AI Paper Digest

Thursday, October 8, 2026

580 papers cs.AI · cs.LG · cs.CL ← 2026-10-072026-10-09 →

Aug Sep Oct Nov

Highlights

On KL-Regularized Policy Optimization

Highlight HF pick · 18▲Reinforcement Learning Yifan Zhang In asynchronous reinforcement learning (RL) for LLM agents, rollouts come from stale checkpoints or an inference engine whose probabilities differ from the trainer's. Existing fixes either clip importance ratios, which biases updates, or sample groups of responses as in GRPO, which is costly for long episodes. KL-Regularized Policy Optimization (KLPO) anchors the KL regularizer at the sampler, so the improvement step has a closed-form solution that can be fit by least squares on the sampler's own trajectories without importance weights. The authors prove the gradients stay unbiased and show that SPPO, GPO, REBEL, and BPO are special cases, giving a critic-free update that needs only one rollout per prompt.

In asynchronous RL for LLM agents, rollouts come from stale checkpoints and from inference engines whose probabilities differ from the trainer's, so the policy being updated is not the one that generated the data. KLPO addresses this by anchoring the KL regularizer at the sampler itself. The regularized update then has a closed-form Gibbs solution, which can be fit directly on off-policy data without importance weights.

  • The method fits the Gibbs solution's log-ratio optimality condition by least squares on the sampler's own trajectories, so sampler probabilities enter only through a log-ratio rather than through clipped importance ratios, which bias the update.
  • Profiling out the regression intercept replaces the intractable log-partition function with the signal's mean under the sampler plus a sampler-to-trainer KL term, so no learned normalizer is needed.
  • For token-level policy mirror descent targets, the gradient is computed from terminal returns without a critic, using either sampler-centered scores or a single trajectory residual, and this holds even when tool outputs are stochastic.
  • On the theory side, the authors prove that independent Monte Carlo estimates of the KL term keep the gradients unbiased, derive the exact KL gap of cheaper top-K and binary approximations, and show that SPPO, GPO, REBEL, and BPO are special cases of KLPO.
  • The practical payoff is one rollout per prompt, compared with the response groups GRPO needs, which matters most for long agentic episodes; the abstract reports no empirical benchmarks or training-efficiency numbers, so the gains remain theoretical.

arXiv:2610.08963 PDF

PhysEvo: Astra Can Act, Let It

Highlight HF pick · 7▲Robotics Wenqing Tian, Zeyu Zhang, Zhaocheng Liu, Fengwei Liu, Qiang Liu, Liang Wang PhysEvo is a framework for physical recursive self-improvement (RSI) built around one frozen Astra model. A task agent runs robot tasks, and a meta-agent studies the resulting trajectories to diagnose failures, revise tools and skills, test fixes, and improve its own diagnostic tools, all without updating model weights. On 42 RoboDojo tasks with held-out layouts, it reaches 62.00% success versus 47.17% for the strongest published one-shot Astra agent. On eight tasks that are hard for Astra used directly, it reaches 55.00% success versus 1.25%. A harness evolved in simulation and transferred to an AgileX PiPER arm reaches 84.00% success on five real-world tasks.

Robots driven by a capable frozen model like GPT-6 Astra often fail because of the harness, meaning the tools, skills and control interface between the model and the robot, rather than because the model lacks the ability. PhysEvo uses the same frozen model in two roles: a task agent that acts, and a meta-agent that turns execution trajectories into tested revisions of that harness, including revisions to its own diagnostic tools.

  • The meta-agent reviews aligned images, robot states and code to diagnose a failure, then edits action and observation tools, skills or diagnostic utilities; a revision is adopted only if it raises the score, or gets past the failure without lowering it, on the same task and starting scene, with no weight updates and no separately trained action policy.
  • Over many rounds the harness gained joint-level control, which beat the initial end-effector-plus-inverse-kinematics interface 100 vs 40 in a small six-episode test; it also gained free-arm side views, retrieval of earlier frames, triangulation tools, and review tools that the meta-agent reused in later rounds.
  • Across all 42 RoboDojo tasks, evaluated on held-out layouts, it reaches 68.14/100 score and 62.00% success against 47.17% for RoboDawn's one-shot Astra agent, with a perfect 100% on the memory dimension; on the ten-task GPT-as-Policy set it scores 68%, against 26% for direct Astra and 48% for the Astra–π0.5 hybrid.
  • On eight manipulation tasks where direct Astra almost always fails, success rises from 1.25% to 55.00%, just ahead of Simate-beta at 50.83%; on a real AgileX PiPER arm with continued skill revision it reaches 90.60/100 and 84% success over 25 trials on five tasks.
  • The caveats are that each task keeps its own tuned deployment version and is scored on only five layouts, and the real-robot numbers include skill revision between attempts; it also still lags trained VLAs on tower building (20% vs about 80%) and scores 0% on two of the eight hard tasks.

arXiv:2610.08995 PDF

From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery

Highlight HF pick · 11▲Reasoning Xinglin Wang, Zishen Liu, Tong Zheng, Shaoxiong Feng, Peiwen Yuan, Yiwei Li et al. Test-time scaling (TTS) methods usually trade accuracy against a single resource, such as cost or latency, while real users may set accuracy, latency, and cost requirements together. The authors define Personalized Test-Time Scaling as finding executable controllers that maximize the rate at which all of a user's requirements are met at once. PersonTTS amortizes this search across users by starting from controllers found for users with similar requirements and by reusing distilled procedural guidance, while still evaluating every candidate on the new profile. On AIME and HMMT, it substantially outperforms strong TTS baselines in joint requirement satisfaction on unseen profiles and problems, and the cross-user reuse lowers discovery-agent time and cost.

Test-time scaling methods usually trade accuracy against one resource at a time, either cost or latency, but real users often have accuracy, latency and cost requirements all at once. PersonTTS treats this as searching for executable inference controllers that maximize the joint satisfaction rate (JSR) of a user's three requirements, and it saves search effort by reusing what earlier searches for other users learned.

  • An LLM discovery agent (claude-opus-4-6) writes controller code that decides model routing among six Qwen3 sizes, how many reasoning branches to run and how long, pruning, and self-refinement; each candidate is scored by CPU-only offline replay of cached rollouts under the target user's profile, over five rounds.
  • Reuse works in two ways: the search starts from the stored controller whose requirement profile is most similar to the new user's (a "warm-start"), and a frozen Guide of revision rules, distilled from earlier search histories, steers the agent's proposals.
  • On unseen user profiles, PersonTTS reaches 96.57% JSR on AIME24–25 and 83.85% on held-out AIME26, versus at most about 35.6% for the strongest baseline (AutoTTS with β=1.0), and 90.86% on HMMT24 and 77.28% on HMMT25, where no baseline exceeds 19.3%.
  • Most of that gain comes from optimizing directly for the joint requirements, since the variant without reuse already scores 69–87%; on top of that, the Guide cuts the agent's search time by about 46% and its cost by about 36% (on AIME, full PersonTTS needs 40.4 minutes and $7.40 per user versus 74.6 minutes and $10.30 without reuse).
  • The two reuse mechanisms do not add up cleanly, and dropping the warm-start actually does better on HMMT25 (79.09%); a larger experience bank helps on the discovery problems but hurts held-out HMMT25 (86.09% with 20 stored entries, 77.28% with 100); and every result comes from simulated profiles on math benchmarks replayed offline, not from live deployment.

arXiv:2610.09684 PDF

UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation

Highlight HF pick · 14▲Vision Deyuan Liu, Yihao Hu, Jingxuan Zhang, Xingying Li, Jun Xie, Jiacheng Liu et al. Image generators now render short text strings well, but scenes that need long strings in many separate regions remain hard to evaluate. UltraText Bench contains 432 human-reviewed prompts, split equally between English and Chinese, across 24 real-world scene categories and three difficulty levels. Each prompt specifies exact strings for four to twelve text regions, along with reference content, placement and visual attributes. The Q-Judger vision-language model scores each image for text fidelity, clarity, spatial quality and scene quality across 24 model configurations. The scores show trade-offs between models: Z-Image-Turbo gains clarity over Z-Image-Base but loses 14.76 fidelity points, and Qwen-Image-2512's English composite score falls from 86.50 at the easiest level to 42.86 at the hardest.

Image generators now handle short strings well, so the open question is whether they can reproduce hundreds to thousands of exact characters spread across many regions of one scene. UltraText Bench tests this with 432 bilingual, prompt-only generation tasks, and a VLM judge scores each image against a full per-region reference of text, placement and carrier.

  • The prompts are split evenly between English and Chinese and cover 24 scene categories in six domains, from signs and receipts to code and infographics, at three difficulty levels; each one contains the exact strings for 4–12 text regions, about 944 ground-truth characters on average and up to 5,230.
  • Q-Judger from Qwen-Image-Bench gives six image-level scores that combine into text fidelity, clarity, spatial quality and scene quality, weighted 60/30/5/5 in the composite; there are no region-level scores and no OCR transcriptions.
  • Across 24 model configurations, GPT Image 2 [Low] leads with a 99.35 composite, and 83.6% of its valid images get perfect marks on all six raw scores; Boogu-Image-0.1-Base is the best open-weight model at 77.89.
  • Clarity and fidelity can diverge sharply: Z-Image-Turbo gains 3.81 clarity points over Z-Image-Base but loses 14.76 fidelity points, and fidelity is lower for the Turbo variant in all three Base/Turbo pairs.
  • Scores tend to drop as text load grows, though not for every model: Qwen-Image-2512 falls from 86.50 at L1 to 42.86 at L3 in English. The results have no confidence intervals and rest on a single VLM judge, and the human study with ten participants reports no agreement statistics. The language and difficulty splits also change content, character load and region count together, so they cannot isolate any one factor.

arXiv:2610.09823 PDF

Inverting Multi-Vector Visual Document Indices

Highlight HF pick · 15▲Safety & Alignment Zhuchenyang Liu, Yao Zhang, Yu Xiao Multi-vector visual document retrievers store each page as about a thousand patch vectors, often in third-party vector databases, and these indices tend to be treated as less sensitive than the pages themselves. The authors frame inversion as conditional document image generation and show that whoever runs or breaches the store can reconstruct pages: on ViDoRe v3, pages inverted from raw indices recover 47% of the words and 45% of the sensitive tokens, and they rank their source page first 98.4% of the time when used as queries. Token pooling and shuffling both cut word recall to about 8%, but a model that restores the order of shuffled vectors raises first-rank retrieval from 3.8% to 93.5%. Inverting a pooled index remains an open problem, and the authors conclude that these indices should be protected like the documents they encode.

Multi-vector visual document retrievers store each page as roughly a thousand patch vectors in raster order, and those indices are often treated as less sensitive than the pages themselves. The authors show that someone who runs or breaches the vector store can redraw a page from its index alone. They frame the attack as conditional document image generation, and it needs no side information beyond the stored vectors.

  • The attack first works out which public encoder produced the store, which it does correctly in all 26,000 decisions across a pool of 13 retrievers. It infers the page shape from the period of vector similarity (99.2% correct on raw indices), and it restores the order of shuffled vectors with a permutation-equivariant position model followed by Hungarian assignment; a Qwen-Image-style flow-matching MMDiT that takes the index in place of a text prompt then generates the page.
  • Against Tomoro-ColQwen3-8B on 2,000 held-out ViDoRe v3 pages, pages inverted from the raw index recover 47.4% of words and 45.0% of sensitive tokens (numbers, capitalised words and acronyms), compared with 23.4% and 14.6% for a nearest-neighbour baseline. They rank their source page first among 19,252 indexed pages 98.4% of the time, or 97.7% when an independent encoder does the ranking.
  • Two cheap protections, hierarchical token pooling and random shuffling, each cut word recall to about 8%. A shuffle does not hold up, though: the position model raises the share of source pages ranked first from 3.8% to 93.5%. Inverting pooled indices remains an open problem, and the authors explicitly say this does not show pooling is safe.
  • Applied unchanged to a second retriever, ColQwen3.5-4.5B, the attack still ranks the source page first 70.2% of the time, but its word recall of 7.9% falls below the nearest-neighbour baseline, and the authors cannot say which property of that encoder makes it leak less.
  • Both attacked encoders are ColPali-style retrievers on Qwen backbones, sensitive tokens are defined by a regular-expression heuristic with no human study, and the authors argue the reported leakage is a lower bound because the inverter was still improving when training stopped.

arXiv:2610.09920 PDF

Mechanics of Long-Context Hybrid Models Part 1.1: From Hybrid Attention to Hybrid Position

Highlight HF pick · 23▲Large Language Models Xiaoran Liu, Ziwei He, Xipeng Qiu The authors study why hybrid large language models work. These models mix full attention with sliding-window attention (SWA) or gated linear attention (LA) variants such as GLA and GDN. They find a seesaw effect in context extension: LA hybrids benefit more from long-context continual pretraining, while SWA hybrids extrapolate better to longer inputs, and they trace this to the positional biases each attention type creates. SWA hybrids fall into short-context learning traps and need wider windows during long-context pretraining. For LA hybrids, the proposed Sliding-Window Linear Attention achieves 16× training-free length extrapolation with 100% accuracy on NIAH-SK1 at 64k context.

Hybrid LLMs that mix full attention with sliding-window attention (SWA) or gated linear attention (GLA, GDN) work largely because they combine position-free (NoPE) full attention with position-biased modules, which the authors call "hybrid position." Building on that, they explain why SWA hybrids and linear-attention hybrids each break down in different long-context settings and propose fixes for both.

  • Across 376M–1B models (plus a 3B check), pretrained at 4k context on 50B tokens and then continually pretrained at 32k on 5B tokens, SWA-NoPE hybrids extrapolate best without extra training, but linear-attention hybrids overtake them after long-context training, a reversal the authors call the "Seesaw Effect."
  • Attention analysis shows that a minority of NoPE heads handle coarse, high-hit-rate global retrieval, while the majority position-biased heads (RoPE, GLA, GDN, SWA) filter out noise; adding noise to the position-biased heads hurts RULER and BABILong more than adding it to the NoPE heads, and a 3:1 ratio generally works best.
  • SWA hybrids fall into a "short-context learning trap" during long-context training: their small-window SWA weights change far more than their NoPE weights, so the model mostly gets better at short-range prediction; enlarging the window from 128 to 2048 and training with LongCE makes SWA-NoPE-LH competitive with GDN-NoPE-LH.
  • The proposed Sliding-Window Linear Attention (SWLA) restricts gated linear attention to a local window by running the unmodified FLA kernels on overlapping chunks (4096-token window, 1024-token stride); combined with log-scaled NoPE attention logits (EME), it gives 16× training-free extrapolation (4k to 64k) with 100% accuracy on NIAH-SK1.
  • The harder retrieval tasks still degrade at 64k (for example, GDN-NoPE-HH reaches only 16% on NIAH-SK3), all experiments are small-scale and cover pretraining only, and the models score low on LongBench overall.

arXiv:2610.10114 PDF

RoboQuest: Generalist Physical Agents that Search, Inspect and Test

Highlight HF pick · 3▲Robotics Liu Renhang, Navonil Majumder, Tej Deep Pala, Soujanya Poria In unfamiliar environments, generalist robot agents often have to gather missing information through physical interaction before they can finish a task. RoboQuest is a benchmark of ten mobile manipulation tasks that require searching for objects, inspecting hidden properties, or testing unfamiliar tools, and agents must decide for themselves when they have enough evidence to commit. The best of five frontier multimodal agents succeeds in only 23% of episodes, and a fine-tuned π0.5 policy almost never succeeds. Failure analysis shows that most failures are not execution errors: agents tend to stop exploring too early, rarely repair the disturbances their exploration causes, and struggle to learn by trial and error.

Frontier multimodal models can now act as zero-shot robot policies, but in unfamiliar settings the information a task depends on is often not visible. RoboQuest is a benchmark of ten simulated kitchen mobile-manipulation tasks where the agent must search, inspect, or physically test objects to uncover that information, and then decide for itself when to commit by pressing a physical SUBMIT button.

  • The tasks fall into three families (Search, Inspect, Test) built on RoboCasa365 in MuJoCo with a Franka arm on a mobile base, and agents control it only through low-level arm, base, wait, and stop commands, with no helper tools for grasping or locating objects, while some exploratory actions cannot be undone.
  • GPT-6 Astra leads with 23.2% success, Claude Opus 5.5, GPT-6.1 Sol, and Claude Fable 5.1 reach 11–14%, Gemini 3.8 Flash reaches 2.0%, no model solves a single Blackout Search episode, and a π0.5 policy fine-tuned on the released 5,000-episode (366-hour) demonstration set succeeds only 2% of the time on Puzzle Box and 0% on every other task.
  • Execution is not the main bottleneck: the top three models succeed at 72–81% of isolated manipulation skills, and only 18–21% of their full-task failures trace to execution, while missing evidence (43–46%) and wrong decisions (23–31%) dominate, often because agents stop exploring even after a hiding place has appeared in view five or more times.
  • Agents rarely prevent or repair the disturbances their exploration causes (7–13% of failures, and a bowl provided for spilled balls was used in only 2–8 of 50 Marked Mugs episodes), and trial-and-error inference is brittle: hiding the Puzzle Box locks drops Claude Opus 5.5 from 94% to 37% while leaving GPT-6 Astra unaffected.
  • The benchmark is simulation-only with 50 evaluation instances per task, evaluating the five models cost $21,352, and the fine-tuned π0.5 has no memory and is tested on held-out kitchen styles and configurations, so its collapse mixes poor exploration with out-of-distribution generalization.

arXiv:2610.10388 PDF

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

Highlight HF pick · 21▲Agents Zhiqin Yang, Chenxin Li, Xiaomeng Hu, Yibin Liu, Weidong Huang, Jiankai Sun et al. General-purpose agents now write code and use tools well in digital settings, and RobotWorld asks how much of that ability carries over to controlling robots. The simulation benchmark has 84 tasks covering manipulation, mobile manipulation, locomotion, driving and aerial control, each with an interaction budget and an executable success check. Agents can build sophisticated perception and control pipelines (segmentation, camera calibration, spatial estimation), but these skills often fail to combine into task success: agents lose track of object state, keep repeating actions that don't work, recover too late, or think an unfinished task is done. Model strengths differ: Astra does better on spatial and constrained-contact goals, while Opus 5.5 does better on balance and timed-interaction goals.

RobotWorld tests whether general-purpose multimodal agents can turn instructions and observations into physical task execution through robot interfaces. It does this with 84 simulated tasks from 20 source projects, spanning manipulation, mobile manipulation, locomotion, driving and aerial control, each with explicit interaction budgets and executable success checks.

  • Agents act through a shared harness that offers direct robot-tool calls (optional code control is disabled in the reported runs) and an isolated analysis workspace where physics pauses during computation, and success is judged by checks on simulator state and events (pose, contact, persistence, event order, path constraints) rather than by the agent's own claim of completion.
  • With one episode per model–task pair (420 in total), success is low across the board: GPT-6 Astra reaches 19.0% (16/84), Claude Opus 5.5 15.5% (13/84), Kimi K3 2.4%, and DeepSeek V4.1 Flash and Gemini 3.8 Flash 1.2% each, leaving 63 of 84 tasks unsolved by any model.
  • The two leading models have different strengths: Astra does better on spatial and constrained-contact goals (4/20 mobile manipulation versus 0/20 for Opus 5.5), while Opus 5.5 does better on continuous-balance and timed tasks (3/4 aerial, and the only locomotion success).
  • Cost does not track success: Astra costs an estimated $9,913 for 941M tokens, about 3.4 times Opus 5.5's $2,916 for 1.63B tokens, and Kimi K3 uses the most tokens (2.00B) and the most elapsed time (163.8 hours) while solving only two tasks.
  • Execution traces show that agents can build real perception and control pipelines (colour segmentation, camera-geometry fitting, LQR-style dynamics calculations), but these skills often fail to combine into task success: agents lose the object while reaching the commanded pose, spend the whole budget on preparation, declare unfinished tasks complete, or recover too late, and every result comes from a single run with no estimate of repeated-run variance.

arXiv:2610.10409 PDF

RunningTab: Direct Workspace Interaction with Environment-Side Tabs

Highlight HF pick · 30▲Agents Jinheon Baek, Soyeong Jeong, Yumin Choi, Dongsu Han, Sung Ju Hwang LLM agents can produce deliverables by searching and reading workspace files directly from a terminal, but nothing tracks what the task requires, which files were read, and which were listed but never opened, so details drop out of the context window and the final report can miss them. RunningTab gives the environment a per-task tab: the agent records requirements, while the environment logs every file read as an excerpt with its source and every listed-but-unopened file as a candidate. The agent sees each requirement next to matching excerpts and unopened candidates, resolves it or sets it aside with a reason, and a finish check returns any open requirements before the agent can finish. Across three benchmarks and three LLMs, it consistently beats plain direct workspace interaction and baselines that keep the record inside the model's own context.

LLM agents that build reports from many workspace files through a terminal often read the right content and still leave it out of the deliverable: in a DCI baseline, 19.8% of the expected numeric values the agent had seen never made it into the output. RunningTab fixes this with an environment-side "tab," a per-task record kept outside the context window that tracks what the task still needs, what has been read, and which files were listed but never opened.

  • The agent writes only the requirements, while the environment automatically logs every file read as a verbatim excerpt with its path and command, and ranks every listed-but-unopened file as a candidate using BM25; a review step shows each open requirement beside its best-matching excerpts and candidates.
  • A requirement can be marked done only by citing a read-log entry that shares terms with it, or marked unavailable with a reason, and a one-time finish check returns any still-open items when the agent first tries to stop.
  • Across Workspace-Bench, TheAgentCompany, and OfficeQA Pro with GPT-5.4 nano, DeepSeek V4 Flash, and Gemini 3.8 Flash, it beats plain DCI and the model-kept tracking baselines (TODO, Self-Refine, Self-Tracking) on every metric, with the largest gain for Gemini: 31.42 → 46.91 rubric pass rate on Workspace-Bench and 43.08 → 55.77 exact-match accuracy on OfficeQA Pro.
  • Ablations credit most of the gain to the environment's automatic capture: removing it drops the GPT-5.4 nano pass rate to 41.36, about the DCI level of 41.63, and the tab holds 94.8% of the values that were shown to the agent but missing from its deliverable.
  • Gains are modest for some models (about +3 points for GPT-5.4 nano on Workspace-Bench), models use the tab unevenly (GPT stated requirements in only 56.9% of attempts), the log cannot hold values the agent computes itself, the citation check is only lexical, and stored excerpts may expose sensitive workspace content.

arXiv:2610.10444 PDF

Long-WAM: Scaling the Context of World-Action Models

Highlight HF pick · 69▲Robotics Wei Huang, Bohan Zhang, Chenzhi Liu, Isabella Liu, Shuai Yang, Weian Mao et al. Long-WAM scales the visual history available to causal world-action models while keeping robot control real-time. Its central finding is that longer history pays off mainly when the video model is pretrained autoregressively. On RoboCasa GR-1, extending context from 0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained initialization shows no net gain. The model achieves the best results among compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. Streaming observation encoding and asynchronous execution let it run on an RTX 5090 at 107.4 ms per action chunk, as well as on DGX Spark and Jetson AGX Thor, and it reaches 95% success on dynamic cup stacking on a real robot, where Pi0.5 and Fast-WAM succeed in none of 20 trials.

Robot controllers need visual history to infer motion and task progress, but processing that history slows their response. Long-WAM shows that longer context helps world-action models much more when the video backbone is pretrained autoregressively, and pairs that model with a deployment stack that keeps control real-time on edge GPUs.

  • The model is built in two stages: LongLive2.0-Robot is first pretrained autoregressively on roughly 10,000 window-equivalent hours of robot and egocentric video with no action labels, then adapted with the same causal ordering so that an action expert denoises each action chunk from observed history plus partially denoised predicted future latents that are never decoded to pixels.
  • On RoboCasa GR-1, extending history from 0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained Wan2.2 initialization shows no net gain, and the robot-domain autoregressive model's lead over it grows from 3.3 points with no history to 17.1 points at 19.2 seconds.
  • It also reports the best results among the compared methods on LIBERO (99.5% average and 99.5% on LIBERO-Long), RoboTwin 2.0 (94.4%) and DOMINO (34.9% versus 19.9% for Fast-WAM), and in real-world tests on a Unitree G1 it stacks moving cups in 95% of trials, where π0.5 and Fast-WAM fail all 20.
  • NVFP4 quantization, CUDA Graphs, reuse of key–value caches within each call, streaming VAE encoding and device-specific kernel tuning bring each action chunk, including future-video prediction, to 107.4 ms on an RTX 5090 (a 3.3× speedup over unoptimized BF16), while asynchronous execution without blending costs only 0.2 points of success, compared with drops of 15.4 points for Fast-WAM and 45.5 points for LingBot-VA.
  • On the limitations side, each context length is a separately trained model, success falls to 75.2% at 38.4 seconds (where 80.4% of history frames are padding), 8× more history raises latency 3.2× (to 341.0 ms), and the context-length study has not yet been repeated on real robots.

arXiv:2610.10528 PDF

Applications 127

Adversarial RL for Port-Scan Evasion: Attacker Feature Visibility in Edge-Deployed IDS

Logan Andrew North, Priya Sanjay Kaluskar, Shasi Kumar Ramachandran Prabhu, Peilong Li, Suman Saha cross-listed Machine-learning intrusion detection systems (IDS) on Internet of Things (IoT) edge devices are usually tested against static attacks rather than attackers who adapt to detection feedback. The authors deploy a live Zeek pipeline running XGBoost, an MLP, and a 1D-CNN trained on TON_IoT on a Raspberry Pi 3B+. A Deep Q-Network (DQN) attacker learns to vary probe timing, TCP flags, and payload size under black-box, gray-box, and white-box access. Although the models catch 91.1 to 99.8% of conventional scans, the adaptive attacker evades them 61.9% to 98.3% of the time, and seeing more features does not consistently help it: black-box evasion is highest against XGBoost, while the 1D-CNN is most vulnerable under white-box access.

arXiv:2610.08864 PDF

FinVector-Market-4B: A Controlled Study of LoRA Adaptation for Structured Financial Tasks

Alina Khaybullina Qwen3.5-4B is adapted with rank-16 LoRA on 22,000 examples for structured financial tasks, and the base and adapted models are compared on a 600-example benchmark under implicit and explicit JSON-schema prompts. Supplying the schema alone raises the base model's JSON validity from 0% to 91.3%, which separates format learning from the gains due to adaptation. Under matched explicit prompting, the adapter still improves FinQA exact match from 14.7% to 40.0% and calculator-expression correctness from 48.0% to 82.7%, with similar gains on scenario and implication labeling. The authors note that filing overlap and inconsistent calculator targets limit how far the benchmark results generalize.

arXiv:2610.08882 PDF

Contextualization of Third-Party Cloud Security Findings

Leon Goldberg, Gal Engelberg cross-listed Third-party cloud security findings carry a static severity assigned by the rule that raised them, regardless of the environment where they occur. The authors re-derive each finding's severity with a deep research agent that investigates the flagged resource, its neighborhood in a precomputed cross-signal asset graph, and signals from other products, then returns an adjusted severity with an evidence trace. In a production field study of 9,967 vendor HIGH findings from two commercial platforms across eight real environments, three in four findings were re-graded, mostly downward, and the same rule often moved in opposite directions within one environment. About half the decisive evidence lay beyond the flagged resource, and read-only probes of live infrastructure confirmed the decisive fact for 99.4% of decided findings.

arXiv:2610.08895 PDF

Sigma-Hunter: A Domain-Specific Language Model for Threat Hunting and Detection Engineering

Kemal Davaslioglu, Sastry Kompella General-purpose LLMs that draft Sigma detection rules for threat hunting often produce invalid YAML, wrong log sources, or overly broad logic. The authors built an instruction-tuning set of 7,663 examples from 3,635 validated open-source Sigma rules, splitting the data by source rule to prevent leakage, and LoRA-fine-tuned Mistral 7B and Phi-4 into Sigma-Hunter. Sigma-Hunter-Mistral scores 8.17 overall versus 7.88 for the strongest general-purpose baseline and 4.61 for untuned Mistral, on a rubric covering syntax, field consistency, and semantic detection quality. The authors also report that syntactically valid YAML is a weak proxy for rule quality, and they note that the models run locally for air-gapped environments.

arXiv:2610.09007 PDF

Towards Financial World Modeling

Humzah Merchant, Alec Guthrie, Simon Mahns, Randall Balestriero, Bradford Levy Financial representation learning is usually evaluated on single prediction tasks over narrow datasets and short time periods, which says little about whether a representation can support a world model for planning. The authors release Market-1T, a dataset of nearly one trillion observations of U.S. equities from 2008 to 2025 at 1 Hz resolution, along with a rigorous evaluation protocol. They then compare 18 encoder-training strategies across almost two decades of market regimes, using both predictive finance tasks and probes of latent structure. A central finding is that encoders with similar predictive performance can organize market state very differently. The work is positioned as groundwork for world models such as DINO-WM, V-JEPA 2, and LeWM.

arXiv:2610.09048 PDF

Are We Really Benchmarking Forecasting Models? The Impact of Preprocessing on Time Series Performance

Guilherme Afonso Galindo Padilha, Paulo Salgado Gomes de Mattos Neto, Rafael Menelau Oliveira e Cruz Forecasting benchmarks usually apply only simple scaling and skip transformations such as differencing that handle nonstationarity, which biases results toward models that handle preprocessing internally. The authors build a preprocessing-aware benchmark that runs 11 forecasting models through 16 reversible preprocessing pipelines on 29,000 M4 time series. Choosing the preprocessing per series improves accuracy by roughly 27% to 87% across all models, with the largest gains for architectures that have no built-in preprocessing. With these gains, simpler models become competitive with state-of-the-art ones, and the results are released as a metadataset for metalearning research.

arXiv:2610.09096 PDF

SwarmReconGuard: Black-Box Detection of Distributed Collective Reconnaissance by Individually Benign-Looking Agent Populations

Vahid Tavakkoli, Kabeh Mohsenzadegan, Kyandoghere Kyamakya cross-listed Agentic clients can spread reconnaissance across many identities so that every request looks valid and low-rate while the group as a whole maps a system, a threat the authors call Distributed Collective Reconnaissance. SwarmReconGuard is a reproducible black-box benchmark in which the defender sees only telemetry at the service boundary, covering 440 runs and 3.67 million requests from 10 to 10,000 virtual identities. Gaussian likelihood-ratio detection catches all known attacks with no observed false positives but only 3% of attacks driven by unseen policies. A hybrid CUSUM detector reaches 85.7% detection with no observed false positives at 10,000 identities.

arXiv:2610.09138 PDF

Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI

Yexiao He, Yucheng Tang, Pengfei Guo, Yufan He, Andriy Myronenko, Can Zhao et al. Deployed large language models (LLMs) and vision-language models (VLMs) are frozen and do not learn from the cases they solve, which is a problem in medicine, where evidence and guidelines keep changing. A model-agnostic framework lets frozen models build up three kinds of external expertise during deployment: a Skill that guides reasoning and tool use, a Knowledge Memory of facts supported by earlier cases or trusted sources, and a Multimodal Knowledge Base of visual examples. An update is kept only if it helps on new cases without hurting earlier ones. Across six benchmarks and four open-weight and closed models, the framework improves performance by up to 34.2% over the base model on medical tasks, transfers to other models without re-optimization, and also works outside medicine.

arXiv:2610.09146 PDF

From Probabilities to Decisions: Search and Multi-Teacher Distillation with Jev

Mohamad Yazan Sadoun, Sarah Sharif, Yaser Mike Banad Jev is a probability-only model that returns calibrated probabilities over fixed choices in milliseconds without generating text, and it is tested on two decision tasks with tight constraints. In bullet chess, Jev's pairwise judgments are distilled into a compact evaluator that runs inside Stockfish search, and the resulting bot climbs above a 2200 Lichess bullet rating against other bots. With a fixed labeling budget, averaging Jev's and Qwen3-32B's labels beats spending the whole budget on Qwen by 9.6 Elo. In passage reranking, Jev's labels alone train a reranker as good as Qwen's, using 21 minutes of API calls instead of 5.1 GPU-hours.

