AI Paper Digest

Thursday, October 22, 2026

442 papers cs.AI · cs.LG · cs.CL ← 2026-10-212026-10-23 →

Aug Sep Oct Nov

Highlights

Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds

Highlight HF pick · 16▲Agents Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen et al. Scaling data and model size helps embodied agents in the physical world, and this controlled study tests whether the same holds across virtual game worlds that differ in visuals, controls, dynamics and goals. The authors run 120 continued pre-training (CPT) configurations covering model sizes from 0.8B to 27B, data budgets from 50 to 1,600 hours, and data mixtures from Minecraft-only to many worlds. More data improves performance in worlds seen during training, and diversity helps transfer to related worlds but adds little for distant ones. They also identify an early phase where gains spread across worlds and a later phase where they narrow, and fit an empirical scaling law relating validation loss to data volume, diversity and model size. The resulting models compete with Minecraft specialists and frontier models, but the persistent gaps on distant worlds suggest imitation-based CPT alone will not produce generalist agents.

Whether continued pre-training (CPT) scales for game-playing agents is hard to judge: virtual worlds differ in visuals, physics, controls and goals, and earlier work changed data, diversity and model size all at once. A controlled grid of 120 configurations separates the three factors, using Qwen3.5 backbones from 0.8B to 27B, 50 to 1,600 hours of keyboard-and-mouse gameplay, and five mixtures (S1–S5) that range from Minecraft-only to no Minecraft at all.

  • Training shows two phases: in an early "interface" phase, the first ~50 hours lower validation loss below the base model on every world, including unseen ones; in the later "world" phase, gains are confined to worlds in the training data and closely related ones, and out-of-distribution loss stalls or rises slightly.
  • A Chinchilla-style scaling law extended with a term for mixture share fits validation loss well (R² = 0.982 on Minecraft and 0.995 on games in the training corpus), but it does not apply to out-of-distribution loss, which stops improving after about 50 hours.
  • Diversity mainly helps related 3D worlds: at 1,600 hours, diverse mixtures score about 2.8× higher than Minecraft-only training on OmniGameArena, and larger models need up to 32× less data to reach the same performance.
  • Against Minecraft specialists, GameScaling-27B-S1 reaches 63.3% average success on MCU, about 13 points above the best one; in real-time play, GameScaling-9B-S3 averages 25.3 across the four benchmarks versus 19.2 for the best frontier model, because frontier models lose ~90% of their score on fast games like ViZDoom while waiting on inference.
  • Limitations: transfer to dissimilar games remains weak, with frontier models scoring about twice as high on GameWorld, 27B-S1 losing ~17 points there relative to its base model, and 27B-S3 gaining 16 points on OmniGameArena while dropping 4 on GameWorld; each configuration was trained with a single seed (~45,000 H200 GPU-hours), and the training objective imitates recorded play with no feedback on task success.

arXiv:2610.28698 PDF

Qwen-Audio-3.1-ASR Technical Report

Highlight HF pick · 3▲Multimodal Chuanmeng Bian, Daren Chen, Peixin Chen, Zhigao Chen, Chong Deng, Zhiyun Fan et al. Qwen-Audio-3.1-ASR is a family of audio-language models that improves automatic speech recognition (ASR) over its predecessor and adds speaker-diarized transcription and general audio understanding. It supports 30 languages and 16 Chinese dialects, plus instruction-controlled features such as conditioning on prior context, hotword customization and transcript polishing. Diarized output labels speakers consistently across turns, with timestamps, and represents overlapping speech, while the audio-understanding variants handle captioning, sound event detection, temporal grounding and audio question answering. The family ships separate variants for file transcription, low-latency streaming, interactive voice input and non-streaming use, and is evaluated across standard, multilingual, dialectal, diarized and general audio benchmarks.

Qwen-Audio-3.1-ASR extends the earlier Qwen-Audio-3.0-ASR beyond plain transcription. One encoder–adapter–LLM family now also answers who spoke and when, and what is happening in the audio beyond the words. It keeps support for 30 languages and 16 Chinese dialects, plus instruction controls for historical context, two-tier hotwords and transcript polishing done inside the same decoding pass.

  • The model pairs an audio encoder (pretrained with BEST-RQ, then trained as an encoder–decoder with attention) with a Mixture-of-Experts Qwen decoder; the decoder is pretrained on ~20M hours of audio–text data, then refined on ~1M hours of higher-quality data, where samples are dropped if a second recognizer (VibeVoice) disagrees with the reference by more than 0.3 WER.
  • Speaker-diarized ASR is produced as serialized [SPK|start-end]text segments that may overlap in time, and two training tweaks substantially improve the diarization error rate (DER) and the speaker-aware word error rate (cpWER): speaker-ID tokens get a 10× loss weight, since they are 2.90% of tokens but only 1.97% of the loss, and timestamps are snapped to an 80 ms grid that matches the encoder's 12.5 Hz frame rate.
  • After fine-tuning, reinforcement learning with GRPO runs on an asynchronous framework called FunVerl; its rewards cover WER, keyword and hotword hits, use of context, hallucination penalties, and for diarization DER, cpWER and the time-constrained tcpWER.
  • A final step, multi-teacher on-policy distillation, uses a reverse-KL estimate on responses the model itself generates to merge separate ASR, diarization and audio-understanding teachers into one model, with a Qwen3.5 teacher preserving text and dialogue skills.
  • The capabilities depend on the variant: diarization comes only with the non-streaming Flash and Next models, general audio understanding only with Next, and the Streaming and Message variants support neither; the provided text is cut off before the detailed results, so the only headline benchmarks visible are summary figures on GigaSpeechBench, AISHELL-4, AliMeeting, MMAU and MMSU, alongside several internal test sets.

arXiv:2610.28711 PDF

Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

Highlight HF pick · 5▲Agents Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski Agent4Sci is a long-horizon testbed that measures how far frontier coding agents can improve the published best models on 20 machine-learning-for-science benchmarks across 11 domains. Today's agents beat the published best on every benchmark, with Claude Opus averaging a 10.4% relative improvement and no run flagged for academic misconduct, but the gains come almost entirely from standard ML-engineering tricks rather than domain insight. Analysis of the agents' trajectories shows three recurring weaknesses: misjudging how novel an idea is, under-using available compute, and quickly collapsing into narrow engineering sweeps. Adding lightweight plug-in skills for research ideation, experiment management and scheduled checklists nearly doubles the gains, lifting Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%.

Frontier coding agents are increasingly pitched as automated researchers, and Agent4Sci tests that claim. It is a human-verified testbed of 20 ML-for-science benchmarks (25 tasks, 11 domains) in which an agent starts from the published best model, works in its codebase, and tries to beat it under a fixed compute budget.

  • Each task is first standardized with a verified reproduction of the baseline, fixed splits and a sequestered test set; a separate evaluation agent then scores the submitted artifacts on held-out data, and an auditing agent checks the trajectory for gaming, after a pilot found evaluation gaming such as selective reporting, test-set contamination and hardcoding on 5 of 12 raw tasks.
  • With 8 H200s and 12 hours per run, every benchmark was pushed past its published best by at least one agent, with mean relative improvements of 10.4% for Claude Opus 4.6, 9.6% for Sonnet 4.6, 6.8% for Gemini 3.1 Pro and 4.4% for Codex (GPT-5.4), and no run flagged for misconduct across cross-family re-audits.
  • The gains came almost entirely from standard ML engineering such as learning-rate schedules, wider or deeper models, backbone swaps and augmentation, and trajectory analysis found three recurring failures: agents present standard tricks as novel ideas, leave GPUs idle, and quickly narrow into local hyperparameter sweeps.
  • Three short markdown plug-in skills for research ideation, experiment management and hourly self-check checklists roughly doubled the gains, lifting Opus from 10.4% to 20.1% (better on 18 of 20 benchmarks) and Codex from 4.4% to 8.1%; on PFDelta, Opus moved from 64.4% to 91.0% by training on the physics-derived power-balance loss and refining its predictions with the AC power-flow equations.
  • Results vary from run to run, so the averages are not meant as a leaderboard, and even with skills most benchmarks see only marginal, engineering-driven gains, with some regressions (for example gRNAde, and LITHOS and PFDelta for Codex).

arXiv:2610.28763 PDF

Building AI Research Teams for Autonomous Discovery of Linear Attention Architectures

Highlight HF pick · 11▲Agents Yujian Liu, Li An, Zelin He, Jacob Helwig, Jiabao Ji, Rohit Jain et al. The authors study how to organize teams of autonomous AI researchers to avoid local minima when experiments are costly. Their team consists of persistent agents, each grounded in a different scientific discipline, that keep private research histories, share experimental results, and receive experiment slots through adaptive allocation that balances exploration and exploitation. Deployed to search for linear attention architectures, the team discovers a design that outperforms GDN2 at 1.3B scale on language modeling, common-sense reasoning, and retrieval tasks. Ablations show that removing disciplinary backgrounds, private histories, or adaptive allocation each yields markedly weaker discoveries.

