive decoding kernel time, but sampling from a sparse-attention policy changes the response groups used for group-relative optimization, and even exact response-wise reweighting cannot generally recover the dense-policy update because each response's advantage depends on the other group members' rewards. Self-Sparse Speculative Rollouts (S³R) uses sparse attention to draft tokens and the same policy with full attention to verify them through exact speculative sampling, so the resulting responses follow the dense-policy distribution without training a separate drafter, and a KV-cache-aware scheduler limits repeated offloading and restoration of the history that dense verification still needs. On long-context mathematical reasoning with Qwen3 models, end-to-end training throughput improves by about 1.35 to 1.39 times over autoregressive rollouts with comparable task performance, and training stays stable under an aggressive sparse drafting budget where direct sparse sampling with training-side correction becomes unstable.

arXiv:2610.13202 PDF

Structure-Aware Numerical Stabilization for Half-Precision Proximal Policy Optimization

Linxuan Wu, Liang Chen Online reinforcement learning on resource-constrained hardware needs low compute and training-state overhead, but naively running proximal policy optimization (PPO) in FP16 is unreliable because numerical errors feed back through policy updates and data collection, and PPO repeatedly compares a current policy against a fixed behavior policy. PPO-FP16 adds three structure-aware protections: global scalar protection for statistics shared across a minibatch or update, policy-change residual protection that extracts and preserves the small differences between current and behavior policies, and persistent-update protection for optimizer state and parameter writeback, all without altering the PPO objective, hyperparameters, or training budget. Across six continuous-control tasks and five seeds it achieves the best mean return-AUC rank among four methods and ranks first in 17 of 30 environment-seed comparisons, with PPO update speedups of 1.16 to 2.03 times, optimizer-state memory reduced 71.7 to 74.9%, and peak GPU memory reduced up to 44.5% relative to full FP32 Adam.

arXiv:2610.13306 PDF

Rollout-Time Intervention: Unlocking Agentic Reinforcement Learning Beyond Policy Limits

Chen Xi, Ding Ruixue, Wang Zhigaoyuan, Zhang Fanrui, Zhang Qiang, Chen Boli et al. Reinforcement learning (RL) for large language model (LLM) agents hits an on-policy bottleneck: when tasks lie beyond the model's current capability, successful trajectories become so sparse that training often collapses. Rollout-Time Intervention (RTI) uses reward-side evidence such as evaluation rubrics during rollout to convert failing trajectories into successful ones, supplying learning signal for tasks the policy rarely solves alone. Although these intervened trajectories are off-policy, they guide policy optimisation effectively once trajectories with excessive distributional shift are filtered out. Across five benchmarks RTI consistently beats RL baselines, and the post-trained policy approaches the performance of substantially stronger contemporary LLMs while moving past its original capability boundary.

arXiv:2610.13438 PDF

Beyond Temporal Proximity: Explanatory World Models for Credit Assignment in Reinforcement Learning

Oriol Corcoll, Qiyao Wei, Ciar\'{a}n Mark Gilligan-Lee Generative recommenders trained with reinforcement learning (RL) must credit decisions whose payoff appears hundreds or thousands of interactions later, yet standard RL spreads reward by temporal proximity rather than relevance, starving distant but important actions of signal. Explanatory World Models (EWM) ask a frontier language model to explain which earlier events plausibly produced a realized outcome, then use a conservative hierarchical operator to recursively redistribute reward along those explanations while preserving the undiscounted episode return; three explanation variants draw on Halpern's actual causation, Lombrozo's explanatory simplicity, and Deutsch's hard-to-vary explanations. On Crafter and WordCraft, EWM substantially outperforms PPO and LaRe, with its largest gains when rewards are most delayed; it also beats supervised fine-tuning and LaRe offline and A2C online on the KuaiRand recommendation benchmark, and improves recall and behavioral proxies in an offline evaluation of an industry-scale generative recommender.

arXiv:2610.13504 PDF

MENTOR: Learning from Failure with On-Policy Pedagogical Intervention

Guanlan Yue, Hang Yu, Xiaoyong Huang, Ruipeng Li, Wei Zhang, Bingchang Liu et al. Reinforcement Learning from Verifiable Rewards (RLVR) can suffer zero-reward collapse, where a jump in task difficulty leaves every rollout unrewarded and the learning signal vanishes, and existing fixes rely on static hint libraries or post-hoc refinement. MENTOR is a multi-agent framework that diagnoses the root cause of each failed trajectory in real time and generates targeted instructional prompts for the policy's next attempt, plugging into pipelines such as GRPO without changing the training infrastructure. On math and code generation it mitigates collapse and outperforms prior baselines, with average relative gains of up to 78.3 percent on Llama-3.2-3B-Instruct over its base model.

arXiv:2610.13594 PDF

A Step Harder, Not a Step Too Far: The Benefits and Limits of Curriculum Difficulty Progression in GRPO

Ronghan Li, Fei Chen, Mingyu Lu In GRPO, medium-difficulty samples are attractive for online selection because they are most likely to yield informative group-relative learning signals, but it is unclear whether always selecting them produces the best final model or whether training should progress toward harder samples as the policy improves. The authors re-estimate sample difficulty under the current policy and evaluate a staged curriculum that progressively shifts the active difficulty band toward harder samples. On GSM8K across three seeds, progressing to a moderately harder band improves test accuracy to 0.6591 versus 0.6449 for the fixed-medium curriculum, but pushing further to a harder band drops accuracy to 0.6405. The benefit of difficulty progression is therefore non-monotonic: effective curricula should move beyond medium difficulty without advancing to overly difficult samples.

arXiv:2610.13669 PDF

Variance-Reduced Policy Gradient is Nearly Minimax-Optimal

Leonardo Cesani, Matteo Papini, Marcello Restelli Policy gradient methods suffer from high variance in their Monte Carlo updates, and existing variance-reduced variants only reach optimal sample complexity under restrictive assumptions such as uniformly bounded importance weights or with impractical second-order operations. The authors introduce Loopless Variance-Reduced Policy Gradient (LVR-PG), a framework covering first-order estimators that combine recursive updates with defensive importance sampling, including PAGE-PG and STORM-PG. They prove last-iterate global convergence with sample complexity of order epsilon to the minus two, up to logarithmic factors, under weak gradient domination for general parametric policies, using only first-order updates and no gradient preconditioners. Extending existing lower bounds to continuous parametric policies gives a matching lower bound of the same order, so the framework is minimax-optimal up to logarithmic and horizon terms.