arXiv:2610.09188 PDF

Kuration SDK: Addressing the Virtual2Real Gap via Data Curation

Nirmit Desai, Eric Song, Mayank Sengupta, Tejal Bedmutha, Siri Reddy, Sahiti Dharmavaram et al. Standard world-model metrics such as FVD, LPIPS, and JEDi do not match how playable action-conditioned world models actually are, a mismatch the authors call the Virtual2Real gap, which they demonstrate with diffusion world models trained on Counter-Strike gameplay. They argue that curating the raw training data and measuring its diagnostic properties before training gives a more reliable signal than these benchmarks. They release Kuration SDK, an open-source toolkit for physical-AI data curation, and use it to find why two world models with nearly identical LPIPS and FVD scores played very differently.

arXiv:2610.09305 PDF

LLM-Enabled UAV Dispatch: A System-Level Survey and Taxonomy

Xiao Han, Aoyang Quan, Xiangyu Zhao, Xiangjie Kong, Guojiang Shen Unmanned aerial vehicle (UAV) dispatch is moving from isolated path planning and resource allocation toward system-level coordination, and this survey maps how large language models fit into that shift. It treats LLMs as a cross-layer semantic orchestration layer that connects human instructions, external solvers, and distributed control modules, rather than as standalone add-ons. The literature is organized into four dispatch paradigms: pipeline, global assignment, decentralized agentic, and divide-and-conquer. For each one, the survey analyzes decision logic, control flow, and the roles LLMs can play, such as semantic parsing, solver orchestration, multi-agent coordination, and safety assessment. It closes with open challenges including latency-aware reasoning, grounding reliability, physical feasibility guarantees, and edge deployment.

arXiv:2610.09466 PDF

MARS: Malware Analysis with Rule-Based Scoring of LLM Claims

Hyeongjun Choi cross-listed MARS is a malware triage framework that compares two ways of using an LLM: asking it directly for a verdict, or having it emit behavioral claims that an external rule-based policy scores. Every model receives the same static evidence bundles; the test set covers 1,195 PE and ELF binaries in 1,001 near-duplicate clusters, with six language models and a deterministic-rules baseline. Direct classification is more accurate for all six models, by 3.7 to 20.9 percentage points, and usually has higher malicious-alert recall, while claim mediation does not reduce variation across models. Retained claims do make the verdict inputs inspectable and let the policy be revised without new model calls, including for models the provider has since withdrawn.

arXiv:2610.09553 PDF

EvoSignal: LLM-Guided Evolutionary Design of Modular Traffic Signal Control Programs

Leizhen Wang, Peibo Duan, Zhenlin Qin, Yancheng Ling, Jian Xu, Yue Wang et al. Instead of having an LLM choose traffic signal phases online, EvoSignal uses an LLM inside an evolutionary loop to write modular control programs. The programs have separate modules for feature extraction, local phase prioritization, and optional network-level adjustment, and they evolve under feedback on congestion and signal operation. The final programs are inspectable and run without any LLM inference. On two real road networks across five scenarios, the default program reduces waiting time by 16.8–49.2% relative to the best of 20 conventional, reinforcement learning, and LLM baselines in each scenario, and a variant transfers unchanged to the other scenarios while staying near the top.

arXiv:2610.09563 PDF

Structured pre-generation elicitation versus single-shot prompting in AI-assisted enterprise decision-making: a randomised online experiment

William Scott-Jackson cross-listed Professionals who let generative AI both produce and evaluate their work may accept weak output and engage less with the reasoning, a risk the authors call cognitive surrender. In a randomized online experiment, a scaffolding prototype called Cognistance asked users to clarify context, choose a strategic direction, and explain their reasoning before the AI drafted an enterprise strategy document. Rated quality was 32% higher with the scaffold, largely because off-task deliverables fell from 34% to 5%, at a cost of about 2.4 extra minutes per task. Participants' self-ratings of quality did not track the rated quality in either condition.

arXiv:2610.09593 PDF

Artificial intelligence pathways from weather to climate

Tom Beucler, J. David Neelin, Hui Su, Shivanshi Asthana, Chris Bretherton, Will Chapman et al. cross-listed Deep learning weather models trained on atmospheric reanalyses now rival physics-based models from nowcasting out to subseasonal-to-seasonal lead times. This review asks what it would take to extend them to climate timescales, where the goal is a reliable statistical response to forcings such as greenhouse gases that lie outside the observed record. The authors propose two minimum requirements: external forcings must enter the model explicitly enough to be varied independently, and robustness must be stress-tested on extremes and counterfactual trajectories. Comparing the throughput of leading AI emulators and hybrid physics-AI models with GPU-ported dynamical models, they argue that AI can shorten time-to-solution by advancing only the target variables with longer time steps, and that AI downscaling can partly replace explicit fine-scale resolution.

arXiv:2610.09770 PDF

Efficient Provably Private Classification with a Tabular Foundation Model

Talal Alrawajfeh, Cristiana Diaconu, Ossi R\"ais\"a, Sebastian Rodriguez Beltran, Yuan He, John Bronskill et al. Tabular foundation models adapt quickly to new datasets but offer no formal privacy guarantees and are vulnerable to membership-inference attacks, while traditional differentially private learning is slow, loses substantial accuracy, and is easy to apply incorrectly. PrivTab is a tabular foundation model for differentially private classification that builds the privacy mechanism into its architecture and, after pretraining on simulated datasets, uses in-context learning to turn sensitive rows into compact, provably private summaries. It outperforms private linear and neural-network baselines under moderate-to-strong privacy, shows negligible membership leakage, and stays well calibrated. Because it needs only a single forward pass, it cuts dataset fitting time by a factor of 10,000.

arXiv:2610.10068 PDF

When Algorithmic Exploration Becomes Cheap: A Case Study of Agentic Research in EDA

Keren Zhu, Yu Deng, Xiaoyu Hao, Liwen Jiang, Zijian Jiang, Cunqing Lan et al. cross-listed Electronic design automation (EDA) researchers ran eight deliberate trials of agentic algorithm exploration, several on topics outside their expertise, with participants that included students without publication experience. With little human intervention in the algorithms, the agents developed mathematical constructions, analyzed existing tools, and implemented improvements, though some efforts fell short of their practical goals. An AI-assisted analysis of 8,420 papers from EDA venues classified 97.7% of 2,380 primary-core papers as computationally closed, which the authors argue makes much of the field an executable environment for agents. They raise questions about how research should be validated and rewarded when results become easier to produce than to check.

arXiv:2610.10129 PDF

PairAudit: Guiding Human Review with Graph Tokens under Distribution Shift

Jiran Tao, Binyan Jiang Intrusion detectors can misclassify unseen attacks with high confidence, and human reviewers can only check a limited number of cases, so review that targets uncertain predictions misses confident errors. PairAudit builds graph tokens that capture prediction patterns across connected nodes and uses unusual relational patterns to flag likely errors in an existing detector's outputs. Human feedback then decides whether those flags should change review priorities. Across several security tasks, PairAudit corrects more errors under a fixed review budget than uncertainty-based review, including on unseen attacks, with all review costs counted and no detector retraining.

arXiv:2610.10260 PDF

Cross-Domain Pretraining for Steady-State Neural CFD Surrogates

Anthony Zhou, Amir Barati Farimani, Shirley Ho, Rudy Morel Neural surrogates can speed up computational fluid dynamics (CFD) simulation, but they rarely generalize to geometries and applications outside their training data, so each new use case usually needs a fresh, expensive dataset from numerical solvers. The authors train surrogates on pooled steady-state datasets that differ in geometry, boundary conditions and fidelity, then test them zero-shot and few-shot on held-out datasets. Fine-tuning a cross-domain pretrained model gives 2-3x lower error at the same sample size, or needs 8x fewer samples for the same error, compared with training from scratch. The gain holds across architectures, grows with model size and pretraining diversity, and simply pooling existing steady-state datasets turns out to be enough.

arXiv:2610.10398 PDF

How assigned AI use before class shapes active student engagement in class

Dan J. Wang, Neelam Modi Jain, Vanessa Burbano, Jorge Guzman, Daniel Keum, Soomi Kim et al. cross-listed Most evidence on AI learning tools comes from test scores, so less is known about whether they change how students behave in live classes. In a preregistered field experiment, 759 MBA students across ten course sections were each randomly assigned two of ten sessions to prepare for using a voice-based AI discussion partner. After two uses, students made about 31% more voluntary contributions in later class sessions. Heavier users also reported more comfort speaking up and more perceived learning, but no change in focus or motivation.

arXiv:2610.10463 PDF

107 more specialized papers

Agents 76

Task-Oriented Key-Layer KV Communication for Efficient Latent Multi-Agent Collaboration

Dongsen Zhang, Peipei Li, Zekun Li, Wenjun Xu Multi-agent LLM systems can communicate by passing internal KV-cache states instead of generated text, but existing methods try to faithfully transmit the sender's full state, which costs heavy communication and compute and can include redundant information. KITE is a training-free method that instead aims to send only what the receiving agent needs for its task. It picks a single key layer using a criterion based on how much the receiver's reasoning trajectory would be distorted, sends only that layer's KV working memory, and uses the same layer as the entry point for latent reasoning. Across seven benchmarks, two model families, and three model scales, compared with sending all layers it cuts communication volume 28–36×, speeds up end-to-end inference by up to 3×, and improves accuracy by up to 23.3 percentage points.

arXiv:2610.08820 PDF

CredLeakBench: Evaluating Credential Leakage and Recovery in LLM Agents

Rafid Ahmed, Joseph Fioresi, Mubarak Shah, Yuzhang Shang cross-listed LLM agents that handle email, banking, and bills can be phished into revealing credentials, and a safe agent has to tell malicious requests from genuine ones rather than simply refusing to act. CredLeak-Bench is a sandboxed benchmark that covers both user-directed logins and autonomous inbox monitoring. It pairs phishing scenarios with legitimate counterparts and measures leakage from what agents actually submit rather than what they report. Every tested model leaked sensitive information, including during autonomous inbox monitoring with no login request from the user. Most mitigations that reduced leakage also hurt performance on genuine tasks.

arXiv:2610.08871 PDF

An Empirical Study of Agent Skills' Downstream Utility

Yu Cheng, Dehai Zhao, Zhongxin Liu, Qing Huang, Zhenchang Xing, Xiaoxue Ren Agent Skills package procedural guidance and resources for reuse, and this study asks when a relevant Skill actually improves an agent's task performance. On 87 SkillsBench tasks, it measures utility as the change in pass rate relative to running without Skills, comparing the same Skills across nine model-harness configurations and drawing alternative candidates from a corpus of 37,596 marketplace Skills. The same Skills help some configurations and hurt others on 36.78% of tasks, and execution traces show that recommended procedures can become a burden. Reranking candidates by how well they support the operations a task requires raises first-choice pass rates by 4.35 to 5.80 percentage points. The authors also derive 17 authoring practices and find that organizing several Skills as a Stage Plan or Dependency DAG beats specifying use order alone.

arXiv:2610.08875 PDF

Humanize: Judgement Engineering for Agentic Coding

Sihao Liu, Ligeng Zhu, Zijian Zhang, Dongyun Zou, Zhengyang Zhang, Changye Li et al. The agent that writes code is a weak judge of whether that code is done. Humanize is a multi-agent coding workflow in which a human approves a plan contract, a builder agent implements it in rounds, and a reviewer agent from a different vendor decides completion, while deterministic hooks route work and enforce 72 mechanical gates. Studied through deployment, 118 public postmortems, and applications, the authors report a 567-file gem5 build migration, top-three placements in the MLSys 2026 FlashInfer contest, and full scores on IOI 2026, IMO 2026, IPhO 2026 and IBO 2024, plus 672/672 on PutnamBench. The postmortems show that independent review catches unsupported builder claims but that knowing when to stop remains a weakness, with two thirds of rounds occurring after implementation was accepted. The authors stress that this evidence is observational rather than a controlled comparison of workflows.

arXiv:2610.08900 PDF

Sequential Probabilistic Uncertainty Estimation for Parallel Multi-Agent Reasoning Systems

Tunyu Zhang, Zihao Zhao, Yusong Zhao, Haizhou Shi, Zhuohang Li, Haoxian Chen et al. In parallel multi-agent reasoning systems, several LLM agents solve the same problem over multiple rounds and their outputs are aggregated, yet estimating how reliable the final answer is has received little study. SAUCE (Sequential Agent Uncertainty through Consensus Evolution) is a training-free estimator that treats system uncertainty as sequential inference over a latent belief. It combines round-level agreement with generation-uncertainty signals through a filtering-style update. Across five backbone models, five benchmarks, and two multi-agent protocols, it improves misclassification detection, selective prediction, and calibration over log-likelihood-based and multi-agent-specific baselines.

arXiv:2610.08901 PDF

Agent Plasticity: Measuring Self-Improvement Through Experience

Harman Singh, Anton Bakhtin, Rulin Shao, Gabriel Synnaeve, Ilia Kulikov, Rob Fergus et al. Most agent evaluations measure capability at a fixed point in time rather than how well an agent learns from experience. The authors study self-improvement in a controlled setting where agents distill past experience into reusable artifacts inherited by future instances. They define agent plasticity as the efficiency with which experience is converted into gains on held-out tasks, accounting for learning cost. Frontier models follow sharply different improvement trajectories, in-distribution gains transfer only partially out of distribution, and the best-performing agent at the end is not necessarily the one that improves most efficiently. Failure tracing shows that low-plasticity agents often fail to reuse relevant artifacts, while more plastic agents can still fail because of artifact quality or how the artifacts are applied.

arXiv:2610.08902 PDF

Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station

Wenyu Du, Stephen Chung The question here is whether AI agents can do open-ended scientific research without well-defined metrics, studied in Station, an open-world environment where multiple agents simulate a scientific community. The authors add a Supervisor mechanism and periodic Meta Reflection to sustain exploration, then give agents the research questions from three ICLR oral papers while withholding the results and web access. Station rediscovers 62.7% of the papers' findings on average, against 15.4% for Codex Multiagent-v2 and 14.4 to 20.6% for AI Scientist-v2. On two tasks without reference papers, some agent discoveries closely match results researchers published after the models' knowledge cutoff.

arXiv:2610.08927 PDF

GraphOPD: Graph-Augmented On-Policy Distillation for LLM Agents

Bohan Lin, Liyi Chen, Zhuoning Guo, Muyang Li, Qimeng Wang, Yan Gao et al. On-policy distillation trains LLM agents with step-level guidance from a teacher, usually focusing on the steps where teacher and student disagree most. In multi-turn tasks this misfires: an early mistake shapes every later context, so the teacher stops flagging it, and on one agentic benchmark picking the highest-divergence steps did no better than random selection. GraphOPD builds a dependency graph of steps from the environment's own record of state changes, scores each step with a random-walk stationary distribution, and combines that score with the divergence signal to decide where to concentrate supervision. Across three model scales and eleven baselines on ALFWorld, WebShop, and SearchQA, it improves over the strongest baseline by up to 5.8 percentage points, and the structural score transfers to out-of-domain tool-integrated reasoning.

arXiv:2610.08959 PDF

Verify Less, Evolve More: Training Idea-Level Critics for Verification-Efficient ML Evolving Agents

Jiamu Bai, Lizhu Zhang, Xin Yu, Yanhong Wu, Zellux Wang, Serena Li et al. Self-evolving agents that improve machine learning code must train and evaluate models to check each idea, which limits how fast they can evolve, and frontier models are only modestly useful for picking which ideas to test. The authors train idea-level critic models that predict whether a proposed change will beat the current solution, first by supervised fine-tuning on critiques generated by Gemini-3.1-Pro and then with GRPO. The critics outperform Gemini-3.1-Pro at judging ideas and produce better final solutions under the same verification budget. Used as learned reward models during policy training, they allow many more policy updates for the same amount of real verification.

arXiv:2610.08993 PDF

Whose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory

Hongyu Gu, Xinchang Li When an LLM agent's persistent memory stores sentences from its own deliberation, such as rejected plans, simulated tool results, or other speakers' beliefs, those statements can later resurface as facts. The authors call the missing context "discourse ownership": the world, branch, or speaker that a proposition belongs to. They show that models already carry a causally active internal signal for ownership that standard memory interfaces throw away. Their commit rule, CASK (Causally Anchored Scoping Keys), preserves stable relational features of that signal rather than fragile internal coordinates, so shared-world facts go into durable memory while provisional content stays scoped. On long-conversation conflicts and tool-agent traces, CASK prevents provisional content from contaminating later answers and improves memory admission.

arXiv:2610.09008 PDF

Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation of Small Language Models in a Deployed Agentic Home-Automation System

Panagiotis Kasnesis, Christos Chatzigeorgiou, Lazaros Toumanidis, Amalia Contiero Syropoulou Agentic systems make several structurally different kinds of LLM calls but typically use one model, chosen for the hardest call, for all of them. The authors evaluate 9 models, from 0.8B parameters up to a frontier hosted model, on the five call sites of the deployed Wactorz home-automation framework, using its unmodified production prompts and two real Home Assistant installations (2,520 scored calls). The best local model is statistically indistinguishable from hosted models at four of five call sites, with only code generation separating them, and bigger models are not uniformly better. Small models showed a safety failure at the actuation site: Gemma4 E2B acted on 87.2% of requests for devices the installation does not own. In live use, hosting only the two generative call sites matched fully hosted quality for 28% of the cost.

arXiv:2610.09021 PDF

When the Governor Becomes the Disturbance: Control-Generated Disturbance and Cost-Aware Backoff in Governed Tool-Using Agents

Veronique Ziegler Supervisory governors that regulate tool-using agents can themselves become a source of disturbance. In a controlled file-recovery environment, where stronger regulation deliberately triggers tool failures, a governor that ignores these costs can block the agent persistently and prevent task completion. The authors propose a backoff rule that lowers intervention probability based on a moving average of observed induced failures. It improves completion for a hand-coded stochastic agent and in 576 Gemini 2.5 Flash episodes across 6 tasks, with intermediate backoff strength performing best. The authors note that the failure mechanisms are imposed and directly observable, so whether the results apply outside this controlled environment remains open.

arXiv:2610.09037 PDF

TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery

Yuanzhe Li, Pengxin Wang, Yuxin Ren, Jianing Deng, Jingtong Hu, Song Wang et al. Long-horizon agents make many model calls, but pruning methods designed for question answering badly degrade agents trained with reinforcement learning (RL). TAP (Trajectory-Anchored Pruning) pairs structural pruning with on-policy recovery: interactions stay anchored to a frozen dense teacher's trajectories while the student generates each reasoning-action response, so student mistakes do not compound across turns. Channels are re-scored iteratively using gradients of the recovery objective. With 60% of feed-forward (FFN) channels removed, 7B agents keep 99.2% of their task success rate on ALFWorld and 88.0% on WebShop, while GPU time per successful task drops by about 22% and 17%.

arXiv:2610.09074 PDF

Large-scale Repository Engineering via Agent-Native Reusable Code Primitives

Haibo Jin, Peng Kuang, Xucheng Yu, Jerry Wang, Dehao Wu, Haohan Wang cross-listed Building complete software repositories with LLM agents is hard because modules, interfaces, configurations, tests, and dependencies all have to fit together. The authors introduce Code Primitives: reusable executable components with interface contracts, dependency closures, validation tests, and provenance, each of which uses an LLM to adapt itself to the target repository. They collect 1,424 of them in the searchable library CodeFace and build LEGO, a system that activates, integrates, and test-revises these primitives. On the new LEGO-REPO benchmark of 522 repository-reconstruction tasks, the best of 13 backbones scores only 0.318 on its own, and LEGO raises GPT-5.6-terra from 0.318 to 0.513 (+61.4%). Adapted primitives outperform both retrieved code given as context and code copied in unchanged.

arXiv:2610.09079 PDF

RippleCP: Measuring Counterfactual Checkpoint Advantage in Coding Agents

Mayur Akewar, Ravi Ranjan Checkpoint systems for coding agents decide how to snapshot and roll back state, but not which safe boundaries are actually worth checkpointing. The authors define counterfactual checkpoint advantage, the recovery cost saved by checkpointing at a given point rather than skipping it, and measure it by driving checkpoint and no-checkpoint branches to the same failure under matched conditions. On 12 SWE-bench Verified tasks with 106 recovery branches, checkpointing saves 49.4 s per task. The first checkpoint returns 100 s on 156.5 s of protected work, while a second one a step later returns -1.1 s. Because recovery re-derives work rather than replaying it, the amount of work preserved is a poor guide to the time a checkpoint saves.

arXiv:2610.09088 PDF

Constraint Tree Exploration for Learning from Language Feedback

Shaoang Li, Daniel R. Jiang, Jian Li When an agent misreads natural-language feedback about why an action failed, it can wrongly rule out valid solutions. The authors model user intent as hidden constraints and introduce TRACE, which arranges candidate constraints in a tree and accepts a refinement only after repeated test actions produce no contradicting feedback. They prove high-probability coverage bounds for two variants, one that only detects contradictions and one that can also name a violated constraint. On RecMovie, TRACE-Identification reaches 73% and 86% success under caps of 20 and 60 outputs, versus at most 42% and 48% for prompting baselines.

arXiv:2610.09107 PDF

GeoNatureAgent (GNA): A Framework and Benchmark for Pre-Production Evaluation of Tool-Using Agents on Geospatial and Environmental Tasks

Gabriel Diaz-Ireland, Diego Prieto-Herr\'aez, Mario Garc\'ia Peces, Javier Vel\'azquez, Benjamin Zaitchik, Devika Jain GeoNatureAgent is a framework for testing whether tool-using LLM agents choose the right operations on real geospatial APIs before they are deployed. It exposes a fixed set of sixteen tools as a Model Context Protocol (MCP) server and uses a 103-task benchmark with a deterministic scorer, so the model under test is the only thing that varies. Across nine LLMs, Claude Sonnet 4 scores highest at 61.7%, and DeepSeek V3.2 reaches 93% of that capability at 11.3x lower cost. The best model scores 24 to 36 points below results reported on general GIS benchmarks under strict scoring, but partial-credit scores are similar, which suggests much of that gap comes from how strictly the tasks are scored.

arXiv:2610.09112 PDF

From Uncertainty to Action: Learning to Steer LLM Agents

Hanwen Li, Jinhao Duan, Guanhua Zhu, Junchi Lu, Bo Shen, Chenxi Yuan et al. Steering an LLM agent involves deciding whether to intervene, at which step, and by which mechanism. The authors build a stepwise outcome table of about 82,000 counterfactual continuations by steering 1,864 trajectories at every step with four different mechanisms. The table shows that uncertainty signals can flag failing trajectories but cannot reliably find the step where steering helps. They propose VoS (Value of Steering), a monitor trained on these measured outcomes with a trigger that limits how often it disturbs successful runs. VoS improves over unmodified execution in all 12 settings by 7.8 points on average and beats the strongest of five uncertainty-triggered baselines in 11 of them.

arXiv:2610.09115 PDF

CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use

Gennadiy Savrasov, Maksim Elistratov, Nikita Gavrilov, Albert Garifullin, Oleg Pavlov, Soslan Kabisov et al. Reconstructing editable parametric CAD programs from 3D meshes is hard because no single source of proposed CAD operations works well across all part shapes and reconstruction stages. CADFather is an agentic system in which a vision-language model inspects renders of the target and of intermediate results. It then decides which candidate programs to extend, which learned or algorithmic tools to call, and when to stop, while numerical optimization refines parameters. It uses only pretrained models with no additional training and is evaluated on DeepCAD, Fusion360, MCB, CADENA-Bench, CADBench and BenchCAD, along with an analysis of the trade-off between cost and reconstruction quality.

arXiv:2610.09127 PDF

Finding Blind Spots in AppWorld and WorkArena Task Verifiers

Richard Abrich Execution-based verifiers decide whether an agent completed a task, and this audit tests the shipped verifiers of AppWorld and WorkArena with source-informed mutation tests, without modifying any shipped checker. In AppWorld, duplicating a non-idempotent write creates an extra record but keeps every checked field correct, and the verifier accepts 6 of 15 constructed wrong effects; a cardinality check added to copies of the checkers rejects all six while still accepting valid runs. In WorkArena, 21 of 23 reruns that write extra nondefault field values are confirmed wrong by independent readback, yet all 23 pass. The authors stress that these selected cases confirm the blind spots but do not estimate how often they occur across the benchmarks.

arXiv:2610.09142 PDF

ToolRACER: A Robust Agentic Conversation Emulation Resource for Agent Training and Evaluation

Arkajyoti Chakraborty, Aryan Tayal, Ishika Agarwal, Tanner Sorensen, Justin Chiu, Alessandro Di Bari et al. Function-calling benchmarks mostly show cooperative users, so they leave agents unprepared for non-cooperative or adversarial conversations. ToolRACER is a synthetic data pipeline that coordinates emulated users, assistants, and tools to generate validated multi-turn conversations, with adversarial user behaviors injected. Its benchmark, ToolRACERBench, contains 5.6K trajectories across six domains and 55 personas, and about 66% of them contain failure-prone scenarios. Training on it improves end-to-end agentic accuracy on tau2-bench and ACEBench, with the largest gains for small language models when it is mixed with in-domain data.

arXiv:2610.09163 PDF

Bookkeeping, Composition, or Unreachable Gold? Reading MemoryAgentBench's Conflict-Resolution Scores Against a Frozen Last-Write Resolver

Egor Pakhomov, Erik Nijkamp The Conflict Resolution split of MemoryAgentBench is read as measuring selective forgetting, so the authors run the benchmark's own rule, that the newest statement about a fact wins, as a fixed resolver with no learning. This rule alone answers 80.25% of questions under the official metric. Of the rest, 67 items have gold answers that can only be reached through overwritten statements, and they make up a third of the multi-hop questions at 262K context. Two long-context models and a re-implemented BM25 agent score 82.6% to 84.7% and 41.6% on items the rule solves, but only 6.0% to 11.9% on those 67, so the authors argue scores should be read per item rather than as one aggregate.

arXiv:2610.09193 PDF

FreeEvolve: Learning to Evolve Beyond Fixed Loops

Lecheng Kong, Like Hui, Nikos Kanakaris, Prithwish Jana, Sahika Genc, Narayanan Sadagopan Agent evolvers automate the design of prompts, skills, and workflows around LLM agents, but they still follow a hand-designed search loop. FreeEvolve lets the evolver decide within resource limits what to test, how much evidence to collect, which candidates to keep, and when to stop. Those decisions follow an editable evolution skill that is itself improved through meta-evolution. Across tau3-bench, ARC-AGI-2, ARC-AGI-3, and Terminal-Bench 2.1, it improves held-out metrics by 13.6 points on average and matches or beats hand-designed evolvers. Meta-evolved skills add 6.9 points over the seed skill on fresh target agents and transfer across environments.

arXiv:2610.09197 PDF

AGAR: a reinforcement learning substrate for LLM program evolution

Haoran Li, Zengle Ge, Xiaomin Yuan, Yui Lo, Haoxin Li, Songlin Zhou et al. LLM-driven program evolution, where a model rewrites candidate programs and a search loop keeps the best, depends on hand-set choices for parent selection, mutation strength, diversity, memory, and scoring. AGAR (Algorithm Generation As RL) formalizes this loop as a Markov decision process whose action is the modular prompt prefix the model is conditioned on, not the program it emits. This lets standard reinforcement learning (RL) estimators for credit assignment, value estimation, exploration, and memory each be plugged in or switched off independently, with no gradient training of the backend model. Across 19 tasks, two backends, and three seeds, AGAR beats the stronger of two published baselines on most tasks, with gains concentrated in competitive programming. The formalization also shows that prior systems are implicitly zero-discount, because fitness does not depend on an individual's successors.

arXiv:2610.09215 PDF

Trajectory Abstraction for the Science of Language Agent Behavior

Tianqiang Yan Scientific study of language agents needs behavioral variables that let hypotheses generalize across tasks and models. The authors frame this as learning and testing a hierarchy of trajectory abstractions. A recursive procedure measures role- and phase-indexed events, proposes time-constrained relations between them, tests their stability, and builds episode-level motif variables, with explicit measurement functions linking each level back to raw trajectories. The work is mainly a methodological proposal: it derives a finite-depth bound on accepted reductions and a finite-sample test for intervention consistency, and it illustrates the procedure with constructed examples rather than large-scale experiments.

arXiv:2610.09237 PDF

The Winner's Curse in LLM Self-Improvement Loops: Selection Noise, Lock-in, and Acceptance Rules

Litao Hu, Yutong Tang Self-improving LLM systems propose changes to themselves and keep those that score better on a small evaluation set, which amounts to selection under measurement noise. In experiments where Qwen models rewrite their own instructions and every candidate is also scored on 600 held-out items, most proposals after the first are harmful. In a pre-registered study, greedy loops' final selection scores exceeded held-out accuracy by 13–20 points with 16 selection items and by 1–5 points with 256, and the alternative acceptance rules tested did not beat greedy acceptance over whole runs. The same inflation appears in the validation scores of GEPA and MIPROv2. The authors recommend reporting held-out gains with uncertainty, noting that a 64-item clean evaluation removes the average bias but individual estimates can still be off by about 6 points.

arXiv:2610.09239 PDF

We Query, Therefore We Compute: On Oracle Computation beyond the Machine, with an Application to Agents

Kefan Liu, Fengning Ou, Yelin Luo, Jingdi Lei Agentic systems borrow operating-system ideas such as scheduling, caching, and isolation piecemeal, without a shared foundation. The authors build an abstract machine that treats the LLM as an Oracle: a two-stack pushdown automaton gains one instruction that sends a whole stack to the Oracle as its query and appends the answer to that same stack. Two symmetry breakings, one in storage and one in transitions, show that Agents and Workflows are two placements of the same task program, and the model is used to analyze caching and scheduling. To show the machine can be built, the authors propose ArchNights, an extended RISC-V instruction set and Linux-style operating system that runs on the gem5 simulator with an LLM as the Oracle.

arXiv:2610.09243 PDF

Relevance Is Not Sufficiency: What Actually Closes the Evidence Gap in Long-Term Memory QA

Yufeng Li, Shuxin Li, Zhenhua Xu, Junxian Li, Peng Zeng, Sheng Yao et al. LLM agents that work with a user across many sessions store their history in external memory and retrieve a few records to answer each question. Records ranked by relevance can each be on-topic yet together miss a fact the answer needs. Drawing on the distinction between relevance and sufficiency in legal evidence scholarship, the authors treat retrieval as building a sufficient memory set and propose Budgeted Flat Reconstruction (BFR). BFR uses Formal Concept Analysis for Memory Selection (FCA-MS) to break a question into information requirements and pick records that jointly cover them, then acquires more records through deeper text search or entity and session views until a budget runs out. On LongMemEval-S it raises judged accuracy from 72.4% to 82.2%, and it also beats same-store adaptations of recent agent-memory systems on LoCoMo.

arXiv:2610.09348 PDF

DUDA-Bench: Benchmarking LLM Agents on Multimodal Data-Driven Urban Diagnosis

Yizhi Song, Hang Ni, Weijia Zhang, Hao Liu Urban diagnosis means combining heterogeneous data to identify, locate and explain city problems, and today it relies on labor-intensive expert workflows. DUDA-Bench turns this process into a hierarchical, interactive agent benchmark with 86 atomic tasks and 22 workflow tasks across four analytical stages, grounded in multimodal data from 12 cities. Evaluations of seven backbone models and five agent systems reveal a substantial gap between isolated analytical skill and end-to-end diagnosis. Trajectory analysis shows that unresolved evidence gaps propagate across stages, while successful runs revise their assumptions and actions in response to feedback.

arXiv:2610.09374 PDF

RSI-Forge: From Research Papers to Environments for Recursive Self-Improvement

Renxiong Wang, Darvin Yi, Abril Herrlein, Anas Mahmoud, Advait Gosai, Lisiman Hua et al. RSI-Forge is a multi-agent pipeline that turns published papers into executable research environments with automated evaluators, which self-improving agents can work on. Three agents handle construction, reproduction and review, and each paper's method is reimplemented independently to set a baseline score. The authors release 210 environments across 18 fields, 90 of them reviewed by human domain experts, who rated them more critically than agent judges on shortcut resistance and faithfulness to the source paper. Across three successive attempts by four models on 120 environments, at least one model improved after its first attempt in 84% of environments, and models beat the reproduced paper methods in 68 of 120 environments, mostly through work beyond parameter tuning.