Autonomous research agents often get stuck exploring a narrow part of the design space, especially when experiments are expensive and only a few ideas can be tested. The proposed fix is a team of 16 persistent AI researchers, each given a different scientific background, who keep private research histories, share experimental evidence, and are picked for each new experiment by an adaptive bandit scheduler; the team is used to search for linear attention architectures starting from GDN2.

  • In each of 80 iterations, a discounted UCB rule picks one researcher (all running GPT-6 Astra) whose proposal must pass a technical review and a novelty review. A separate agent implements it, and it must train no slower than 1.2× GDN2 per step. It is then trained FLOP-matched on FineWeb-Edu, and the reward is its gain over the best model so far on a normalized average of language modeling, common-sense reasoning and needle-in-a-haystack retrieval.
  • The best discovery, CFDN (Context-First DeltaNet), adds a recurrent vector alongside GDN2's matrix memory: the vector absorbs the prediction error first, and the matrix then fits what remains. At 150M parameters it reaches a search utility of 1.04 vs. 0.00 for GDN2, with NIAH accuracy of 29.76 vs. 25.79 and WikiText perplexity of 33.84 vs. 34.93, though its common-sense reasoning accuracy is slightly lower (43.29 vs. 43.60).
  • Scaled to 1.3B parameters, CFDN beats GDN, KDA and GDN2 in every category, including common-sense reasoning at 54.53 vs. 52.65 and real-world retrieval at 33.41 vs. 32.52, but some margins are small (NIAH 67.29 vs. 66.86, LAMBADA perplexity 11.04 vs. 11.22).
  • Removing any one organizational component weakens the best result found: utility drops to 0.54 without backgrounds (which stops improving after iteration 22), 0.56 without private memory, and 0.76 when researchers are taken in a fixed rotation. A search that optimizes language modeling alone gets better LAMBADA perplexity (46.40) but does worse on common-sense reasoning and retrieval.
  • Each condition is a single search run with no reported variance, and the 80-iteration budget yielded only 63 trained candidates. During search, retrieval was scored with a teacher-forced proxy rather than actual generation, which leaves open how much of the ablation gap is noise.

arXiv:2610.28887 PDF

SpatialBrain: Spatial Intelligence Growing from Experience

Highlight HF pick · 4▲Multimodal Yifan Zhou, Ziyang Gong, Xiaojun Ye, Yifei Liu, Donggang Li, Xiaolong Zhou et al. Spatial reasoning requires deciding when direct observation is enough and when to reconstruct a scene, take measurements, or switch reference frames. SpatialBrain accumulates experience about these decisions while keeping its vision-language backbone frozen. It combines three parts: 3DFusionMemory, which turns scene observations into queryable spatial relations; a Spatial Skill Library of reusable tool procedures; and a Spatial Cognition module that links task conditions to past experience. Completed trajectories are abstracted into process records between task blocks, and the framework improves over the same backbone's direct answers on four of five spatial-reasoning benchmarks and ties on the fifth.

Spatial reasoning agents have access to reconstruction and geometry tools but no principled way to decide when to use them. The proposed framework keeps a GPT-5.6-sol backbone frozen and builds up external experience about when and how to use memory and tools, kept separate from the evidence about any particular scene.

  • The design has three parts: 3DFusionMemory is rebuilt for each scene and fuses multi-view object, pose, camera and route observations into queryable relations with uncertainty; a Spatial Skill Library holds parameterized tool procedures; and Spatial Cognition maps label-free task signatures to choices of reasoning route, memory access and skills.
  • After each block of tasks, completed trajectories are abstracted into records that keep resource choices, process checks and costs but drop answers and scene coordinates, and these records are consolidated between blocks without any weight updates or reference labels.
  • Compared with direct answers from the same backbone, the framework improves on four of five benchmarks and ties on SAT (91.3), raising the unweighted mean from 67.5 to 72.6, with the largest gain on MindCube (67.2 → 87.6, +20.4 points; errors fall from 344 to 130) and a smaller one on MMSI (53.8 → 57.1).
  • On MindCube, removing a component drops accuracy to 75.2–78.8, but blank answers inflate these gaps: a matched full-versus-no-memory comparison differs by only 0.67 points, and the authors say the earlier MMSI component runs were not verified as complete removals.
  • The self-evolution claim itself is weakly supported, since a fixed-policy control scores 57.4 on MMSI against SpatialBrain's 57.1 and the authors concede the experience curves show accumulation rather than improvement; there are also category-level regressions (OmniSpatial egocentric, SpaceDG distance and counting, MMSI multi-step errors up from 83 to 94).

arXiv:2610.29317 PDF

Memory-Centric Continual Learning for AI Agents

Highlight HF pick · 15▲Agents Dongfang Li, Zhuoen Chen, Bonian Jia, Zixuan Liu, Zhengxuan Lu, Can Deng et al. The authors argue that external memory for foundation-model agents should be judged by whether it improves future actions, not just by what it can retrieve. Their Memory-Centric Continual Learning (MCCL) framework has three parts. The first is a three-axis taxonomy covering where memory state is stored, what form it takes, and how it is managed, with an inner loop that updates memory and an outer loop that refines management policies from task feedback. The second is an evaluation protocol for retention under interference, adaptation to change, and transfer across tasks, built to separate real behavioral gains from simple context expansion. The third is a scaling framework for how agent performance grows with experience, storage, and compute, which together recast agent memory from passive lookup to continual learning.

Agent memory is usually judged by whether it can recall stored facts. MCCL (Memory-Centric Continual Learning) argues it should be judged by whether it improves the agent's future actions. Memory becomes the place where an agent learns during deployment, so the base model's weights can stay frozen.

  • A three-axis taxonomy separates substrates, meaning where memory lives (context, activations, external stores, or adaptive parameters), from assets, meaning what form it takes (episodic, semantic, or procedural), and from policies, meaning how it is managed (prescribed rules, LLM-deliberated decisions, or learned controllers).
  • Continual learning is framed as two coupled loops: an inner loop that forms, consolidates, retrieves, revises, and forgets memory content, and an outer loop that uses delayed task feedback to update the memory-management policy itself.
  • The proposed evaluation uses a four-gate retention cascade (a memory is kept, retrieved, used, and actually changes the outcome) and matched comparisons of agents with accumulated memory against reset agents (ΔP_t), testing retention under interference, adaptation to drift, and cross-task transfer.
  • A memory scaling framework models how performance depends on experience, storage capacity, and compute for both non-parametric and parametric memory, arguing that the two share a resource-limited frontier but differ in how new writes interfere with existing content.
  • This is a conceptual position and survey paper with no new empirical results: its scaling curves are explicitly synthetic illustrations, and it leaves open problems such as credit assignment for memory decisions and policy drift in self-revising agents.

arXiv:2610.29357 PDF

Harness Scaling for Generalizable Agentic Reinforcement Learning

Highlight HF pick · 5▲Agents Fei Bai, Huatong Song, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai et al. Argues that agentic reinforcement learning (RL) through a single fixed agent harness, the scaffolding that governs tool use and context management, leads to models that specialize and transfer poorly to other harness configurations. Harness Scaling builds a factorized space of harness configurations, selects a compact training subset with balanced coverage, and trains across harnesses with controlled variation on the same tasks. Across benchmarks, model scales, and RL algorithms, this improves generalization to unseen configurations. Models trained on structured white-box variation transfer to unseen external black-box harnesses better than models trained on a single complex black-box harness.