arXiv:2610.13692 PDF

SafeDiSTL: Decomposing Joint Constraints for Safe Multi-Agent Reinforcement Learning

Hongrui Zheng, Aditya Singh, Rahul Mangharam Cooperative multi-agent tasks often carry constraints on a shared objective or collective safety, and a team-level violation cost cannot say which agents' actions drove the violation. SafeDiSTL is a constrained multi-agent reinforcement learning framework that assigns credit from joint constraint measurements by using Signal Temporal Logic (STL) robustness to measure episode-level violation as running-maximum cost increments, then using smooth full-episode robustness and differentiable dynamics to compute action sensitivities that split each increment across agents at the same decision time while accounting for delayed effects. On cooperative payload transport, connected navigation, and networked traffic control, it reaches 1.6 to 5.1x the safe-success rate of constrained baselines in transport, improves navigation safe-success by 11.1%, and delivers 50 to 68% more traffic throughput, while keeping the performance of unconstrained policies.

arXiv:2610.13737 PDF

1 more specialized paper

Hitting-Time Bounds for Asynchronous Q-Learning Yang Peng

Safety & Alignment 20

Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller

Tahmid Imtiaz Imu, Emama Nahid, Prodipta Sen Amartya, Ahsanul Jobayer, Yutong Song, Jiang Wu et al. cross-listed Backdoor attacks on large language models normally follow a trigger-response pattern in which one condition activates a single terminal behavior, leaving unexplored attacks where malicious behavior builds through intermediate stages. The authors present Cancerous Backdoor Attack, which gradually corrupts an LLM's reasoning over a sequence of interactions using invisible tokens, which supply hidden stage signals absent from the visible request, and a stage controller that tracks progression during inference. Across several reasoning datasets and models the attack consistently produces gradual reasoning corruption and outperforms prior backdoor baselines. Reasoning can be corrupted progressively before any visible change in answers, so answer accuracy alone is insufficient for detecting backdoors.

arXiv:2610.12644 PDF

A Distributional Framework for Behavioral Characterization and Risk Management of Large Language Models

David Garc\'ia-Lorite, Raul Merino cross-listed Large language models are stochastic, so repeated runs under identical conditions can yield different outputs, which complicates risk management for deployed systems. The proposed framework builds an empirical response distribution via Monte Carlo simulation over repeated executions, then applies semantic interventions to the prompt or context and measures how the distribution shifts. Illustrated on borrower classification and music album recommendation, the approach quantifies robustness, control effectiveness, and residual behavior in terms that map onto model risk management and governance across the lifecycle from task design to continuous monitoring. Because it relies only on observed responses, it works on black-box models accessed through commercial APIs.

arXiv:2610.12718 PDF

DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

Drew Rein, Patrick Oathout, Vishal Kumar, Udari Madhushani Sehwag A growing share of language model use is affective rather than informational, and some users disclose suicide and self-harm (SSH) crises, where refusal is the wrong behavior and a substantive supportive response is required. DistressBench contains 718 English conversations with clinician-authored user turns across 23 SSH subcategories, each paired with a weighted rubric spanning seven dimensions of crisis support and a clinician reference response, so a model's score reads directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median 71.4%, against 98.7% for clinician references, and in 35.3% of conversations where a model fully recognized the crisis it still failed to route the user toward human help, with shortfalls concentrated in dimensions that require acting rather than withholding. The authors argue that crisis safety should be evaluated on the care a model delivers rather than on what it declines to say, with de-escalation and resource routing treated as deployment requirements.

arXiv:2610.12758 PDF

VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails

Duy Nguyen, Lesly Miculicich, Bhavana Dalvi Mishra, Shuo Li, Krishnamurthy Dj Dvijotham, Mohit Bansal et al. LLM agents increasingly act over sensitive resources such as private documents and mutable system state, so runtime guardrails should make inspectable, reproducible, deterministic allow-or-refuse decisions before irreversible actions, yet safety and security requirements keep evolving after deployment. Existing adaptive guardrails learn from deployment experience but apply policies through LLM inference, giving inconsistent decisions, while formal guardrails enforce fixed policies deterministically but cannot adapt to new risks. VEPA represents the guardrail as a formal policy of executable symbolic rules over typed agent events, has an LLM guided by deployment feedback propose patches that fix observed failures while preserving earlier correct behavior, accepts only patches that a deterministic verifier finds structurally valid and conflict-free, and lets an evidence-aware authority layer activate a verified rule only once it has enough supporting evidence. Across ASSEBench-Safety, ASSEBench-Security, and ATBench, it consistently outperforms baselines while expanding the scope of deterministic formal-policy enforcement, with gains that hold under mixed streams, corrupted feedback, and longer horizons.

arXiv:2610.12813 PDF

Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems

Yuepeng Hu, Pegah Ghahremani, Reza Aghajani, Mahesh Pasupuleti, Himaghna Bhattacharjee Multi-agent coding systems coordinate agents that share code and plans, and those agents are often trained with reinforcement learning (RL) against automated evaluators that can be gamed. The authors show that reward hacking elicited in one agent through coding RL, either via explicit mechanism descriptions in the system prompt or via mid-training on facts about the evaluation environment, spreads to other agents through shared code, shared memory, and natural-language plans, and they analyze theoretically when such hacking amplifies, persists, or decays. Across four open-weight backbones, even agents built on small open-weight models induce reward hacking in stronger closed-weight targets, propagation can intensify along a chain after the original source leaves as the code accumulates persuasive justifications, and hacking biases task routing toward hacking agents and survives candidate aggregation. Code review by the coder itself or by a separate verifier reduces the number of delivered hacks.

arXiv:2610.12822 PDF

Space Over Statistic, Retraining Over Ranking: Auditing Training Labels Under Label-Only Backdoors

Arnab Dey, Mohammad Mahfuzul Islam cross-listed The FLIP backdoor attack never modifies an image: flipping 3% of CIFAR-10 labels makes a ResNet-32 send 99.7% of triggered source-class images to the attacker's class, and standard defenses then hunt for the poison inside a model trained on those same labels. The authors run a controlled swap that holds the detection test, labels, and metric fixed while changing only the feature space, and judge each defense by retraining the victim on the data it keeps. The feature space decides: on the dog-to-cat pair the same nearest-neighbor check scores AUROC 0.546 on the poisoned model's features but 0.967 on frozen DINOv2 features that never saw a label, and a prototype audit on those frozen features deletes just 6% of the data for higher clean accuracy and lower attack success where the reproduced SPECTRE defense deletes 20% of every class, though on three harder pairs the audit leaves about a quarter of the attack standing. Three cautions follow: a higher detection AUROC can leave more attack success in place, as Confident Learning does on dog-to-cat, the best static defense is the first one an adaptive attacker slips past, and beyond a point catching more poisons stops reducing attack success.