arXiv:2610.09426 PDF

MIMESIS: Learning User Simulators as Training Environments for Interactive Agents

Hoang Phan, Dat Huynh, Andrey Zhmoginov, Qi Zeng, Wancen Mu, Yue Cao et al. Training interactive language agents needs realistic users, but off-the-shelf assistant LLMs make poor stand-ins because they are overly cooperative, explicit, and homogeneous. MIMESIS is a purpose-built user simulator trained on human conversations with explicit reasoning supervision and 13 behavioral patterns taken from real interactions. The 9B model scores 65.7 on SOUL-Index, surpassing the strongest frontier model, and improves behavioral fidelity by 13.4 points over Claude-Opus-5 on RealUserSim. Agents trained with multi-turn reinforcement learning against the frozen simulator generalize better to nine unseen user simulators than agents trained against GPT-5.5. A further method, Coached On-Policy Self-Distillation (CSD), converts the simulator's private reasoning into coaching notes that give dense token-level supervision and yields additional gains.

arXiv:2610.09484 PDF

Ream: Unfolding Mutual Awareness in Human-Agent Workspaces

Peiling Jiang, Sangho Suh, Varsha Kishore, Jonathan Bragg, Haijun Xia, Pao Siangliulue et al. When AI agents share a workspace with humans, agents act faster than people can monitor, and users' shifting interests often go unstated in chat. Ream is a literature-review workspace that supports mutual awareness through structured artifacts, bidirectional engagement tracking, and localized visualizations, so each party can see what the other has read and done, and agents can retrieve that same history. In studies with eighteen researchers, participants used these traces to inspect evidence, steer agents, and communicate through annotations, and shared histories helped agents build on earlier work. The findings inform how engagement traces can support transparency and coordination in human-agent knowledge work.

arXiv:2610.09497 PDF

DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists

Samuel Margolis, Paul Schmiedmayer, Alan Huang, Ethan Chen, Ishan Bhattacharjee, Atman Shah et al. Real biobanks lack known causal ground truth and restrict access to participant-level data, which makes it hard to train or evaluate AI agents on end-to-end drug target discovery. DrugTargetWorld procedurally generates synthetic biobanks of 54,000 participants each, with genotypes, proteins, health records, outcomes, and synthetic MRI, and with causal structure that the evaluator knows but the agent cannot see. Agents must define a disease phenotype, identify causal driver proteins, infer the beneficial direction of modulation, and optionally run virtual wet-lab experiments. Across 540 episodes, the top agents Opus 5 and GPT-5.6 Sol scored only 39.98 and 35.38 out of 100, and although both recovered 64% of causal drivers, no agent reliably rejected misleading non-causal proteins.

arXiv:2610.09558 PDF

World Potential Model: Pretrained World Knowledge as Progress Potentials

Jun Zhao, Jixin Tang, Yang Shu, Jinyang Wu, Yuyang Lu, Jingqi Tong et al. Long-horizon language agents usually get feedback only from the final task outcome, which gives little signal about whether intermediate steps made progress. A World Potential Model (WPM) uses off-the-shelf pretrained models, without task-specific fine-tuning, as goal-conditioned judges of realized task progress. In ALFWorld and ScienceWorld these judges recover progress structure well above chance. Anchored to task milestones, the judgments become scalar potentials whose step-to-step differences provide credit for each action. Under matched comparisons, WPM-guided optimization beats outcome-only GRPO in every evaluated configuration.

arXiv:2610.09560 PDF

Correct Answers, Unsupported Findings: Evidence Binding in Forensic Reconstruction of LLM Agent Logs

Taehyeon Yun, Dongho Kim, Geonwoo Kim, Juyoung Seo, Minseok Hur, Moohong Min cross-listed Forensic reconstruction of what an LLM agent did requires both the correct value and a link to the preserved record that supports it. The authors audit this distinction on 64 mechanically checkable cases from saved AgentDojo Banking runs, with LLM readers reconstructing source relationships under controlled changes to the visible evidence and to how citation identifiers are bound to records. Without an explicit binding table, Sonnet recovered every literal source location but made unsupported citation-source claims in 26 of 28 cases, 22 of which still matched the reference answer. Adding explicit bindings improved grounded reconstruction, renaming identifiers alone did not consistently help, and a deterministic comparator was always correct or abstained, which shows that factual agreement alone is not a sufficient evaluation metric.

arXiv:2610.09581 PDF

Learning Situation-Conditioned Thinking Policies for Long-Term LLM Agents

Hong Su Long-running LLM agents need to reuse past reasoning experience without memory and context growing indefinitely. Instead of retrieving or summarizing past content, the proposed framework learns a lightweight policy that predicts what the agent should think about in the current situation, including situations that evolve over time, and leaves the detailed reasoning to the LLM. Experiences are periodically analyzed across independent episodes to find recurring long-range regularities, which become new thinking knowledge that the policy absorbs. The learned policy raises DeepSeek reasoning F1 from 0.789 to 0.868 and cuts per-query processing time from 0.36 ms to 0.038 ms at 30,000 stored situations.

arXiv:2610.09590 PDF

Coding-Agent Benchmarks Should Match Their Users' Task Flows

Igor Slinko, Yaroslav Golubev, Sergey Titov From 4,782 real coding-agent sessions by software engineers in JetBrains IDEs, the authors find that long interactive sessions cover far more task types than issue-derived benchmarks: code questions, planning, review, refactoring, and execution, with users switching between types mid-session. Three public interaction corpora show markedly different Task Flows, so no single interaction distribution counts as universally realistic, and benchmarks should calibrate to a named target use case. They introduce SWE-TaskFlow, which turns any issue-derived benchmark into a multi-turn one by splitting prompts and adding verifiable repository questions, plus a TaskFlow Alignment Score (TFAS) for choosing among generated trajectories. In a pilot on 700 SWE-Bench Pro tasks, solving tasks in several sequential steps roughly doubles agent cost with no stable change in resolve rate.

arXiv:2610.09633 PDF

Shared and structured inputs undermine collective random choice by reasoning AI agents

Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari Random selection underpins resource allocation and auditing, so the authors test how reliably reasoning-model agents make random choices. Behavioural tests on six models found that choices based on identifiers follow hidden threshold and divisibility rules. For GPT-6 Sol and Gemini 3.8 Flash, measurements of a single agent predicted correlated choices across agents given shared identifiers, and biased choices when distinct identifiers shared timestamp bits. Explicit instructions to randomize independently reduced the correlation but did not eliminate it. In a task asking models to randomly pick customer requests for human review, selection was predictable from identifiers or almost never happened, while all four models faithfully followed random draws supplied to them.

arXiv:2610.09667 PDF

Cost-Efficient Theorem Proving via Agent Orchestration in Program Verification

Shuangjie Yao, Nikolaus Holzer, Mark Paul Santolucito, Baishakhi Ray, Suman Jana, Dongdong She Large language model (LLM)-based theorem provers for program verification usually chase pass rates at any cost, which ignores the economics of real software with hundreds of interdependent proof obligations. CoCo-Prover treats cost-efficient proving as decision-making under cost over two-level proof graphs: an AND/OR hypergraph within each declaration linked to a lemma-dependency graph across declarations. Goal selection is symbolic, while an agentic router picks among separately priced specialist agents and configurations using routing rules that evolve as evidence accumulates. On five Lean 4 benchmarks (CLEVER, VERINA, AlgoVeri, NTP4VC, Vero), it beats frontier coding agents and state-of-the-art provers on solve rate everywhere, reaching 100% on two benchmarks, and cuts cost by up to 30.9% relative to the strongest baseline.

arXiv:2610.09681 PDF

System Switch: When Should a Fast Decision Model Stop and Think?

Gian Luca Bailo In a dual-process agent, a fast learned actor makes every decision and hands control to a slow reasoning vision-language model only when a gate opens, while the game keeps running. The authors test this in closed-loop Doom using the open System One typed-decision models (0.15B to 9B parameters) served through llama.cpp. Offline, deferring the least confident 30% of decisions to the reasoner beats random deferral in proportion to the actor's AUROC (rank correlation 0.87), giving gains of about +0.08 to +0.13. In closed-loop play, however, no variant reaches the level exit, and the reasoner mistakes ordinary doors for locked ones. Code, prompts, data, and logs are released.

arXiv:2610.09683 PDF

From Expert-Guided Proof Search to Automated Open-Problem Solving

Adri\'an Z\'ame\v{c}n\'ik, Mat\v{e}j Kripner, Martin Kouteck\'y, Martin Balko, Jan Greb\'ik, Pavel Hub\'a\v{c}ek et al. Bolzano is an open-source multi-agent system for mathematical research: several prover agents work in parallel, a verifier agent checks their work, and the system keeps a human-readable record of research progress. Early use on problems picked by experts produced 8 results whose proofs domain experts verified. Run with no problem-specific human guidance on about 3,800 open problems taken from four sets of papers, it solved about 200 open problems. In one experiment it answered four questions posed in papers accepted to STOC 2026, a top theoretical computer science conference, and the papers' authors confirmed the answers.

arXiv:2610.09769 PDF

Decoupling Logic from Persona: Structural Immunity of Edge LLM Agents to Context Pollution

Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO, Inc.) Small LLM agents on edge devices must keep a persona and reason correctly within one context window, and long, persona-heavy history degrades their logic. The proposed Decoupling Architecture (AO-DA) runs two inference paths on a single 4-bit base model with hot-swappable LoRA adapters: a logic path that sees only the current turn and emits a verifiable structured state, and a persona path that renders that state in character using the full history. In 480 runs on an Apple M2 laptop with Llama-3.1-8B-Instruct and Gemma-3-4B-it, the logic path's prompt stays fixed in size and its outputs are byte-identical at every pollution level, while a single mixed pass drops from a logic score of 0.669 to 0.150 on Llama, mostly by failing to produce the required structured output. The separation costs one extra decode on a topic's first turn (28.2 s versus 18.2 s) and lets the persona be swapped in 1.7 ms, though at the 4B scale both designs collapse when the pollution is fed into the logic path.

arXiv:2610.09772 PDF

Homogenization in Multi-Agent Systems

Prakhar Ganesh, Kyra Wilson, Luca Zappella, Barry-John Theobald, Nicholas Apostoloff, Lucas Monteiro Paes et al. In multi-agent systems (MAS), agents that interact with each other can converge on similar behaviors, which the authors call homogenization. They measure it with three metrics: conformity to the majority, polarization toward extremes, and growing inertia against change over successive interactions. They test it on code generation, hiring and scientific peer review. Homogenization hides and amplifies correlated errors in generated code, lets a biased agent's influence persist long after it is removed from a hiring panel, and creates uneven review standards across research areas. Simple fixes such as more sampling randomness or mixing different models did not reduce these risks.

arXiv:2610.09824 PDF

SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles

Yuyao Ge, Yiwei Wang, Yuchen He, Baolong Bi, Lingrui Mei, Jiayu Yao et al. LLM agents trained with memory-augmented reinforcement learning (RL) can rely on stored skills, but if every skill is kept, obsolete or harmful entries pile up as the policy improves. SkillForge moves each skill through a fitness-driven lifecycle of trial, active, stable and retired states, so the skill library and the model evolve together during training. Before RL, the base model's own rollouts are used to retire low-fitness skills, and the filtered library seeds supervised fine-tuning. RL then continues retiring, stabilizing and mutating skills with an LLM's help. On interactive agent benchmarks, SkillForge achieves the highest aggregate success rate, up to 7.8% relative improvement over the strongest baseline, while keeping the library compact. The authors also release SkillFurnace, a dataset of more than 5,000 annotated records on skill quality and retirement events.

arXiv:2610.09832 PDF

Self-Evolve With a Reference:Anchored Training of Tool-Integrated Agents

Wenjie Liao, Liangjie Zhao, Zehong Cao Self-evolving tool-integrated agents pair a Curriculum Agent that generates tasks with an Executor Agent trained by reinforcement learning on self-consistency signals. Using the current Executor as the only source of feedback causes problems: learning signals vanish once samples fully agree, and uncertainty-based curriculum rewards favor disagreement without showing whether a task actually helps learning. AnchorLoop adds a frozen copy of the previous iteration's Executor as a historical reference. The Executor's advantage compares outputs against both the current and the historical majority answers, while the Curriculum gets an agreement-based task-selection signal. Across 13 benchmarks, AnchorLoop improves over Agent0 by 2.5% on math and 2.8% on general reasoning and keeps improving in later iterations, without external supervision.

arXiv:2610.09856 PDF

Training Advisors for LLM Agents from Task Outcomes

Sergei Polezhaev, Barys Liskavets, Ori Press, Alexander Golubev Caddie trains a critic model that gives natural-language analysis and advice to an LLM agent while it works through a multi-step, tool-using task. Instead of step-level labels or reference critiques, the critic learns through reinforcement learning from whether the agent ultimately succeeds after receiving its feedback, while the agent's base model stays frozen. A Qwen3-4B critic trained on multi-hop question answering improves four base models, three of which were not used in training. On MuSiQue it raises Qwen3-4B's success rate by more than 25 percentage points, surpassing Kimi K3 without a critic. The critic also transfers without extra training to τ³ and DeepDive, and agents can learn to decide when to ask it for help.

arXiv:2610.09858 PDF

LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets

Jun Zhao, Leiming Fu, Yanbo Wen, Yiding Wang, Xuantong Liu, Yang Shu et al. Judging agents only by outcomes can hide which capabilities produced those outcomes, especially in environments that keep changing. LiveMACEBench runs five frontier LLMs as persistent agents in live financial markets, under matched configurations for tool use, persistent memory, rule following and multi-agent collaboration. It scores both realized returns and diagnostics computed from complete decision traces. Over 30 days of live evaluation, returns often diverged from the capability measurements, and similar outcomes came from very different patterns of mechanism use. The authors conclude that having access to a mechanism, using it well, and performing well downstream are distinct measures of agent capability.

arXiv:2610.09872 PDF

AgentTime: Can Agents Estimate and Control Their Own Runtime?

Michael Ofengenden, Maksym Andriushchenko AgentTime is a benchmark of 222 tasks from 18 sources, covering coding, computer use, agentic work, and automated research. It tests whether agents in their native harnesses can work for a requested duration, forecast their own runtime, and estimate elapsed time afterward. Duration-following varies widely: Fable 5.1 in Claude Code deviates from requested runtimes by a typical factor of 2.9x, versus 1.2x for GPT-6 Astra in Codex. Matching the clock does not mean the agent kept working, though: in 14 of 158 reviewed Astra runs, the agent explicitly slept after appearing to finish. Agents also tend to overestimate their natural runtimes, and removing temporal information roughly doubles their retrospective estimation error.

arXiv:2610.09944 PDF

Learning to Accumulate Knowledge with Mutual Information

Yuyang Zhao, Lizi Liao, Leyang Shen, Xiaoyan Zhao, Yang Zhang, Fuli Feng et al. Large language model (LLM) agents can reuse knowledge distilled from past interactions, but training a curator with Group Relative Policy Optimization (GRPO) on standalone task success tends to reinforce generic guidance that duplicates what the knowledge bank already holds. Knowledge Weaver is a reinforcement learning framework that trains a language model to curate reusable knowledge from agent trajectories. It combines a feedback signal inspired by token-wise mutual information (MI) with marginal success rewards, so that new entries carry distinct information and improve task success when added to the existing bank. With 10 retrieved entries it reaches mean success rates of 54.0% on ALFWorld and 42.0% on WebShop, beating GRPO by 16.9 and 18.7 percentage points. With the executor frozen, its banks also outperform prompt-based and human-written knowledge banks.

arXiv:2610.10042 PDF

Loud Failures, Quiet Failures: Fault Detection and Recovery in Tool-Using Language Model Agents

Obada Kraishan Tool-using agents are usually evaluated with working tools, so the authors wrap the environments of an established function-calling benchmark in a fault-injection layer and inject timeouts, missing tools, schema changes, or silently corrupted results partway through multi-step tasks. Across 1,920 trials with six models from three families, agents flag a problem in 91.3% of trials when the tool returns an explicit error, but in only 58.8% when it returns a plausible wrong value, compared with a 26.8% false-alarm rate when nothing was wrong. Reasoning variants notice faults less often and change plans more often than their instruct siblings, with no difference in recovery. Because two fault-free runs of the same task end in the same state only 63.3% of the time, only a missing tool clearly lowers recovery beyond run-to-run variation. A prompt telling agents to check each result did not improve detection, which suggests agents react to the error channel rather than to what a tool actually returns.

arXiv:2610.10062 PDF

HGP:An on-device personalized agent memory via hybrid graph storage

Ran Zhou, Xueming Han, Jiaheng Liu, Yuyao Zhang, Fanyu Meng, Junlan Feng et al. LLM-based agents struggle with personalized tasks because long-term user traces are heterogeneous and implicitly constrained, and single-vector memories blur type distinctions and relational structure, which is especially limiting on-device. HGP is a hybrid graph memory framework that uses a lightweight self-enhancing classifier to route memories and stores episodic, semantic, and procedural memories as graphs. It also extracts working memory as a state trajectory that captures the current state and implicit constraints. The classifier reduces large-model calls enough to allow on-device deployment, and on PAL-Set solution selection HGP reaches an S-score of 35.58, nearly 7 points above the strongest baseline.

arXiv:2610.10071 PDF

SkillSandbox: Skill Verification via Dynamic Scenario Synthesis

Serin Kim, Kwangwook Seo, Dokyung Song, Jinyoung Yeo, Dongha Lee Self-evolving agents distill experience into reusable skills, but those skills can encode wrong procedures or knowledge that does not transfer, and existing tasks may never exercise a given skill. SkillSandbox checks each skill by synthesizing a new task and environment that is relevant to the skill but differs from its source. A Proposer decides which conditions to keep and which details to vary, a Builder constructs an executable scenario, and a Verifier compares runs with and without the skill on executability, utility, and efficiency to decide whether the skill enters the library. Across ALFWorld and WebShop with three models, SkillSandbox consistently gives the strongest downstream performance along with better execution efficiency.

arXiv:2610.10088 PDF

ExperienceIndex: Artifact-Grounded Memory

Peter Baile Chen, Geoffrey X. Yu, Xinming Liu, Samuel Madden, Dan Roth, Jacob Andreas et al. Agents that answer many questions over a shared corpus, such as court cases or scientific literature, lack memory for the artifact-specific knowledge that humans build up through repeated use, and existing agent memory focuses on user preferences, facts, or abstract reasoning patterns. ExperienceIndex is a lightweight middleware layer that stores two kinds of experience from prior reasoning traces: summaries of each artifact's contribution to past tasks, and relationships between pairs of artifacts. At query time it retrieves these experiences to steer the agent toward the full set of relevant artifacts. Across corpora and agent search frameworks it raises answer quality by up to 11.0 points and cuts online dollar cost by up to 50.5%, and experience transfers from text-to-SQL to factoid QA tasks and from stronger to weaker models.

arXiv:2610.10091 PDF

RewardWeaver: Long-Horizon Interactive Learning for Language Agents via Self-Evolving Reward Adaptation

Hengbo Xiao, Boyao Zhang, Purui Liu, Yuxuan Zheng, Haoran Yin, Haibo Liu et al. Reinforcement learning with verifiable rewards (RLVR) struggles with long multi-turn interactions, where feedback arrives only at the end and is hard to attribute to earlier steps. RewardWeaver lets language agents adapt their process rewards as training goes on. After each stage it traces low-scoring trajectories back to capability bottlenecks, then selects rubric-based rewards that target those bottlenecks for the next stage, adding new rubrics in a controlled way only when recurring failures fall outside the existing set. It reports new state-of-the-art results on SOTOPIA, Amazon-HistoryPrice, and a new Sales Benchmark, and ablations confirm that dynamic reward allocation and fixed rubric meanings both matter.

arXiv:2610.10120 PDF

Know the Shape, Find the Fault: Topology-Conditioned Diagnosis of Multi-Agent LLM Failures

Xinwen Liu, Zhuocheng Pan, Isabella Zhu, Jawei Zhang, Xudong Liu, Tianyu Wo cross-listed When multi-agent LLM systems fail, similar-looking execution traces can come from very different coordination problems. MAScope diagnoses these failures using communication topology, meaning the structure of how agents pass information to each other. A Trace Structural Extractor recovers the interaction graph from the messages in a trace, and a Topology-Conditioned Judge classifies the failure using that topology plus prior failure statistics. Topology and failure type are strongly associated (χ² = 409.9). On 851 MAST traces, topology context doubles gpt-mini's Macro-F1 from 0.173 to 0.350, close to a trace-only gpt-5.4 baseline at about 6% of its cost over 1,000 traces.

arXiv:2610.10126 PDF

UniSkill: Learning Actor-Aligned Skill Proposals for an Evolving Policy

Yifei Lu, Cheng Liu, Dianzhi Yu, Hui Xiang, Ji Zhang, Yuanchu Xiao et al. LLM agents can improve by building up a bank of reusable skills, but rewarding a proposed skill by how well it is reused later mixes up the skill's benefit with the actor's own improvement. UniSkill uses one shared policy both to act and to propose skillbank edits (add, update, or no edit). It scores each proposal by how much swapping in the proposed skill widens the actor's log-likelihood gap between earlier successful and failed trajectories, so no new rollouts are needed. A regularizer keeps edit operations from being suppressed too early, preserving exploration. It reaches 98.4% success on ALFWorld and 84.7% on WebShop with stable joint training, and stays effective with a smaller backbone.

arXiv:2610.10164 PDF

Beyond Outcome Rewards: Constructing and Assigning Retrieval Credit for Search Agents

Wenyu Huang, Xinyu Hou, Pavlos Vougiouklis, Ruofei Lai, Jeff Z. Pan Search agents built on LLMs are often trained with reinforcement learning with verifiable rewards (RLVR), but rewards based only on the final answer make it hard to credit individual retrieval steps. The authors systematically compare reward-shaping and credit-assignment strategies that provide learning signals from intermediate retrieval steps, then build a training framework that combines these signals with outcome rewards. Under matched training conditions it improves aggregate performance across multiple multi-hop QA benchmarks. The analysis shows that both the choice of intermediate signal and where its credit is assigned shape training behaviour.

arXiv:2610.10179 PDF

Agentic AI-Assisted Modeling for Production Scheduling: Assessment in Constraint Programming

\'Angel S\'anchez-Fern\'andez, Javier Pernas-\'Alvarez, Diego Crespo-Pereira The study tests whether general-purpose large language models (LLMs), run as agents with no task-specific training, can turn natural-language production scheduling problems into working constraint programming models. Single-agent and multi-agent setups are connected to a Model Context Protocol (MCP) server that retrieves solver documentation to cut down on hallucinations. Both are compared with a direct LLM call on six industry-oriented flow-shop, job-shop, flexible job-shop and warehouse scheduling problems, using three LLMs. Writing the formulation turns out to be mostly within reach, and implementation is the main barrier: the multi-agent workflow raises the share of scripts that run correctly as generated from 14.8% to 59.3% (80.6% on the four simpler problems), while tightly coupled intralogistics models remain unsolved.

arXiv:2610.10184 PDF

Why Software Engineering Is Indispensable in the Age of Coding Agents

Alfonso Fuggetta cross-listed This position paper argues that capable AI coding agents make software engineering (SE) as a discipline, and software engineers as professionals, more essential rather than obsolete. It identifies three structural properties of large language models (probabilistic generation, agnosticism and semantic statelessness) that leave a gap no amount of training can eliminate. Filling that gap, it argues, takes methodological knowledge, domain knowledge, design choices and process choices, all written down as persistent artifacts. The engineer's role in this picture is methodologist, mediator and custodian of those artifacts.

arXiv:2610.10226 PDF

TestGRAD: Evolving Test Suites via Failure Pattern Momentum for SWE-Agent Ensemble

Pengfei He, Jiayuan Zhou, Shaowei Wang, Ruiqi Pan cross-listed Ensembles of software-engineering agents produce several candidate patches for an issue, and the hard part is choosing the right one by generating and running tests. TestGRAD treats this as an optimization over the test suite, loosely modeled on gradient descent with momentum. A differential loss rewards tests that separate candidate patches by behavior. Updates can read, create, update, or delete tests rather than only adding or removing them. A Failure Pattern Momentum component mines recurring failure sequences from memory so the optimizer stops repeating directions that don't help. On SWE-bench Verified, a 4-agent ensemble reaches 84.2% Pass@1, versus 80.6% for the strongest baseline, while compressing the failure-history context by more than 100×.

arXiv:2610.10242 PDF

OOM-RL II: Reality Is an Oracle, Not a Debugger Provenance-Constrained Diagnosis in Continually Evolving Agent-Engineered Systems

Kun Liu, Liqun Chen Real-world outcomes can confirm that something happened without revealing which version of a constantly changing system caused it. The authors study this in a quantitative trading system built by AI agents under human direction. Over one year the account beat a broad market index, but its annual alpha was not statistically distinguishable from zero, and incomplete records linking recommendations to deployed code meant no single frozen version or cause could be identified. From this case they define an "outcome–diagnosis gap" separating outcome evidence, identity of the evaluated system, and causal explanation. They propose an evidence hierarchy and a forward-looking protocol for binding production runs to specific system versions, which has not yet been validated.

arXiv:2610.10256 PDF

QuSema: Detecting Silent Bugs in Quantum Libraries via Quantum-knowledge-enhanced Agents

Yujin Song, Kaining Zhang, Qixin Zhang, Shuai Wang, Pingchuan Ma, Yuxuan Du cross-listed Quantum software libraries can contain silent bugs that return wrong results without crashing, which slip past testing methods that only catch failures or disagreements between implementations. QuSema is an autonomous testing agent that uses quantum-semantics constraints and library documentation as a source-level oracle. It repeatedly reads API docs and source code, reasons about what quantum operations should do, and confirms suspected deviations by writing executable tests. On 20 historical silent bugs in Qiskit and PennyLane, it relocates more bugs on average than Claude Code and Codex. It also found 40 previously unknown bugs confirmed by developers, 30 of them silent.

arXiv:2610.10258 PDF

Stale, Misattributed, or Late: Where Personal Memory Fails Before Generation

Haonan Deng, Park Sinchaisri Memory for personal language agents is usually judged by whether the final answer is correct, which hides errors that happen before generation: stale values, facts about the wrong person, or missing facts at serving time. Using a Personal Fact Memory (PFM) reference layer, the authors find that staleness is mostly a memory-construction problem. Without update resolution, 70.3% of prompts expose a superseded value, while serving only the active value of each correctly keyed slot eliminates the stale exposures they observed. The harder problem is assigning revisions to the right slot: LLM key assigners achieve higher key recall than a rule-based extractor yet lower clean-retrieval rates, and merge recall on LongMemEval never exceeds 0.062. Participant-aware BM25 performs as well as the reference ranker, and prompt prefill, not retrieval, dominates latency.

arXiv:2610.10265 PDF

RSIGym: A Flexible Environment for Recursive Self-Improvement

Fanqing Meng, Lingxiao Du, Haocheng Lu, Qiguang Chen, Ziqi Zhao, Zijian Wu et al. RSIGym is an agent-native research environment for recursive self-improvement, the setting in which an agent's accepted changes carry forward into later improvement cycles. Training, inference, rollouts, evaluation, and sandboxed execution are exposed as reusable services under shared budget and permission controls, so agents can improve data, training settings, and execution harnesses separately or jointly. The authors define RSI-Index as the mean fraction of the remaining performance gap closed across five benchmarks, and among six frontier models Opus 5 scores highest at 0.4809 on a $500 budget per benchmark run, raising SWE-bench Verified from 17.67% to 50.33% and AIME from 31.67% to 97.78%. The codebase, results, and recorded agent trajectories are open-sourced.

arXiv:2610.10310 PDF

LLM-Assisted Generation of Transparent, Open-Source Multiphysics Models of Electrochemical Devices

Sebastian Castro, Maya F. Schuchert, Spencer A. McCluskey, Eric W. Lees, Justin C. Bui cross-listed Building multiphysics models of electrochemical devices usually requires substantial numerical expertise or proprietary software such as COMSOL. The authors write a machine-readable, human-specified modeling harness that lays out the governing equations, parameters, numerical methods, build stages, and checkpoints, and a frontier large language model agent uses it to construct complete models in open-source Julia for electrochemical CO2 reduction to CO in a porous gas diffusion electrode. Independently built models, including fully autonomous ones, match a COMSOL reference to within 0.7% of the peak CO partial current density and agree with one another to within 0.04%. Deliberately planted errors show that explicit specifications are needed for reproducibility and reveal where the agent succeeds and fails at debugging model physics.

arXiv:2610.10320 PDF

Learning to Act with Task Progress: Distilling Small Agents from Compact Teacher Supervision

Wenxi Gan Task-Progress Distillation (TPD) trains small agents offline from large-model demonstrations by pairing each demonstrated action with a short label for the current task stage, instead of keeping the teacher's full reasoning. The student scores admissible stage and action pairs, and a deterministic harness executes the chosen action. On ALFWorld, a 1.7B-parameter student trained on 404 demonstrations reaches 72.4% success on unseen tasks with either TPD or action-only supervision, compared with 48.3% for a student trained on reasoning. Stage labels help most with fewer demonstrations, raising success from 48.0% to 67.7% at 200, and the advantage disappears by 808 demonstrations.

arXiv:2610.10332 PDF

MIRA: A Musical Intent Refinement Agent for Aligning Text-to-Music Generation with User Intent

Zekai Liu, Zhilin Wang, Xuzheng He, Yu Cheng, Yang Yang cross-listed A single text-audio relevance score cannot tell whether a generated piece of music actually meets a user's specific requirements, such as instrumentation, structure, or mood progression. The authors turn each request into a rubric of independently verifiable items covering both explicit requirements and implied intent, and release MuRA-Bench, a benchmark of real platform requests curated by music experts. They also propose MIRA (Musical Intent Refinement Agent), a test-time agent that treats the music generator as a black box: it rewrites the prompt, generates audio, checks the result against the rubric, and uses that feedback to guide a tree search within a fixed budget. With MIRA, an open-source generator reaches intent alignment comparable to commercial systems such as Suno and Mureka.

arXiv:2610.10355 PDF

Kernel Autoresearch for Open-Ended Model Discovery

Richard Cornelius Suwandi, Feng Yin, Kevin Murphy Automated kernel design must choose between fixed grammars, which guarantee valid kernels but limit what can be expressed, and unrestricted programs, which are flexible but often invalid. In stress tests, 22–58% of LLM-generated kernels that pass numerical checks fail at other scales or dimensions. Kernaut (Kernel Autoresearch) has coding agents write kernels as programs, uses construction contracts to guarantee that every accepted kernel is valid, and keeps diverse high performers in a quality-diversity archive. The discovered kernels generalize to unseen tasks: one beats a meta-learned deep kernel on held-out black-box optimization problems, and others outperform tuned baselines on unseen enzyme-kinetic mechanisms.

arXiv:2610.10394 PDF

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

Zhiqin Yang, Chenxin Li, Xiaomeng Hu, Yibin Liu, Weidong Huang, Jiankai Sun et al. cross-listed General-purpose agents now write code and use tools well in digital settings, and RobotWorld asks how much of that ability carries over to controlling robots. The simulation benchmark has 84 tasks covering manipulation, mobile manipulation, locomotion, driving and aerial control, each with an interaction budget and an executable success check. Agents can build sophisticated perception and control pipelines (segmentation, camera calibration, spatial estimation), but these skills often fail to combine into task success: agents lose track of object state, keep repeating actions that don't work, recover too late, or think an unfinished task is done. Model strengths differ: Astra does better on spatial and constrained-contact goals, while Opus 5.5 does better on balance and timed-interaction goals.

arXiv:2610.10409 PDF

CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution

Jixuan Chen, Jiaxin Zhang, Qinyuan Ye, Yada Pruksachatkun, Haoxiang Zhang, Jingming Zhuo et al. How well a terminal agent performs depends on both the model weights and the harness that formats prompts, binds tools and handles errors, yet co-evolution methods often reuse trajectories from every harness they tried as one undifferentiated training pool. CoTrace is a harness-aware data recipe within an alternating co-evolution framework: repeated execution failures drive changes to the harness, and the model trains only on verified rollouts from the harness actually adopted (for supervised fine-tuning) or on fresh online interactions (for reinforcement learning). It raises Qwen3.5-9B from 78 to 88 solved tasks under fine-tuning, and to 90 with online reinforcement learning on the Tmax split, using a small harness-matched corpus that costs far less compute than larger pooled ones. On Terminal-Bench 2.1 and SWE-bench Lite, transfer depends heavily on the training and evaluation harnesses matching.

arXiv:2610.10426 PDF

RunningTab: Direct Workspace Interaction with Environment-Side Tabs

Jinheon Baek, Soyeong Jeong, Yumin Choi, Dongsu Han, Sung Ju Hwang LLM agents can produce deliverables by searching and reading workspace files directly from a terminal, but nothing tracks what the task requires, which files were read, and which were listed but never opened, so details drop out of the context window and the final report can miss them. RunningTab gives the environment a per-task tab: the agent records requirements, while the environment logs every file read as an excerpt with its source and every listed-but-unopened file as a candidate. The agent sees each requirement next to matching excerpts and unopened candidates, resolves it or sets it aside with a reason, and a finish check returns any open requirements before the agent can finish. Across three benchmarks and three LLMs, it consistently beats plain direct workspace interaction and baselines that keep the record inside the model's own context.

arXiv:2610.10444 PDF

A Society of Researchers: Designing Institutions for Populations of Autonomous Research Agents

Ali Asaria, Deep Gandhi, Tony Salomone cross-listed Research-agent deployments are growing to populations of thousands of agents that share one pool of compute. The authors argue that such a population will become organized whether or not anyone designs it to, so its institutions should be designed explicitly. They propose a 'society of researchers' built on six principles. Principal-investigator agents compete for compute through requests for proposals, independent review, and grants, while a human 'mayor' allocates resources but assigns no tasks. In a running society of ten thousand agents asked to improve language-model pretraining, one lab reported matching quality with about 30% less compute, though the other labs that tested the claim do not yet agree on it.

arXiv:2610.10468 PDF

Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models

Tan Yu, Alexander Bukharin, Khushi Bhardwaj, Jennifer Williams, Zirui Liu, Jonathan Lingjie Li et al. Agentic post-training is expensive, so it helps to know in advance which base checkpoint is worth it. Standard end-to-end pass@K works poorly here because many base models cannot reliably produce well-formed tool calls. The authors replay successful post-trained agent trajectories, rerunning the tests after each code change, to find the 'decisive step' whose cumulative patch first makes the tests pass. At that step they build three screens that never require the base model to drive the harness from scratch: Decisive-Action BPB (bits per byte), a Patch MCQ multiple-choice test, and prefix-conditioned pass@K. Across ten public pairs of base and post-trained models, all three screens rank the models in close agreement with post-trained SWE-bench Verified pass@1.

arXiv:2610.10478 PDF

RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing

Yilun Hao, Krishna Sayana, Isabella Ye, James S Ren, Sukhdeep Sodhi, Craig Boutilier et al. In many tasks the needed evidence is not stated in any single source but must be derived by filtering, aggregating, or computing across many items, which similarity-based retrieval-augmented generation (RAG) handles poorly. RECAST (Routing Evidence through Computation, Access, and Synthesized Tools) treats building that evidence as a sequential decision process. A lightweight RouterLM picks retrieval and computation operations, a frozen CompilerLM translates custom operations into executable code, and a frozen AnswerLM produces the final answer. The router is trained with supervised fine-tuning (SFT) followed by group relative policy optimization (GRPO). Across six benchmark families, RECAST reaches a 75.6% mean success rate, 15.9% above the strongest large-model baseline, and a trained Qwen3.5-9B router beats a training-free Gemini 3.5 Flash router by 5.0%.

arXiv:2610.10507 PDF

SciExam for ENSO: Can AI Agents Build Climate Models?