Agentic RL usually trains a policy through one fixed harness, meaning the scaffolding of prompts, tools, observations, context handling and workflow around the model, and the resulting policy can overfit to that setup. Harness Scaling treats harness diversity as a new scaling axis: it splits the harness into five components, picks a compact training set of configurations that is balanced across those components and their pairs, and trains each task on several closely related harnesses.

  • Method: Each component (system prompt, tool interface, observation, context management, workflow) has three model-generated variants, giving 3^5 = 243 configurations. A strength-2 orthogonal design picks just 18 of them so that every pair of variants appears exactly twice, and each task is paired with a random anchor harness plus its nearest neighbor by Hamming distance (K=2).
  • Unseen in-space configurations: Trained on ClawGym-Task with Qwen3-4B-Instruct and Qwen3-30B-A3B-Instruct, the method beats single-harness RL with both GRPO and PPO. The 4B GRPO model's overall average across ClawGym-Bench, JobBench-Easy and PinchBench rises from 29.81 to 39.04, and its in-space gain is +9.80.
  • Transfer to unseen black-box harnesses: On Claude Code, models trained on the white-box harness variants average 37.22 (4B) and 41.20 (30B). That beats both supervised fine-tuning (35.08 / 38.08) and RL (33.38 / 36.28) run through the more complex OpenClaw harness, and the same trend holds when OpenClaw itself is the unseen evaluation harness for the white-box-trained models.
  • Ablations: How the training set is chosen matters more than how many harnesses it has. Under the same budget, the structured 18-harness set beats both a random set of 18 and the full 243-configuration space, and switching from nearest-neighbor to random assignment costs 2–3 points. Raising K from 2 to 4 brings no further gain.
  • Limitations: Gains on the base training harness are essentially flat (−0.19) while in-space performance rises +10.48. Absolute scores on JobBench-Easy stay low (roughly 13–19), the harness variants were generated by Opus-4.7 and may not reflect real-world harness diversity, and only Qwen3 backbones were tested.

arXiv:2610.29458 PDF

TraceForge: Distilling with Your Own Thinking

Highlight HF pick · 13▲Large Language Models Yan Wang Some frontier model providers return only answers or short summaries, not full reasoning traces, which limits distillation, and existing recovery tricks break terms of service. TraceForge instead trains a generator that writes a reasoning trace from a context and the teacher's answer in the style of the student's own model family, first with supervised training and then adversarially against a critic that judges how real a trace looks. An independent judge rates the generated traces about as authentic as real ones (0.68 versus 0.66). With the same 5,000 questions and teacher answers, training on the generated traces beats training on Fable 5's own thinking content on coding, by 3.7 points on HumanEval+ and 7.1 on MBPP+.

Some frontier labs now return only answers or short summaries, which hides the detailed reasoning traces that distillation normally trains on. Rather than trying to recover the teacher's hidden reasoning, TraceForge writes a new trace that arrives at the teacher's answer, in the style of the student's own model family; the authors call this a "family-native" trace.

  • A Qwen-family generator is fine-tuned on about 10,000 Qwen3.8-Max conversations to produce the reasoning trace from the system prompt, user turns, tool calls and final answer, while a critic at least as large learns to tell real traces from generated ones; five rounds of adversarial co-training then turn the critic's verdicts into DPO preference pairs.
  • An outside judge (GPT-6.1-SOL) rates generated traces as real with probability 0.68, against 0.66 for genuine family traces and up from 0.54 after supervised training alone, and selecting the critic's best of four candidates raises this to 0.82.
  • With Qwen3.5-9B-Base as the student and Fable 5's answers held fixed, replacing Fable 5's own reasoning with family-native traces improves HumanEval+ by 3.7 points (68.29 vs 64.63) and MBPP+ by 7.1 points (59.52 vs 52.38), and the MBPP+ lead holds at every checkpoint; on the non-reasoning benchmarks (ARC-E, ARC-C, BoolQ), all corpora land within 1.6 points of each other.
  • Training on Qwen3.8-Max's own traces and answers trails by 21 and 35 points on the two code benchmarks, and TraceInversion traces over GPT-6.1-SOL answers collapse into repetitive dead loops and finish at 3.66 and 6.61, whereas family-native traces over the same answers reach 59.15 and 60.05.
  • Each cell is a single run, there is only one 9B student, the data pool is confidential, the abstract (3,500 questions) and setup (5,000 trajectories) disagree, and the judge measures stylistic resemblance rather than correctness; the authors also concede that the comparisons do not separate family style from other pipeline differences, and their explanation that smaller catastrophic forgetting drives the gains is never tested.

arXiv:2610.29680 PDF

Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use

Highlight HF pick · 12▲Agents Yanzhe Chen, Ziyi Yang, Jifeng Zhu, Qiming Huang, Ruihe An, Peiyao Xu et al. This survey asks what changes when multimodal agents move from operating computers to operating robots, and it defines multimodal embodied agents as systems that couple multimodal reasoning with physical action and feedback. It introduces PAPAV, a framework of five capabilities defined by function rather than architecture: Perceive, Anticipate, Plan, Act, and Verify. It uses this framework to compare computer-use agents, robotic systems, and embodied agents under five physical constraints, such as partial observability and outcomes that cannot be verified. A review of 64 benchmarks finds that evaluation centers on acting, with only 5 benchmarks assessing anticipation and 3 assessing verification.

Moving a multimodal agent from computer-use to robot-use changes more than the action interface, and existing surveys tend to describe each setting in its own terms without comparing them. The authors propose PAPAV, a capability-centric framework of five functions defined by their role rather than their architecture: Perceive, Anticipate, Plan, Act, and Verify. It is used to compare multimodal agents, robotic systems, and their intersection, which the authors call multimodal embodied agents (MMEAs).

  • Each capability is a role within a task-level loop, so one model can serve several capabilities and one capability can be spread across several modules; the framework separates Anticipate (predicting what a candidate action will cause) and Verify (judging the outcome against independently specified criteria) from a plain observe-decide-act loop.
  • Five physical constraints explain what changes once an agent acts in the physical world: partial observability, limited simulatability, irreversibility, real-time coupling, and unverifiable outcomes; together they leave fewer choices fixed in advance and less room to undo mistakes.
  • A system counts as an MMEA only if it meets three criteria: it runs a persistent embodied task loop, each claimed capability produces an output that feeds a later task-level decision, and its decisions take physical feasibility, timing, and control limits into account.
  • An analysis of 64 benchmarks finds that explicit evaluation centers on Act, with only 5 assessing Anticipate and 3 assessing Verify, so both capabilities are largely hidden behind overall task success.
  • The authors propose reporting standards such as forecasts with stated applicability limits, two-way execution records between planner and controller, and disclosure of what evidence a verifier shares with the policy it checks, but these are analysis criteria rather than validated tools, and the paper reports no new experiments.

arXiv:2610.30108 PDF

Think Harder Where It Matters: Item-Level Adaptive Recursion for Sequential Recommendation

Highlight HF pick · 4▲Applications Zhanshuo Cao, Jiakai Tang, Yuqi Zhou, Sunhao Dai, Jun Xu Sequential recommenders usually spend the same amount of computation on every item in a user's interaction history. The authors show that most items settle after shallow processing while a small minority need full depth, and that an oracle assigning different depths per item beats any fixed depth. Their method, LoopRec, applies a single weight-shared block recursively, and a lightweight gate decides for each item at every recursion whether to exit or keep processing, without needing depth labels. Across public benchmarks and four backbone families, LoopRec consistently outperforms matched fixed-depth counterparts, and the learned allocation gives deeper computation to lower-frequency items and more recent interactions.

Sequential recommenders usually give every item in a user's history the same number of layers. Diagnostics on a three-layer SASRec suggest this wastes compute: 76.14% of items settle after one layer, 6.72% need all three, and an oracle that knows the target prefers the fixed allocation in only 4.6% of sequences. LoopRec replaces fixed depth with item-level adaptive recursion. It applies one weight-shared block repeatedly, and a lightweight gate decides for each item whether to stop or take another pass.