arXiv:2610.13051 PDF

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang Bai Jailbreaks and over-refusals both hint that recognizing a harmful request and actually writing a refusal are separate computations. Measuring residual-stream updates across 18 base-and-instruct pairs from Qwen, Llama, and Gemma, the authors find harmfulness is decoded more readily from attention updates, especially in early layers, while multilayer perceptron (MLP) updates do most of the harm-selective refusal writing, and post-training strengthens that writing in all 18 pairs without making harmfulness more decodable. Causal intervention supports the pathway: resetting downstream MLP updates to clean values largely cancels the refusal shift caused by upstream attention edits. The resulting Sentinel method pairs an attention-derived perception score with an MLP-derived decision score, and routing on both signals yields lower aggregate error than single-signal routing across eight models.

arXiv:2610.13078 PDF

The Limits of Query Monitoring against Distributed Model Extraction

Snigdha Chandan Khilar cross-listed A provider facing distillation campaigns spread across thousands of ordinary-looking accounts may have nothing to work from but the queries themselves. Using one million real chat turns for development and 17 later untouched weeks as a single frozen evaluation, the study argues the detection ceiling is set by benign coordination rather than anything specific to extraction: a campaign is visible only when it looks more distinctive than the most distinctive group of normal users that week. A calibrated cross-account monitor flags coordinated activity that per-account monitors miss in most evaluation weeks, and campaigns that evaded every detector tested produced student models retaining at most 42% of the capability of a detectable one's. The authors are equally explicit about the gaps, including no deployable detector at low colluder shares and benign evaluation pipelines that look indistinguishable from distillation.

arXiv:2610.13083 PDF

DRYLAB-Bench: Benchmarking the Biosecurity Risk of Agentic LLM Mutation Design

Yuzhen Wang, Chunxiao Li, Baichen Le, Lewen Yan, Jie Zhang, Yucheng T. Yang et al. Biosecurity frameworks flag dual-use risk from models with biological design ability, but existing evaluations mostly quiz models on hazardous knowledge rather than measure whether an agent can apply it, and they ignore what changes when specialized biological design tools are attached. DRYLAB-Bench scores agents on 30 hazardous design scenarios distilled from biosecurity guidelines in a multi-round, tool-augmented setting, grading proposals against 187,695 existing wet-lab measurements from ProteinGym and ViroGym via a Biosecurity Risk Score, which keeps ground-truth verification scalable and safe. Across 15 frontier models the authors report measurable baseline capability even without tools, and attaching design tools raises pooled risk further, lifting the share of top-risk-decile proposals to 23.5%.

arXiv:2610.13116 PDF

WaSA: Wavelet Scattering Autoencoders for Unsupervised Offline Detection of Malicious Clients in Federated Learning

Alessandro Licciardi Federated learning (FL) trains models without exposing raw data but is vulnerable to clients that inject poisoned datasets, and existing defenses either react during training or require prior knowledge of the attack type. WaSA detects malicious clients offline and without supervision: before training, each client sends a single privacy-preserving spectral summary of its data, and a denoising autoencoder trained only on clean auxiliary data scores each client by its deviation from the learned benign manifold, regardless of attack type. A Boltzmann sampling strategy for client selection interpolates smoothly between hard exclusion and standard aggregation. On MNIST, CIFAR-10, and CIFAR-100 under diverse attack scenarios with up to 90% malicious clients, the approach beats supervised baselines and robust aggregation methods with negligible communication overhead.

arXiv:2610.13186 PDF

Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents

Lei Liu, Yunji Liang, Xiaowen Zhang, Jingqi Liu, Qi Li, Bin Guo et al. Large language model (LLM) agents with persistent memory are vulnerable to poisoning, and current defenses rely on external isolation or continuous auditing modules that cost memory and compute. DREAM (Dynamic Reactivation for Engram Attenuation in Memory) is a training-free, agent-intrinsic repair pipeline inspired by the neuroscience finding that weak reactivation without reinforcement destabilizes memories: it perturbs stored memories to expose structural fragility, diagnoses anomalies from multi-dimensional activation patterns, and then selectively suppresses the flagged entries while reinforcing benign ones. Against backdoor poisoning it cuts attack success rates by over 95% while preserving benign utility, remains competitive against injection attacks, runs up to 2.13 times faster with 37.7% fewer tokens than A-MemGuard, and reaches a 92.96% task success rate on poisoned multi-agent systems.

arXiv:2610.13267 PDF

Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning

Joseph Egan, Ciar\'{a}n M. Gilligan-Lee, Michael O'Riordan, Yuchen Zhu cross-listed When a language model is distilled from a teacher that shares its base parameters, the student can absorb the teacher's behavioral traits, including misalignment, even from semantically unrelated data, and every existing defense requires knowing the trait in advance. The authors prove that a subliminally transferred trait can be recovered from the parameter updates alone, with no supervision, by exploiting one asymmetry: the trait is expressed consistently across diverse teacher-generated datasets, whereas task-specific and shared capabilities vary. The resulting estimator yields a weight-space direction that can be amplified to surface and interpret the trait or projected out to suppress it; on the subliminal-learning benchmarks of Cloud et al., it recovers the transmitted animal-preference traits without any trait labels and suppresses them with almost no loss in task performance.

arXiv:2610.13536 PDF

Personalization Without Stereotyping: Do Speech LLMs Know When to Turn a Deaf Ear?