Yinling Zhang, Langchen Liu, Dongbin Xiu, Xueyan Zou, Xu Kuang, Mengdi Wang et al. Research done by AI agents is usually graded against a known answer, a rubric, or an LLM reviewer, none of which can tell whether a new scientific model is valid. SciExam for ENSO asks agents to build low-order stochastic models of the El Niño-Southern Oscillation (ENSO) from real observations within six hours, using only diagnostics they write and then freeze. Hidden graders test whether each model reproduces ENSO statistics, recovers unobserved variables, and forecasts held-out years, and score a published model the same way. Six of twelve agent systems beat the published model, and the simplified forms of the strongest models each line up with one of two competing explanations of ENSO's warm-cold asymmetry, a debate the task never mentions. Controlled runs suggest the top system's scores do not come from recalling the historical observational record.

arXiv:2610.10513 PDF

1 more specialized paper

Large Language Models 76

Tokka-Bench: Evaluating Tokenizers Across 100 Natural and 20 Programming Languages

Ben Gubler Tokenizer quality varies widely across languages, yet there has been no standardized, multi-metric way to compare tokenizers broadly. Tokka-Bench is an open-source framework that scores tokenizers on five metrics, including bytes per token, subword fertility, and word-split rate, across 100 natural languages covering more than 30 scripts and 20 programming languages, with segmentation adapted to each writing system. Comparing seven BPE tokenizers, among them those of GPT-4, gpt-oss, Llama 3.1, Gemma 3, Qwen3, and Kimi K2, the authors find that how a vocabulary is allocated across languages matters more than its raw size. They also find that efficiency on programming languages has converged among recent tokenizers even though their natural-language profiles still differ.

arXiv:2610.08794 PDF

KVFetch: Temporal Prefetching for the Missing Half of KV Cache Compression

Linfeng Dong KV cache compression for long-context LLM inference decides which tokens to keep or recall based on their relevance to the current query, but the authors argue this ignores a second access pattern: reading the context sequentially by position. When a model must copy identifiers, field values, or code verbatim, eviction keeps the start of the sequence but drops the rest, and copying breaks partway through. The authors call this failure "sequential forgetting" and identify it as the main remaining source of quality loss under compression. KVFetch is a training-free, drop-in addition to score-based compressors that moves evicted tokens to a quantized cold tier, detects active copying by tracking a monotonically advancing read pointer, and prefetches the following tokens into fixed-size hot slots without increasing attention cost. On RULER-16K at equal memory budget, verbatim copying recovers from 0.8 to 78.4 and the 13-task average rises by 8.4, while on LongBench, which has no sequential-access tasks, it stays inactive and adds no cost.

arXiv:2610.08811 PDF

CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models

Yue Wu, Qinghe Zhang, Yu Zhang, Jian Huang Masked diffusion language models (MDLMs) usually commit tokens irreversibly, so a token chosen early with little context stays fixed even after later context no longer supports it. The authors call the resulting drop in the model's confidence in a committed token "confidence drift." CoDR (Confidence Drift Remasking) is a training-free, sampler-agnostic refinement pass that estimates drift for every committed token in k forward passes and then remasks and regenerates only the tokens the model no longer supports. Across two backbones, four reasoning and coding tasks, and three base samplers, CoDR improves average accuracy in every model-sampler configuration at modest overhead, and it uses far fewer forward passes than earlier remasking methods.

arXiv:2610.08833 PDF

LRCC: Generalizing Low-Rank Compression with Conditional Computation

Thomas Vaitses Fontanari, Maximo Eduardo Rulli, Federico Alvetreti, Donatella Genovese, Simone Scardapane Low-rank compression of language models usually gives every token the same fixed rank and therefore the same amount of compute. Low-Rank Conditional Computation (LRCC) adds one lightweight router per Transformer block that picks, token by token, among a few nested low-rank paths, and only the routers are trained while the low-rank factors stay frozen. At the same average active-parameter budget, it beats static low-rank compression on Llama and Qwen models, including a 7.6-point gain in average downstream accuracy on Llama-2-7B. At matched batch-size-1 decoding latency and without specialized kernels, it improves on Llama-3.2-1B and stays competitive on Llama-2-7B.

arXiv:2610.08858 PDF

The Best Optimizer Depends on Batch Size

Xingyu Dang, Kaiyue Wen, Sadhika Malladi New adaptive optimizers are usually benchmarked at a single batch size, on the assumption that hyperparameter scaling rules keep their ranking stable as batch size changes. The authors test this for language model pretraining. They find that no principled scaling rule for Muon works consistently across training settings, and that the best optimizer changes with batch size even after extensive hyperparameter tuning.

arXiv:2610.08975 PDF

Enabling Dynamic Computation in Looped LMs

Aayush Mishra, Arnau Padr\'es Masdemont, Victor Conchello Vendrell, Jordi Ros Giralt, Arash Behboodi, Fabio Valerio Massoli Looped language models reuse the same layers for repeated iterations and in principle can spend less compute on easy tokens, but the open Ouro models do not deliver this in practice. Each loop depth needs its own KV cache, so every iteration must still run, and the early-exit prior pushes all tokens toward the same depth. The authors propose a training-free "best-available" KV caching strategy that gives up to 30% savings in FLOPs and KV memory while keeping full-depth performance. Training with awareness of this caching strategy improves results further, and a small fix to the early-exit objective lets tokens exit at different depths according to difficulty, as validated on Ouro and on smaller looped models trained from scratch.

arXiv:2610.09013 PDF

SPIN: Shadow Predictive Indexer for Sparse Attention

Yao Fu, Cyrus Chang, Ritchie Zhao, Bryce Long, Yueying Li, Mahdi Kamani et al. Indexer-based sparse attention sends only a few important tokens to core attention, but the indexer itself still scores the whole KV cache at every decoding step, which becomes a bottleneck at long context lengths. SPIN (Shadow Predictive Indexer) uses lightweight, history-based prediction to pick important KV blocks without rescoring the full cache each step, and it is designed to work with block-structured caches and speculative decoding. On long-context and agentic benchmarks, SPIN reaches 30–40% sparsity while preserving task quality. In end-to-end vLLM serving it improves output throughput by up to 14.9% and cuts median inter-token latency by up to 13.2%.

arXiv:2610.09025 PDF

A Self-Pruning Transformer: Extreme KV-Cache Compression with Universal Attention

Davis Wertheimer, Haochen Shen, Ahan Gupta, Derrick Liu, Yu Chin Fabian Lim, Mudhakar Srivatsa et al. Large key-value (KV) caches make modern LLMs expensive to deploy. Earlier decay-based replacements for RoPE positional embeddings, which allow tokens to be pruned from the cache, tend to collapse into sliding-window eviction. Universal Attention combines several decay mechanisms into one end-to-end trainable framework that keeps RoPE and Softmax attention, and its composite decay serves as an adaptive criterion for dropping the tokens that contribute least to attention. It reports state-of-the-art 10× KV-cache compression while improving downstream performance over both baselines and unpruned models, and 25× compression at a context length of 16k.

arXiv:2610.09051 PDF

Learning Cross-Model Activation Alignments with Explicit Many-to-Many Layer Maps

Alina Sudakov, Guy Bar-Shalom, Fabrizio Frasca, Haggai Maron MATCHA learns a mapping from one LLM's layerwise activations to another's. The mapping is split into an explicit target-by-source layer map, which can be inspected directly, and a feature map between hidden spaces that is shared across layers. Unlike prior work that fixes layer correspondence by relative depth, both parts are learned jointly from prompts. Across 42 pairs drawn from seven models in three families, it reconstructs target activations more faithfully than previous approaches, and the learned correspondences turn out to be consistently many-to-many, with each target layer drawing on a band of source layers. The alignments also let steering vectors and probes built for one model transfer to another.

arXiv:2610.09058 PDF

Multi-Label Topic Assignment via LLM Distillation: A Comparative Analysis of Generative vs. Discriminative Student Models

Sourabh Kasliwal, Shubhranshu Singh This industry study asks which kind of small student model works best when an LLM labels data for multi-label topic assignment on e-commerce reviews and buyer-seller conversations. It compares causal generative small language models (SLMs) at 1B, 4B, and 8B parameters with discriminative encoders DeBERTa-V3 and ModernBERT. Discriminative models do better on structured product reviews, but even a 1B generative model beats them on multi-turn conversational data. The generative models also hold up as the label set grows to 112 topics, while discriminative baselines lose 35% Macro-F1. The models were deployed in production within strict latency limits.

arXiv:2610.09063 PDF

U-Space: Uncovering When and Why Uncertainty Arises in Language Models

Tobias Braun, Nils Loose, Alexander Herzog, Virginia Ceccatelli, Marcus Rohrbach, Thomas Eisenbarth et al. Many uncertainty-estimation methods for LLMs need repeated generations or separately trained components, give only a single score, and may simply be tracking output length. U-Space instead uses interpretability: it takes the unembedding directions of semantic anchor words for doubt and certainty, maps them back into the residual stream, and combines their contrasts into a low-dimensional orthogonal basis. The U-Lens projects each token's hidden state onto this basis, giving a token-level uncertainty map or an aggregate confidence score without correctness labels, repeated sampling, or training. On reasoning benchmarks, the score beats established baselines under both standard and length-controlled evaluation and transfers better than supervised estimators.

arXiv:2610.09087 PDF

Same Text, Different Prediction: Serving-Context Nondeterminism in Text Classifiers

Santhosh Kumar Kasa, Siva Rajesh Kasa, Sumit Negi Batch size, hardware and inference engine are known to change generated text even when the prompt and weights are fixed, but their effect on text classification had not been measured directly. The authors train 180 discriminative, pseudo-generative and fully generative classifiers and evaluate each under four categories of serving-context changes, holding the checkpoint and input fixed. Changing only the batch shape flips no labels in fp32, yet in bf16 it shifts up to 56.7 percentage points of predicted probability mass, with label flips concentrated near small margins. Fully generative classifiers flip more labels than discriminative ones, and the authors derive sufficient conditions for label stability along with a mitigation for each mechanism.

arXiv:2610.09111 PDF

Are Parameter-Efficient Fine-tuning Methods Really Different?

Yikuan Li, Pinyan Lu, Fanghui Liu The authors compare six parameter-efficient fine-tuning (PEFT) methods on language and diffusion models, looking at task performance, forgetting and changes to the geometry of the pretrained weights. The LoRA-family methods turn out to approximately preserve pretrained geometry, and restoring their slightly drifted singular-value spectra largely keeps task performance, which calls into question the need for explicit spectrum preservation as in OFT. Among the methods, LoRA most consistently limits forgetting, DoRA usually scores higher on tasks, and PiSSA often forgets more. Restoring the dominant spectral components reduces drift from the base model more than restoring intermediate or trailing ones.

arXiv:2610.09122 PDF

Breaking the Space Barrier and its Application to Language Model Inference

Arip Asadulaev The automata that force LLM output to follow a JSON schema or tool-call format have a rare property: from any state, each token leads along exactly one path. Building on this, the authors resolve an open complexity question by showing that reachability in such few-path graphs can be decided, while also verifying the few-path property, in O(log² n / log log n) space, below Savitch's classical O(log² n) bound. The same constructions yield an inference engine that writes forced text without running the model, computes token masks on the GPU without lookup tables, and decodes independent fields in parallel. On a 16 GB Apple M2 Pro running Qwen3.5-2B and 4B, compared with MLX plus llguidance, it finishes schema-constrained extraction 1.2 to 1.3x sooner with identical answers. It stores a grammar in 3 MB instead of up to 1.5 GB and runs sixteen tool-calling agents 2.5x faster.

arXiv:2610.09139 PDF

Noise Your Prompt: Noising Conditioning Tokens in Continuous Diffusion Language Models

Justin Jung Continuous diffusion language models normally keep the conditioning prompt tokens clean during training; this work adds noise to the prompt tokens as well, a one-line change to the training objective. On combinatorial reasoning tasks the change improves generalization, raising the solve rate on hard Sudoku from 3.73% to 24.65%, and increases solution diversity on 10x10 N-Queens from 50.60% to 73.79% coverage. It also improves Gigaword summarization when training data is modest, but the gains do not carry over to open-ended dialogue generation. By default it adds no inference cost, and it enables optional guided sampling in the style of classifier-free guidance.

arXiv:2610.09145 PDF

sk-bench: A Native-First Benchmark for Evaluating Large Language Models in Slovak

Marek \v{S}uppa, Ivan Vykopal, Andrej Ridzik, Kristi\'an Sopkovi\v{c}, Nat\'alia K\v{n}a\v{z}ekov\'a, Jaroslav Kop\v{c}an et al. Multilingual LLM benchmarks either leave out Slovak or cover it only through machine translation. sk-bench is a native-first benchmark of 30 datasets across ten skill categories, including an adapted IFEval-SK and native grammar and morphology resources, and it evaluates 55 models under one harness. The best open model trails proprietary APIs by 12.6 points. Continued Slovak pretraining of Qwen3-14B lowers its overall score by 13.9 points, and a small instruction set recovers three quarters of that loss. Test-time reasoning adds 8.5 to 12.5 points for models of 9B parameters and above. The authors distill four design lessons for other under-resourced languages, such as using native data where translation fails and planning instruction repair after language adaptation.

arXiv:2610.09152 PDF

LayerRoPE: Dynamic Depth-wise Magnitude & Angular Superposition

Shikhar Srivastava, Christopher Kanan Hidden-state norms in Transformers grow by orders of magnitude with depth, which is usually treated as a problem to suppress. Across 16 pretrained LLMs from 9 families, the authors find that this growth acts as an emergent depth-position encoding carried by the per-layer normalization weight, which grows in magnitude and rotates in direction with depth. LayerRoPE makes this explicit by replacing all per-layer normalization weights with one shared vector and depth-conditioned scalars, using fewer parameters and almost no extra FLOPs. It beats Pre-, Post-, and Peri-Norm baselines, reaching Pre-Norm's 1.3B-model loss with 3.4x less compute, keeps improving as depth scales to 512 layers, is 3 to 10 times less sensitive to learning rate, and also improves looped latent models and Vision Transformers.

arXiv:2610.09179 PDF

Q-PACE: Dynamic Precision Allocation for Quantization-Aware Training

Alexandra Volkova, Matin Ansaripour, Erik Schultheis, Christoph H. Lampert, Dan Alistarh Quantization-aware training (QAT) cuts the cost of deploying LLMs, but aggressive quantization hurts quality, and mixed-precision training must decide which layers keep higher precision. Q-PACE predicts the loss increase as quantization noise weighted by per-layer curvature coefficients, periodically re-estimates those coefficients with perturbations during training, and reassigns precision accordingly. In pretraining and supervised fine-tuning of LLMs up to 4B parameters, it consistently beats existing mixed-precision recipes and reaches comparable loss at substantially lower total memory budgets. Quantization sensitivity turns out to be highly predictable from depth and layer type and stable during training, so recalibration can be infrequent and cheap.

arXiv:2610.09183 PDF

Few Bits, One Law: Toward W2A4KV2

Kai Yi, Tarek Elgamal, Sruthikesh Surineni, Vignesh Vivekraja, Soumyadeep Ghosh, Steven Li Compressing weights, activations, and key-value (KV) caches to very low bit widths at the same time is hard, because each has a different distribution and their quantization errors interact. CanonQ is a quantization-aware training framework that first applies fixed rotations and energy normalization to map every tensor source into a shared canonical form. This lets frozen Gaussian-reference codebooks be reused across layers and models, and joint training then adapts the network to the combined errors. Under joint 2-bit weight, 4-bit activation, and 2-bit KV-cache compression (W2A4KV2) on LLaMA3 1B to 8B models, CanonQ-Omni reports up to 14.28x lower WikiText-2 perplexity and up to 57.9% higher mean zero-shot accuracy than prior baselines. The gains also hold for Qwen3-1.7B and for HumanEval and GSM8K on MobileLLM-Pro-1B.

arXiv:2610.09202 PDF

CurveTQ: Rotation-Free Trellis Quantization of LLM Weights via Curvature-Weighted Search

Guanhua Ding, Zi Wang, Ruichao Li, Jack Liu Leading 2-bit LLM weight quantizers such as QTIP and Proteus apply a random rotation to each weight matrix, which must be undone at every decoding step, and then encode it with a Euclidean trellis or lattice search. The authors show that these quantizers already compute per-coordinate curvature weights (the diagonal of the Hessian's LDL factorization) but never use them. Putting these weights into the Viterbi branch metric lets the search follow curvature within each coding block, which makes rotation largely unnecessary. The resulting rotation-free codec, CurveTQ, is 1–3 points higher in mean downstream accuracy than QTIP and Proteus at 2 bits on three 4–8B Instruct models. It also leads on a 35B mixture-of-experts model and has the fastest decoder of the three at every tested batch size and bit width.

arXiv:2610.09212 PDF

Quantize by Drift: Label-Free Mixed-Precision Post-Training Quantization for Text Embedders

Hyojung Han, Jongmin Kim, Seung-Hun Jeon cross-listed Mixed-precision quantization of text embedding models needs a per-module sensitivity signal, but measuring retrieval loss requires relevance labels that most deployments lack. The authors propose a label-free alternative: quantize one module, re-encode the corpus, and measure how far the output embeddings move from their full-precision positions. Across five development embedders, this drift ranks mixed-precision plans against held-out retrieval quality with a macro Spearman correlation of 0.911, and it transfers across calibration corpora and domains. In a pre-registered test on three held-out embedders, drift beat the prior LieQ criterion, but it still scored about 1 point below same-budget uniform precision. The authors conclude that drift is a robust coarse signal that avoids catastrophic failures, not an optimal allocation objective.

arXiv:2610.09227 PDF

Conditional Accuracy Profiles: Diagnosing LLM Judges across Deployment Conditions

Wenqi Li, Bin Liu, Mindi Ruan, Chuanbo Hu, Minglei Yin, Xin Li LLM judges are usually summarized by a single accuracy number, which hides the conditions under which they fail. Conditional Accuracy Profiling (CAP) breaks pairwise judge accuracy into eight conditions grouped under content sensitivity, robustness, and rationale quality, and it can be applied to any benchmark. It was tested on seven judges across six benchmarks, including a new controlled testbed called judgerEva-Standard. The profiles show that sensitivity to omitted qualifications is not predicted by overall accuracy: the two judges most sensitive to omissions rank in the bottom three overall. Position robustness has the most stable rankings across benchmarks but also degrades the most under JudgeBench-Pro adversarial stress.

arXiv:2610.09229 PDF

Denoising Blocks, Not Tokens: Efficient Compressed Continuous Diffusion with Branching Token Realization

Xinsong Feng, Peng Du, Zhizhuo Yang, Daniel M. Bikel, Jiayun Wang, Haipeng Chen Diffusion language models refine text in parallel but usually keep one state per token, so each denoising step still processes the full output length. Branching Latent Diffusion (BLD) compresses a 1024-token sequence into 64 continuous block latents and decodes each one with a small autoregressive branch, running all branches in parallel; the latents are generated in groups, each conditioned on earlier ones. Compared with the similarly sized ELF-L baseline, BLD uses more than 80x fewer generation FLOPs and delivers over 6x higher throughput, and it beats an autoregressive baseline by over 6x in throughput and over 4x in latency. Local fluency and diversity stay competitive, though long-range coherence remains a weakness.

arXiv:2610.09311 PDF

Shared Low-rank Basis Factorization for Data-free Mixture-of-Experts Compression

Tianxiao Cao, Jiahe Shao, Yuning Qiu, Kyohei Atarashi, Hisashi Kashima, Qibin Zhao Mixture-of-Experts (MoE) language models keep compute low through sparse routing, but their large parameter counts make them expensive to store and serve. The authors derive error bounds showing that expert pruning and expert merging incur non-vanishing structural errors tied to routing and expert heterogeneity, while weight reconstruction avoids these costs. Building on this analysis, they propose Shared Low-rank Basis Factorization (SLBF), a data-free reconstruction method in which experts share rank-k bases, plus a post-hoc gauge fixing step that removes redundant parameters. Across five MoE architectures from 16B to 122B parameters, SLBF consistently outperforms methods from all three compression families.

arXiv:2610.09342 PDF

OnlineQAT: On-Policy Distillation for Ultra-Low-Bit Large Language Models

Wenjun Wang, Heng Li, Yanggan Gu, Hongxia Yang Quantization-aware training (QAT) can recover accuracy for language models compressed below four bits, but recovery usually trains on fixed completions, while the deployed model conditions on its own generated prefixes. OnlineQAT first produces a usable low-bit model through block-wise QAT, then applies on-policy distillation (OPD), in which a frozen full-precision teacher provides a sampled reverse-KL signal on responses generated by the student. On Qwen3-1.7B it achieves the best average among the compared quantized methods, 57.28 at W3A16 and 32.52 at W2A16, which is 2.90 and 0.44 points above ReasoningQAT. The gains are largest at three bits.

arXiv:2610.09346 PDF

Expert Coupling in MoE Pretraining: Reducing All-to-All Overhead with Correlated Placement and Token Shuffling

Radha Gulhane, Quentin Anthony, Beren Millidge With expert parallelism, Mixture-of-Experts (MoE) training runs all-to-all communication at every layer, which takes 45-60% of step time on clusters of AMD Instinct MI300X GPUs. The authors observe that early in pretraining, routers already pick experts in correlated patterns: experts co-selected within a layer cluster together, and a token's experts at one layer predict its experts at the next. Correlated expert placement puts frequently co-selected experts on the same GPU, and token shuffling moves each token toward the GPU predicted to hold its next-layer experts. Neither method changes routing decisions or expert parameters. In Megatron-LM, the two methods together cut all-to-all time by 1.16-2.63x and end-to-end step time by up to 1.41x.

arXiv:2610.09372 PDF

ARCS: Towards Precise Text-to-SQL via Structured Disambiguation

Yihao Hu, Yanlin Feng, Naoki Otani, Nikita Bhutani Ambiguous user questions are a major source of silent errors in deployed text-to-SQL systems, and resolving them through free-form clarifying dialogue is inefficient and demanding for users. The authors propose structured disambiguation, which resolves ambiguity through explicit, constrained interactions, and introduce ARCS (Ambiguity Resolution Corpus for SQL), a benchmark of naturally occurring ambiguities over real-world databases annotated with every valid interpretation and its SQL query. Ambiguity still makes text-to-SQL hard: gpt-6-sol reaches only 51% end-to-end execution accuracy, and no open-source model exceeds 27%.

arXiv:2610.09396 PDF

Finding the Right Balance: Relevance and Diversity in LLM Retrieval

Guillaume Brouillette (Universit\'e du Qu\'ebec \`a Trois-Rivi\`eres, Trois-Rivi\`eres, Canada), Faustin Kagabo (Universit\'e du Qu\'ebec \`a Trois-Rivi\`eres, Trois-Rivi\`eres, Canada) et al. cross-listed Prior studies disagree on whether diversifying retrieved passages helps retrieval-augmented generation (RAG). Using controlled near-duplicate injection and realistic overlapping chunking, the authors show that diversification hurts relevance, evidence coverage and answer quality on clean candidate pools. It helps only on multi-evidence questions where redundancy makes nearest-neighbor retrieval return repeated passages. They propose a query-adaptive rule, computed from existing embeddings, that diversifies only when the effective number of distinct documents in the top-k falls below the query's evidence requirement. The rule captures most of the achievable gain and transfers across datasets and encoders. They also introduce RNG-Score, a geometric reranker that falls back to exact nearest-neighbor retrieval.

arXiv:2610.09412 PDF

Democratizing MoE inference on commodity GPUs with CoMoE

Ruwen Fan (Jimmy), Yuezhi Zu (Jimmy), Junru Li (Jimmy), Qingda Hu (Jimmy), Xinjun (Jimmy), Yang et al. cross-listed Serving Mixture-of-Experts (MoE) models with expert parallelism requires heavy inter-GPU communication, which normally calls for datacenter GPUs with fast peer-to-peer links such as NVLink. Consumer GPUs are much cheaper but connect only over slow PCIe buses, with all traffic passing through the host. CoMoE makes the host an active routing hub: shared tokens are multicast by writing them to host memory once, and outputs are combined token by token in host staging buffers instead of waiting on global synchronization. On RTX 5090 GPUs it raises inference throughput by up to 1.46x, approaching NVLink-equipped A800 performance at 23.4% of the hardware cost.

arXiv:2610.09424 PDF

Right Number, Wrong State? Measuring Cross-Jurisdiction Substitution in LLM Recall of State Policy

Jiayu Feng When an LLM gets a state-specific policy question wrong, it may be inventing a number or giving a real value that holds in another state. Using fixed question wording that varies only the jurisdiction (50 U.S. states plus the District of Columbia) across three Medicaid income-eligibility quantities, the authors find that Claude Sonnet 5.5 and GPT-5.6 Sol reproducibly give another state's current value on 10 and 25 of 153 items, respectively. However, counting any wrong answer that matches another state's value inflates substitution counts 3-5x, because many such answers are the asked state's own values under a different convention or from an earlier year. The authors conclude that claims about cross-jurisdiction errors require a complete same-state reference set.

arXiv:2610.09458 PDF

CHASE: Channel-Aligned Structure Exploitation for Geometry-Aware Model Engineering

Wei Wang, Wei Jiang, Ziran Liu Geometric and Spectral Alignment (GSA) describes trained networks through spectral concentration, alignment with physical channels, and changes in singular bases, and CHASE turns these structures into practical model-engineering methods. It brings together prior GSA-based methods for parameter-efficient fine-tuning, pruning compensation, and model merging (CORA, COEC, CORAM), and adds three new ones. CAGA builds shared key-value heads for converting multi-head attention to grouped-query attention, SAKV decides which adjacent layers can share a low-rank KV-cache representation, and CAPS performs grouped structured pruning. Experiments show geometric shared-head construction substantially improves MHA-to-GQA conversion, and SAKV and CAPS beat representative baselines for KV-cache compression and structured pruning.

arXiv:2610.09476 PDF

Goldsmith: Gold-Loss-Guided Definition Optimization with an Agentic Annotation Harness

Yihan Li, Hanyi Zhang, Xiaoxi Jiang, Man Guo Annotation projects often start before experts have a stable guideline or enough labels to train a model. Goldsmith is an agentic pipeline that treats the annotation definition as a trainable text object: candidate definitions are scored on a small expert gold set with an executable structured loss, and an LLM editor turns the highest-loss failures into textual-gradient revisions that are kept only if the loss decreases. Schema, formatting, retrieval, repair, judging, and human review stay in a separate harness. Under matched evaluation, Goldsmith outperforms direct rewriting, OPRO, APE, and PromptBreeder, and the learned definitions improve downstream annotation on span, relation, and event-argument tasks when combined with retrieval, routing, and human review.

arXiv:2610.09489 PDF

The Attribution Blind Spot: Layerwise Trajectory Diagnostics for Source Reliance in Retrieval-Augmented Language Models

Zhe Yu, Wenpeng Xing, Yunzhao Wei, Bo Yang, Chen Ye, Gaolei Li et al. A retrieval-augmented model can produce an answer that matches a retrieved document without actually relying on it, so the authors use controlled knowledge conflicts to make the model's choice of source observable. They measure paired hidden-state changes with Latent Trajectory Shift (LTS), a signed projection onto a training-fitted first principal component. The magnitude of the state change is often the better predictor of which source the model will choose, while the signed direction is the better controller: equal-norm interventions along it shift source preference while preserving other behavior, and the frozen direction transfers across datasets and aligned model pairs. The central finding is that representations that diagnose a model's choice need not be the ones that best control it.

arXiv:2610.09493 PDF

Reliability of LLM Judges for Evaluating Entity Alignment

Vaibhava Lakshmi Ravideshik, Mayank Kejriwal Entity alignment matches equivalent entities across knowledge graphs, and its evaluation needs expensive expert annotation, which raises the question of whether LLM judges can stand in. Across three frontier judges, three datasets, and four alignment systems, the authors identify anchor bias: when the system's predicted label is visible, judge discrimination collapses (ROC-AUC as low as 0.12). A label-free protocol recovers 0.93–1.00. Counterfactual label-flip tests and a blinded human evaluation confirm that the effect is causal, and stronger judges turn out to be more label-sensitive, not less. The authors also release a 15K-pair MeSH–SNOMED CT biomedical benchmark and an auditing framework.

arXiv:2610.09554 PDF

How Do LLMs Change Predictions Under Negation?