  • At each recursion boundary, a router scores each still-active item's current state plus a learned recursion-index embedding and makes a hard exit-or-continue decision; only the continuing items go through the shared block again, so items that exit at any depth share one feature space.
  • Training needs no depth labels. A depth-dependent readout weight keeps the router differentiable, and the setup borrows tools from mixture-of-experts training: a load-balancing loss against depth collapse, a router z-loss against premature gate saturation, and a warm-up phase at full depth.
  • Across three Amazon datasets and four backbones (SASRec, BSARec, NextItNet, LRURec), all 48 comparisons improve and 44 are statistically significant, with relative gains of 0.69% to 49.07%; the largest are on the sparsest dataset, Patio, where NextItNet's NDCG@20 rises from 0.0192 to 0.0263.
  • The learned routes are interpretable and match the fixed-depth diagnostics: recent items and rare items get deeper computation (mean depth 2.16 for items seen once in training vs 1.33 for items seen 100+ times), and LoopRec also beats a sequence-level looped baseline with no per-item routing.
  • The gains on strong attention backbones are modest (for example, 2–6% NDCG@20 on SASRec), the evaluation covers only three Amazon categories, and cutting compute to 0.41–0.76× of the fixed-depth model only reduces encoder latency to 0.87–0.995×, while the maximum recursion budget must still be fixed before training.

arXiv:2610.30128 PDF

Applications 84

VHDLBench: Benchmarking and Fine-Tuning Large Language Models for VHDL-Based Chip Design

Patrik Drazic, Benaoumeur Senouci, Boualem Benatallah cross-listed LLMs are rarely applied to generating the VHDL hardware description language. The authors build a dataset of more than 18,500 syntax-verified VHDL examples, collected from open-source repositories and converted from Verilog, and fine-tune five 7B-parameter models with QLoRA so they can run on limited hardware. Evaluation has three levels: syntax and synthesis checks, functional simulation with testbenches, and deployment on Zynq FPGA boards. The fine-tuned models reach up to 92% syntax correctness and 73% functional accuracy, DeepSeek-Coder gains the most, and the generated designs ran on real hardware.

arXiv:2610.28676 PDF

Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books

Valentin Mohl, Alexandre Bismuth, Aramis Marti-Shahandeh, George Nigmatulin, Kang Li, Stefan Zohren et al. cross-listed Autoregressive foundation models can simulate markets by generating the message flow of limit order books (LOBs), but little is known about what they learn internally. Probing an 8.2M-parameter Mamba-3 state-space model trained on tokenized message streams for eight large-cap US stocks, the authors find an internal representation of market state, with order-flow imbalance decodable at R² ≈ 0.59, far above controls and baselines. The representation carries over to unseen stocks and also covers traded volume and volatility. Pushing the model's state along a feature's direction shifts its next-message predictions in proportion to the push, which is consistent with an emergent world model.

arXiv:2610.28753 PDF

CALLIOPE: Private Insights into AI Use, Without Embeddings

Edith Cohen, Vadym Doroshenko, Badih Ghazi, Charlie Harrison, Pritish Kamath, Alexander Knop et al. CALLIOPE builds topic hierarchies over large unstructured corpora, such as logs of LLM conversations, with support for differential privacy (DP). Most existing methods, including recent LLM-augmented ones, ultimately cluster vector embeddings. CALLIOPE instead uses an LLM directly for the whole top-down pipeline. Experiments show better utility than embedding-based methods under DP, especially in the high-privacy regime, measured with standard clustering metrics and new utility metrics the authors introduce.

arXiv:2610.28797 PDF

AutoFraming: Synthesizing Dynamic Frames of C Code with LLMs

Han Zhou, Yu Luo, Dianxiang Xu LLM-based specification synthesis for C code has focused on functional contracts and loop invariants, largely ignoring frame conditions, which state which memory locations a function may or may not modify. The proposed approach extracts frame-relevant information from code property graphs (CPGs) and uses it to guide an LLM toward accurate ANSI/ISO C Specification Language (ACSL) frame specifications that capture region separation and data dependencies. On 84 real-world functions, it produces error-free frame specifications for 92.9% of them, a 179% relative improvement over prior techniques. Plugged into existing ACSL synthesis frameworks, it improves frame-specification generation by up to 61.9%.

arXiv:2610.29046 PDF

VeeReview: Measuring Build-Verify Pairs in Agent-Authored Pull Requests from a V-Model Perspective

Gabriel Dautovi, Jie JW Wu cross-listed VeeReview tests whether the build-and-verify pairs of the V-Model can measure verification evidence in pull requests, and whether they can be used to compare agent-authored with human-authored changes. It classifies four pairs using repository signals and a language-model judge, with one pair decided purely by a rule over the diff and repository tree, and is applied to 1,454 matched agent/human pull-request pairs from 741 repositories. Contrary to the preregistered hypothesis, agent-authored pull requests had lower build-only rates, but the instrument failed its preregistered calibration criterion: agreement with adjudicated labels ranged from 4% to 60%, and inter-rater kappa was 0.154. The authors conclude that LLM-judged process measures partly reflect how a pull request is presented, and that such measures need a mechanically decided component and a validated human reference.

arXiv:2610.29143 PDF

ORSMT: Unleashing the Potential of Satisfiability Modulo Theories for Semantic Verification and Repair in LLM-Based Optimization

Rui Zhang, Chenyu Zhou, Jianghao Lin, Yaoxin Wu, Dongdong Ge, Wen Song cross-listed When LLMs turn natural-language problems into optimization models, checking that the math, the solver code and the reported results agree requires semantic comparison, because equivalent models can look different and inconsistent ones can produce the same objective. ORSMT uses exact satisfiability modulo theories (SMT) queries to align decision variables, project away auxiliary variables, prove or refute that feasible regions and objectives match, localize mismatches, and then reconstruct one artifact from another. Across 3,034 problems and six generator sizes, protected selection raises the model-code match rate from 43.2% to 88.2%, and certified voting beats ReLoop and Opt-Verifier on refuted outputs.

arXiv:2610.29145 PDF

AI-Based End-to-End Automotive Design and Software Development for SDVs

Nenad Petrovic, Alois Knoll cross-listed Building on the Central Car Server (CeCaS) and follow-up EU projects, the authors present a toolchain for Software-Defined Vehicles (SDVs) that combines LLMs with Model-Driven Engineering (MDE). The toolchain turns natural-language requirements into system architectures, executable software and validation artifacts while preserving traceability and safety-critical compliance. It is validated on a centralized vehicle computer running adaptive cruise control and emergency braking in Hardware-in-the-Loop (HiL) and Software-in-the-Loop (SiL) setups. The paper also discusses current limitations and the path toward more automated automotive software development.

arXiv:2610.29160 PDF

Resuming Vision-Language Inference across Onboard Interruptions

Xiaolong Huang, Yinyuan Zhang, Daliang Xu, Yun Ma, Qing Li, Mengwei Xu et al. cross-listed Running vision-language models (VLMs) on commercial satellite payloads is hard because power limits, thermal limits and reboots can cut off inference mid-request, which wipes activations and KV caches. OrbitWeaver checkpoints progress at boundaries chosen per phase (vision encoding, prefill, decoding) and schedules recovery from partially completed states, overlapping checkpoint and weight reads with recomputation. On satellite payload hardware replaying on-orbit telemetry traces, it achieves up to a 4.07-fold end-to-end speedup and a 3.56-fold reduction in SSD writes over the baselines. A 72-hour deployment on two low-Earth-orbit satellites showed it resuming inference across repeated shutdowns forced by the satellite bus.

arXiv:2610.29183 PDF

BRIDGE: An Acceptance Gate to Block False Improvements in Model Updates, with a Retrospective Clinical Evaluation

Dwi Sunaryono A model update can look better on validation data and still fail to improve on new data, which is risky in clinical settings. BRIDGE is a model-agnostic acceptance gate that judges each update on data folds not used to choose it (cross-fitting) and accepts it only when the lower bootstrap bound of the gain exceeds a margin. Across synthetic tasks, public dataset collections, and 12 retrospective clinical tasks, it lowers the false-improvement rate from roughly 12–26% under naive acceptance to about 0–1.7%, removing 97–100% of false improvements. Its mean gain per deployed update is 1.8 to 3.1 times that of the naive rule.