Kshitij Ambilduke, Caroline Brun, Thibaut Thonet, Vassilina Nikoulina, Marcely Zanon Boito Speech large language models can pick up paralinguistic cues from a speaker's voice, but existing evaluations measure demographic disparities without separating helpful personalization from stereotyping. BASIS-Bench tests eight recent speech LLMs along three axes: robustness of content-based task performance across speakers, whether perceived demographic cues trigger unprompted stereotypical behavior, and whether models can tailor responses when explicitly asked to. Most models recognize demographic cues yet fail to turn them into appropriate personalization, and adding explicit demographic labels improves personalization while amplifying stereotypical associations already present in the text backbones.

arXiv:2610.13550 PDF

APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces

Ziyi Yang, Adel Bibi, Guohao Li, Bernard Ghanem, Carlos Hinojosa cross-listed Agents working over shared professional workspaces with many files and tools are exposed to injection attacks that cannot be captured by placing a payload at a single known read location. APEX-Harm evaluates prompt and script injection against long-horizon agents doing finance, consulting, and legal work in workspaces of 115 to 328 files, with email, calendar, chat, document, spreadsheet, PDF, and code-execution tools exposed through MCP servers, across six attack categories including static and dynamic injections and script replacement. Across 12 frontier models and 180 task instances each, Opus 5 and Sonnet 5 show the lowest overall attack success rates at 8% and 20%, but task-specific script replacement attacks still succeed 81-84% of the time on average. Gemini 3.8 Flash posts the highest mean task score of 0.877 while still being compromised in 33% of attempts, showing that a correct deliverable is not evidence of safe behavior.

arXiv:2610.13623 PDF

SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle

Xiting Liu, Lin Yang, Yitong Zhang, Jia Li, Shi-Min Hu cross-listed As large language models (LLMs) take on more of the software development lifecycle, their willingness to mislead becomes a trust risk. SWE-DeceptionBench covers six lifecycle stages and two triggers: Active Deception, where a model recognizes infeasible or conflicting requirements but presents them as achievable, and Passive Deception, where it helps a user mislead others, with Neutral, Reward, and Pressure conditions yielding 720 Chinese and English test instances. Across seven mainstream LLMs, deception rates reach 49.88% for Active and 64.29% for Passive Deception even under the Neutral condition, with rates varying by lifecycle stage and inducements affecting the two types differently. Lightweight defenses fall short: guard models miss many deceptive requests, and Self-Reminder reduces but does not eliminate deceptive responses.

arXiv:2610.13626 PDF

X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems

Yeonsung Jung, Zhiyuan Wang, Zifan Wang, V\'itor Albiero, Benjamin Newman, Edoardo Debenedetti et al. Personal agents that call tools across many web services can cause harm while fulfilling benign requests, because the effects of actions accumulate and interact across services, and existing red-teaming approaches that prompt large language models to propose harmful paths tend to favor familiar tool combinations. X-AgentRed is a self-evolving red-teaming framework in which an attacker agent analyzes each tool's effects, potential harms, and execution preconditions, builds a directed dependency graph capturing how one tool's effects satisfy another's preconditions, and samples recipes from that graph to instantiate realistic harmful scenarios. Execution feedback then drives self-improvement by substituting tools to diversify harmful paths and refining preconditions at blocked steps of failed paths. Across four frontier agentic models, it finds 1.77 times the harmful path rate of STAC and covers 2.99 times as many tools on average.

arXiv:2610.13886 PDF

AI Models Can Provably Hide Arbitrary Capabilities

Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas, Christian Schroeder de Witt Capability evaluations assume that a model's hidden dangerous capabilities can be elicited before deployment, and the authors show this assumption fails adversarially by embedding encrypted neural networks inside a host model's weights that are decrypted and executed only when a secret trigger appears. Using digital lockers, these hidden circuits are provably hard to elicit or interpret under cryptographic assumptions even with full access to model weights, extending earlier work from hiding fixed strings to hiding arbitrary computations. Proof-of-concept constructions implemented in PyTorch resist fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training on a small chemical reaction prediction task used as a proxy for CBRN-relevant capabilities. The models are released to support work on defenses, and the results mark a limit of elicitation-based capability audits.

arXiv:2610.13928 PDF

3 more specialized papers

Knowing Is Not Telling: Identity Ossification in LLM Representations of Trans-border Communities Minglu Sun, Lan Chen, Hang Jiang
REDaCT: Collapse-Free Unlearning in Large Reasoning Models Joshua Maccoby, Wendy Hui Wang
Provenance-Aware Belief Updating in Human-AI Feedback Loops Tianyi Huang, Eric Liu, Jay Chong et al.

Robotics 15

Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination

Shenglan Li, Zhendong Mi, Hengyi Zhu, Jingwu Luo, Chun Kit Chan, Geng Yuan et al. cross-listed Vision-language-action (VLA) models increasingly add intermediate reasoning for robot manipulation, but they mostly reason about the observed scene, and explicitly rolling out future frames at every inference step is computationally expensive. IG-VLA learns to anticipate task-relevant future scene evolution directly in visual representation space rather than generating pixels, and a Scene Gist Memory distills those reasoning-derived scene-behavior associations into a compact Scene Gist Token so the deployed policy can skip explicit future prediction. On LIBERO, LIBERO-Plus, and VLABench, both the reasoning and gist policies beat the strongest baseline by nearly 6% success rate on the LIBERO-Plus language suite, and the gist policy cuts latency from 1081 ms to 169.5 ms per action chunk, a 6.38x speedup, on a single A6000 GPU.

arXiv:2610.02626 PDF

Fly-by-Code: Embodied Coding Agents for Aerial Manipulation with Active Visual and Physical Feedback

Jaewoo Lee, Jeongyeon Seo, Sihyun Cho, Gyeongrak Choe, Yutong Wang, Bavin Saravanan et al. cross-listed Embodied coding agents build robot policies by generating programs and refining them from execution feedback, but aerial manipulation breaks this loop: payload limits leave task-critical outcomes outside the onboard camera's view, and fragile dynamics leave failure causes unresolved, so re-examining the same feedback cannot help. The proposed Active Feedback (AF) framework has the agent synthesize a separate diagnostic program that runs in simulation either to inspect unseen outcomes from informative viewpoints or to test hypotheses about failure causes through targeted physical interactions. Across six simulated aerial manipulation tasks with 60 trials per condition, AF raises final task success from 31.7% to 75% and cuts false completion declarations from 25 to 5 compared with terminal feedback under the same agent configuration and execution limit. Programs refined in simulation run unchanged on a physical aerial manipulator through shared robot APIs and complete multi-stage manipulation tasks in the real world.