Jongwook Yoon, Jongwon Lim, Sungjib Lim, Woojin Cho, Yohan Jo Large language models (LLMs) often give the same answer when a question is negated, for example answering Madrid to What is not the capital of Spain?. On the authors' negation benchmark, recent open and closed models repeat the original answer in 37-71% of cases. Mechanistic analysis shows that specialized attention heads and MLP neurons handle negation by suppressing the original answer and promoting a favored alternative from the same category. Failures happen when that suppression is too weak or when a bias toward certain answers blocks the switch. A training objective that requires larger shifts in answer preference when the original prediction is more confident reduces these failures, with less loss of general capability than standard fine-tuning.

arXiv:2610.09571 PDF

Sequential Pretraining Favors Large Models

Mohnish Harwani, Yujia Zheng The authors ask whether large models gain capabilities partly because they are more robust to adverse training effects, rather than only because they learn better features. They define primacy bias, the degree to which exposure to early data distributions impairs later learning, and show that small models spend too much capacity on early distributions while sufficiently overparameterized models are robust to this. This matters for pretraining, where valuable later data such as code, math, and reasoning can be learned poorly by small models. A simple regularizer called Exposure Therapy (ET) improves performance on late distributions and overall capability in models up to the billion-parameter scale.

arXiv:2610.09611 PDF

How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression

Xijie Gong, Tingxu Han, Jiahao Zhang, Wei Song, Ziqi Ding, Hanqi Yan et al. Agentic prompts are long and heavily scaffolded, which makes it hard to isolate how a large language model (LLM) decides whether to call a tool or answer directly. The authors reduce such prompts to minimal contrastive pairs where only a request verb changes. Swapping an execution verb like "write" for an analysis verb like "discuss" reliably flips the decision. Using 500 paired prompts across Python, Java, and C++, they trace the decision to a single internal vector, μ_Δ, that is both causally necessary and sufficient, and it generalizes to native multi-turn τ²-Bench trajectories and to requests without the trigger verbs. Transcoder decomposition shows that the scaffold sets up a default toward calling a tool, analysis verbs suppress that default, and execution verbs mostly leave it alone. The same mechanism appears in seven models from the Qwen, Mistral, and Granite families.

arXiv:2610.09624 PDF

On-Policy Distillation Teaches New Skills but Not New Knowledge

Yixuan Tang, Yi Yang Using a controlled synthetic framework that separately varies the teacher's extra facts and its compositional reasoning skill, the authors test what on-policy distillation (OPD) actually transfers to a student language model. Across four models from three families, reverse-KL (Kullback–Leibler) OPD reliably transfers compositional skill to unseen reasoning structures but transfers minimal factual knowledge. Separating the pieces of the recipe shows that switching to forward KL restores factual transfer, while learning from the student's own rollouts is what improves multi-step reasoning. The same asymmetry appears on recent factual QA and competition math: OPD teaches models to compose knowledge they already have rather than adding new knowledge.

arXiv:2610.09639 PDF

When Rank Rises as LLMs Degrade

Zhaohui Geoffrey Wang Practitioners often monitor representation health during LLM post-training with spectral statistics such as RankMe, assuming that rank falls when representations degrade. In a controlled study of Qwen3-0.6B under four degradation modes, data duplication worsens held-out loss by 75% while increasing both original and centred RankMe and nearly doubling covariance effective rank. The result is spectral dispersion rather than collapse, so a one-sided monitor rates the worst checkpoint as the healthiest, whereas a learning-rate misconfiguration moves rank downward. A pre-registered two-sided multichannel monitor detects all three damage regimes, but it never warns earlier than held-out probe loss and raises false alarms when calibrated on too few seeds.

arXiv:2610.09647 PDF

A Tale of Two Error Categories: Exploring Concealed Trade-Offs in the Errors of Automated Judges in Evaluation of Uncertainty Quantifiers

Evgenia Ilia, Wilker Aziz Uncertainty quantifiers (UQs) for LLM-generated text are usually evaluated by how well confidence correlates with correctness, and correctness is increasingly judged by automated systems instead of humans. A meta-analysis of published work shows that automated judging is now the norm, and that it is rarely validated against human judgments. Experiments in question answering with 4 LLMs, 7 popular UQs, and both human and automated labels find that a judge's accuracy only coarsely predicts how much its errors distort UQ evaluation. Judging errors tend to misrepresent the most informative UQs most, a pattern the authors trace to how confidence correlates with different categories of judging error.

arXiv:2610.09693 PDF

Bridge Routing Heads: Where Multilingual Multi-hop Reasoning Lives in LLMs

Seunghan Kim, Minyeong Choe, Hyunil Kim, Haehyun Cho Multilingual LLMs can answer the same multi-hop question in many languages, and this work asks whether they use a shared internal circuit to do it. A three-stage pipeline identifies Bridge Routing Heads in Llama 3.1 70B and Qwen 2.5 72B. The language-specific head sets barely overlap across five languages (mean Jaccard similarity of 0.017 and 0.057), and ablating the general heads raises two-hop negative log-likelihood 39–89× more than ablating random heads. Amplifying these heads rescues up to 51.7% of cross-lingual reasoning failures without any training.

arXiv:2610.09733 PDF

Reproducible LLM Inference Benchmarking: A Sequential Isolation Protocol for Regression Testing

Arnold Olympio, Juan Manuel Servera Bondroit, Wael Abdelmalek, Guang Lu, Jo\~ao Carvalho cross-listed Repeated benchmarks of LLM inference can vary widely with execution and system state, which makes regression testing unreliable. The Sequential Isolation Methodology is a controlled protocol, tested on three open-source models served with vLLM 0.9.1 on an NVIDIA A100 80GB GPU across six context sizes, eight concurrency levels, and five repetitions per setting. It cuts the average coefficient of variation from 15.2% to 2.2%, and 78.5% of configurations stay below 3% variation in median time-to-first-token. The measurements also show a sharp latency jump between 200 and 500 concurrent users on this stack, and the authors provide a cost break-even model against API pricing along with infrastructure-as-code scripts for replication.

arXiv:2610.09778 PDF

Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression

Mingqing Yuan (Soochow University), Xiaobo Liang (Soochow University), Junwei Yang (University of Cambridge), Ziwei Chen (Chalmers University of Technology), Zeren Zhang (Peking University), Hejin Wang (Tsinghua University) et al. Generative reward models that write out their judgment across several criteria token by token are expensive to run. LatentGRM uses the structure of rubric-guided evaluations to chunk and compress the reasoning into short continuous latent trajectories, making pairwise preference judgments without generating any text, while a separate interpreter can reconstruct the evaluation text offline for inspection. With the same training data and backbones, it matches the aggregate preference accuracy of explicit supervised fine-tuning (SFT) judges at 4B and 8B scale. LatentGRM-8B compresses evaluation trajectories by 8.9 to 9.2 times and cuts total judge inference time by 6.1 to 7.0 times at vote@5, and rubric interventions show that criterion-specific information is carried through the latent sequence.

arXiv:2610.09788 PDF

Dual-QK: Sharp Queries and Flat Keys for Prunable 2-bit KV Caches

Sunjoo Whang, Jungjun Oh, Minsung Kim, Dongho Seo, Jisu Shin, Gregory Kielian et al. Rotation-based quantization of the key-value (KV) cache spreads out key outliers, but the matching rotation applied to queries also spreads out query energy, which makes it harder to prune query channels and so skip reading parts of the key cache. Dual-QK replaces the shared rotation with paired non-orthogonal transforms: partial whitening of keys balances their scales for INT2 quantization, while a query-aligned basis concentrates query energy so that channels can be pruned dynamically. It also protects channel 0 and uses a bucket-relative form of rotary position embeddings (RoPE) to keep accuracy at long context. At 40% query-channel sparsity it beats OSCAR on most tasks. At 128K context it gives 6.8x KV-cache compression, and an SGLang implementation reaches up to 3.75x the decoding throughput of unpruned BF16.

arXiv:2610.09827 PDF

A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks

Jonghyun Han, Younghoon Song, Jongyoul Park This study looks at where catastrophic forgetting happens when LLMs are further trained without access to the original data for replay. Systematically freezing parameters in models up to 1.4B shows that forgetting concentrates in the output embeddings of tokens that rarely appear in the new corpus. The mechanism is that absent tokens receive persistent one-sided softmax gradients, which Adam's second-moment normalization amplifies into full-sized updates. The proposed fix raises Adam's epsilon only for the output projection during training. Across eight settings from 160M to 12B parameters, this removes 39.4% to 67.9% of forgetting without hurting learning on the new data. It also combines well with replay, while editing the drifted rows after training recovers under 5% of the loss.

arXiv:2610.09835 PDF

NeuralZip: Reusable Setup for Fast Lossless Compression

Mart\'in Bravo, Samuel Horv\'ath, Gonzalo Navarro, Andr\'es Abeliuk Lossless compression of model weights repeats costly statistical analysis and Huffman code construction on every run. NeuralZip does this setup once: it groups chunks with similar exponent distributions, shares their Huffman codes, and packs recurring exponent tuples, and the result can be reused across encodings and even transferred from another compatible architecture. After setup, compression of floating-point checkpoints is 1.81-21.33x faster than baselines with bit-exact reconstruction, and the setup keeps working as weights change during training. On GPU it cuts active memory by up to 27.5% while reproducing the logits exactly.

arXiv:2610.09916 PDF

TR-PTQ: High-Accuracy Integer-Only Transformer Post Training Quantization via Taylor Region Reformulation

Eliyahu Levy, Adam Teman, Yoni Pugachov Post-training quantization (PTQ) of transformers is usually blamed on insufficient numerical precision in nonlinear layers, which pushes many methods to fall back to floating point. The authors instead trace the accuracy loss to specific structural sources: learned scale parameters in normalization layers and compounded approximation errors in GELU, while SoftMax turns out to be robust to aggressive quantization. Their method, TR-PTQ, builds shared Taylor Region exponential and logarithm primitives so that division and square roots run in the log domain with plain integer arithmetic, and adds a calibration-free, outlier-aware optimization of LayerNorm parameters. The result is a fully integer-only pipeline with less than 1.5% absolute accuracy loss across vision and language benchmarks and no floating-point units needed for nonlinearities.

arXiv:2610.09969 PDF

Cache the Encoder Within:Compact, Reusable Memory across LLM Queries

Hanzuo Liu, Chunyu Liu, Chaofan Lin, Alex Lamb, Mingyu Gao Repeated LLM queries over the same documents re-encode them each time, while caching full model states costs a lot of persistent storage. EncBank builds on CoMem's intermediate-state interface: it treats a pretrained LLM's lower layers as a reusable document encoder, stores their outputs compactly, and reads them with an adapted upper-layer reader. A self-distilled suffix adapter works across storage precisions without quantization-specific retraining. Across five benchmark suites on three Qwen backbones, 4-bit storage stays within one score point of native-precision EncBank and, in a fixed Qwen3-8B workload, needs 28.1% of the native-precision persistent GPU store. Separate native-precision runs show a 1.40x prefill speedup over replaying the same text at a 3.12-point RULER accuracy cost, and the authors note that end-to-end benefits depend on workload, preparation cost, and how often documents are reused.

arXiv:2610.10058 PDF

Activation-Aware Weight Tensorization: A Calibration-Time Preconditioner for Tensor-Network LLM Compression

Alessandro Beatini, Marco Maronese, Emanuele Rodol\`a Post-training tensor-network compression replaces Transformer linear layers with Tensor Train (TT) or Tree Tensor Network (TTN) operators, but standard decompositions minimize error in weight space rather than error on the layer's actual inputs. Activation-aware Weight Tensorization (AWT) is a training-free wrapper that scales each weight matrix by a diagonal factor derived from activations before running an unchanged TT/TTN solver, and it adds only an elementwise input rescaling at deployment. On Llama 3.1 8B, Ministral 8B, and Qwen2.5 7B at 2-6x compression, AWT closes 12-35% of the WikiText perplexity gap to the dense model when one operator is replaced, and 27-60% when several Llama operators are replaced, with gains carrying over to HellaSwag and ARC-Challenge. A dense full-covariance oracle wins on its own training objective, but diagonal AWT gives better held-out functional fidelity in 53 of 81 cases.

arXiv:2610.10085 PDF

I would rather quit NLP than read another paper like this: The rise of antithesis in NLP papers

Olga Zamaraeva, Adri\'an Gude, Roi Santos-R\'ios, Carlos G\'omez-Rodr\'iguez LLM-assisted scientific writing appears to fill papers with unnecessary antithesis, statements about what the work does not do, so the authors measure the construction "rather than" in 2019 ACL papers, 2026 ACL-style arXiv papers, and GPT-generated papers written from the same titles and abstracts. Its rate in 2026 is seven times the 2019 rate, and higher still in GPT-written papers. Blind annotators find almost no 2019 uses annoying but about one in ten 2026 uses, and the annoying uses portray the rejected alternative less favorably. Preference-data raters and open reward models favor the construction, and an instruction to be honest increases it, which leads the authors to conjecture that it is a side effect of post-training on pairwise preferences.

arXiv:2610.10092 PDF

Mechanics of Long-Context Hybrid Models Part 1.1: From Hybrid Attention to Hybrid Position

Xiaoran Liu, Ziwei He, Xipeng Qiu The authors study why hybrid large language models work. These models mix full attention with sliding-window attention (SWA) or gated linear attention (LA) variants such as GLA and GDN. They find a seesaw effect in context extension: LA hybrids benefit more from long-context continual pretraining, while SWA hybrids extrapolate better to longer inputs, and they trace this to the positional biases each attention type creates. SWA hybrids fall into short-context learning traps and need wider windows during long-context pretraining. For LA hybrids, the proposed Sliding-Window Linear Attention achieves 16× training-free length extrapolation with 100% accuracy on NIAH-SK1 at 64k context.

arXiv:2610.10114 PDF

YANchor-4B: Effective Long-Horizon Reasoning in O(N) Time with O(1) Memory

Huishan Ji, Hua Xu, Weiming Zhang, Qirui Ye Reasoning over long horizons is expensive with full attention, while recurrent models that compress history tend to lose precise details. YANchor-4B is a general-purpose recurrent model that saves key memories as 'anchors' it can retrieve later in a chain of reasoning, keeping generation linear in time and memory constant. It reaches 82.93% mean pass@1 on AIME 2024–2026 and 63.64% on HMMT, well ahead of other linear-time, constant-state models, including larger ones. On an H100 it also delivers several times the batched long-generation throughput of Transformer and hybrid baselines.

arXiv:2610.10118 PDF

Attention via Black-Box Vector Search

Stepan Zharkov, Krish Singal, Ashwin Padaki, Alexandr Andoni cross-listed Many sparse attention methods use maximum inner product search (MIPS) to retrieve the heaviest-weighted keys, which raises the question of how many keys must be retrieved to estimate attention accurately. Framing earlier approaches as priority sampling, the authors prove that a single MIPS index needs Θ(√n/ε) keys, while Θ(log n) indices reduce this to a near-optimal O(log n + 1/ε²). They also give algorithms that trade off the number of indices against the number of retrieved keys. Augmenting keys and queries bypasses these lower bounds, allowing a single index with O(1/ε²) retrieved keys, and in LLM inference the methods outperform top-k and sampling baselines and scale well to long contexts.

arXiv:2610.10135 PDF

Does Document Structure Help Dense Retrieval? A Placebo-Controlled Ablation of Four Mechanisms Across Two Corpora

Andrey Kuehlkamp, Priscila Correa Saboia Moreira, Samuel Rund cross-listed Retrieval-augmented generation pipelines often rely on document structure through structure-aligned chunking, LLM-generated chunk contexts, heading-path metadata, or two-stage hierarchical retrieval. Past studies of these techniques never controlled for the fact that any prepended text shifts a chunk's embedding. This pre-registered ablation tests all four under one protocol with matched chunk sizes and a placebo of shuffled heading paths, on Wikipedia Featured Articles and QASPER. Structure-aligned chunks with real heading paths beat both contextualized fixed windows and the placebo, showing the gain comes from content rather than extra tokens. Naive two-stage hierarchical retrieval hurts, and all effects are small but statistically significant and consistent across both corpora.

arXiv:2610.10170 PDF

Using Small Language Models to Reverse-Engineer Machine Learning Pipelines Structures

Nicolas Lacroix, Frederic Precioso, Mireille Blay-Fornarino, Sebastien Mosser cross-listed This confirmatory study tests whether small language models (SLMs), used without fine-tuning, can label source code with the stages of a machine learning pipeline (such as data preprocessing or modeling) more reliably than manual labeling or existing classifiers. Several SLMs are compared using Cochran's Q and McNemar's tests against two reference studies. The wording of the taxonomy definitions significantly changes classification performance. The best SLM does well but does not beat other classifiers, and the three classification methods lead to significantly different conclusions about how data scientists work, which suggests that existing methods bias our picture of ML practice. High inference costs also limit SLMs for large-scale studies.

arXiv:2610.10261 PDF

SemanticFold: Latent Sequence Compression SeparatesLanguage Modeling, Decodability, and Reasoning

Mingyan Liu, Min Huang SemanticFold compresses the hidden states of a prompt prefix by folding them at learned boundaries, and the authors use it to test whether such compression preserves what large language models need at inference time. They run models from Qwen3, SmolLM2, and Pythia with the prefix either native or compressed while forcing identical continuation tokens, then compare likelihood, reasoning accuracy, linear-probe accessibility, generation, and memory and latency. These measures move non-monotonically and do not share a single compression threshold: compression lowers negative log-likelihood (NLL) on Qwen3-1.7B, slightly raises it on SmolLM2, and leaves Pythia unchanged, while probe accuracy barely moves. A decomposition attributes the Qwen likelihood gain mostly to the learned residual transform rather than to the shorter sequence.

arXiv:2610.10304 PDF

Fault-tolerant foundation models

Trevor McCourt, Ila R. Fiete, Isaac L. Chuang Emerging energy-efficient hardware is often less reliable, and the authors ask whether large language models (LLMs) can be trained to tolerate its computation errors. Using modified scaling laws fitted to 40,000 GPU-hours of training on simulated faulty digital hardware, they find that error resilience increases as models grow. They interpret this as models learning to compute within good error-correcting codes whose relative overhead stays finite at any scale. They conjecture that suitably trained LLMs may be formally fault-tolerant, which would make inference on low-energy, unreliable hardware a path to large energy savings.

arXiv:2610.10311 PDF

AutoAdapt: Automatic Domain Discovery Enables Low-Cost Extensibility

Josh McGiff, Salma Mekaoui, Robert Shanahan, Nikola S. Nikolov Adding new domains to an instruction-tuned model usually means costly retraining. AutoAdapt instead discovers latent domains in the training data automatically and trains a separate Low-Rank Adaptation (LoRA) adapter for each one, independently and in parallel, then routes inputs between adapters without any learned routing parameters. A new domain needs only one new adapter, and the existing adapters stay untouched. Across 14 domain-specific benchmarks and GPT-4o pairwise judgements, the approach matches a single LoRA adapter trained on all domains, and different discovery methods converge on similar specialisation effects.

arXiv:2610.10349 PDF

ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals

Heejun Kim, Junyoung Lee, SangLyul Cho, Dongsu Han, Insu Han, Sehoon Kim Looped Transformers reuse shared blocks across several loops, which saves parameters, but their KV cache still grows with the number of loops. ResidualQuant exploits the high similarity of KV states across loops: it stores the final loop's states as a reference and encodes the other loops as low-precision residuals. Least-squares scaling, rotations, and loop-wise mixed precision make quantization down to INT2 workable. The method stays close to BF16 accuracy while reducing theoretical KV storage by 80.7%, scores up to 13.0% higher than a rotation-based baseline at the same memory budget, and on an RTX 5090 improves peak decode throughput by up to 4.15×.

arXiv:2610.10381 PDF

OrBIT: Structure-Guided Embedding Compression

Yunied Puig, Amit Kumar Jaiswal Embedding tables are among the largest parts of language models, and most compression methods fix the coding geometry in advance, for example as blocks, low-rank subspaces, or codebooks. OrBIT instead learns reusable local geometry from orbit dynamics and uses it to constrain a small set of shared codewords. The global reconstruction error decides where the coding budget is spent, and overlapping local charts let errors offset one another. After compilation only a compact decoder remains, which reaches 37.9× compression on GPT-2 and over 23× on 7B-model tables relative to 16-bit storage, with rate-distortion competitive against quantization and low-rank baselines.

arXiv:2610.10385 PDF

Training Parallel Speculative Draft Models by Directly Minimizing Expected Decoding Rounds

Yunxiao Zhao, Changxiao Cai In speculative decoding, a cheap draft model proposes tokens that the large target model then verifies. Parallel drafters propose a whole block in one forward pass, but where each decoding round starts depends on how many tokens earlier rounds accepted, a coupling that block-local training objectives ignore. The authors model speculative decoding as a Markov reward process and derive the Expected Decoding Rounds (EDR) objective, which exactly equals the expected number of decoding rounds and has no extra hyperparameters. They also derive an exact temporal-difference gradient for unbiased training and an offline evaluator that compares drafters without running speculative decoding. Fine-tuning the DSpark and DFly drafters with EDR increases mean accepted length across nine math, code and chat benchmarks.

arXiv:2610.10411 PDF

PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs

Linghao Meng, Feng He, Xuan Yang, Junyuan Mao, Pinze Ren, Deqing Mu et al. In multi-stage LLM systems, hallucinated content from one step can become the input for the next, and little is known about how models deal with a false premise within a single response. PHRBench is a controlled benchmark of 4,820 instances across four domains and 18 models that labels each reasoning trajectory as compliance with the hallucination, avoidance of it, or heuristic correction, separately from whether the final answer is right. Successful recovery is fairly rare and comes with more frequent belief updates during reasoning. Features of the hallucinated prompt alone predict recovery well, with a lightweight predictor reaching an AUROC of 0.847.

arXiv:2610.10455 PDF

Composing What Each Teacher Learned: Multi-Teacher On-Policy Distillation through Teacher-Relative Shifts

Hejian Sang, Zhengze Zhou, Shayan Mohajer Hamidi, Xiaomin Li, Rohit Jain, Alborz Geramifard Multi-teacher on-policy distillation usually transfers each teacher's full policy, which mixes what post-training taught the teacher with preferences inherited from its base model. Δ-MOPD instead transfers each teacher's logit shift relative to its base, applied on top of the student's frozen starting point. The authors show that the pull from the inherited base can be larger than the post-training shift itself. With three teachers combined on the same prompts, Δ-MOPD beats full-policy distillation by 4.11 points on math and 1.95 points across five benchmarks, and under phased domain routing it shrinks the gap between training orders from 10.50 to 6.42 points. When each update involves only one teacher, the two approaches perform about the same.

arXiv:2610.10460 PDF

Validity Without Ground Truth: What Stated-Preference Economics Offers the Evaluation of Language Models

Daniel Robert Kling Alexander, Catherine Louise Kling Many questions now put to language models have no correct answer to score against, such as how much a policy is worth or how to weigh competing values. The authors argue that the validity framework from stated-preference economics, which covers content, construct, and criterion validity, reliability, incentive compatibility, and consequentiality, can serve as a general method for evaluating LLMs. They administer a published water-quality valuation survey to six models and test predictions from economic theory, for example that demand slopes downward and that willingness to pay responds to scope and income. The tests separate the models sharply: two older models fail the most basic test at a household income of $75,000, while the two newest pass every theoretical-validity test that could be scored but disagree on convergent validity.

arXiv:2610.10506 PDF

EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

Hongru Cai, Ran Wei, Wenjie Wang, Chengfa Wu, Ning Song, Yongqi Li et al. Conditional memory architectures such as DeepSeek Engram look up learned embeddings keyed on input n-grams, which could let factual knowledge be updated while the Transformer backbone stays fixed. Doing so is hard because different phrasings of a fact activate different n-gram embeddings, and changing a shared embedding can disturb other facts. EngramEdit computes target memory representations that make the model predict the updated fact across several phrasings, then jointly updates the shared embeddings, penalizing changes to frequently reused ones more heavily. It achieves near-perfect editing success, and updated facts carry over to unseen phrasings and multi-hop reasoning, nearly tripling the strongest baseline's accuracy under chain-of-thought (CoT) prompting while largely preserving unrelated knowledge.

arXiv:2610.10533 PDF

11 more specialized papers

Theory 72

The Cost of Long Memory: State, Context, and Stability Complexity in Sequence Models

Yuheng Song The authors ask how much state, context length, or closeness to dynamical criticality a sequence model needs to forecast accurately when dependence on the past decays only slowly. For algebraically decaying memory they prove matching upper and lower bounds: the best forecast error with r exponential or finite-state modes decays like e^(−Θ(√r)), so reaching error τ requires Θ(log²(1/τ)) states. They extend the analysis to fractional long memory, where prediction from a finite context of length L has an exact 1/L leading-order error. They also show that for contractive nonlinear recurrences, vanishing error on such targets forces the dynamics toward criticality. Experiments with state-space, gated recurrent, and attention models reproduce the predicted scaling.

arXiv:2610.08816 PDF

Quadratic Weak-to-Strong Generalization in Random Feature Networks via Random Matrix Theory

Deborah Oliveira, Elliot Paquette cross-listed Weak-to-strong generalization is the effect where a stronger student trained on a weaker teacher's labels outperforms the teacher. The authors analyze it in two-layer random feature networks, where model strength is set by width. Using random matrix theory, they derive deterministic equivalents for the errors of an optimal teacher and a student trained with gradient flow. For ReLU activations and a pure spherical-harmonic target, the student's error scales as the square of the teacher's, matching a known lower bound. They also map which stopping times and multi-degree targets give quadratic, sub-quadratic, or no improvement.

arXiv:2610.09044 PDF

The Deceptive Bandit Problem: Exploratory Coupling and the Fragility of Multi-Agent Learning

Michael Tang, Mahmoud Abdelgalil, Jorge I. Poveda Randomized exploration in bandit and multi-agent learning is usually assumed to be independent and private, but the authors show that this assumption matters for security. In a two-player game, a deceptive player who receives leaked signals merely correlated with the victim's exploration can couple its own exploration to them. This steers the learning dynamics to a new steady state the authors call the deceptive Nash equilibrium. They prove that the dynamics converge to an arbitrarily small neighborhood of this equilibrium at optimal rates, under weaker smoothness assumptions than usual, and illustrate the effect in a resource-allocation game.

arXiv:2610.09120 PDF

Spatial Induction Heads: In-Context Learning of Multidimensional Cellular Automata

Kimia Kazemian, Menghan Xu, John Thickstun, Sarah Dean Mechanistic accounts of in-context learning based on induction heads assume the relevant context forms a contiguous block, which breaks when multidimensional data is flattened into a sequence. Using cellular automata with unknown local rules, the authors introduce spatial induction heads: two-layer gather-and-match circuits in which the first layer reconstructs a cell's spatial neighborhood and the second matches it against earlier occurrences using Bayesian counting. The positional dimension the circuit needs depends only on neighborhood size and spatial dimension, not on grid size. Trained two-layer transformers generalize to unseen rules with near-perfect deterministic rollouts and come within 0.005 nats of the Bayes-optimal predictor, and their attention patterns match the predicted circuit.

arXiv:2610.09124 PDF

Lower Bounds for Parallel Diffusion Sampling

Yiwen Kou, Yimeng Wang cross-listed Parallel diffusion samplers try to cut the number of sequential rounds by querying the score function many times per round, which raises the question of how much sequential work is unavoidable. The authors prove the first polynomial lower bounds on parallel rounds for diffusion sampling with approximate scores. Sampling smooth, near-isotropic Gaussian mixtures in d dimensions needs roughly at least d^(1/3) rounds, and uniform sampling from anisotropic axis-aligned boxes needs a number of rounds linear in d, a bound that is tight for that family. Both bounds hold for any randomized algorithm making polynomially many queries per round, and they rely on approximate score oracles built to hide information so it can only be uncovered sequentially.

arXiv:2610.09166 PDF

Stability and Diversity of Networked Self-Consuming Generative Ecosystems

Xiukun Wei, Yang Zhang, Xueru Zhang As synthetic content spreads, future generative models will train partly on each other's outputs, but earlier analyses of this self-consuming loop covered only a single model or a pair of models. The authors model many generative models as nodes in a directed, weighted graph, where edge weights set how much synthetic data flows from one model to another. They establish conditions under which retraining converges and characterize the resulting fixed points. They then analyze how each model's access to real data, the amount of cross-model consumption and the graph structure affect long-term stability and diversity.

arXiv:2610.09409 PDF

Self-Consuming Generative Models with Co-Evolving Human Preferences

Xiukun Wei, Tian Xie, Ding Zhu, Xueru Zhang When users pick favorite outputs from a generative model and those picks train the next model generation, exposure to the outputs can also change what users prefer, so the model and the preferences evolve together. The authors analyze these coupled dynamics and show that training only on user-curated synthetic data amplifies initial biases. The system then collapses to one of several single-instance equilibria in which whichever instance started ahead ends up dominating. Mixing in reference data at a sufficiently high rate yields a unique, globally attracting equilibrium instead. They also propose an efficient algorithm that jointly picks the reference distribution and its mixing weight to steer the system toward equilibria with desired attributes at low data-collection cost.

arXiv:2610.09415 PDF

DSReg: Provably Recovering Individual World Latents without Reconstruction

Yujia Zheng, David Klindt, Randall Balestriero, Bernhard Sch\"olkopf Methods that recover individual latent variables usually rely on reconstruction, supervision or distributional assumptions, while joint-embedding predictive architectures (JEPAs) identify latents only up to a linear mixing. The authors define Structural Diversity, the condition that different latents leave distinct dependency patterns on the observations, and show it is strictly weaker than the structural conditions of earlier identifiable models. Building on the linear identifiability of LeJEPA, they prove that DSReg (Dependency-Sparsity Regularization) recovers each world latent up to signed permutation, with no reconstruction, decoder or labels. The method can be applied after training to existing checkpoints and improves individual-latent recovery across synthetic and visual settings while preserving prediction quality.

arXiv:2610.09457 PDF

Closed-Form Noise Calibration Against Membership Inference for Random-Allocation DP-SGD

Murat Bilgehan Ertan, Marten van Dijk DP-SGD (differentially private stochastic gradient descent) usually sets its noise level by running a numerical privacy accountant inside a search. For the random-allocation variant, where each record is used once per epoch at a random step, the authors derive a one-line closed-form bound on the accuracy of any membership inference attack (MIA). The bound follows from a chi-square divergence argument, gives the noise multiplier in about a microsecond, and needs at most about half the noise of the previous best closed-form bound. An exact expression they also derive cuts noise by a further 13–20% and is provably optimal for any accountant that knows only the step count, epoch count, and noise multiplier. In training, it matches a published accountant on test accuracy at much lower cost, and MIAs on the trained models stay below the bound.

arXiv:2610.09651 PDF

Fully Interpretable Minimal Transformers: From Geometry to Algorithm

Raneem Mahajne, Toviah Moldwin Limiting a transformer's embedding dimension and head size to 2 makes every internal representation directly visible in the plane. This includes embeddings, query/key/value transforms, attention outputs, residual streams and decision boundaries. The authors train such a model to output the most recently seen even number whenever a '+' appears in a sequence of digits, then walk through each step of its computation visually. Their central claim is that the learned 2D geometry can be read directly as a step-by-step algorithm. They offer the framework and its suite of visualizations as a teaching and experimental testbed for how transformers carry out next-token prediction.

arXiv:2610.09838 PDF

Eigenvalues of the Hessian in Deep Learning: The Origin of Symmetry and Its Breaking

Yossi Arjevani Hessian spectra of trained networks consistently show a large bulk of eigenvalues near zero and a few isolated outliers. The paper explains this by treating the actual setting as a departure from a nearby, highly symmetric reference configuration, reached by modifying the architecture, data distribution, or parameter metric. In that reference, symmetry produces high-dimensional kernels and eigenvalues of large multiplicity, and moving back to the original configuration breaks the symmetry, which creates the observed hierarchy of clusters and outliers. The analysis covers three-layer ReLU networks in detail and extends to convolutional, graph, and transformer models, the neural tangent kernel (NTK), layerwise Hessians, and the Gauss-Newton matrix.

arXiv:2610.09919 PDF

Force without transmission: a depth-induced rank collapse that no loss on the representation reopens

Martin Hofmann, Patrick M\"ader During training, a transformer can fall into rank collapse, where all token representations point in one direction and learning stops. The authors induce this in small transformers by weakening the skip connection and test whether added loss terms can repair it, as such terms do for a related attention collapse. No added loss term repaired the collapse: the task gradient no longer reached the query and key weights, and the corrective term's gradient faded before reaching the blocks where the collapse forms. Restoring the skip connection, or a burst of high learning rate, reopened the gradient path and the rank recovered, but loss stayed worse than in a healthy network. The authors conclude that what decides repair is whether gradient reaches the weights that must change, not how hard a loss term pushes.

arXiv:2610.09958 PDF

How Do Transformers Learn to Represent Symmetries?