arXiv:2610.29515 PDF

What Prediction-Market Contracts Reveal at Launch: Trading Volume and Dispute Risk

Ahmed Mahrous, Roberto Di Pietro cross-listed Using 802,863 resolved Polymarket markets from January 2023 to May 2026, the authors predict each market's lifetime trading volume and whether its settlement will be disputed, using only information available at launch. Contract information predicts log volume well (out-of-sample R² of 0.747), mostly through interpretable features such as market topic, and neural text models add little beyond those features. That accuracy breaks down across a period of platform change: AUC for predicting whether a market reaches $1,000 drops from 0.940 to 0.577 when models trained on past quarters are tested on later ones, largely because of question formats that were absent from training. Dispute risk can be ranked well above chance (AUC 0.78), but with a 0.31% base rate the precision is too low to forecast individual disputes.

arXiv:2610.29578 PDF

Leveraging Time-Series Foundation Models for Financial Crime Detection

Danae Natalie Savvidi, H. \c{C}a\u{g}r{\i} Bilgi, Kubilay Atasu, Arie van Deursen Graph-based methods for detecting financial crime capture relationships between accounts but make little use of how each account behaves over time. TCEE (Temporal-Context Edge Enrichment) summarizes each account's transaction history with a frozen time-series foundation model (TSFM) and attaches that context to each transaction. The enriched transactions can be scored directly or fed to a graph neural network (GNN). On AMLWorld, it improves PR-AUC by a relative 10.8% on average over the strongest baseline on five of six datasets while training 6x faster. On an Ethereum phishing dataset, TSFM-based account representations fed to a GNN raise PR-AUC by 14.5%.

arXiv:2610.29667 PDF

SAKURA: Serving Attention Key-Value State Under Rate Adaptation for Private Image Editing on Phones

Shuncheng Jie, Desen Sun, Ri Hong, Yuyang Tian, Hong Zhang, Sihang Liu cross-listed On-device image editing with diffusion transformers protects privacy and can reuse precomputed attention key-value (KV) state for common templates such as face swapping. That state is too large to store on phones, and fetching it from the cloud over mobile networks can be slower than recomputing it. SAKURA compresses the cached state with token-importance-aware compression, picks a cache representation per shard based on current bandwidth, and overlaps transfer and reconstruction with on-device execution. On a Snapdragon 8 Elite Gen 5 phone it cuts latency by 1.82x versus an adapted CacheGen and 3.27x versus full on-device generation, and reduces cache storage by 3.06x while preserving image quality.

arXiv:2610.29733 PDF

Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home

Marcos Junca (Senecta Inc.), Mauro Junca (Senecta Inc.) cross-listed The authors test whether a production multimodal LLM, Gemini 3.8 Flash at temperature 0, can serve as the decision stage of a privacy-preserving fall screener for older people living alone. It receives one request per short, low-resolution thermal clip, with sound when available, and its predecessor Gemini 3.7 Flash answers the same requests. On a public 32x24 thermal dataset, the model caught 69% of held-out subjects' falls and raised no alarm on any of 499 non-fall windows. Over 76 hours of ordinary life in an occupied home it made no false calls but caught only 4 of 9 visible staged falls. Removing one cautionary sentence from the prompt raised held-out sensitivity from 69% to 83% and home detection to 9 of 9, at the cost of flagging deliberate lie-downs, while two conventionally trained detectors produced 14.2 and 23.9 false alarms per day in the home.

arXiv:2610.29913 PDF

Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority

M. Mikail Demir The authors test whether legal AI abstains when the correct legal authority is missing, comparing System One decision models, which return a calibrated probability for each option instead of generating text, with six generative LLMs. They extend CaseHOLD so that the correct holding is removed in 1,670 of 5,314 items. When no option is correct, generative LLMs still pick one in 38-86% of items, against 30% for the most cautious decision model. On standard CaseHOLD a decision model matches the best generative LLM (88.1% vs. 88.2%) at 1/55 of the cost; accepting the LLM's answer only when a decision model agrees halves its wrong citations, and distilling decision-model probabilities into Qwen3-8B without labels yields a well-calibrated local model.

arXiv:2610.29961 PDF

FaultGene: LLM-Assisted Fault-Augmented Test Generation for Hardware-in-the-Loop Validation of Autonomous Driving Systems

Chaoran Yuan, Fadi Kurdahi cross-listed Hardware-in-the-loop (HIL) testing of autonomous driving systems runs the driving stack on real embedded hardware inside a simulator, but preparing fault-injection scenarios for it is still largely manual. FaultGene uses an LLM to turn a baseline Scenic/CARLA scenario and a structured fault specification into an executable, fault-augmented test case. During offline refinement, semantic-consistency checks and execution feedback correct candidates that lose the scenario context or fail to produce the intended degradation, and the corrected outputs become fine-tuning data. The authors claim that the fine-tuned generator then produces fault-augmented tests directly at deployment time, with no manually engineered reference executions or repeated refinement loops.

arXiv:2610.29969 PDF

HiddenLens: A Multi-Layer LLM Framework for Detecting and Resolving Hidden Assumptions in Software Requirements

Tanu Singh, Aashna Suman, Vinod Patidar cross-listed Natural-language software requirements often contain unstated assumptions about performance, operating conditions, user behaviour, and domain constraints that look reasonable but become costly when they surface late in development. HiddenLens is a six-layer pipeline that combines BERT and LLMs: it locates unresolved spans via attention entropy, uses chain-of-thought (CoT) prompting to articulate each assumption and its risk, classifies it into one of six non-functional requirement (NFR) categories, proposes ranked clarifications, collects human validation, and rewrites the requirement into a testable form. The authors also release the HiddenLens Corpus, 312 requirements annotated with 847 hidden assumptions. The system achieves an end-to-end F1-score of 0.79, outperforming keyword-based and QuARS-adapted baselines by 71% and 61%, and its BERT NFR classifier reaches 0.847 F1.

arXiv:2610.29996 PDF

Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories

Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao, Martin Klitte, Diarmuid Corcoran cross-listed Large, fast-changing codebases are hard for engineers to understand and hard to feed to an LLM, which needs a method's surrounding repository context as well as the method itself. Code Comprehensor combines a code dependency call graph, CodeScribe (code-aware LLMs that write hierarchical documentation across the call graph), and CodeContrast, a CLIP-style dual encoder trained on that documentation to align text and code embeddings. Retrieved code neighborhoods then ground an MCP-based (Model Context Protocol) LLM orchestrator that answers domain-specific questions. On OpenCV and a proprietary 5M-line repository, richer documentation-derived supervision improves text-to-code retrieval, and RAGAS LLM-as-a-judge evaluation gives above 90% faithfulness.

arXiv:2610.30016 PDF

Orbitol: Crystal Generation on Wyckoff Orbits

Ivan \v{Z}ugec, Joshua Rosenthal, Ivor Lon\v{c}ari\'c, Jan-Willem van de Meent, Tin Had\v{z}i Veljkovi\'c Generative models for crystals usually cannot be told which space group to produce, yet the space group determines whether properties such as piezoelectricity are symmetry-allowed. Orbitol builds on the Crystalite diffusion Transformer but replaces atom tokens with Wyckoff-orbit tokens, so every valid structure it generates is invariant under the requested symmetry group, and it runs two to three orders of magnitude faster than symmetry-aware baselines. On MP-20, 13.8% of attempts yield stable or metastable, unique, novel structures that keep the requested space group after relaxation, compared with 10.8% for the next-best model, SGFM. Among screened piezoelectric candidates, 89% passed density functional theory checks, and their median figure of merit was nine times that of piezoelectric materials computed in the Materials Project.

arXiv:2610.30086 PDF

VX-One: Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift

Juan S. Santillana cross-listed Network perimeter triage must choose between fast signature engines that cannot see new attacks and generative models that are slow and produce unparseable output. VX-One is a small multi-head network (100K to 10M parameters) that outputs typed, calibrated decisions in about 1 ms on a CPU: a binary gate, a closed-set choice that always includes an explicit unknown option, and an ordinal score. It is evaluated on UNSW-NB15, CIC-IDS2017, and EMBER2024 and as a second stage behind Suricata. It wins on detection under temporal drift, while XGBoost wins on multi-class attribution. Its unknown-attack detection works on some benchmarks but drops below chance on others, and calibration collapses under temporal drift for every method tested (expected calibration error rises from 0.04 to 0.40).