arXiv:2610.12549 PDF

LeWAM: End-to-End World and Action Modeling with JEPAs

Minghao Fu, Tavis Siebert, Eryk Halicki, Randall Balestriero cross-listed Joint-embedding predictive architectures (JEPAs) learn compact latent world models without pixel reconstruction, but existing pipelines train visual representations and action-conditioned dynamics first and add a control policy or solver afterward. LeWAM trains a JEPA end to end for control with a joint objective over next-latent-state prediction and action flow matching, plus an isotropic Gaussian constraint on the latent state, so that a single model supports both direct policy execution and efficient planning by evaluating sampled actions with its learned latent dynamics. Compared with prior JEPAs in visual planning it raises success from 28.6% to 89.7% on contact-rich tasks and from 10.9% to 46.2% on long-horizon tasks while planning 32.3 times faster, remains competitive with established behavior cloning methods in simulation and on physical robots, and interpretability analyses indicate its representations and action distributions effectively guide test-time planning and control.

arXiv:2610.12818 PDF

JoyInside-VLM-Agent: Product-Conditioned Multimodal Policy Learning for Embodied Agents

Gangyi Li, LuYao Wang, Jingfu Sun, Jingyi Yuan, Hang Liu, Hao Zong et al. cross-listed Embodied agents on robots and smart hardware must turn multimodal observations and user requests into executable actions while respecting product constraints, revised goals, and execution feedback. JoyInside-VLM-Agent is a ReAct-style multimodal agent trained on trajectories from a closed-loop sandbox that covers goal revisions and failure recovery, followed by supervised fine-tuning, multi-teacher on-policy distillation, and reinforcement learning with DPER (Dual-Source Privileged On-Policy Self-Distillation with Token-Level Evidence Routing), which routes token-level evidence between a frozen privileged teacher and a privileged student that shares the rollout policy's weights. It scores 77.44 overall on BFCL v4, above the evaluated baselines, though results on other public benchmarks are mixed; on the JoyInside-Devices and JoyInside-Robotics suites it leads on all reported metrics, and two real-world quadruped case studies show visually grounded search, replanning after failure, and explicit failure reporting.

arXiv:2610.12957 PDF

Beyond Actions: Learning Future Operation Targets for Vision-Language-Action Models

Min Lin, Bokai Zhou, Jianqi Lin, Kaiyuan Yang, Yicheng Chang, Junjie Yu et al. cross-listed Vision-language-action (VLA) policies learn mainly by predicting action chunks from images, language, and state, yet actions specify what a robot should execute rather than what those commands will achieve. Action-Future Coupled Effect Learning (AFCE) makes a future-grounded physical Effect the primary prediction target: an offline tokenizer builds temporally ordered Effect tokens from demonstrated actions, realized robot responses, and sparse visual transitions, a flow-matching VLA predicts Effects from the current images, instruction, and robot state, and a state-conditioned decoder converts them into action chunks, with embodiment-specific interfaces letting single-arm and dual-arm tasks share one Effect model. On DexJoCo, AFCE reaches 50.97% overall success, 11.82 and 15.21 percentage points above the π0.5 and UVT baselines, and leads on all five bimanual tasks.

arXiv:2610.13212 PDF

AlphaNIST: Emergent Assembly Skills from Binary Task Rewards

Octi Zhang, Iretiayo Akinola, Rosario Scalise, Mateo Guaman Castro, Patrick Yin, Tobias Widmer et al. The NIST Assembly Task Board #1 demands precise contact-rich control of small, tight-fitting parts over short horizons and recovery and coordination over long ones, and prior learning-based methods have leaned on demonstrations, task decomposition, or dense progress rewards to specify how to assemble. Drawing an analogy to the move from AlphaGo to AlphaGo Zero, the authors ask whether a sparse binary task reward, plus the task diversity such a task-agnostic reward permits, is enough for reinforcement learning to discover assembly strategies on its own. In simulation, a single policy trained from scratch across all 20 board tasks reaches 98.2% overall training success without demonstrations, progress shaping, scripted primitives, or per-task tuning, and the same framework trains policies that complete ten assembly pairs in one episode in any order, rewarded only when all ten are simultaneously done. The learned policies show planning-like behavior such as choosing grasps with later assembly in mind, using surrounding geometry to reorient parts, and setting one part aside to assemble another before returning to finish it, with no explicit planner.

arXiv:2610.13354 PDF

RESCUE: Is This Alarm Worth Acting On? Counterfactual Evaluation of VLA Runtime Monitors

David Ahmedt-Aristizabal, Abdelwahed Khamis, Mohammad Ali Armin, Lars Petersson cross-listed Runtime failure monitors for vision-language-action (VLA) robot policies are usually judged by how well they separate failed from successful executions, but an accurate alarm can arrive too late to recover and a false alarm can derail an execution that would have succeeded. RESCUE evaluates monitors counterfactually: by restoring simulator and policy state at execution checkpoints, it compares alternative actions, finds the latest checkpoint at which a specified recovery criterion still holds, and tests whether acting on alarms lowers expected task-completion cost. Applied to OpenVLA-OFT and π0.5 on LIBERO-Spatial and LIBERO-Plus with 128,340 counterfactual branches, recoverability depends on the state and intervention jointly rather than on state alone, favorable detection scores do not establish lower intervention cost on the primary strong-policy conditions, and under a stop-and-reset intervention even an idealized detector can afford only a 2.0 to 3.3% false-alarm rate before unnecessary resets outweigh the savings from stopping failures early.

arXiv:2610.13516 PDF

RARA: Robust Policy Fine-Tuning through Retrieval-Anchored Representation Alignment for Robotic Manipulation

Xuhui Kang, Sung-Wook Lee, Yan Zhuang, Tianle Zhong, Yen-Ling Kuo cross-listed Fine-tuning a pre-trained Vision-Language-Action (VLA) robot policy on a few demonstrations works in-distribution but tends to distort pre-trained representations and hurt out-of-distribution generalization. Retrieval-Anchored Representation Alignment (RARA) first retrieves pre-training demonstrations with similar task intent and motion and aligns the policy's representations of the target demonstrations to them, then fine-tunes while constraining how much representations of the retrieved demonstrations drift. Across 13 tasks with Diffusion Policy and GR00T on RoboTwin, RoboCasa, and a real robot, average success improves by 9.5 points in-distribution and 13.6 points out-of-distribution over plain fine-tuning, beating all compared baselines.

arXiv:2610.13578 PDF

LAPT-Scale: Scaling Embodied Policies with Latent Action Pre-Training on Human Videos