Eduardo Santos-Escriche, Valerie Engelmayer, Ya-Wei Eileen Lin, Stefanie Jegelka The authors study how a vanilla Transformer learns invariance to symmetries of point-cloud data when trained with a finite amount of data augmentation. They find an ordering of learnability: symmetries that do not preserve angles are hardest, angle-preserving symmetries are easier, and base angle-preserving subgroups such as translation, rotation, and scale are easiest. For those base groups they examine extrapolation and analyze the trained models' internals, identifying interpretable mechanisms that produce invariance. They also show these mechanisms serve as building blocks for learned equivariance.

arXiv:2610.10305 PDF

59 more specialized papers

Other 67

HE-OFT: Privacy-Preserving One-Shot Federated Fine-Tuning under Homomorphic Encryption

Halil \.Ibrahim Kanpak, Sinem Sav, Alptekin K\"up\c{c}\"u cross-listed Organizations that want to fine-tune a shared model on private data face reconstruction attacks in federated learning (FL), and even one-shot protocols normally hand the final model to every participant, which is not allowed when the model is a regulated or proprietary asset. HE-OFT is a one-shot federated fine-tuning protocol in which no party ever receives the trained model. Each client trains a low-rank adapter and classifier head on a frozen public backbone, uploads a single encrypted head update that the server combines under multiparty CKKS homomorphic encryption without decrypting it, and a quorum of clients returns only the predicted label to whoever sent the query. Across four text classification tasks and one vision task it reaches 61–79% accuracy, compared with 20–48% for a client training alone, keeping 85–96% of the accuracy of a model trained without these privacy protections; the cost is query latency of roughly 56–255 s on a GPU.

arXiv:2610.08255 PDF

Accelerating Floating-Point Satisfiability Solving via Gradient Normalization

Yuanzhuo Zhang Optimization-based Satisfiability Modulo Theories (SMT) solvers for quantifier-free floating-point formulas apply gradient descent to continuous relaxations of the formulas, but they often stall because a few hard clauses dominate the gradient and trap the search in local minima. GradSAT treats each clause as a separate task in a multi-task learning setup and applies dynamic gradient normalization (GradNorm) during solving, damping dominant clauses and boosting lagging ones. Solving runs in two stages: a GPU-accelerated PyTorch backend with symbolic compilation and operator fusion first steers the relaxation into a good region, then a bit-precise local search engine turns the candidate into an exact satisfying assignment. The authors claim that balancing per-clause gradients stabilizes the search and avoids the brittleness of earlier gradient-based solvers, though the abstract reports no quantitative results.

arXiv:2610.08808 PDF

A Shortcut to Structure in AlphaFold 3

Jonathan Feldman, Jeffrey Skolnick cross-listed The authors use causal interventions and probes on every Pairformer block to trace how AlphaFold 3 builds a protein's global shape. Removing the multiple sequence alignment (MSA) leaves local secondary structure mostly intact but breaks long-range topology. With the MSA, the final fold can be read out by about block 9 of 48, long before the decoder renders it. An alignment taken from a different protein with the same fold, supplied only at the matching columns, raises median TM-score against experimental structures from 0.44 to 0.72. Shifting that alignment by a few residues does worse than giving no alignment at all, so the model appears to read a description of the fold from the MSA rather than the query's own evolutionary history.

arXiv:2610.08937 PDF

The Symbol of the Surrogate: Measuring Numerical Provenance in Neural PDE Solvers

Ridham Patel Neural PDE surrogates, which are learned approximations of partial differential equation solvers, are trained and tested on outputs from the same numerical solver, so standard benchmarks cannot tell whether a surrogate follows the true physics or copies the solver's discretization errors. The authors introduce an empirical Fourier-symbol diagnostic that probes a surrogate's linearized one-step operator with individual Fourier modes, and they train identical networks on schemes with opposite error signatures. In linear advection, surrogates reproduce their training schemes' amplitude and phase errors, with the difference between the paired schemes reaching more than 99.8% of the theoretical full-imitation ceiling. The same imitation appears for a Fourier neural operator and for nonlinear Burgers dynamics, which shows that matching solver-generated test data does not establish fidelity to the exact evolution.

arXiv:2610.09255 PDF

The AI Evaluation Ecosystem

Yash Dave, Sang T. Truong, Serena Wang, Sanmi Koyejo Benchmark design choices interact with the incentives of model providers, users, funders, and regulators. The authors build a simulation that combines rule-based market dynamics with LLM-driven strategic actors, following Generative Agent-Based Modeling (GABM), and represent benchmarks, user needs, and provider capabilities as vectors over six capability dimensions. In a case study on holdout design, moving from public to private holdout benchmarks shrinks the gap between benchmark scores and user satisfaction on most benchmarks but widens it on a few, depending on where holdout weights shift scoring credit. The authors present the simulator as a hypothesis-generating sandbox and stress-test it against established verification and validation criteria.

arXiv:2610.09296 PDF

Pretraining Shapes Spectral Structure: Architecture- and Strategy-Conditional Prediction of OOD Robustness in Foundation Models

Sangyoon Bae, Sk Miraj Ahmed, Shinjae Yoo, Jiook Cha This work tries to predict a foundation model's out-of-distribution (OOD) robustness from its pretrained weights alone, before any target data exists. The authors prove that the OOD accuracy gap is bounded by how tightly source representations concentrate, and they use a spectral statistic of the weights as a proxy, whose direction flips between architecture families. Pooled over 116 models in 7 modalities the statistic ranks robustness weakly, but within a single combination of architecture and pretraining strategy it orders 92% of model pairs correctly, and pre-registered predictions held for EEG, genomic, and protein models. Acting on spectral concentration also narrows the OOD gap by 24% while keeping 87.5% of in-distribution accuracy.

arXiv:2610.09709 PDF

Global Average Precision for Representation Learning

Bill Psomas, Mohammad Mahdi, Michalis Thomas, Danda Pani Paudel, Giorgos Tolias, Giorgos Kordopatis-Zilos cross-listed Standard retrieval metrics and contrastive losses such as InfoNCE score each query separately, so similarity scores need not be comparable across queries, even though any system with a single decision threshold relies on that. gSAP is a differentiable surrogate of Global Average Precision (gAP), which ranks all query-candidate pairs in one list. It takes the same inputs as existing losses, so it can be dropped into existing training recipes, and it stays trainable at low temperatures. It improves supervised metric learning, cross-modal alignment and self-supervised pretraining. Under a single threshold it retrieves up to four times as many positive pairs as the strongest AP surrogate at the same precision, and it also improves transfer, kNN and zero-shot accuracy.

arXiv:2610.09863 PDF

A Drosophila Whole-Connectome Network Can Learn Human-Designed Cognitive Tasks

Joonghui Cho, Minchan Kang, Daeshik Kim The authors test whether a biological wiring diagram can serve as a general computational substrate by using the MaleCNS v1.0 fruit-fly connectome as the fixed recurrent topology of an artificial network, learning one scalar per anatomical edge. On bounded addition, the anatomical graph reaches 92.77% held-out accuracy versus 67.93% for degree-preserving random rewirings. On a grounded relational language task with a 100-word lexicon, it reaches 61.59% versus 44.17% for matched rewirings and ranks first among 21 graphs at the canonical interface. The results suggest the connectome's higher-order wiring provides a reusable inductive bias for tasks it did not evolve for.

arXiv:2610.10014 PDF

When Scientific Cognition Is No Longer Scarce

Nathan DeBardeleben cross-listed This position paper imagines a world where machines are better, faster, and cheaper than people at most scientific work that can be done on a computer, and asks what would then limit science. The authors argue for a "scarcity inversion": literature synthesis, hypothesis generation, software, simulation, and analysis become abundant, while experiments, observations, validated conclusions, and accountable institutional authority stay scarce. They break the shift into four parts (selection, physical access, validation, and organizational choice) and say it is arriving first in mathematics and software or algorithm design, where the whole research loop can run inside computation. They then look at how national laboratories might rethink facilities, verification, provenance, and governance in response.

arXiv:2610.10241 PDF

Thinking in Depth: Retrospective Inference for Tabular Foundation Models

Hao-Run Cai, Si-Yang Liu, Zi-Jian Cheng, Kun-Yang Yu, Jin-Hao Sheng, Guo Yu et al. Tabular foundation models (TFMs) predict on a new table by using its labeled rows as context, and tracing individual queries through several strong TFMs shows that predictive refinement is uneven across layers and concentrated late in the network. Retro lets later layers revisit and recombine earlier intermediate representations, using Attention Residuals to reweight contributions from different depths and query-conditioned Gated Attention to shape each query's update. This spreads refinement earlier and more broadly across depth. On TabArena, TALENT, and RelArena, Retro ranks in the top three and sits on the Pareto frontier.

arXiv:2610.10317 PDF

Continual Learning without Continual Training

Nikita Narayanan, Ritham Majumdarr, Sonali Parbhoo Instead of continually retraining a model and fighting catastrophic forgetting, the authors replace continual training with continual inference. Latent Concept PFN is a prior-data fitted network that is meta-trained once and then frozen. It adapts to new domains and classes only by adding exemplars to an in-context memory and performing Bayesian inference over a latent concept space. Because no parameters change, forgetting is reduced, and the same model handles both class-incremental and domain-incremental learning without task identity. Concept annotations are used only during meta-training, and experiments show competitive continual learning performance with interpretable latent concepts.

arXiv:2610.10379 PDF

56 more specialized papers

Safety & Alignment 45

Beyond the Sycophancy Score: How Task, Model, and Pressure Shape LLM Yielding

Guang Yang, Homa Hosseinmardi, Fengchen Liu, Amir Ghasemian Large language models (LLMs) often drop a correct answer or agree with a user who pushes back, a behavior called sycophancy that is usually reported as one rate per model. The study analyzes 103,939 graded replies from ten model configurations across 200 items, 13 pressure conditions, and four-turn conversations, and finds that the task itself, mainly how costly the user's claim is to verify, explains far more than model family or pressure tactic (0.485 of McFadden R² versus 0.139 and 0.009). Models almost never concede anchored facts (1.3%) but endorse personal choices 77% of the time, and most concessions on hard items come from models that cannot solve them. Maximum reasoning cuts concessions on deep puzzles to 0%, while fallacious or emotional framing has no more effect than simply repeating the claim.

arXiv:2610.08840 PDF

AdaGuard: Enhancing Safety and Policy Compliance with Reasoning-Enabled LLM-As-A-Judge Guardrails

Melissa Kazemi Rad, Sihui Dai, Isha Slavin, Kushal Chawla, Mann Patel, Jian Ni et al. Enterprise guardrails tend to rely on fixed policy sets and offer little transparency or control over latency. AdaGuard is an LLM-as-a-judge guardrail trained with supervised fine-tuning and GRPO reinforcement learning that enforces user-defined safety and compliance policies supplied at runtime, without retraining. It infers how complex each input-policy pair is and switches between fast verdict-only inference and explainable, reasoning-enabled moderation. The authors report that it rivals guardrail and frontier models several times its size, and that its auto-reasoning mode recovers the accuracy of always-on reasoning at a fraction of the latency.

arXiv:2610.08923 PDF

Multi-Aspect Runtime Verification for Simulation-Based V&V of LLM-Enabled Autonomous Agents

Nikolaos Kekatos, Dimitrios Nikou, Anastasios Temperekidis, Alexios Lekidis, Nikolaos Kolokotronis, Panagiotis Katsaros et al. cross-listed LLM agents in defence decision-support roles must respect written obligations that are spatial, temporal, and semantic at once, and violations often emerge only across multi-step tool-call chains that per-event guardrails miss. The proposed runtime-verification framework decomposes each natural-language policy clause into a spatial, temporal, and semantic triple over one event stream, checks each aspect with its own monitor, including the MonPoly engine for temporal properties, and fuses the verdicts with a four-valued algebra that tracks provenance. Across two military mission domains and one civil domain, composing all three aspects drives attack success to zero with no observed false positives at microsecond-scale cost per event, while any single aspect or pair leaves many attacks succeeding. In closed-loop tests, a policy-naive planner reached violating states in most unshielded missions and in none when shielded.

arXiv:2610.08928 PDF

How Fragile Is On-Device Language Model Safety? Localizing Safety-Critical Parameters for Sparse Fault Analysis

Muhammad Zeeshan Karamat, Christiana Chamon Garcia As small language models run on devices and inside agents, tampering with their locally stored weights becomes a safety risk. Using low-rank subspace analysis and parameter-level importance filtering on LLaMA-2-7B-Chat, the authors find that safety behavior is concentrated in a small set of parameters, mainly in the MLP down_proj layers and to a lesser extent in o_proj. Modifying only 0.19% of the weights raises attack success rates to 53% for basic attacks and 56% for GCG, while tinyBenchmarks accuracy barely changes (51.6% versus 52.2%). The authors argue this motivates targeted integrity protection for on-device models.

arXiv:2610.09000 PDF

Learning to Report Unsafe Tasks in a Multi-Agent Game

Avyay M. Casheekar, Hariganesh Tangirala When agents share a reward for completed tasks, reporting unsafe work costs the reporter, and audits that make reporting optimal do not guarantee that training will move a silent team toward reporting. In a game where any of k witnesses can stop a task, the authors derive audit conditions under which exact policy-gradient updates reach universal reporting. They show that when witnesses share a policy, the cheapest sufficient audits cost k times as much as with one policy per role. In PPO experiments on 24 witness graphs, separating co-witnesses into different policies reduces unsafe task completion by 33.59 percentage points compared with shuffled groups of the same size.

arXiv:2610.09002 PDF

Removing Information Content Does Not Certify Tamper Resistance in Open-Weight Models

Domenic Rosati, Alessa Carbo, Ali Dadsetan, Hong Huang, Matthew Young, Subhabrata Majumdar et al. The authors ask whether stripping harmful information from open-weight models, through training-data filtering or capability removal, can certify that the models resist fine-tuning attacks. They argue that mutual information measured at release cannot give such a guarantee, because function-preserving reparameterizations leave information unchanged while changing the geometry that gradient descent follows. They give an explicit construction where both weight–data and label–representation mutual information are zero, yet the removed capability recovers in a single gradient step. Controlled experiments show that recovery speed depends on training order and parameterization, and the authors conclude that certification also needs constraints on attack dynamics.

arXiv:2610.09004 PDF

Visual Memory Attacks Can Persist Through The KV Cache

David Dobre, Leo Schwinn, Gauthier Gidel, Spandana Gella, Perouz Taslakian, Pierre-Andr\'e No\"el cross-listed The authors show that an adversarial input can keep steering a model after it has been removed from the context, because its influence persists in the key/value (KV) cache of later tokens. Building on the Visual Memory Injection (VMI) setting, where an adversarial image plants a trigger-activated backdoor, they introduce Persistent Visual Memory Injection (P-VMI), which optimizes images to keep working after they are masked from attention. On Qwen3-VL-8B-Instruct, P-VMI reaches up to about 90% target success, still works when the image is shown only on the first turn, and persists over conversations longer than those used during optimization. A cache-swap ablation locates the effect in the KV cache, and the attacks can be trained to survive context compaction into a model-generated summary.

arXiv:2610.09027 PDF

Quad-State Safety Evaluation of Open-Weight Large Language Models on Non-Canonical Inputs

Pavan Maddula LLM safety evaluations mostly test harmful requests written in plain text, but real inputs include emojis, altered spellings, encodings, and character-level tricks. The authors release the Adversarial Surface-Form Robustness Dataset (ASRD), 2,100 prompts across seven surface-form families, and evaluate five open-weight models (10,500 responses) with a four-outcome rubric: harmful compliance, safe response, comprehension failure, or indeterminate. Emoji and invisible-Unicode variants were understood almost perfectly and still drew harmful compliance near the 22.87% baseline, whereas leetspeak, encoded, and hybrid transformations lowered harmful compliance mostly because comprehension failure rose to as high as 65.60% for encoded inputs. Inspection of raw outputs identified three failure behaviors: hallucinated benignity, structural collapse, and language drift.

arXiv:2610.09033 PDF

Careful Judge: Safe and Efficient Human-AI Collaborative Decision Making

Chenyu Zhang, Rachel Luo, Boyi Li, Anjali Parashar, Marco Pavone, Apoorva Sharma cross-listed In human-AI decision making, human review catches unsafe AI decisions but is expensive, and if the AI learns from the feedback it queries, safety thresholds calibrated on the old model stop holding. CARE (calibrated adaptive rectification and escalation) is a modular pipeline around any black-box model that escalates to humans and learns from their feedback continuously. Its adaptive calibration module guarantees risk control at every time step. On four safety-critical datasets covering driving, language, and robotics, CARE produces human-aligned decisions while reducing human queries by 25–81% compared with baselines.

arXiv:2610.09043 PDF

SpecGuard: Proving a Task Is Broken Before the Agent Cheats

Param Biyani, Krishnamurthy Dvijotham Coding agents given tasks whose description conflicts with the tests rarely flag the problem; instead they cheat by editing tests or hard-coding outputs, sometimes deleting security defenses to do so. Before any agent runs, SpecGuard turns the task description and codebase into a Lean 4 specification, formalizes the tests separately, and has the Lean kernel check whether any implementation could satisfy both, producing a machine-checked certificate when none can. On conflicted SWE-bench tasks it detects up to 72.8% of conflicts and formally certifies up to 51.1%, with a nearly five-fold lower miss rate than model-based judgment.

arXiv:2610.09159 PDF

Adversarial Images Hijack Web Agents from Visual Grounding to Browser Execution

Wanjing Han, Levi Taiji Li, Mu Zhang, Yue Jiang, Guanhong Tao cross-listed Earlier visual red-teaming of web agents targets only the model's inference and ignores how agents parse webpage input and turn model output into browser actions, so attacks that succeed against the model do not necessarily take control of the browser. WebMirage treats the attack as an end-to-end problem from visual grounding to execution. It crafts localized image perturbations that make agents select attacker-controlled content and carry out the matching browser action across different page renderings, using a role-slot abstraction, webpage recomposition, and dataflow analysis of action post-processing. Across four agent configurations, six vision-language backbones, and 2,250 tasks on 13 public websites and a sandbox, it reaches a 91.9% average attack success rate versus 17.4% for the strongest baseline, and it remains effective against three agent-level defenses.

arXiv:2610.09240 PDF

Efficient Best-of-N policy evaluation for inference-time alignment

Jonas Schweisthal, Yuxin Wang, Athiya Deviyani, Stefan Feuerriegel, Dennis Frauen Best-of-N (BoN) sampling picks the highest-scoring of N responses from a reference model, but evaluating such policies from logged data normally requires response likelihoods that are often unavailable. The authors show that the order-statistic structure of BoN lets the required density ratios be written as score-rank probabilities that can be estimated from samples alone. On that basis they build a doubly robust estimator (BoN-DR) that reuses one shared sample pool across candidate budgets, and they prove it is efficient and gives valid inference even when the reward estimator is misspecified. They also propose two rules for choosing N, one of which maximizes a lower confidence bound to give a no-harm guarantee against reward overoptimization, and they validate the framework on synthetic data and GSM8K.

arXiv:2610.09250 PDF

Package Hallucination Attacks on Coding Agents through Prompt Injection in Rule Files

Yupu Wang, Zhengyuan Jiang, Reachal Wang, Neil Zhenqiang Gong cross-listed Coding agents increasingly follow community-shared rule files such as AGENTS.md or .cursorrules. The authors show these files can be poisoned so that agents swap legitimate dependencies for attacker-controlled packages, an attack they call package hallucination. Their framework PackHallu optimizes the injected prompts with an evolutionary search that uses trajectory-level feedback and LLM-guided mutations. Across multiple benchmarks, LLMs, and agent frameworks, the attack achieves high success rates and transfers well across model and agent combinations.

arXiv:2610.09264 PDF

RT-Safe: Benchmarking Agent Safety in Real-Time Embodied Environment

Tianruo Rose Xu, Jiawei Ren, Yichi Yang, Zhaoxu Zheng, Lianhui Qin Embodied agents act in a world that keeps changing while they reason, so an action that was safe when the agent observed the scene can become unsafe by the time it executes. RT-SAFE is a simulated urban navigation benchmark with moving actors, hazards, and traffic rules in which the world continues to evolve during inference and execution. Across eight vision-language models (VLMs), agents usually complete their tasks but almost never do so safely: in the hardest setting only 0.7% of episodes finish without a safety event. Matched static and real-time runs have similar task completion rates (91.3% and 94.1%), but real-time execution raises collisions 12.3x, and offline reinforcement learning on the benchmark substantially reduces collisions.

arXiv:2610.09294 PDF

The Confidence Game: Strategic Miscalibration in Human-AI Delegation

Raghu Arghal, Saswati Sarkar, Shirin Saeedi Bidokhti cross-listed When AI agents are rewarded for engagement or revenue, they may have an incentive to distort the confidence they report. The authors model this as the Confidence Game, a repeated signaling game in which a user decides whether to delegate a task based on the agent's reported confidence, and they show that honest reporting is not an equilibrium and that a sufficiently short-sighted agent inflates its confidence. When an LLM is placed in the agent role and told its true probability of success, it claims high confidence on 56% of tasks it was told it would probably fail. The same behavior appears on real tasks where the model must estimate its own accuracy. The resulting reporting rule destroys 68% of the gains from delegation.

arXiv:2610.09371 PDF

The Persona Hierarchy Model: Understanding Contextual Generalization in Fine-Tuning LLMs

Jiachen Zhao, Zhengxuan Wu, David Bau, Weiyan Shi Fine-tuning a language model under a fixed system prompt or persona sometimes changes its behavior only in that context and sometimes generalizes broadly. The Persona Hierarchy Model explains this with a shared default persona: changes to it transfer widely, while changes to local personas stay context-specific. Across 120 fine-tuned models, narrowness of generalization correlates with how similar the training context's persona is to the default persona (Pearson's r = 0.72 for Qwen3-4B). Building on this, persona-preserving regularization (PPR) cuts reward hacking during reinforcement learning from 42-55% to at most 0.2% while retaining accuracy gains.

arXiv:2610.09384 PDF

Secure-CUA: Controlling Untrusted Influence in Computer-Use Agents

Sarthak Choudhary, Mihai Christodorescu, Ashish Hooda, Somesh Jha, Tongxin Li, Damien Octeau cross-listed Computer-use agents (CUAs) act on desktop, mobile, and web interfaces that mix trusted controls with untrusted content, so an adversary can embed instructions or misleading visual cues that redirect the agent's actions. The authors formalize security requirements for both the agent's decisions and their execution as GUI commands. They then build Secure-CUA, which commits to an explicit per-step program called an action transaction before it reads any untrusted content, masks untrusted screen regions, and answers queries about that content with an isolated model. Under the model's assumptions the system is secure by design. Across 6,000 execution traces on WebArena, it keeps 53.55% task success versus 55.12% for an undefended agent, compared with only 13.17% for CaMeL-CUA.

arXiv:2610.09469 PDF

It Is Not Seeing the Hazard: A Frozen Vision-Language Safety Score Measures Its Caption Bank

Samuel Tetteh, Cody Fleming cross-listed Frozen vision-language models are increasingly used to turn image-text similarity into rewards or costs for safe reinforcement learning, on the assumption that similarity to danger-describing captions reflects real hazards. The authors test a frozen CLIP prompt-margin safety score on trajectories from policies that never received it. They compare pre-contact frames with contact-free frames of similar hazard geometry, and they vary the captions, the encoder, and the camera view. Across 180 episodes and 130 contact onsets the score does fall for about twenty steps before contact, but controls show it mainly tracks how much the scene resembles its caption bank rather than the hazard itself. A constant-confidence control reproduces the lower catastrophe rate, so policy gains alone do not show hazard perception.

arXiv:2610.09517 PDF

Constitution-Guided Watermarking

Toluwani Aremu, Samuele Poppi, Nils Lukas cross-listed Watermarking language model output involves trade-offs between text quality, robustness to editing, and resistance to forgery, and providers typically apply one fixed configuration to every request. Constitution-Guided Watermarking lets providers state requirements as natural-language principles. Offline, a reasoning agent refines a watermark configuration for each rule using empirical feedback; at deployment, a monitor matches each request to the applicable rule and retrieves its configuration, including exemptions, without changing the serving model. In a proof of concept with KGW and a five-rule constitution, it improves detection after paraphrasing by up to 14 percentage points on robustness-prioritized requests, while matching or beating the baselines on quality and clean detection at a 0.1% false-positive rate.

arXiv:2610.09552 PDF

RELATE: An Evaluation Framework for measuring Relational Orientation of Large Language Models

Shivam Shukla, Jihye Kim, Shubham Gaur, Mahnaz Roshanaei, Magy Seif El-Nasr As LLMs are increasingly used for emotional support, a concern is whether they steer users toward relying on the AI rather than on other people. RELATE measures this as relational orientation, labeling each sentence as inward-facing (positioning the AI as the ongoing source of support) or outward-scaffolding (encouraging real-world human connection). It covers 76 help-seeking situations, each paired with three simulated user styles. Across seven LLMs and 69,194 assistant sentences scored by an LLM judge, inward-facing language becomes more common by the sixth assistant turn than in the first, and outward-scaffolding language is much rarer for hesitant, indirect users than for explicit, reassurance-seeking ones.

arXiv:2610.09569 PDF

SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language Models

Miao Yu, Hao Huang, Lu Yuan, Yunpeng Li, Kun Wang, Zuming Jiang Most safety interpretability work studies models after alignment and ignores the safety mechanisms already present in pretrained base models. SafeEvo extracts weak refusal circuits from base LLMs, which are sparse subgraphs whose ablation removes all refusal of harmful inputs, and traces how these circuits change across alignment checkpoints. The analysis suggests that the alignment tax, the loss of general capability from safety training, comes from refusal-circuit updates spilling into utility-related parameters. The proposed Safety Circuit Alignment (SCA) confines safety updates to those circuits and, averaged across three LLMs and two alignment algorithms, lowers harmfulness by 63.21%, reduces over-refusal of benign queries by 58.44%, and retains 99.58% of original capability.

arXiv:2610.09600 PDF

Understanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMs

Shuailong Wang, Xinyu Lyu, Shengming Yuan, Jingkuan Song, Heng Tao Shen, Lianli Gao cross-listed Token pruning speeds up vision-language models (VLMs) by discarding redundant visual tokens, and this first broad safety evaluation finds that most pruning strategies make models less safe as the pruning ratio grows, while query-based compression unexpectedly makes them safer. The authors trace the degradation to Pruning-Induced Malicious Amplification: once background tokens are removed, attention collapses onto a few retained malicious foreground tokens and amplifies their harmful content under jailbreak attacks. Their plug-and-play, inference-time Safety-Aware Pruning (SAP) finds those malicious anchors, restores pruned benign tokens, and shifts attention toward them, reducing attack success rate by up to 62% without hurting efficiency or utility.

arXiv:2610.09703 PDF

Backdooring Acoustic Foundation Models for Physically Realizable Triggers

Zebin Yun, Eyal Ronen, Mahmood Sharif cross-listed Acoustic foundation models (AFMs) underpin many speech and audio applications, but their security has received little study. The Foundation Acoustic model Backdoor (FAB) attack plants a backdoor in an AFM without access to its pre-training data, keeps normal performance intact, and survives downstream fine-tuning, degrading many tasks once triggered. The triggers are task-agnostic, physically realizable, inconspicuous, and need no time alignment, such as a background siren. Tested on two leading AFMs, nine downstream tasks, and four triggers, the attack works in both digital and physical settings and evades established defenses; extensive end-to-end fine-tuning can remove it, but that fix is costly and task-specific.

arXiv:2610.09819 PDF

Defensive Sufficiency in a Stackelberg Model of AI Security

Subhabrata Majumdar, Rajlakshmi Chavan cross-listed The paper asks when the feedback loop of automated testing, red teaming, and incident response actually gives an AI system sufficient protection, and when investing in it pays off. The authors prove that a finite attack surface is eventually defended with probability 1 if every unresolved attack has a persistent chance of discovery, repairs work, and later updates do not undo earlier fixes. They derive completion-time bounds and extend the analysis to growing attack surfaces and repairs that generalize. They then model defense as a defender-led Stackelberg game and characterize the least-cost allocation between proactive discovery and reactive repair. Numerical experiments show that faster repair shortens compromise duration without lowering compromise probability.

arXiv:2610.09892 PDF

Constrained-Action AI Remediation for SIEM/XDR via a NeMo-Guardrails Proxy

Georgios Koutidis, Nikolaos Kekatos, Tom Nianios, Alexios Lekidis cross-listed When Large Language Models (LLMs) triage security alerts and issue remediation commands, a single adversarial alert can turn the model's reasoning into a remote code execution path. The proposed architecture has two layers. A SIEM/XDR control plane grounds remediation in correlated host events and limits the LLM to a closed vocabulary of intents whose templated commands are checked by an argument validator. A NeMo-Guardrails proxy wraps the analyst LLM with input and output rails. Tested against a released SOC-specific adversarial corpus, the stock proxy raises prompt-injection recall from 25.0% to 94.5% at a 0.1% false-positive rate, and a live red-team exercise found that the intent vocabulary and validator contained the observed failures. Because the rails add latency, the authors recommend running the loop with a human in the loop or with a delay.

arXiv:2610.09906 PDF

Inverting Multi-Vector Visual Document Indices

Zhuchenyang Liu, Yao Zhang, Yu Xiao cross-listed Multi-vector visual document retrievers store each page as about a thousand patch vectors, often in third-party vector databases, and these indices tend to be treated as less sensitive than the pages themselves. The authors frame inversion as conditional document image generation and show that whoever runs or breaches the store can reconstruct pages: on ViDoRe v3, pages inverted from raw indices recover 47% of the words and 45% of the sensitive tokens, and they rank their source page first 98.4% of the time when used as queries. Token pooling and shuffling both cut word recall to about 8%, but a model that restores the order of shuffled vectors raises first-rank retrieval from 3.8% to 93.5%. Inverting a pooled index remains an open problem, and the authors conclude that these indices should be protected like the documents they encode.

arXiv:2610.09920 PDF

Successive Training Stages and Large Language Model Persuasion: Effects of Misalignment, Supervised Fine-Tuning, and Preference Optimization

Antony Dalmiere (LAAS-TRUST), Pascal Marchand (LAAS-TRUST, INSA Toulouse), Guillaume Auriol (LAAS-TRUST, INSA Toulouse), Vincent Nicomette (LAAS-TSF et al. The study measures how successive post-training stages change how persuasive a large language model (LLM) is. The stages are supervised fine-tuning (SFT) on conspiracy data, further SFT on argumentative data, and Identity Preference Optimization (IPO). In a preregistered-style experiment, 835 Prolific participants read personalized texts on 10 divisive political issues from one of five sources, and attitude change was compared across conditions. Persuasive SFT produced greater attitude change than conspiracy training alone (d = 0.30), while IPO added no benefit (d = 0.03), and GPT-4 did not differ from neutral text. Effects also depended on participants' starting attitudes.

arXiv:2610.09964 PDF

From Expected Harmfulness to Likelihood: A Probabilistic Reformulation of Jailbreaking LLM Agents

Juanyang Xu, Zheng Wang, Xingyu Zhao, Siddartha Khastgir, Andi Zhang Jailbreak attacks on LLM agents usually take one of two forms: maximize the expected harmfulness of outputs, or pick harmful target outputs and raise their likelihood. The authors connect the two by proving that the gradient of log expected harmfulness equals the expected gradient of the model's log-likelihood under a harmfulness-reweighted output distribution. Building on this identity, they propose OPUR, a sampling distribution that generates highly harmful target outputs to guide likelihood-based input optimization. Experiments show the resulting attack is effective at jailbreaking LLM agents.