arXiv:2610.30090 PDF

65 more specialized papers

Large Language Models 71

First Passage Attention

Yuxuan Liu cross-listed Softmax attention with forget gates adds up decay along the path between a key and a query. Because that sum only grows, such mechanisms cannot erase intermediate context while keeping older tokens that still matter. First Passage Attention (FPA) instead lowers a key's logit according to how far a learned scalar field dips below that key's level over the span in between. It includes the Forgetting Transformer (FoX) as a special case and can represent stack-like sets of surviving keys with a constant head dimension, whereas standard approaches need a dimension that grows with the number of keys. In a 110M-parameter language model, FPA matches FoX in perplexity and achieves perfect passkey retrieval at 8x the training length, even with its key-value cache compressed to about 25%.

arXiv:2610.22714 PDF

To Keep or Not to Keep: Learning KV Cache Retention in Disaggregated LLM Serving Systems

Dong Liu, Yanxuan Yu, Eric Jiang, Shu Wang, Ying Nian Wu In disaggregated LLM serving, prefill and decode run on separate node pools. A KV-cache miss therefore costs a network transfer as well as recomputation, and admitting a block to the global cache costs bandwidth too. KVLearn makes retention cost-aware with three parts: a lightweight Prefix Reuse Predictor that estimates how likely a prefix is to be reused, a Cost-Aware Retention Score that weighs recomputation, transfer, and storage costs, and an Adaptive Threshold Controller that tunes admission thresholds online. On text and multimodal workloads, it cuts time-to-first-token by up to 56% compared with no cache and by 33% compared with Mooncake-style baselines, and reduces inter-node KV transfer volume by up to 53% compared with an LRU pool.

arXiv:2610.28637 PDF

A Small Model Asks Better: Fine-Tuning Query Clarification for Retrieval-Augmented Generation

Lihong He, Lucian Popa, Marina Danilevsky Retrieval-augmented generation (RAG) systems need to recognize when a user's query needs clarification. That includes queries that are ambiguous given the earlier conversation and queries that look well-formed but match several interpretations in the retrieved documents. The authors release ClariEdge, about 3,900 examples that include hard negatives: queries that look ambiguous but can be resolved from context. A lightweight LoRA fine-tune reaches 96.6% accuracy at deciding whether clarification is needed and beats prompted larger models both at that decision and at writing the clarification question. Hard negatives proved essential, and the same adapter works both before and after retrieval in a RAG pipeline.

arXiv:2610.28686 PDF

Big Claims, Small Margins: Do Reported GraphLLM Gains Survive Simple Baselines?

Michal Podstawski LLM-based methods for node classification on text-attributed graphs routinely claim state-of-the-art accuracy, so the authors re-ran cheap, carefully tuned baselines on the exact splits used in three recent studies. Correct & Smooth on TF-IDF features trails the best reported LLM result by a median of only 2.3 percentage points, and an MLP on frozen e5-large-v2 embeddings by 2.8 points, though the LLM methods remain significantly better on 12 of 22 testable tasks. Reading the released code alongside the papers turned up evaluation problems: hyperparameters or checkpoints chosen on the test set, test-node labels leaking into training, undisclosed reuse of baseline numbers, and a text-label misalignment in a released Cora loader. Results reported for the same method on the same split also differ by up to 12.5 points across papers.

arXiv:2610.28714 PDF

SkillEval: Learning Interpretable Ability Profiles of LLMs via Cognitive Diagnosis Models

Berke Arda, Peng Cui, Qiaoyuan Zheng, Rudolf Debelak, Mubashara Akhtar, Mrinmaya Sachan Benchmark scores collapse a model's performance into one number and hide which skills it has or lacks. SkillEval uses LLMs to automatically identify interpretable skills from benchmark questions, then applies cognitive diagnosis models (CDMs) from psychometrics to estimate each model's ability on each skill. Applied to 3,811 LLMs on 9,523 items from MATH, BBH, GPQA, MuSR and IFEval, the resulting profiles are better calibrated and predict performance on unseen items more accurately than average scores. The profiles also enable LLM routing that cuts inference cost by over 60% while improving performance, and computerized adaptive testing that evaluates a new model efficiently.

arXiv:2610.28767 PDF

Building a Nested LADDER for LLM Unstructured Post-Training Pruning

Nicol\'as Mauricio Cuadrado \'Avila, Ivo Gollini Navarrete, Martin Tak\'a\v{c}, Samuel Horv\'ath Post-training pruning of large language models tends to lose quality sharply at high sparsity, even when sparsity is spread unevenly across layers. LADDER (Layer-Wise Allocation via Damage-Driven Evaluation and Refinement) assigns sparsity step by step. At each step it tries pruning a further fraction of weights in every eligible layer, then keeps the layer update that causes the least end-to-end degradation per removed parameter, which yields exactly nested masks across sparsity budgets. Tested with SparseGPT as the backend on five OPT, Llama, and Qwen models, it is comparable or slightly weaker at moderate sparsity but substantially more robust from 70% sparsity onward, and this advantage largely survives distillation on four of the five models.

arXiv:2610.28773 PDF

LIFT: Scoped Constraints for Evaluating and Training Instruction Following

Eran Goldstein, Yuval Belfer, Yonatan Belinkov, Daniel Gissin Instruction-following benchmarks such as IFEval use simple, programmatically checkable constraints that also serve as reinforcement learning rewards, which invites reward hacking at the expense of the main task. Lift-Bench asks models to draft a document from a topic and a structured outline, with verifiable and LLM-judged constraints scoped to specific sections, and it scores both the main task and constraint adherence. Frontier models struggle on its harder subsets, and training on the companion Lift-Train set is much less prone to reward hacking than training on IFTrain. The benchmark, training data, and generation pipeline are all public.

arXiv:2610.28819 PDF

How to Use Parallel Data for Multilingual Pretraining

Zhihong Huang, Yao Lu, Derek F. Wong, Hamid Alinejad-Rokny, Min Yang Standard multilingual pretraining samples documents and their translations independently, so it ignores the parallel structure of machine-translated web corpora. The authors compare parallel pretraining (PPT), which concatenates each document with its translation in the same training sequence, against standard pretraining (SPT), training models from scratch at matched token budgets. PPT alone builds a stronger cross-lingual foundation but does not consistently improve general language understanding, which the authors attribute to a mismatch between parallel training contexts and single-language task inputs. A two-stage recipe of PPT followed by SPT consistently beats standard pretraining on both general understanding and cross-lingual tasks, in high-resource and low-resource settings.

arXiv:2610.28838 PDF

How Much Gradient Does Truncated Backprop Lose in Looped Transformers?

Aur Shalev Merin Looped transformers reuse one block of layers many times and are often trained by backpropagating through only the last k loops; the authors measure exactly how much weight gradient this discards, for every k, across nine looped language models. On trained models the lost fraction is set near the output and barely changes with loop count, and a simple geometric-decay estimate overstates it at long loop counts; on Ouro, where the backward signal grows, keeping two of four loops still loses 81% of the final output's gradient. The lost fraction drops by 0.07 to 0.22 within 300 fine-tuning steps on three of four models, so a trained model's value overstates what truncation removes later. The authors report that they could not show the measurement predicts the training cost of truncation better than measurement-free rules.

arXiv:2610.28871 PDF

SANE: Automated Scorer-Specific Metamorphic Testing for LLM Judges

Ethan Ludwick, Yu-Chung Hsiao LLM judges are now used for benchmarking, model selection, and training, but existing reliability tests rely on human-written perturbation recipes with assumed scorer types. SANE is an agent-based metamorphic testing framework that, given a scorer's specification, implementation, and sample inputs, identifies the capabilities that scorer should have, derives perturbation recipes and expected score changes, and validates the generated tests against scorer intent. Across five domains it produces 3.7 times as many validated test families as the Judge Reliability Harness (JRH) and finds scorer deficiencies in every domain, including two where JRH finds none. On faithfulness, its test scores rank 11 judge models consistently with human-labeled FaithBench performance (Spearman ρ = 0.852), versus 0.100 for JRH.