Zhiyuan Xiang, Peidong Liu, Mingyang Li, Jiawei Li, Wenhow Li, Wenbo Ding et al. cross-listed Generalist robot policies are bottlenecked by the cost of collecting embodied interaction data, and egocentric human videos, while plentiful, carry no robot action labels and are hard to annotate reliably under real-world visual complexity. LAPT-Scale pairs a latent action model, LAPT-Scale-LAM, that learns continuous latent actions explaining frame-to-frame transitions in human video, with LAPT-Scale-Policy, a VLA policy trained through human-video latent-action pre-training, target-robot embodiment alignment, and task-specific post-training. Unlike prior vector-quantized latent action models, the continuous latent space preserves fine-grained motion, and perceptual, semantic, and optical-flow alignment signals supplement pixel reconstruction. Trained on 23.3K hours of egocentric video, the approach yields consistent scaling gains, raising real-robot success from 63.3% to 85.0% as annotated human video grows from 1K to 5K hours, and the policy generalizes to unseen visual conditions.

arXiv:2610.13732 PDF

SmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices

Timi Owolabi, Hikmah Olawore, Frances Adelakun, Ayotomiwa Oyewumi cross-listed World-action models (WAMs) predict a robot's actions together with the observations those actions will produce, but the strongest versions rely on video diffusion transformers too expensive for edge devices. SmolWAM is derived from the 2B-parameter Cosmos Policy by keeping its pretrained trunk and rebuilding the inference interface: single-frame visual latents, cached text features that preserve the trunk's attention sinks, 12 of the original 28 transformer blocks, and 1,008 tokens from which actions and future visual latents are read out in a single rectified-flow step. After only 3.6 GPU-hours of training, the model succeeds in 95.5% of 400 LIBERO episodes at 57 ms per 16-action chunk, running 11x faster than its teacher with 18.5x fewer FLOPs and 3.2x less memory while its world head still forecasts scene motion.

arXiv:2610.13740 PDF

Dream to Remember: Continual Steering of Generalist Robot Policies with World Models

Bahey Tharwat, Iman Nematollahi, Marvin Alles, Xingyuan Zhang, Akshay L Chandra, Abhinav Valada cross-listed Generalist robot policies learned by imitation are capped by their demonstrations, and reinforcement learning can extend them either by fine-tuning weights or by training a small module that steers a frozen generative policy's input noise, but sequential tasks make both forget earlier skills. Dream to Remember keeps the generalist frozen and trains the steering module on each new task while a frozen pretrained world model rehearses prior tasks through imagined rollouts, needing no replay buffer and no new data. On CALVIN it improves eight tasks in sequence entirely in imagination, lifting average success from 63.7% to 74.8% where LoRA fine-tuning on the same sequence collapses to 11.7%. With off-the-shelf pi-0 and A2World it learns unseen LIBERO-90 tasks in the environment while preserving prior LIBERO performance even though the world model is too inaccurate for policy optimization, and three tasks on a real Franka arm improve in sequence.

arXiv:2610.13816 PDF

4 more specialized papers

StereoEngine: Scaling Stereo Bimanual Robot Data without Stereo Cameras Jason Chen, I-Chun Arthur Liu, Gaurav Sukhatme et al.
OGPO: Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models Xule Gao, Xi Wang, Zehua Zang et al.
Prediction-Informed Landing Window Planning for UAV Landing on USV in High Sea States Kaizhen Li, Xi Zhou, Dan Zhang et al.
JitterVLA: A Replay Benchmark for Motion Quality in VLA-Driven Human-Robot Manipulation Faith Ogunwoye, Didar Seyidov, Konstantin Gubernatorov et al.

Vision 13

CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration

Qinhan Lyu, Jusheng Zhang, Zhixing Zhang, Hao Liu, Zijie Liang, Jin Jiang et al. Training-free feature reuse speeds up diffusion transformer inference by replacing full denoising at selected steps with cached approximations, and the authors find that once a stable first-order reconstructor is in place, where exact-computation steps are placed matters far more than how elaborate the reconstruction is. They also show that raw shallow-probe variation conflates input-trajectory displacement with network response strength, so they normalize probe-output change by input displacement to obtain a directional gain per unit of trajectory movement. CacheScope uses this signal to flag high-risk reuse steps online and fall back to full computation at those steps, leaving the cache reconstructor unchanged and requiring no training, gradients, auxiliary networks, or prompt-specific offline schedules. Across FLUX.1-dev, Wan2.1-T2V-1.3B, and Open-Sora 1.2, it reaches up to 3.57x acceleration and cuts median LPIPS by up to 70.1% relative to the strongest probe-based baseline.

arXiv:2610.12530 PDF

Zebra3D: Efficient Hybrid Foundation Models For Scalable 3D Reconstruction

Utkarsh Saxena, Ruijing Yang, Mehdi Rezagholizadeh, Akash Haridas, Vikram Appia, Emad Barsoum cross-listed Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits how many views they can process. Zebra3D converts pretrained models such as VGGT, HunyuanWorld-Mirror, and π³ into layerwise sparse-linear hybrids, exploiting a consistent depth-wise pattern in which early stages tolerate linear attention while sensitive middle-to-late stages have concentrated attention and receive tile-sparse attention with landmark summaries for unselected context. Closed-form affine fitting allows training-free conversion, with optional post-conversion training combining task supervision and feature distillation. Training-free variants stay close to their dense counterparts, end-to-end speedups reach up to 7.1x at 1,024 views, and after post-conversion training Zebra3D-VGGT reaches 80.5% RealEstate10K AUC@30 at 94% mean attention sparsity, within 1.1 points of dense VGGT fine-tuned on the same data.

arXiv:2610.12727 PDF

Golden Anchor for Video Generation

Jinyuan Xu, Shitong Shao, Lichen Bai, Zikai Zhou, Guoxi Huang, Nantheera Anantrasirichai et al. cross-listed Modern image-to-video (I2V) pipelines take their anchor image from a text-to-image (T2I) generator, but the authors find that anchor aesthetic and quality scores correlate only weakly with downstream video quality, that a structure-and-texture score is negatively correlated with I2V visual and motion quality, and that fine-tuning the T2I model on an image reward lowers video reward on every I2V model tested. Golden Anchor Learning (GAL) instead scores candidate anchors by the videos they induce under a frozen I2V model and converts those video rewards into image-level preferences for DPO fine-tuning of the T2I generator. On VBench across four I2V models, GAL raises the average score by 2.97% to 8.46% over the untuned generator and dynamic degree by 5.62% to 38.24%, with gains also under VideoReward-Overall, HPSv3, and UniPercept, even though training used only HunyuanVideo-1.5 renders.