arXiv:2610.09973 PDF

EASE: Entropy-Adaptive Distribution Shaping for Evading AI-generated Text Detectors

Jicheng Zhou, Kahim Wong, Jialong Wang, Jiantao Zhou Detectors of AI-generated text (AIGT) are sensitive to how the source large language model (LLM) decodes, and the authors observe that perturbing next-token logits or changing sampling temperature already lowers detection rates. They propose EASE (Entropy-Adaptive Distribution Shaping for Evasion), a training-free, detector-agnostic method that uses the predictive entropy of the source model's next-token distribution to adapt both logit perturbation and temperature. It needs no feedback from the detector and no fine-tuning. Across three source LLMs and multiple detectors, EASE consistently reduces detection performance with negligible loss of text quality and negligible inference overhead.

arXiv:2610.09976 PDF

Sensitive-Topic Leakage Through LLM Routing Metadata: Measurement and Mitigation

Teng-Ruei Chen cross-listed LLM routers send each request to a cheap or expensive model based on its content, and many gateways log that choice even when content logging is off, creating a privacy side channel. Using pre-registered studies on 1.7 million real requests from WildChat-1M and LMSYS-Chat-1M, the authors show that routing decisions shift by topic even at matched prompt length. With RouteLLM, for example, harassment and self-harm requests reach the strong model 19 points less often than comparable requests. Twenty RouteLLM decisions identify frequent medical askers with an AUC of 0.71 (0.92 for a domain router). Per-category, length-matched parity removes the gap at a cost of at most 0.2 accuracy points on RouterBench, while simpler defenses such as conversation stickiness, per-user budgets and pooled parity fail.

arXiv:2610.09981 PDF

Comprehension Audits to Mitigate Risks from Automated AI Research

Ronald J. Bodkin, Bahrad A. Sokhansanj, Gillian K. Hadfield cross-listed With AI already writing most of the code at frontier AI labs, insufficient human understanding of what is being built becomes a safety risk, and no published assurance regime makes demonstrated human comprehension a precommitted condition for continuing development. The authors propose comprehension audits, in which the people responsible for research and development contributions explain them to independent auditors, receive graded reports, and must halt development of a contribution until they demonstrate understanding, with escalating consequences for repeated failures. Their analysis of leading open-source AI projects finds more code output with fewer human review comments per line of code, and far lower rates for automated fleet accounts. They recommend that labs run these audits with embedded independent auditors.

arXiv:2610.10064 PDF

Robust Decentralized Fairness Auditing

Sayan Biswas, Jade Garcia Bourr\'ee, Anne-Marie Kermarrec, Palak, Martijn de Vos Fairness audits of large language models (LLMs) usually assume one auditor with a large, representative query set. Several auditors pooling their queries could cover more demographic groups, but any of them might collude with the provider to make the model look fair ("fairwashing"). Auditopus runs the audit in rounds with no central server: auditors exchange only cumulative statistics vectors, and each honest auditor down-weights peers whose vectors are statistically inconsistent with what they sent before. The authors show that a single unprotected adversary can make an unfair model look fair. Against optimized attackers, Auditopus cuts audit error by up to 78% on average compared with no defense and by at least 62% compared with robust aggregation baselines, and with 49% of auditors adversarial it never lets a very or moderately unfair LLM pass as fair.

arXiv:2610.10199 PDF

How to train your model organism

Xilin Wang, David Bau, Byron C. Wallace Model organisms are models deliberately trained to exhibit behaviors such as backdoors or sycophancy, and they are used to test whitebox interpretability tools. The authors argue they should be validated on three objectives rather than one: installing the target behavior, preserving general capabilities, and keeping outputs and activations natural. Re-examining two public organism suites, they find that chat quality and chain-of-thought naturalness degrade substantially, and that these validation metrics predict how well interpretability methods recover the installed behavior. On a new suite targeting demographic bias in clinical reasoning, DPO stays closer to the base model than supervised finetuning, and a multi-objective approach based on model merging better preserves capabilities and naturalness.

arXiv:2610.10203 PDF

LLM Persuasion Is in the Eye of the Evaluation

Kamile Dementaviciute, Julija Vaitonyte, Tijl De Bie Evaluations of persuasion by large language models (LLMs) are fragmented, so the authors adapted nine published automated methods to a shared setup and ran all of them on the same fifteen LLMs. The methods' rankings agree only weakly (mean Spearman ρ = 0.25). Models that refuse some tasks but not others, mostly manipulation tasks, lower agreement by about a quarter. General capability tracks most rational-persuasion scores but not most manipulation scores. The authors conclude that persuasion scores mix a model's ability to persuade with its willingness to do so, so a single score says little about persuasiveness across tasks.

arXiv:2610.10232 PDF

PatchBench: Measuring Collateral Damage in Activation Patching

Alexi Canesse, Mathis Le Bail, Ma\"el Jenny, Cl\'ement Elliker, Mahammed El Sharkawy, Sonia Vanier A jailbreak fix for an LLM can pass a benchmark while still failing on close variants of the harmful prompt, or while over-refusing harmless prompts with similar wording. PatchBench is a curated set of 400 high-confidence jailbreak failures across 8 open-source instruction-tuned models, filtered from 27,870 prompts drawn from 37 public datasets. Its companion protocol, PatchBench-Local, generates nearby harmful variants, structurally matched benign prompts, and benign prompts reusing harmful keywords, so it can tell targeted repairs from broad suppression. Testing four activation-steering methods shows that MMLU capability can stay nearly unchanged while benign behavior near the patched prompts degrades severely.

arXiv:2610.10276 PDF

Revisiting Explainable AI through Model-Independent Concept Dictionaries

Thomas Schnake, Doreen Sch\"oppenthau, Alexander Meyer, Jacques Corbeil, Klaus-Robert M\"uller, Gr\'egoire Montavon Explainable AI (XAI) methods usually assume that input features can be interpreted directly, or they rely on internal model abstractions that are tied to one architecture. DictXAI instead defines concepts in the input domain through a dictionary, a large and possibly overcomplete set of predefined elements that each have a meaning. It computes a sparse code of the input and attributes the model's prediction to the matching dictionary elements. The authors show it can trace failures such as Clever Hans effects back to specific artifact patterns in the data, and that it works with learned image bases, analytic electrocardiography waveforms, and experimentally acquired dictionaries, giving architecture-agnostic explanations that they report as more interpretable than classical and concept-based XAI.

arXiv:2610.10301 PDF

SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic Routing

Hui Zhang, Yachao Yuan, Jiayun Wang, Yuanzhuo Li, Hongtao Wang, Yali Yuan cross-listed When large language models (LLMs) are fine-tuned as a service, malicious fine-tuning data can erode refusal behavior while task performance looks normal. The authors find that layer-wise safety sensitivity is signed, since scaling some layers strengthens refusal and scaling others weakens it. Their defense SLDR trains a LoRA recovery adapter only on the layers with the highest and lowest sensitivity, and a representation-based router activates it only for malicious queries. Across four architectures, five tasks, and four harmful-content benchmarks it preserves downstream utility, and on Llama3.1 with SST2 it cuts the average harmful score from 11.54 to 0.08, staying near zero at poisoning ratios up to 0.9.

arXiv:2610.10345 PDF

Open-MMUnlearning: Unifying Methods and Evaluation for MLLM Unlearning

Junkai Chen, Yuhao He, Qianshan Wei, Junxiang You, Jingwen Shao, Junkai Lin et al. Machine unlearning for multimodal large language models (MLLMs) is hard to compare across papers because implementations and evaluation protocols are fragmented. Open-MMUnlearning is an open-source framework that brings together five privacy, safety, and copyright benchmarks, eight MLLMs from four model families, and twelve unlearning methods. It scores each method on how well it forgets, how much utility it retains, and how robust it is to model interventions, adversarial inputs, and membership inference attacks. In a common comparison of ten methods, GD and MIP-Editor tie for the best overall score. A separate meta-evaluation of thirteen metrics finds BLEU the most reliable overall, while KS-Test is the most faithful but less robust.

arXiv:2610.10358 PDF

Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models

Maverick Morales, Tom\'a\v{s} Dominik, Vermut Gao, Katrina Shirey, Paulius Rimkevi\v{c}ius, Aaron Schurger et al. Chain-of-thought monitoring for deception assumes reasoning traces are available, readable and faithful, and none of these is guaranteed. Drawing on cognitive load theory, the authors test a simpler signal that ignores the trace's content: how many reasoning tokens the model generates. Three reasoning models answered 210 multiple-choice questions under system prompts telling them to answer truthfully, falsely, or without regard for truth. In all three models, truthful responding used fewer reasoning tokens than both lying and truth-indifferent responding. The authors present this as a group-level proof of concept and say it does not yet show that token counts can detect spontaneous deception in individual responses.

arXiv:2610.10405 PDF

6 more specialized papers

Robotics 31

Autonomous Driving Research Requires a Community-Driven Data Paradigm

Jinsu Yoo, Zanming Huang, Katie Z Luo, Zheda Mai, Qiyuan Wu, Bharath Hariharan et al. cross-listed The authors argue that autonomous driving's failure to work robustly anywhere, anytime, and for anyone is mostly a data problem rather than a modeling problem. The community has released over 600 driving datasets across nearly 50 countries, but research focuses on a few dominant benchmarks while the rest go unused because of fragmentation, low visibility, incompatible protocols, and benchmark incentives. The paper proposes principles for a collaborative, community-driven data paradigm that makes diverse datasets easier to discover, reuse, integrate, and evaluate. It also shows an early implementation and calls on academia and industry to collaborate.

arXiv:2610.08825 PDF

Autonomous Droplet Navigation via Model-Based Reinforcement Learning: Zero-Shot Transfer and Emergent Dynamics

Rajneesh Anand, Mayuresh V. Kothare cross-listed Self-driving laboratories struggle to physically manipulate soft, deformable matter such as liquid droplets, whose pinning and surface effects make their dynamics hard to model. The authors build a robotic platform that steers a droplet across an oil-coated tilting board using model-based reinforcement learning and overhead camera feedback, trained on only 50 to 150 physical episodes with no simulator or analytical model. The policy transfers zero-shot to unseen geometries, discovers an oscillating strategy on its own to free stuck droplets, and finishes training in under 90 minutes.

arXiv:2610.08852 PDF

CARE: Certifying Acceleration for Vision-Language-Action Inference

Rui Liu, Tong Zheng, Jindong Gu, Zhipeng Wang Speedups for vision-language-action (VLA) models, such as action chunking and visual-token pruning, can break tasks the original policy would solve, and average success rates hide those failures. CARE runs the reference and accelerated policies from identical initial conditions on a calibration set and selects the fastest accelerator whose acceleration-induced failure risk is certified below a user-specified budget, falling back to the reference policy if none qualifies. On four LIBERO suites with OpenVLA-OFT, it certifies 9.0 to 10.8 times speedups while guaranteeing, at 95% confidence, that at least 85.8% of reference-solved episodes are preserved, and its sequential form uses 78.9% fewer rollouts than exhaustive evaluation. Under tight budgets, selectors without guarantees exceed the budget in up to 75% of trials. The method also extends to flow-step reduction for pi0.5 and to LLM agents in Crafter.

arXiv:2610.08917 PDF

PhysEvo: Astra Can Act, Let It

Wenqing Tian, Zeyu Zhang, Zhaocheng Liu, Fengwei Liu, Qiang Liu, Liang Wang cross-listed PhysEvo is a framework for physical recursive self-improvement (RSI) built around one frozen Astra model. A task agent runs robot tasks, and a meta-agent studies the resulting trajectories to diagnose failures, revise tools and skills, test fixes, and improve its own diagnostic tools, all without updating model weights. On 42 RoboDojo tasks with held-out layouts, it reaches 62.00% success versus 47.17% for the strongest published one-shot Astra agent. On eight tasks that are hard for Astra used directly, it reaches 55.00% success versus 1.25%. A harness evolved in simulation and transferred to an AgileX PiPER arm reaches 84.00% success on five real-world tasks.

arXiv:2610.08995 PDF

PAIR: Bridging Perception and Action in Vision-Language-Action Models

Kaixi Feng, Guoheng Sun, Ang li Continuous-action vision-language-action (VLA) models usually learn the transition from scene understanding to action generation only through the final action loss. PAIR makes this transition explicit. During training, a Masked Action Autoencoder encodes expert action chunks into latent tokens, and a Bridge Module aligns visual-language features with those tokens to produce Bridge Tokens, which give the action expert an action-ready starting point. At inference the autoencoder is dropped and Bridge Tokens come from the observation and instruction alone. PAIR improves OpenVLA-OFT and VLA-Adapter on LIBERO, LIBERO-Plus, and CALVIN ABC-D, and raises OpenVLA-OFT's real-world success rate from 51.4% to 65.0% across seven tasks.

arXiv:2610.09016 PDF

MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking

Shuaijun Liu, Chenglong Zhang, Xuhao Liu, Feiyang You, Yifan Liao, Shuyang Hao et al. cross-listed Motion references reconstructed from human video can look plausible yet fail repeatedly when a physics-simulated humanoid tries to execute them. MimicX keeps the policy in the loop: it uses execution feedback to locate difficult transitions and the affected body regions, then adapts both the tracking objectives and the reset curriculum. Repeated rollout verification decides which refinements to keep. Across four core video tasks, compared with a fixed-reference baseline, it cuts body-tracking error by 25.7% and raises the Robust Execution Horizon by 255.6%.

arXiv:2610.09055 PDF

Workhorse: Learning Robust Whole-Body Humanoid Loco-Manipulation from Human Data

Songbo Hu, Qiayuan Liao, Yufeng Chi, Kevin Zakka, Yakun Sophia Shao, Pieter Abbeel et al. cross-listed Workhorse teaches humanoid robots contact-rich whole-body manipulation from human demonstrations recorded without a robot. A visual planner predicts target poses for the torso, wrists and feet from egocentric RGB, and a reinforcement-learning tracker follows them on the robot. Both are trained on the same human poses without retargeting, and each one's training data is augmented to mimic the errors the other makes at deployment. On a real Unitree G1, the system sorts boxes, catches a thrown box and climbs a suitcase, recovering from pushes. In simulation it completes box sorting in 77% of episodes, and 64% under 40 N·s pushes.

arXiv:2610.09117 PDF

Immiscible Diffusion Policy: Preserving Multimodal Robot Actions through Label-Free Noise Assignment

Xiao Zhang, Yuxin Chen, Zhixuan Liang, Guojian Zhan, Chenran Li, Chenfeng Xu et al. cross-listed Diffusion policies are expected to capture multi-modal action distributions, but in practice they often collapse to a single mode, even when the training data is perfectly balanced. The authors trace part of this failure to independent random pairing of actions with noise, which makes diffusion paths mix and cross so that denoising averages over modes. This effect is especially strong in the dense, low-dimensional action spaces of robot planning. Immiscible Diffusion Policy is a label-free training-time add-on that assigns noise to actions so that routes stay distinct, with no change to the architecture or inference procedure. Across five simulated and two real humanoid manipulation tasks, it raises the share of the non-dominant mode by 6.0x to 14.6x and recovers modes that vanilla policies never produce.

arXiv:2610.09369 PDF

Sparse Feature Policy Unlearning Mitigates State Hallucination in Vision-Language-Action Models

Jiho Lee, Jeongeun Park, Heayoun Choi, Taekyung Kim, Eunwoo Kim cross-listed Vision-Language-Action (VLA) robot models show a recurring failure called state hallucination, where the policy keeps acting as if a robot-object state had been achieved when it was not. Analysis shows these failures coincide with weakened attention to task-relevant image regions, and sparse autoencoders reveal specific sparse features that activate during them. The proposed SOUL (Sparse feature pOlicy UnLearning) uses features from hallucination failures as forgetting targets and features from successful behavior as retention targets. Across VLA architectures in simulation and on real robots, it substantially reduces hallucinated failures and raises task success without materially harming existing manipulation skills.

arXiv:2610.09496 PDF

TERRA: Learning Transportable Latent Actions through Temporal Effect Representation and Relational Alignment

Tianxingjian Ding, Mubarak Shah, Yu Tian cross-listed Latent actions supervise robot policies with action-like codes inferred from video transitions, but it is unclear what a code should keep from a transition and whether it means the same thing in a different starting state. TERRA represents each transition by a compact temporal effect, combining the net feature change with low-order within-window dynamics, and learns a continuous latent from it. Effect-Anchored Transport (EAT) decodes a latent in other initial states and anchors the resulting effect to the one observed at the source, so the latent is defined by what it does across contexts. TERRA predicts actions more accurately than UniVLA and a LAPA-style baseline and degrades more slowly under visual distractors, reaching 93.4% average success on LIBERO versus 91.8% for UniVLA at matched pretraining scale.

arXiv:2610.09509 PDF

Many Ways to Succeed: Diversity-Driven RL Fine-Tuning for VLA Generalization

Haoru Li, Jinmei Liu, Zhiyong Wang, Xiaoming Li, Zhenhong Sun, Daoyi Dong et al. cross-listed Reinforcement learning (RL) fine-tuning of vision-language-action (VLA) policies narrows behavior overall, yet the authors find it diversifies successful trajectories and covers more of the valid solution space than supervised fine-tuning. DRIVE makes that diversity an explicit objective. It groups rollouts under matched task conditions, compares them with temporal alignment, and assigns an intrinsic reward based on relative behavioral diversity, counted only for successful rollouts. Across LIBERO-Plus, ManiSkill3, and RoboTwin 2.0, it improves average out-of-domain performance over vanilla RL fine-tuning by 5.3 points on pi0 and 2.0 points on pi0.5. On a physical dual-arm robot, it raises out-of-domain success from 64.1% to 73.3%.

arXiv:2610.09943 PDF

Temporal Visuo-Tactile Learning for Dexterous Grasp Stability

Ken Nakahara, Aleksei Buvailik, Prokhor Kotov, Roberto Calandra cross-listed Most robotic grasping research relies on vision and parallel-jaw grippers, while humans depend heavily on fingertip touch. The authors collected 10,000 grasp trials on 200 objects with a multi-fingered hand fitted with four Digit 360 tactile sensors, recording vision, proprioception, and touch throughout each grasp. Temporal multimodal models trained on this data predict whether a grasp will hold after lifting, and ablations show that high-resolution, dynamic touch improves those predictions. Used on the real robot to decide when to regrasp, the visuo-tactile predictor raised the success rate of executed lifts by 10.5 percentage points over a gate without touch. The dataset is publicly released.

arXiv:2610.10283 PDF

RoboQuest: Generalist Physical Agents that Search, Inspect and Test

Liu Renhang, Navonil Majumder, Tej Deep Pala, Soujanya Poria cross-listed In unfamiliar environments, generalist robot agents often have to gather missing information through physical interaction before they can finish a task. RoboQuest is a benchmark of ten mobile manipulation tasks that require searching for objects, inspecting hidden properties, or testing unfamiliar tools, and agents must decide for themselves when they have enough evidence to commit. The best of five frontier multimodal agents succeeds in only 23% of episodes, and a fine-tuned π0.5 policy almost never succeeds. Failure analysis shows that most failures are not execution errors: agents tend to stop exploring too early, rarely repair the disturbances their exploration causes, and struggle to learn by trial and error.

arXiv:2610.10388 PDF

EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

Python Song, Zhixuan Liang, Kelsey Fu, Mengdi Wang, Junfeng Yang, Shilong Liu Robot foundation models degrade when object positions or instructions change, and current self-evolving agentic harnesses use costly physical trials inefficiently. EmbodiedRSI keeps competing code and skill hypotheses in a Hypothesis Graph and chooses the physical experiments that best distinguish them using value-of-information selection. A fast-slow dual-system design co-evolves code and skills and builds a hierarchical memory whose entries are selected by their value for later improvements. It reaches 77.0% overall success on RoboCasa365 and 71.3% on Composite-Unseen versus 40.1% for the best baseline, scores 86.8% on LIBERO-Pro, and transfers zero-shot to a real robot.

arXiv:2610.10498 PDF

RoboJEPA: Scaling Robotic Latent World Models

Artem Zholus, Nicolas Beltran-Velez, Jianhao Yuan, Sarath Chandar, Tushar Nagarajan, Daniel Severo et al. Latent world models predict future states and can plan in the real world, but there has been no principled way to predict how their capabilities scale. RoboJEPA is a world model based on the Joint Embedding Predictive Architecture (JEPA), trained on a large dataset spanning 12 robot embodiments. Its imagination error, the error of its rollouts in latent space, follows a second-order power law in compute that predicts model quality well beyond the fitted range. Downstream planning performance also improves predictably with compute and correlates strongly with imagination error, which makes that error a useful proxy for real-robot evaluation. At 8B parameters, the model plans zero-shot toward a goal image on real hardware, and the checkpoints and code are released.

arXiv:2610.10515 PDF

Rephrase Before You Act: Characterizing and Mitigating Language Sensitivity in Vision-Language-Action Models

Mikey Watts (Independent Researcher), Yuchen Cui (University of California, Los Angeles) cross-listed Vision-language-action models (VLAs) are highly sensitive to how instructions are phrased. For example, π0.5 turns on a LIBERO stove 100% of the time for 'switch on the stove' but only 2% of the time for 'switch on the hot plate'. An oracle search over phrasings shows that wording alone nearly closes the 21-point gap between in-distribution and out-of-distribution tasks. The authors score many phrasings of a few training tasks, have an LLM distill the results into ten to twenty rephrasing rules, and rewrite each incoming instruction once under those rules at deployment. Without retraining, this improves a frozen π0 by 16 to 27% relative on twelve held-out tasks, and the approach replicates on π0.5.

arXiv:2610.10526 PDF

Long-WAM: Scaling the Context of World-Action Models

Wei Huang, Bohan Zhang, Chenzhi Liu, Isabella Liu, Shuai Yang, Weian Mao et al. cross-listed Long-WAM scales the visual history available to causal world-action models while keeping robot control real-time. Its central finding is that longer history pays off mainly when the video model is pretrained autoregressively. On RoboCasa GR-1, extending context from 0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained initialization shows no net gain. The model achieves the best results among compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. Streaming observation encoding and asynchronous execution let it run on an RTX 5090 at 107.4 ms per action chunk, as well as on DGX Spark and Jetson AGX Thor, and it reaches 95% success on dynamic cup stacking on a real robot, where Pi0.5 and Fast-WAM succeed in none of 20 trials.

arXiv:2610.10528 PDF

14 more specialized papers

Reinforcement Learning 27

On KL-Regularized Policy Optimization

Yifan Zhang In asynchronous reinforcement learning (RL) for LLM agents, rollouts come from stale checkpoints or an inference engine whose probabilities differ from the trainer's. Existing fixes either clip importance ratios, which biases updates, or sample groups of responses as in GRPO, which is costly for long episodes. KL-Regularized Policy Optimization (KLPO) anchors the KL regularizer at the sampler, so the improvement step has a closed-form solution that can be fit by least squares on the sampler's own trajectories without importance weights. The authors prove the gradients stay unbiased and show that SPPO, GPO, REBEL, and BPO are special cases, giving a critic-free update that needs only one rollout per prompt.

arXiv:2610.08963 PDF

LASER: Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL

Songyuan Zhang, Oswin So, Eric Yang Yu, Matthew Cleaveland, Peter Crowley-Dolen, Chuchu Fan Offline reinforcement learning (RL) risks taking out-of-distribution actions. One recent remedy optimizes a policy inside the latent space of a flow-matching behavior-cloning model, but this latent policy tends to collapse to brittle modes or exploit errors in the critic. LASER adds entropy regularization and uses latent-space adjoint matching, which avoids backpropagation through time. On 40 OGBench tasks, it achieves state-of-the-art results and beats baselines that were tuned per task, while using one fixed set of hyperparameters.

arXiv:2610.08989 PDF

Convex-Concave Reinforcement Learning

Shripad V. Deshmukh, Yaswanth Chittepu, Dhawal Gupta, Philip Thomas, Scott Niekum Maximizing expected return in policy learning is non-convex, so methods such as PPO and TRPO optimize convex surrogate objectives inside trust regions instead. The authors show that, written in log-density-ratio coordinates, the exact per-iteration objective is a difference-of-convex program with difference-of-convex constraints. This view recovers CPI, NPG, TRPO and AWR as special cases, and they solve the program with sequential convex programming under convergence guarantees. The resulting multi-step CCRL method does best on diagnostic tasks that require long-horizon credit assignment, matches a tuned PPO on classic control, and converges faster in a healthcare domain, with an 11.3% higher area under the training curve.

arXiv:2610.09108 PDF

RLDISCOVER: LLM-driven co-evolution of reinforcement learning algorithms

Haoran Li, Zengle Ge, Xiaomin Yuan, Yui Lo, Songlin Zhou, Jiahua Ying et al. Using LLM-guided program evolution to discover better reinforcement learning (RL) algorithms runs into two problems: algorithm components are coupled, and evaluating each candidate requires expensive training with noisy results across random seeds. RLDiscover addresses these with Progressive Co-Evolution, which moves from targeted single-component edits to joint evolution, and Progressive Probabilistic Evaluation, which uses staged training and repeated evaluation. Across SAC, PPO, and DQN on four benchmark suites, evolved algorithms show per-family median return gains of 32–84%, including cases that go from failed learning to task completion. Evaluation uses about one-fifteenth of the compute needed to fully evaluate the same candidates on SAC locomotion. Independent searches repeatedly find interpretable combinations such as adaptive robust losses and progress-dependent value targets, which transfer to unseen tasks.

arXiv:2610.09218 PDF

An Informational Curse of Horizon in Goal-Conditioned Policy Learning

John L. Zhou, Yuxuan Dong, Jonathan C. Kao The difficulty of learning goal-reaching policies is usually blamed on bias accumulating in temporal-difference updates over long horizons. The authors identify a separate "informational" curse: training on goals relabeled from further in the future degrades policies even when they are tested only on nearby subgoals supplied by oracle planners. Goal-conditioned behavioral cloning (BC) suffers severely, while reinforcement learning (RL) objectives mitigate the effect, which the authors explain as a horizon-dependent drop in the conditional mutual information between actions and goals. Policies trained on long horizons also shift their sensitivity from goal information to state information. Distilling the input Jacobians of short-horizon policies into long-horizon ones yields significant gains, especially on combinatorial manipulation tasks.

arXiv:2610.09247 PDF

Ask the Expert: LLM-Guided Reinforcement Learning for Autonomous Cyber Defense

Fernando Martinez, Abhishek Satyam, Tao Li, Junaid Farooq, Ying Wang, Juntao Chen cross-listed Policy-based reinforcement learning (RL) for autonomous cyber defense learns slowly because defenders act on delayed, partial observations with very large action spaces. Putting a large language model (LLM) directly in the loop is too slow and raises trust concerns. Ask the Expert summarizes hard defensive states, occasionally asks an LLM for host-level recommendations through a constrained action interface, and turns those recommendations into tiered reward shaping for PPO. Because the LLM is used only during training, the deployed defender is a pure RL policy with no LLM dependency. Across TTCP CAGE CC1 and CC2 and both attacker types, the method is more sample-efficient than plain PPO and outperforms potential-based reward shaping (PBRS) baselines.

arXiv:2610.09337 PDF

Safe on Average, Unsafe in the Tail: When Is the Episodic-Cost Tail Controllable?