arXiv:2610.28886 PDF

Combating Catastrophic Forgetting in Continual Domain Adaptation via Knowledge Recasting

Zijie Liu, Zhen Tan, Charles Fleming, Tianlong Chen Most methods against catastrophic forgetting during domain adaptation constrain how far a language model drifts from its previous state. Skill Schema Transport (SST) instead recasts new-domain training examples as instances of abstract skill schemas that keep essential operations, dependencies, and verification steps while dropping domain-specific surface forms, so the model learns new knowledge through procedures it already knows. On two 3B instruction-tuned backbones in continual domain adaptation, SST improves average final performance by 2.53 to 5.75 points over sequential post-training and lowers forgetting in every setting, cutting the maximum drop from 42.41 to 7.10 points on the Qwen-3B MATH stream.

arXiv:2610.28892 PDF

CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer

Mingyu Yang, Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge, Nima Tajbakhsh, Ashwath Aithal, Pavlo Molchanov Distilling knowledge between models with different tokenizers is hard because their next-token distributions are defined over different vocabularies and segmentations, and existing methods rely on shared-vocabulary subsets or pair-specific heuristics. Common-Support Distillation (CSD) builds a shared support at each aligned input chunk, using common-token pairs for one-to-one alignments and a virtual support computed with the autoregressive chain rule otherwise, plus a residual event that keeps both distributions normalized, and matches them with a single objective. Across Llama-3.2-1B and Gemma3-1B-PT students with Qwen and Phi teachers, CSD achieves the best average performance among cross-tokenizer methods, beating the strongest alternative by up to 1.8 points and also surpassing continued pretraining and same-tokenizer distillation. For Llama-3.1-8B it improves over the strongest baseline by up to 4.2 points.

arXiv:2610.28895 PDF

Auditing Conformal Prediction for Language Models Under Inference-Time Model Shift

Achref Doula, Max M\"uhlh\"auser Conformal prediction (CP) for language models is usually evaluated with the same inference settings at calibration and deployment, yet real deployments often change quantization or sampling temperature to meet memory or throughput limits. The authors formalize this as inference-time model shift and show theoretically that it breaks the exchangeability assumption that split-conformal guarantees rely on, even when the task and trained model are unchanged. Across 8 models, 4 benchmarks, and 12 inference configurations, matched evaluation can substantially overstate the coverage actually achieved at deployment, and the effect varies widely by model and dataset. They recommend reporting inference configuration as a standard axis in CP evaluations.

arXiv:2610.28937 PDF

Is 2:4 Sparsity Still Worth It? Semi-Structured Sparsity versus Quantization in LLM Inference

Xiuyi Qin NVIDIA tensor cores support 2:4 semi-structured sparsity, which can be combined with quantization, so practitioners must choose between adding sparsity and dropping to a lower precision. Comparing both at equal bits (counting sparsity metadata) across all six matrix multiplications in a transformer layer on three GPU generations, the authors find that while a lower precision is still available, adding 2:4 sparsity is both less accurate and slower. Once the lowest supported precision is reached, 2:4 still speeds up linear layers by 1.3x to 1.7x, and a retrained 2:4 model served in FP8 with vLLM on an H100 gets 27% higher prefill throughput with GSM8K accuracy on par. Recovering quality requires retraining, and on Blackwell's sparse FP4 the accuracy loss was not recovered within the authors' retraining budget.

arXiv:2610.28947 PDF

Scan or Index? large Evidence Coverage, Answer Quality, and Cost in Long-Document RAG

Juan Terven, Diana-Margarita C\'ordova-Esparza, Julio-Alejandro Romero-Gonz\'alez, Edgar Chavez-Urbiola, Juan Irving Vasquez, Julian Andres Ramirez-Bautista et al. The study asks whether scanning every passage of a long document actually beats index-based retrieval in retrieval-augmented generation (RAG). It compares 11 configurations covering lexical, dense, hybrid-reranked and full-scan approaches on 396 QASPER questions and 80 questions drawn from eight full textbooks, with a fixed answer generator. On books, Hybrid + Jev recovers the most evidence (94.1%, versus 88.7% for a full scan), since later reduction steps in full scans can discard relevant passages. Higher evidence coverage does not reliably produce better answers: Hybrid + Decisions gets the best answer accuracy (0.892) despite recovering less evidence, while full scanning adds substantial latency and reusable dense retrieval is cheaper as queries accumulate.

arXiv:2610.28957 PDF

Target Regeneration Shrinks the Gains of Draft-Training Objectives in Speculative Decoding

Jungseob Lee, Hogun Seong, Euijun Choi, Sugyeong Eo New draft-training objectives for speculative decoding are usually measured against cross-entropy training on externally written responses, but many pipelines have the target model regenerate that data first. Using a vision-language target and a block drafter, the authors compare cross-entropy, token-level distillation, and a verification-gated distillation variant on four versions of the same corpus. Both distillation methods show large gains on external responses, but after target regeneration, distillation's gain shrinks and the gated variant's nearly disappears. The effect holds across data orders, verifiers, temperatures, a second vision corpus, and a text-only replication, so the authors recommend measuring objectives against a target-regenerated baseline and across training lengths.

arXiv:2610.29013 PDF

Our Own Jev: Typed Decisions in One Forward Pass, from an In-House Judge to Open LoRA Models

Yuchuan Weng, Chi Liu, Xin Chen, Srinivasan Manoharan A 'typed decision' returns a probability for each option in a fixed set from a single forward pass, read from a language model's logits or a small head, rather than by generating and parsing text. In the first part of the paper, a self-hosted frozen backbone with a rank-16 LoRA adapter replaces a frontier LLM judge for triaging about 148K agentic coding sessions. It judges 1,000 sessions in about three GPU-minutes at roughly 1% of the frontier judge's cost, and the frontier judge's stored labels turned out to match a fresh consensus on only 26.9% of sessions. In the second part, the public Eikos recipe applied to 1B to 35B backbones yields a 35B model scoring 79.3% on the hard tier of JevBench, statistically on par with the commercial reference, though retraining on a later data release shifted one small model's score by 11.7 points.

arXiv:2610.29052 PDF

Does Every Instance Count Equally? Order Effects in LLM Multi-Instance Processing

Jingxuan Chen, Mohammad Taher Pilehvar, Jose Camacho-Collados When large language models (LLMs) aggregate many independent inputs into one answer, such as labeling a set of reviews and reporting the most frequent label, the result should not depend on the order of those inputs. The authors test this multi-instance processing (MIP) setting on fifteen LLMs, five datasets and seven ordering patterns, keeping each instance set fixed while only the order changes. Accuracy can swing by about 90 percentage points between the best and worst orderings, and models do best when majority-label instances come first. Reasoning traces mention early instances disproportionately, and gradient-based attribution scores are higher at the start and end of the input than in the middle.

arXiv:2610.29057 PDF

A Large-Scale Evaluation of Role-Enhanced Zero-Shot Chain-of-Thought Prompting

Ibrahim Khan, Thai Van Nguyen, Yi Xia, Pratch Suntichaikul, Mustafa Can Gursesli, Daniel Fernandez Galeote et al. Role-Enhanced Zero-Shot Chain-of-Thought (RE-ZS-CoT) is a prompting framework that tells a large language model (LLM) to adopt a domain-expert role before reasoning step by step, without any task examples. Roles are either picked from eight predefined experts (static selection) or chosen by the model itself for each benchmark (dynamic selection). Across 20 LLMs from 2B to 14B parameters and seven arithmetic, commonsense and reading-comprehension benchmarks, the method often beats zero-shot and zero-shot chain-of-thought baselines, with gains of up to 44 percentage points. Static and dynamic selection tend to help on the same benchmarks, static roles reach the highest peaks, and roles that do not fit the task hurt performance.

arXiv:2610.29073 PDF

Same Answer, Different Reasons: Trace-Level Divergence in Multilingual Large Language Models