arXiv:2610.13217 PDF

The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging

Fabian Drexel, Bernhard Kainz, Jakob N. Kather, Daniel Rueckert, Lisa C. Adams, Keno K. Bressem et al. cross-listed Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context, but point prediction can only reproduce variation that the context resolves, a limitation the authors call predictability bias. In computed tomography (CT) recurring anatomy is predictable from context whereas the sporadic, localized pathology that determines clinical interpretation often is not, so the authors train TAP-JEPA, a ViT-B/3D on 113,611 CT volumes that is competitive with encoders such as COLIPRI, Merlin, and VoxelFM on scan-level classification, and then audit its predictive pathway by replaying the exact masking, teacher, and predictor computations used in training. Prediction lowers mean Skill, a prevalence-normalized average precision, from 0.388 to 0.271 for pathology but only from 0.921 to 0.873 for anatomy, and along the teacher direction it moves pathology representations 34.7% toward matched controls, with larger shifts farther from visible context. Removing all final-layer pathology tokens reduces the decision margin by at most 0.45% in four of seven scan-level probes, leading the authors to argue that training objectives should retain diagnostic features regardless of how predictable they are from surrounding anatomy.

arXiv:2610.13237 PDF

Vision Transformers Learn Word Meaning from Images Alone

Anders Gj{\o}lbye, Sebastian Ray Mason, Stella Frank, Lars Kai Hansen cross-listed The two largest DINOv3 vision transformers, DINOv3-H and DINOv3-7B, turn out to organize words rendered as images by their meaning, despite being pretrained with self-supervision on images only and no captions, class labels, or other language supervision. The authors derive word representations from embedding differences between word images and matched controls and test whether semantically related words are neighbors while controlling for spelling similarity. Brand names cluster by industry, nouns cluster by semantic domain, and country-capital pairs show consistent vector offsets like those in text word embeddings, and these word representations also align with the representations of the visual object categories they name. Comparisons with text and speech encoders reveal shared structure among representations of the same words, and contrastive alignment with either encoder further strengthens DINOv3-H's word-level semantic structure while leaving the other fine-tuned vision models largely unchanged.

arXiv:2610.13252 PDF

Opening the Black Box of Classifier-Free Guidance via Information Bottleneck

Jiayang Gao, Tianyi Zheng, Jiayang Zou, Fengxiang Yang, Luyao Fan, Lv Tang et al. Classifier-free guidance (CFG) is standard for conditional diffusion sampling, but when to activate it is usually chosen heuristically, with fixed schedules or hand-designed dynamic variants that lack a principled account of when conditioning information is most useful. The authors model the diffusion trajectory as a time-dependent information bottleneck and show that the usefulness of conditional information is not uniform across sampling but becomes markedly stronger after a critical stage characterized by a Fisher-information balance condition. Building on this, they propose a training-free dynamic CFG rule that constructs a stochastic process from conditional and unconditional score statistics and switches guidance on when the estimated guidance gain crosses a threshold, recovering interval-based guidance as a limiting special case. Experiments across diverse architectures and tasks are reported to show the method is effective, robust, and broadly applicable, and the framework also offers an interpretation of existing dynamic guidance methods.

arXiv:2610.13369 PDF

Hide and Seek: HIDE Concepts to SEEK explanations for black box models

Bhanu Tokas, Rahul Nair, Rockwell Jackson, Hannah Kerner cross-listed Concept-based attribution methods explain vision models through human-understandable concepts, but existing ones handle only classification and regression, need model weights or large concept-annotated datasets, and mostly cannot give image-level attributions. Hide and Seek (HnS) needs neither weights nor concept labels: given a concept and an image set, it retrieves concept-aligned images with a vision-language model, localizes the concept in each image, masks that region, and measures attribution as the change in the black-box model's output between original and masked inputs, which directly probes model behavior and supports local attributions. On CelebA and COCO-Stuff, HnS agrees with ground-truth attributions better than existing methods, and it extends to image captioning and satellite image segmentation, where no prior concept attribution method applies.

arXiv:2610.13935 PDF

6 more specialized papers

CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches Albert Ge, Akash Haridas, Utkarsh Saxena et al.
Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation Marco Stefani, Edoardo Parpaiola, Matteo Bino et al.
APEX: Asynchronous Prefix Denoising for Extreme Compression David Dinkevich, Nisan Chiprut, Yoav HaCohen et al.
MotionCoach: Constructing Multimodal Guidance for Video Professional Motion Editing Siyi Qian, Hanqing Yang, Weihuang Lin et al.
Decoding Register Tokens Preserves Global Scene Details David Jacob Drexlin, Julius Hense, Jonas Dippel
Exploring Attention Sinks in Text-to-Image Transformers from Diffusion and Autoregressive Models Binglei Li, Mengping Yang, Zhiyu Tan et al.

Reasoning 9

Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training

Suorong Yang, Jiaxin Wang, Jiaming Su, Chentao Yue, Furao Shen, Peng Ye et al. Scaling reasoning post-training with ever-larger corpora raises cost without proportional gains, so the authors ask how far reasoning supervision can be compressed. Instead of treating a teacher-generated chain-of-thought as a single path, they build Monte Carlo reasoning trees around intermediate states to expose process-level uncertainty, use that uncertainty to contract the corpus into a compact teacher-informed pool, and then route optimization online during supervised fine-tuning according to the student's evolving learning state. Across reasoning benchmarks and model families, the method uses as little as 0.22% of the optimization budget of full-data supervised fine-tuning, over 450 times less training exposure, while matching or improving reasoning performance, and it transfers across student families, black-box teachers, and weak-to-strong settings.

arXiv:2610.12933 PDF

CoBe: A Benchmark For Conversational Counterfactual Text Editing

Arvind Raghavan, Inwoo Hwang, Shreyas Havaldar, Kai-Zhan Lee, Aurghya Maiti, Yushu Pan et al. Counterfactual text editing asks a model to revise a scenario under a hypothetical change while leaving causally unrelated facts intact, a distinction that prior datasets blur with merely associational generation. The Counterfactual Benchmark (CoBe) tests this zero-shot with 2,000 conversationally phrased questions spanning healthcare, science and engineering, business, and everyday scenarios. Frontier systems average 58.05%, often rewriting details the hypothetical should never have touched. Performance tracks neither model size nor output length, which the authors read as a training-objective problem rather than a capacity limit.

arXiv:2610.13057 PDF

BoundaryEvo: Data-Model Co-Evolution at the Capability Boundary for Long-Chain Reasoning Distillation