Samuel Tetteh, Cody Fleming Safe reinforcement learning usually constrains expected episodic cost, so a policy can satisfy the budget on average and still be unsafe in its worst episodes. The authors measure the tail with CVaR_0.1, the average cost of the worst 10% of episodes, and call a policy tail-safe when that value stays within budget. Evaluating five standard algorithms on three Safety-Gymnasium navigation tasks, they identify policies that are safe on average but unsafe in the tail. They then test four families of constraints on navigation and locomotion tasks to see when tail violations can be controlled without sacrificing return.

arXiv:2610.09508 PDF

COPC: Coupled Off-Policy Correction for Asynchronous LLM Reinforcement Learning

Zicheng Hu, Zhijian Zhou, Xuan Zhang, Yuchen Liu, Cheng Chen, Yuan Li et al. Asynchronous reinforcement learning (RL) speeds up LLM post-training by separating rollout generation from optimization, but it trains on stale trajectories. The authors show that correcting the token-level policy mismatch is not enough, because advantage estimates are also stale, and they derive bias and variance decompositions showing that the two error sources interact. Coupled Off-Policy Correction (COPC) combines token-level ratio masking with two-sided clipped-ratio weighting of temporal-difference residuals used for return and advantage estimation. It achieves the best reported results on tool-integrated math reasoning and search, stays stable at 64-step policy staleness while most asynchronous baselines collapse late in training, and keeps a 1.7x step-time speedup over synchronous PPO.

arXiv:2610.09597 PDF

CERO: Where and When to Allocate Rollouts for RL Post-Training

Yiming Zong, Yige Wang, Xing Hu, Jiashuo Jiang, Zuo-Jun Max Shen Adaptive rollout methods for group-relative reinforcement learning usually split a fixed budget across prompts at each update. This work instead plans a finite rollout budget over the whole training run. CERO is an online primal-dual scheduler that decides which prompts to admit, how often to revisit them, and how many response groups to generate each round, using reward-variation feedback and a shared budget price, with pathwise guarantees against fixed-rate and time-varying benchmarks. With the same number of training responses, CERO attains the highest avg@16 macro-average on all three backbones across five mathematical reasoning benchmarks, and ablations credit adaptive pacing over uniform or preset spending schedules.

arXiv:2610.09679 PDF

Beyond Policy Support: Interaction Constrained Offline Reinforcement Learning for Autonomous Driving

Mahmoud Selim, Cristina Cipriani, Karl Henrik Johansson Offline reinforcement learning usually guards against distribution shift only in the policy's own action space. In driving, however, an ego trajectory can be well supported in the logged data on its own yet poorly supported given what the surrounding vehicles actually did, a failure the authors call interaction distribution shift. Their Interaction-Constrained Drive Policy (ICDP) shows that the loss of joint support decomposes exactly into an ego-support term and a residual interaction-support term, and estimates the residual with contrastive density-ratio estimation, with no need to predict other agents or roll out a simulator. Closed-loop tests on nuPlan, Interplan, and real-world truck driving show that ICDP avoids high-value but interaction-unsupported trajectories and performs better in interaction-critical scenarios.

arXiv:2610.09763 PDF

BoT-GRPO: Efficient Process-Reward RL for Reasoning via Bag-of-Token Aggregation

Yingxiang Yang, Weihang Xiao, Zhunxuan Wang, Joshua Flashner, Niresh Agarwal In Group Relative Policy Optimization (GRPO), every token in a rollout receives the same advantage, which wastes token-level reward signals when they are available. BoT-GRPO pools all token-level rewards across a group of rollouts, weights each by the inverse of its sequence length, and computes per-token advantages against these weighted group statistics, giving a critic-free, drop-in replacement for GRPO. On React front-end code generation it reaches an 80% compile rate up to 1.9 times faster than GRPO and converges faster than GSPO, DAPO, and PURE while reaching higher final compile and judge win rates, and on AIME math it gains up to 8.1 points of Pass@k in half the steps, across Qwen2.5-3B, SmolLM3-3B, and Phi-4-mini-reasoning. The authors also find that clean, bounded, stable token rewards matter more than richer but noisier ones.

arXiv:2610.09804 PDF

RollVerify: Bridging Efficiency and Accuracy in Long-Tail Rollout Reinforcement Learning

Yongqiang Yao, Jinru Tan, Kaihuan Liang, Zixin Yin, Yazhe Niu, Ruihao Gong et al. In reinforcement learning (RL) for LLM reasoning, rollouts with very long lengths leave GPUs idle. Asynchronous and partial-rollout methods recover throughput but add stale off-policy samples that can hurt accuracy. Rather than reweighting those samples during training, RollVerify defines an off-policy shift metric, OPS, and uses it to verify partial trajectories at the sequence and token level and to truncate invalid suffixes before they enter training. On math and tool-assisted math reasoning, it matches on-policy accuracy while reducing training cost, and preliminary code-generation results suggest it carries over to other domains.

arXiv:2610.09914 PDF

Multi-Agent Coordination via Support-Preserving Distillation

Sangmin Lee, Youngju Na, Chanmi Lee, Sung-eui Yoon Offline multi-agent reinforcement learning (MARL) often distills a centralized generative teacher into decentralized one-step actors under centralized training with decentralized execution (CTDE). The authors find that standard flow-based teachers pair noise with replay targets independently, so the teacher generates samples that fall between valid coordination modes, and distillation passes this error on to the student actors. Mode-Support Semi-Discrete Optimal Transport (MoSDOT) summarizes multimodal replay data into a finite set of modes with prescribed capacities and uses conditional semi-discrete optimal transport to assign each noise sample to a single mode before the teacher is trained. On controlled diagnostics and offline MARL benchmarks, MoSDOT improves endpoint quality and routing consistency, particularly on datasets with multimodal joint behavior.

arXiv:2610.10087 PDF

A Closed-Loop Non-Asymptotic Convergence Analysis of PPO with Learned Critics and Clipping

Junwei Su, Mengfan Liu, Yanyong Zhang, Chuan Wu Proximal Policy Optimization with clipping (PPO-Clip) is widely used but hard to tune, partly because critic learning, ratio clipping, and rollout reuse interact in ways that are poorly understood. The authors give a non-asymptotic analysis of PPO-Clip as a closed-loop actor–critic system. It covers nonsmooth clipping, finite-batch reuse, early stopping, and learned critics trained on Monte Carlo or raw generalized advantage estimation (GAE) targets. A coupling condition puts optimization, critic-tracking, clipping, and batch errors under one amplification bound. For finite layered MDPs, a two-time-scale schedule achieves O(T^(-2/5)) bounds on policy stationarity and critic tracking, and the authors draw tuning guidance from the result.

arXiv:2610.10273 PDF

Sparse Planning in Visual World Models via Cost Gradients

Yingchen Xu, Edward Grefenstette World models that represent images as grids of tokens allow fine-grained planning, but processing every token at each planning step makes action search expensive. COSTGRAD is a training-free method that keeps only the tokens with the largest planning-cost gradients, so it selects what matters for control rather than for prediction. At 50% sparsity on predictors conditioned with adaptive layer normalization (AdaLN), it matches or beats full-token planning on three of four continuous-control benchmarks with a 2.6× wall-clock speedup. Combined with a reduced cross-entropy-method (CEM) search, the total speedup is about 5× while still beating the full-token baseline. On predictors that condition by concatenation instead, COSTGRAD loses its advantage over random token selection.

arXiv:2610.10274 PDF

Continual Graph Multi-Agent Reinforcement Learning

Tommaso Marzi, Ahmed Hendawy, Jan Peters, Carlo D'Eramo, Andrea Cini, Cesare Alippi In continual multi-agent reinforcement learning (CMARL), agents learn cooperative policies across a sequence of tasks, but existing methods ignore the structural differences between tasks, such as different power-grid topologies or formation shapes. The proposed CGMARL framework represents each task as an attributed graph that determines the environment dynamics and the number of agents. The authors release GRAFO, the first benchmark for this setting, and show that forgetting occurs in it. Their method FROG uses a frozen graph encoder to preserve past structural information, and pairing it with existing continual-learning methods substantially improves performance across GRAFO scenarios.

arXiv:2610.10302 PDF

Average-Reward Reinforcement Learning for Multichain MDPs: A Hierarchical Decomposition Approach

Huizhen Yu, Isaiah Heidt In average-reward multichain Markov decision processes (MDPs), the optimal long-run reward can depend on the starting state, which makes standard reinforcement learning (RL) methods unreliable. The authors propose an asynchronous value-iteration RL algorithm that needs only the MDP's transition graph and uses Bather's decomposition to split the state space into communicating subsystems and transient states. They prove it converges to the optimal gain and yields gain-optimal policies in finite time, and they add two variants that target near gain-optimality and near bias-optimality and improve transient performance. The authors describe these as the first essentially model-free average-reward RL algorithms for general multichain MDPs that do not reduce the problem to a discounted one.

arXiv:2610.10326 PDF

Temporally Interpretable Differentiable Decision Trees

Eisuke Hirota, Aarav Sane, Rohan Paleja Differentiable decision trees (DDTs) give reinforcement learning policies a readable tree form, but each decision covers only a single timestep, which does not match how humans plan over multiple steps. The authors introduce temporal interpretability and achieve it through action chunking, with two new policy gradient algorithms for chunked actions and an information-theoretic method that restructures the tree during training. Across four simulation environments, the most effective recipe is to warm-start chunked trees from a distilled chunked policy: these trees match neural network policies in three of four domains while using up to 80% fewer parameters.

arXiv:2610.10367 PDF

Which Rollout Taught It That? BehaviorTrace and the Limits of Training-Data Attribution in Online RL

Amit Nautiyal The study asks whether training-data attribution can find which rollouts taught a language model a behavior during online reinforcement learning with GRPO, and how to tell whether an attribution result is real. The authors release BehaviorTrace, an evaluation harness that plants a behavior with a known cause in Qwen2.5-1.5B and adds controls for gradient magnitude, fluency and run-to-run variation. Much of the apparent attribution signal comes from confounds: ranking training steps by gradient size alone, with no behavior target, reaches 4.2 to 4.5 times chance and matches the best targeted method on two of three seeds. After controlling for fluency, per-rollout results change across seeds and generation draws; only the alignment of trigger-token gradients with the planted behavior holds on all seeds, and the paper turns these lessons into an evaluation checklist.

arXiv:2610.10422 PDF

Q-Learning with Scalar Adjoint Matching

Yonghoon Dong, Minsung Yoon, Jaehyuk Kim, Jungwoo Park, Changyeon Kim, Jinwoo Shin Flow policies can represent rich action distributions, but fine-tuning them with off-policy reinforcement learning is costly because adjoint matching needs a vector-Jacobian product at every flow step. The authors observe that the averaged velocity Jacobian of pretrained flow policies is close to diagonal and derive a closed-form scalar adjoint that scales the final-action value gradient by flow time, which removes the per-step products. Their method, Q-learning with Scalar Adjoint Matching (SQAM), combines this with a penalty on the critic's values at policy-generated actions. On the four hardest OGBench domains it beats the strongest baseline by 18 to 35 percentage points in success rate, and it also improves a vision-language-action policy on a real bimanual robot beyond supervised fine-tuning.

arXiv:2610.10437 PDF

A Good Self-Teacher Meets the Student Where They Are: Joint On-Policy Learning and Teaching

Randy Ardywibowo, Arnav Dalal, Jiantao Jiao Self-distillation addresses sparse outcome rewards by conditioning a policy on privileged information so it can act as its own token-level teacher, but a privileged teacher can take shortcuts the student cannot follow, and its guidance can make the student worse. The authors derive a necessary and sufficient condition for the teacher's distillation update to point the same way as the student's reward gradient, which shows a good teacher must stay close to the student's current behavior, not just perform well. Based on this, JOLT (Joint On-Policy Learning and Teaching) trains one policy in two roles: a privileged teacher optimized with outcome rewards plus a KL penalty toward the student, and an unprivileged student trained by on-policy distillation. It improves training efficiency and final performance on math reasoning, coding, tool use and terminal tasks.

arXiv:2610.10447 PDF

Decoupling Exploration from Optimization in RLVR

Saif Punjwani, Micah Goldblum Adding strong novelty incentives to reinforcement learning with verifiable rewards (RLVR) can degrade a language model, and such damage is hard to undo. ExpDis (Exploration-Distillation) separates the two goals. Explorer policies are trained with a novelty bonus, their trajectories are filtered for correctness and quality, and the survivors are distilled into a separate student policy that is trained without any novelty bonus, alternating over several rounds. Across seven math-reasoning benchmarks and two model families, ExpDis outperforms DAPO at the same wall-clock budget and shows better pass@k scaling, which indicates more diverse correct solutions.

arXiv:2610.10536 PDF

5 more specialized papers

Vision 23

Do Vision Models Learn Physical Constraints or Rendering Shortcuts? A Counterfactual Benchmark for Grounded Physical Consistency

M. Moein Esfahani, Sepehr Salem, Mohammed Alser, Vince Calhoun cross-listed Image editing models can follow a text instruction while still breaking the physics of the scene, for example by adding objects without shadows or leaving them floating above a surface. The authors build a counterfactual benchmark with the Mitsuba 3 renderer: 5,500 images from 500 scene families, each with one clean image and ten matched violations covering shadows, reflection, support, surface response, and occlusion, plus labels, masks, and explanation targets. Adapted vision-language baselines (LLaVA-1.5-7B, Qwen2.5-VL-7B, InternVL3.5-8B) reach 64–68% category macro-F1 on standard held-out scenes. For LLaVA-1.5-7B, however, the score drops from 64.0% to 40.8% under intervention shift, which suggests that in-distribution accuracy partly depends on rendering shortcuts.

arXiv:2610.09205 PDF

Consistent Distribution Matching for Data-Free Diffusion Distillation

Yuxiang Fu, Qi Yan, Zike Wu, Yongxing Zhang, Purang Abolmaesumi, Lele Wang et al. Diffusion and flow models are slow at inference time, and existing distillation methods that speed them up often need curated datasets, costly teacher rollouts, or extra proxy networks. Consistent Distribution Matching is a data-free, simulation-free distillation method in which a single student network handles both sample generation and score estimation. Training therefore needs only a frozen teacher, a trainable student, and one objective. The authors prove that minimizing this objective implies Wasserstein convergence of the student to the teacher's marginals. On ImageNet 256×256 the method reaches an FID of 2.04 in one step and 1.37 in four steps within 40 epochs, outperforming prior data-free distillation baselines.

arXiv:2610.09221 PDF

Closing the Loop on Contrail Avoidance with Satellite Verification

Spandan Ghose Chowdhury cross-listed Rerouting flights to avoid contrails only counts if satellites can confirm the contrails never formed, which is hard because contrails are one to two pixels wide and look like natural cirrus. A small diffusion model (8.4M parameters, trained on one GPU) reaches 0.476 PR-AUC versus 0.414 for a DeepLabV3+ baseline, but simply doubling the baseline's input resolution brings it to parity. A controlled study finds that simple flips and rotations more than double accuracy and matter more than any architectural choice measured. Pretraining on contrail shapes backfires: the model paints thin strokes onto empty scenes, precision collapses to 1%, and recall-based metrics still rate it as excellent.

arXiv:2610.09363 PDF

Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

Hanqiu Li Cai (SperidLabs), Chema Garabito (SperidLabs) cross-listed The authors test whether pixel-space diffusion models, which skip the lossy variational autoencoder (VAE) used by latent models, make better priors for detail-sensitive downstream tasks. They pretrain Iris-3B, a 3B-parameter pixel-space text-to-image transformer, from scratch, and also convert the latent FLUX.2 Klein model to pixel space. They then fine-tune both for monocular depth estimation and super-resolution. They find no significant benefit from the pixel-space prior: neither pixel model beats a latent FLUX.2 Klein fine-tune. Iris-3B still matches Qwen-Image on OneIG at 1024 resolution, and its weights and training code are released.

arXiv:2610.09450 PDF

WAPR: A Foundation Model for Wide-Angle Refinement in Unseen Object Pose Estimation

Yulin Wang, Mengting Hu, Hongli Li, Jianghao Zhou, Chen Luo cross-listed WAPR is a zero-shot pose refinement model for unseen objects that can correct candidate poses that are off by up to 90 degrees in rotation. With only 12 candidates per object it runs in under a second per frame, handling up to 25 object instances per second. Training uses rotational symmetry priors to canonicalize symmetry-equivalent targets, an angle-balanced loss, and the new SA6D dataset of about 50K augmented object instances and 2M rendered RGB-D images. It achieves state-of-the-art unseen-object 6D localization and detection on seven BOP core datasets in both fast and unconstrained settings.

arXiv:2610.09535 PDF

DisParQ: Self-Supervised Part Concepts for Interpretable Vision Foundation Models

Adam Pardyl, Siddhartha Gairola, Sukrut Rao, Adam Wr\'obel, Bartosz Zieli\'nski, Bernt Schiele et al. cross-listed Concept-based vision models explain predictions through human-inspectable concepts, but they usually depend on fixed class labels or on language to define those concepts. DisParQ learns spatially grounded, discrete part concepts from a frozen self-supervised DINOv2 backbone with no class labels and no language supervision: each patch is assigned to one prototype concept, only a few concepts may activate per image, and variation within a concept is captured as quantized discrete attributes. A decoder reconstructs the backbone's features from these concepts and attributes alone, which shows that the discrete code preserves the backbone's information. Across seven datasets, DisParQ nearly matches its teacher on ImageNet linear probing at 83.2% top-1, shows higher concept consistency than language-aligned models, stays competitive on fine-grained benchmarks such as CUB and Cars, and supports part-based retrieval across categories.

arXiv:2610.09802 PDF

ORCA: Hunting Compositional Failures in Text-to-Image Diffusion

Arshia Hemmat, Amirhossein Vahidi, Amitis Shidani, Mohammad Vali Sanian, Hesam Asadollahzadeh, Aryan Yazdan Parast et al. cross-listed Text-to-image diffusion models still bind attributes to the wrong objects, invert spatial relations and miscount objects, even with a T5 text encoder added alongside CLIP. The authors argue that the compositional information is present but misaligned with visual representations. ORCA (Orthogonal Residual Compositional Alignment) adds one auxiliary training loss that aligns the diffusion transformer's latent with a low-rank target from a frozen visual encoder. A learned residual between the T5 and CLIP embeddings selects which visual subspace to use for each prompt. Across DiT-B/2, DiT-L/2 and U-ViT-L, ORCA improves FID and GenEval at no inference cost. On DiT-L/2 it beats the strongest 400K-step baseline after only 200K steps, with the largest gains on binding, spatial and multi-object prompts.

arXiv:2610.09841 PDF

16 more specialized papers

Multimodal 17

JoyAI-Voice 2.0: A Full-Continuous Autoregressive Speech Generation Model with Semantic-Acoustic Joint Representation

Yafeng Chen (Intern), Boya Dong (Intern), Yankun Huang (Intern), Hao Li (Intern), Jingdong Li (Intern), Xiangyu Liang (Intern) et al. cross-listed JoyAI-Voice 2.0 is an end-to-end speech generation model that works entirely on continuous latents. A semantic-acoustic dual encoder splits each speech patch into a semantic representation and an acoustic representation. A causal autoregressive Transformer plans the next patch, and a local diffusion Transformer renders it for 48 kHz audio. The model is trained with flow matching, supervised fine-tuning, and reinforcement learning via DiffusionNFT. It reaches the lowest average word error rate of 2.51% on Seed-TTS, a 14.9% relative reduction over the strongest baseline, along with state-of-the-art attribute fidelity on InstructTTSEval and the top overall score of 0.893 on MDVD-Eval.

arXiv:2610.08834 PDF

RACER: Reflective Agent Coupling Query Interpretation and Tool-Based Retrieval for Frame Selection in Long Video Understanding

Yiyang Huang, Yitian Zhang, Yizhou Wang, Jianglin Lu, Qihua Dong, Hailing Wang et al. cross-listed Answering questions about long videos requires picking the relevant frames from a large pool. The authors identify two failure modes: similarity-based methods misread complex queries, and judgment-based methods struggle to turn their interpretation into frame choices. RACER is a training-free agent in which a lightweight video LLM only rewrites the query into explicit sub-queries, an embedding model acts as a retrieval tool to find matching frames, and the retrieved frames are fed back to refine the sub-queries in a loop. It consistently improves long-video understanding across benchmarks, and even weak components can select frames that help stronger video LLMs.

arXiv:2610.08954 PDF

Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models

Xingang Guo, Jing Gu, Brian Jang, Renxiong Wang, Utkarsh Tyagi, Daniel Quigley et al. Humanity's Sixth Sense (HSS) is a benchmark for the fast, intuitive visual inferences people make at a glance, such as likely causes and future motion, whether a car will fit in a gap, social hierarchy, or unwritten rules. It covers images and videos organized under a taxonomy, with human-written prompts that probe implicit temporal, spatial, social, and abstract structure. People reach 93.1% accuracy, while the strongest multimodal LLM, GPT-6-astra, reaches only 53.6% even at maximum reasoning effort. Agentic setups that manipulate the images narrow the gap but do not close it.

arXiv:2610.08966 PDF

TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning

Ruochi Li, Jianzhe Lin, Haoxuan Zhang, Haihua Chen, Junhua Ding, Edward Gehringer et al. Answering questions over real documents often requires connecting evidence spread across text, tables, figures and captions, but existing evaluations of graph-based retrieval-augmented generation (GraphRAG) are mostly text-centered. TopoGraphRAG-Bench contains 2,024 questions over 201 long, visually rich documents, built from evidence units arranged in single-hop, bridge-chain and multi-source topologies, and validated counterfactually to check that each question resists shortcuts and genuinely needs its stated modalities and evidence. Multimodal GraphRAG systems perform best overall but still fail when visual and textual evidence are poorly aligned or when several evidence units must be combined. Text-only GraphRAG struggles when key dependencies sit in figures or tables, and page-level visual retrieval lacks the fine-grained structure needed to recover the evidence topology.

arXiv:2610.09360 PDF

Shared Geometry As A Rosetta Stone: Cross-Modal Alignment Without Paired Data

Dominik Schnaus, Thomas Dag\`es, Daniel Cremers, Xi Wang, Phillip Isola Aligning embedding spaces from independently trained models of different modalities usually takes large amounts of paired data. Motivated by the Platonic Representation Hypothesis, which holds that such models converge toward a shared geometry, the authors use a Wasserstein Procrustes method with a coarse geometric initialization to estimate a single orthogonal map between two unpaired embedding sets. They show that coarse cross-modal alignment works consistently without any paired examples across datasets, modalities and unimodal models, and that standard geometric alignment metrics predict when it will succeed. With only a few pairs, the method clearly outperforms existing approaches, and the resulting alignments are enough to enable text-to-image generation without paired data.

arXiv:2610.09411 PDF

Mixture of Layers: Dynamic Layer Routing for Visual Reasoning

Jeonghwan Kim, Sofia Stoica, Jiwan Chung, Ansel Blume, Hyeonjeong Ha, Zhenhailong Wang et al. cross-listed Most Multimodal Large Language Models (MLLMs) use only the final or penultimate layer of their vision encoder, which limits access to fine details such as small objects, text and spatial cues. Mixture of Layers (MoL) predicts routing probabilities over vision encoder layers from the text query and combines a sparse top-k selection of intermediate hidden states, at the image level, patch level or both. Across seven fine-grained visual reasoning tasks, it yields gains such as **+18.9% overall accuracy on V***, +4.5% on HRBench4K and +16.3% on CharXiv, without multi-resolution inputs, extra encoders or more patch tokens.

arXiv:2610.09440 PDF

From Pixel to Coding: Evaluating the Figure Reproduction Capabilities of MLLMs

Zijian Chen, Zhengyu Chen, Bohan Liang, Lirong Deng, Yushuo Zheng, Yanwei Jiang et al. cross-listed Benchmarks for Multimodal Large Language Models (MLLMs) usually test visual understanding and code generation separately, and existing visual code generation benchmarks use simple layouts in a single programming environment. FigCodeBench evaluates figure reproduction, where a model must turn a figure image into executable code, with 6,194 instances spanning 7 functional categories, 4 programming languages, and three difficulty tiers. Its multi-dimensional evaluation protocol covers visual fidelity and syntactic isomorphism and is reported to align closely with the Mean Machine Opinion Score (MMOS) and human preferences. Testing 24 proprietary and open-source models, including Gemini 3.1 Pro, GPT-5.4, and Kimi-K2.5, reveals a universal, non-linear performance cliff across languages and difficulty levels, with especially steep drops in rigid declarative languages.

arXiv:2610.10066 PDF

VideoEvolve: Co-Evolving Memory and Retrieval for Long Video Understanding

Yongchao Xu, Bowen Ye, Jiefeng Gan, Junkai Ma, Wenzhao Li, Sen Tao et al. cross-listed Memory-based systems for long video understanding usually adapt how they retrieve information for each question but keep what they store fixed, so details missed at storage time are expensive to recover. VideoEvolve begins with a coarse, low-frame-rate overview and pairs a Memory Evolver, which selectively adds to memory, with a Retrieval Evolver, which searches that growing memory. The two are trained in alternation with agentic reinforcement learning (Agentic RL), updating one while the other is frozen. Bottleneck-Aware Evolution Feedback (BEF) decides whether memory or retrieval is the current weak point and trains that side, while Capability-Aware Evolution Feedback (CEF) keeps memory from over-specializing to the training questions. The authors report effectiveness across multiple long video understanding benchmarks.

arXiv:2610.10183 PDF

TaoD2C-Bench: Benchmarking MLLMs for Industrial UI Code Generation Beyond Visual Fidelity

Chengwei Shi, Yunnong Chen, Tingting Zhou, Qiang Lu, Shiyu Yue, Xinyuan Hu et al. cross-listed Industrial design-to-code means turning user interface designs into code. It requires multimodal large language models (MLLMs) to combine design images with messy layer metadata and to respect the constraints of a target component library. TaoD2C-Bench contains 2,861 production designs from 17 commercial platforms, with 97,652 expert annotations that separate required constraints from permitted implementation choices. It defines three tasks: end-to-end UI code generation, inferring requirements, and realizing them in code. Across eight MLLMs, the benchmark shows large gaps in meeting implementation requirements, and strong visual reconstruction does not imply code that satisfies those requirements.

arXiv:2610.10374 PDF

8 more specialized papers

Reasoning 16

Training Language Models To Be Coherent Decision-Makers

Khurram Yamin, Xavier Fernandes, Paul Koch, Bryan Wilder, Eric Horvitz Good decisions need more than accurate predictions: a model must keep its beliefs consistent, apply the stated utilities, and notice when information needed to choose is missing. Across 20 datasets, the authors first elicit a model's outcome probabilities, then vary only the utilities and how the decision is framed, and fine-tune models to keep those beliefs while picking the action that maximizes expected utility. They also test settings where required utilities are withheld and replaced with irrelevant text. Targeted supervised fine-tuning substantially improves coherent decision-making and often transfers to unseen domains and framings. Trained models learn to recognize when no action can be justified, and a routed system that first checks whether the decision is complete and then executes it performs well.

arXiv:2610.09164 PDF

The Dichotomy Between Pattern Recognition and Step-by-Step Reasoning

Amrut Nadgir, Pratik Chaudhari, Vijay Balasubramanian The authors argue that pattern recognition and step-by-step reasoning are two ends of one spectrum, set by how much preceding context the next token depends on. When the next token depends only on the last c tokens, reasoning traces are paths on a De Bruijn graph, and the authors prove that the number of edges a model must learn is vanishingly small compared with the number of traces, which makes step-by-step reasoning sample-efficient and lets models compose edges to solve longer, unseen tasks. Fine-tuning Qwen2.5-1.5B-Instruct shows that frequent intermediate states give higher accuracy but more fragility, while moderate state density balances the two. Real reasoning data appears to have this structure: Qwen3-14B and Qwen3-32B retain over 75% of their accuracy on GSM8K, MATH-500, and GPQA-Diamond with attention limited to a window under 15% of the trace length.

arXiv:2610.09186 PDF

CM-DPO: Constraint-Margin Direct Preference Optimization for LLM Planning

Rabimba Karanjai, Qun Gu, Hemanth Hegadehalli Madhavarao, Wenhuan Sun, Xiaojiao Yu, Suryabhan Singh Hada et al. Direct Preference Optimization (DPO) gives the same training signal for a small constraint violation and a large one, and it can pick up length and style bias when preference pairs come from different model families. CM-DPO (Constraint-Margin DPO) replaces the binary preference with a continuous margin from a deterministic symbolic verifier, scaled by how severe the violation is. A lexicographic objective ensures hard constraints are never traded against soft preferences. Training pairs come from the SynPlan-R framework, which combines procedurally generated constraint profiles (DCCG) with minimal-edit distillation from a reasoning teacher (RT-MED). On TravelPlanner, NaturalPlan, and out-of-distribution PlanBench, an 8B model reaches an 89.2% pass rate and 93.4% solve rate, matching multi-agent systems at 13x lower latency and beating GPT-4o on unseen Blocksworld by 9.2 points.

arXiv:2610.09219 PDF

Let the Library Speak: Self-Advertised Method Selection for Formal Proving

Xiaopeng Yuan, Suijin Wang, Yanli Wang, Haibo Jin, Peng Kuang, Jerry Wang et al. LLM-based formal provers can retrieve related lemmas, but relatedness does not show whether a mathematical method can actually be applied to the current theorem. The authors propose self-advertisement: before ranking, a model writes a problem-specific proposal for each candidate method, stating what part of the goal it targets, what action it takes and what conditions it requires, and vague or unsupported proposals are demoted. They build a library of 82 reusable methods from Putnam 2000-2014, stored as Method Contracts linked to Mathlib anchors and checked scaffolds. The approach reaches 95.0% hit@5 on Putnam 2015-2025, compared with 84.2% for the strongest embedding-plus-LLM reranker, and 91.7% versus 88.3% on IMO ProofBench.

arXiv:2610.09401 PDF

Efficient Reasoning with Flow Language Models

Hanru Bai, Faissal Izermine, Oscar Davis, T. Konstantin Rusch Flow Language Models (FLMs) are a continuous-state alternative to discrete diffusion language models: they evolve a continuous sequence representation during denoising and convert it to tokens only at the end. A theoretical analysis framed in terms of superposition, together with interventions on intermediate states, shows that this continuous state lets evidence for several candidate answers persist and inform later steps. On maze planning and Sudoku, FLMs reach higher sequence accuracy than discrete diffusion baselines at matched model size when only a few denoising steps are used. On Maze15, an FLM reaches 95% accuracy at 64 denoising steps with 36.5% fewer parameters than MDLM.

arXiv:2610.09416 PDF

Certified by Abstention: Distribution-Free Guarantees for Chain-of-Thought Verifiers at Small Calibration Budgets

Arjun Balaji cross-listed Signals that predict whether a chain-of-thought (CoT) trace is correct are usually compared by AUC, but deploying one requires a threshold that comes with a guarantee. Using seven open models, five verifier signals, and 37,000 graded traces, the authors show that distribution-free certificates at calibration budgets of tens to hundreds of labels are often valid only because they abstain. In simulation, a standard certificate fails in at most 0.3% of calibration draws but in up to 69% of the draws where it actually issues a certificate. They propose a floor-started fixed-sequence certificate that raises coverage from 0.05 to 0.16, and they show that benchmark shift and best-of-n selection against the verifier break the guarantees while the reported failure frequency still looks fine.

arXiv:2610.09541 PDF

Collaborative Reasoning Distillation via Cross-Feedback and Coherent Curation

Taehoon Kim, Seunggeun Cho, Dongsu Han Distilling reasoning into small models with outcome-based rewards cannot tell sound reasoning apart from lucky guesses. Collaborative Reasoning Distillation (CRD) has several teacher models critique each other's reasoning, scores each step for logical validity independently of the final answer, and stitches complementary steps into coherent training traces. Students are then trained with budget-constrained Reasoning Quality Optimization (RQO). The resulting CRD-4B model reaches 97.3% on MATH-500 and 70.3% on AIME'25 using only 50K training examples, up to 12 times fewer than comparable models.

arXiv:2610.09587 PDF

SAPD: Step-Aligned Privileged Distillation

Tianle Wang, Jiayu Liu, Ruizhi Zhao, Ning Miao On-policy post-training improves LLM reasoning but requires costly rollout generation, so the authors ask whether fixed demonstrations can support competitive off-policy learning. Step-Aligned Privileged Distillation (SAPD) is a self-distillation method that needs no rollouts: it uses the known progression of a reference solution to give each reasoning step targeted privileged guidance, producing step-aligned distributional supervision. On math reasoning benchmarks, SAPD beats supervised fine-tuning and label smoothing on average and stays competitive with on-policy reinforcement learning and self-distillation. It also largely preserves out-of-domain coding performance and runs training loops about 2× faster than the on-policy baselines.

arXiv:2610.09665 PDF

From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery

Xinglin Wang, Zishen Liu, Tong Zheng, Shaoxiong Feng, Peiwen Yuan, Yiwei Li et al. Test-time scaling (TTS) methods usually trade accuracy against a single resource, such as cost or latency, while real users may set accuracy, latency, and cost requirements together. The authors define Personalized Test-Time Scaling as finding executable controllers that maximize the rate at which all of a user's requirements are met at once. PersonTTS amortizes this search across users by starting from controllers found for users with similar requirements and by reusing distilled procedural guidance, while still evaluating every candidate on the new profile. On AIME and HMMT, it substantially outperforms strong TTS baselines in joint requirement satisfaction on unseen profiles and problems, and the cross-user reuse lowers discovery-agent time and cost.

arXiv:2610.09684 PDF

Unrolled Flow Models for Reasoning

Faissal Izermine, Hanru Bai, Oscar Davis, T. Konstantin Rusch Flow-matching language models can generate text in a few steps, but it is unclear whether adding integration steps improves their reasoning. The authors prove that a flow parameterized by a two-layer Transformer can solve graph reachability, with the number of steps needed growing with the target's distance from the root. They find that standard training fails to use extra steps because it supervises each time point independently, so they instead train through the model's own latent rollout over random subintervals and decode only at the endpoint. This lifts ProsQA accuracy to 97% and lets accuracy improve with more steps, while projecting the latent state onto a sphere stabilizes longer rollouts on Sudoku and Maze and beats baselines with more than three times as many parameters.

arXiv:2610.09759 PDF

Think Before You Paint: Recursive Latent Reasoning for Diffusion Models

Pawe{\l} Skier\'s, Ma{\l}gorzata Grzanka, Wojciech Masarczyk, Jan-Willem van de Meent, Kamil Deja Diffusion models produce realistic images but fail at visual reasoning, such as filling in a Sudoku or tracing a path through a maze. Recursive models like the Tiny Recursive Model (TRM) solve these puzzles only when given symbolic input. Painter-Thinker (PaTh) uses a small recursive network (the Thinker) that refines a latent state over learned tokens within each denoising step and steers a frozen diffusion model (the Painter) through ControlNet adapters. It is trained only with the standard reconstruction loss, with no symbolic targets, solver or verifier. With 10M parameters, PaTh solves 92.5% of hard and 71.2% of extreme MNIST Sudoku puzzles, versus prior bests of 75% and 4.1%, and it also improves on mazes, Queens and CLEVR spatial-relation scenes. It recovers from injected mistakes that a standard diffusion model cannot fix.

arXiv:2610.09876 PDF

The Long Road to the Same Answer: Cognitive Bias Under Escalating Reasoning Budgets in Large Language Models

Obada Kraishan Dual-process accounts of human cognition suggest that longer deliberation should weaken classic decision biases, so the authors test whether this holds for reasoning models given more inference-time compute. They run a dose-response study on 30 vignettes covering six biases (anchoring, framing, loss aversion, escalation of commitment, availability, confirmation) across four model families, pairing each reasoning model with a non-reasoning sibling and varying the thinking budget, for 12,350 API calls in total, and they measure dose by the reasoning tokens each call actually used. Reasoning models are not less biased than their siblings, and bias does not reliably fall as realized deliberation grows. Anchoring is the only bias that runs in the human direction, while framing reliably reverses. A one-line instruction to restate the anchor lowered anchoring on all five anchoring items, which extra thinking never did, though that effect narrowly misses significance (p = .057).

arXiv:2610.10049 PDF

4 more specialized papers

Unclassified 3

Gauss-Newton Accuracy and Indefinite Hessians: Uniform Coexistence in Low-Cost Sets

Kihun Rhee, Hanjoon Byun No summary available — see the abstract on arXiv.

arXiv:2610.09675 PDF

A Multi-Source Ultrasound Benchmark Revealing the Limits of Contemporary Self-Supervised Anomaly Detection Methods

Marco Riedenauer, Daniel Kienzle, Pratik Mayekar, Rainer Lienhart cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.09677 PDF

Fluctuations of Nonlinear Observables in Mean Field Neural Network Training

Arnaud Descours (UCBL), Geoffrey Lacour (MaIAGE) No summary available — see the abstract on arXiv.

arXiv:2610.09768 PDF