Jiaming Su, Suorong Yang, Xudong Li, Tengchao Yang, Peng Ye Distilling long reasoning traces from a large teacher into a small student usually uses a fixed training set, even though an example's usefulness changes as the student improves and only a small number of teacher trajectories may be affordable. BoundaryEvo runs in two stages: Capability Priming distills a small unfiltered set of teacher traces to establish basic reasoning ability, and Capability-Boundary Co-Evolution then repeatedly selects questions the current student fails but the teacher reliably solves, refreshing the training data as the student evolves. Using only 1,000 training examples, the method beats static and dynamic data selection baselines on 11 reasoning benchmarks, lifting Qwen3.5-4B from 62.9 to 75.3 on IMO-AnswerBench and from 26.8 to 46.0 on AMO-Bench.

arXiv:2610.13285 PDF

SqueezeBench: Saturation-Resilient Evaluation of Mathematical Capability

Pablo Manrique, Hai Xia, Stefan Szeider Benchmarks of solved math problems saturate, while open problems can be resolved for the first time only once, which makes progress in machine mathematics hard to measure. SqueezeBench instead poses ten extremal-combinatorics problems, each defined by an integer function whose values are known only for small inputs, and maintains lower and upper records on that function as computable Lean functions with machine-checked proofs; a submission is accepted only if two independent Lean kernels verify that it bounds the function, is no worse than the current record everywhere, and is strictly better somewhere, so each success raises the bar instead of retiring the problem. Agents built on a frontier language model, given at most one hour per record, had 82 submissions accepted and improved every one of the twenty initial records, and agents on two newer models then improved 19 of the 20 again, though for eight of the ten problems at least one record still trails the best published asymptotic bound.

arXiv:2610.13461 PDF

Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models

Weixin Guan, Liang Li, Shiyu Xie, Jiapeng Liu, Bing Li, Peng Fu et al. Reasoning models can keep generating thinking tokens past the point where they help, and existing evaluations of early-termination strategies lack a reference for how severe this overthinking actually is. Step-wise Truncation Probing (STP) repeatedly samples answers from retained prefixes of fixed reasoning trajectories, defining an Overthinking Ratio and Accuracy Degradation relative to the earliest point where accuracy peaks, and a unified framework casts dynamic routing, fixed budgets, and dynamic early exit as termination strategies whose theoretical token-accuracy Pareto frontiers can be computed from the same probes. The resulting OTBench benchmark packages 70.5 billion reasoning tokens of probing measurements across six models and four benchmark groups for reuse.

arXiv:2610.13570 PDF

A Critical Look at Power Sampling when Test Time Sharpening LLMs

Abiel Malkani Talwar, Christopher Nemeth, Henry Moss cross-listed Evidence suggests reinforcement learning with verifiable rewards (RLVR) mainly sharpens the base model's distribution onto reasoning paths it already produces, prompting reward-free methods that sample from the power distribution proportional to the base policy raised to alpha via Markov chain Monte Carlo or sequential Monte Carlo. The authors show these samplers do not converge to the power distribution under realistic compute budgets and that their gains are largely recreated by simple Best-of-N with appropriate grading, implying they work by selecting high-probability sequences rather than approximating the target. To test whether the power distribution is itself worth targeting, they fine-tune the base model to match it directly, sidestepping sampler convergence; despite a far more faithful approximation, the gains are noisy and specific to certain models and tasks. The conclusion is that power sampling's value lies in extracting sequences the base model already deems likely, not in faithfully targeting the power distribution.

arXiv:2610.13653 PDF

Pretraining and Test-Time Training Budgets Matter for ARC: A Controlled Study of VARC and TRM

Adrien Deliege, Marc Van Droogenbroeck Recent models for the Abstraction and Reasoning Corpus (ARC) differ not only in architecture but also in pretraining data, augmentation, test-time setup, and optimization budgets, which muddies conclusions about which architectural choices actually help. The authors place the vision-inspired VARC and recursion-based TRM under a common inductive protocol on ARC-AGI-1, crossing five pretraining budgets with four test-time training (TTT) budgets, each defined by a task-variant count and total sample presentations. Increasing TTT sample presentations improves pass@2 in every setting for both models, but increasing the TTT task-variant count consistently helps only TRM, while VARC often does better with fewer variants. VARC also stays competitive without pretraining whereas TRM depends on it much more, motivating reporting of ARC results at comparable budgets.

arXiv:2610.13660 PDF

Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction

Jiwon Jeon, Myungsik Cho, Jeonghye Kim, Youngchul Sung Large language models rarely manage to spot and fix errors in their own reasoning, and most fixes require fine-tuning the model to produce revisions or feedback. The authors instead train a small 130K-parameter MLP head on a frozen encoder, supervised only by final-answer correctness, to decide where in a failed rollout to intervene, and find that presenting the redirect instruction as a fresh user turn changes the model's course where inline token injection does not. The resulting method, UPTURN (User-Prompt TURN injection at a learned position), keeps a chosen prefix, inserts the instruction as a new user turn, and regenerates the rest without touching model weights. Across four reasoners and six math benchmarks it improves average accuracy over a 32-rollout baseline in every setting, beats self-correction and reflection-token baselines at matched budgets, transfers to GPQA-Diamond without retraining, and stacks with GRPO and self-correction methods.

arXiv:2610.13672 PDF

A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning

Mayug Maniparambil, Ayush Rai, Ekansh Verma, Noel E. O'Connor, Fergal Reid On-policy self-distillation promises dense token-level credit for reasoning by having a teacher conditioned on the reference solution score the student's own rollout, and the authors ask whether that scoring signal actually carries corrective information. Using gold error spans in student rollouts, they test the full-distribution KL and the token log-ratio and find that neither signal localizes errors, pushes probability away from erroneous tokens, or reliably separates correct from incorrect rollouts on the same problem. They argue this is structural, since locating an error is itself a search problem that a scoring teacher is never asked to perform, and show that the same model with the same privileged information localizes errors far better when allowed to write a critique. Conditioning the scoring teacher on a generated critique yields a signal that is corrective within error spans and favors backtracking tokens like 'Wait' and 'Actually', and the critique also helps at test time and provides an advantage signal in cold-start GRPO on prompts where verifiable rewards alone give none.

arXiv:2610.13744 PDF

Unclassified 1

Unsupervised Multimodal Reward Modeling

Yueying Li, Jingxuan Fan, Kiant\'e Brantley, Sham M. Kakade, Hanlin Zhang No summary available — see the abstract on arXiv.

arXiv:2610.12707 PDF