Rui Sun, Peizhao Mei, Xiewei Cheng, Kaizhuo Chen, Taoran Li, Zhiyuan Yu et al. Multilingual large language models are usually checked for whether they give the same final answer across languages, not whether their visible reasoning is comparable. The authors compare reasoning traces as structured step sequences, scoring structural similarity, order consistency and holistic similarity, with calibration for each language pair. Across three reasoning benchmarks, seven languages and six frontier models, the common shortcut of embedding each whole response and taking cosine similarity overstates agreement: 52.1%–67.3% of examples in the top quartile of holistic similarity are not in the top quartile of structural similarity. Blind bilingual annotators agree with the structural measure, and two English traces for paraphrased questions agree more closely than an English trace and its translated counterparts.

arXiv:2610.29092 PDF

GRAM: Group-wise Rank-Aware Model Merging via Subspace Alignment

Di Wu, Shuaidong Pan Model merging combines several fine-tuned models into one, which is useful when privacy rules mean only the fine-tuned weights are available, but merging everything uniformly loses task-relevant directions when task updates live in misaligned subspaces. The authors show that each task's projection error after merging depends on the principal angles between that task's subspace and the shared one, so grouping compatible tasks provably lowers the error. Group-wise Rank-Aware Model Merging (GRAM) uses this geometry to cluster tasks and then runs a standard SVD-based merge within each group. With just one additional model (K=2), it consistently beats prior merging methods on GLUE with Flan-T5, an 8-task CLIP vision suite, and medical LoRAs on Qwen2-7B-Instruct.

arXiv:2610.29110 PDF

NormTune: Norm-Guided Parameter-Efficient Fine-Tuning for Reasoning-Intensive Tasks

Zhangming Li, Qinghao Hu, Weifan Guan, Yifan Zhang, Peisong Wang, Jian Cheng Parameter-efficient fine-tuning (PEFT) lags well behind full fine-tuning on reasoning-heavy tasks. The authors trace much of this gap to under-adaptation: PEFT produces much smaller weight-update norms than full fine-tuning. NormTune adaptively amplifies the updates of layers whose update norms are too small, and it works with several PEFT methods. Trained on S1K-1.1, it improves rank-8 LoRA by 24.24% average accuracy across AIME 2024, MATH 500, GPQA and AIME 2025, with gains of 11.33% for Bone and 16.21% for MoRA, and rank-256 variants match or slightly exceed S1 full fine-tuning.

arXiv:2610.29134 PDF

GRAFT: Restoring the Missing Sampling Interface in Continuous Diffusion Language Models

Haoning Yang, Lili Pan Continuous diffusion language models such as ELF never sample discrete words at intermediate steps. The authors show that this design collapses the lexical long tail: ELF's rarest words appear at only 0.30x the human rate, a gap that persists from 105M to 652M parameters and that no exposed sampler setting closes. They attribute this to each step averaging over candidate tokens, which pulls rare words toward common competitors. GRAFT is a training-free fix that commits rare words while the model is still undecided and then regenerates the sequence with those commitments fixed, raising the rare-word share from 0.42-0.77x to 0.82-1.10x the human rate across ELF scales and LangFlow.

arXiv:2610.29136 PDF

Why is AI-generated language so hard to understand?

Thomas Hikaru Clark, Matias B. Martin, Kyle Mahowald The authors measure psycholinguistic features known to make text harder for humans to process, such as object relative clauses, long syntactic dependencies and low-frequency n-grams, in a new cross-domain corpus of AI-generated text. All nine frontier and near-frontier models use more inanimate subjects with animate verbs and more rare n-grams than human writers, and most also use longer dependencies and more stacked nouns, with substantial variation across model families. The authors hypothesize that training for math, code and reasoning creates pressure toward information density at the expense of communication. As a proof of concept, they show that fine-tuning a model under a length penalty pushes its language away from human-like patterns.

arXiv:2610.29162 PDF

LTEM: LLM-Friendly Table-Enhanced Markdown for Table Understanding and Generation

Eunbin Kim, Hyeonseok Kang, Sangkeun Jung The lightweight table formats LLMs handle best, such as Markdown and CSV, cannot express merged cells or multi-level headers, while HTML and LaTeX can but cost many tokens and are harder to parse. The authors break table structure into six atomic features with matching complexity levels and build STB (Synthetic Table Benchmark), a controlled set of 1,089 tables. They then propose LTEM, a formally specified minimal extension of Markdown that encodes cell spans and hierarchical headers. Across five LLMs, LTEM matches HTML's understanding accuracy with fewer tokens and produces far fewer invalid responses than LaTeX. For table generation it is comparable to MultiMarkdown, though HTML remains best.

arXiv:2610.29188 PDF

Understanding Target Probability Reuse in Speculative Decoding

Jungseob Lee, Giyoon Jang, Chailin Nam, Jiwoo Kum, Sihyeon Lee, Kyojun Choo et al. In speculative decoding, a large target model verifies a tree of tokens proposed by a smaller draft model in parallel. The authors ask whether the target's next-token distributions can also guide selection of the next tree. They prove that guaranteed exact reuse requires a distribution computed at the next round's full context, and characterize when keeping it costs accepted tokens. Across six target models, reranking first-token branches with a separately computed target distribution improves expected acceptance over draft-based ranking, especially at small tree budgets. Both selection rules also decode faster than draft ranking when the extra target pass overlaps with drafting.

arXiv:2610.29210 PDF

When Multiple Choices Are Correct: Asymmetric Probability Allocation in Jev and Other Decision Models

Rushab Munot Decision models return normalized probabilities over candidate choices, but calibration on single-answer tasks says little about how they behave when several choices are correct. Testing Jev, Liquid d1 and GPT-6 Luna Decisions on MetaTool, HotpotQA and ECtHR, the authors find that even when every correct choice is ranked above every distractor, 28% to 50% of such lists give at least one correct choice less than uniform probability. Given five identical copies of a correct choice, the models assigned them probabilities that differed by 17 to 79 percentage points on average. The probabilities can rank correct choices, but a probability threshold does not recover the set of correct ones.

arXiv:2610.29217 PDF

Hierarchical Masked Diffusion Language Models

Parul Gupta, Satwik, Avni Aggarwal, Bhavesh Gurnani, Arnab Kumar Mondal, Dinesh Khandelwal et al. Masked diffusion language models generate tokens in parallel, but they apply one noise schedule to every position and have no explicit notion of which positions should be resolved first, such as headings before body text or easy Sudoku cells before hard ones. HMDLM (Hierarchical Masked Diffusion Language Model) attaches a hierarchy level to each position from the training data, splits generation into a pass that predicts hierarchy levels and a pass that predicts tokens conditioned on them, and gives each level its own noise schedule. This creates a soft preference for revealing higher-priority positions earlier while keeping decoding fully parallel, and the authors re-derive the forward and reverse processes and the training bound for this setup. It lowers perplexity by over 1.5 points versus ADLM on LM1B and by up to 7 points versus MDLM on WikiHow, and improves Sudoku solution accuracy by up to 8.6 absolute points over diffusion baselines.

arXiv:2610.29255 PDF

Compressing Onboard Language Models for Space Robots under a Behavioral-Fidelity Requirement

Jia Deng, Zizhe Wang, Xin Lu, Yi Huang, Zehao Jia cross-listed Language models running onboard space robots must be compressed to low bit-widths while staying behaviorally close to a full-precision model validated on the ground. QSelect builds a single static low-bit model by assigning each linear module to one of three quantizers (AWQ, GPTQ, or AutoRound). It accepts a substitution only when the KL divergence from the BF16 reference's next-token distribution decreases, which requires no labels. On DeepSeek-R1-Distill-Qwen-7B, Llama-3.1-8B, and Llama-3.2-3B at 3 and 4 bits, the composed models achieve lower held-out perplexity than every single-quantizer model in all six settings. An optional mode that uses a small labeled set adds 4.3–4.5 accuracy points at 3 bits; all experiments use general-domain benchmarks on a ground GPU rather than space tasks or flight hardware.

arXiv:2610.29302 PDF

Seeing the Forest for the Trees: Emergent Agreement from Individual Conversations

Adi Simhi, Yaniv Nikankin, Jonathan Herzig, Idan Szpektor, Yonatan Belinkov, Shay B. Cohen The authors study how two LLMs reach agreement over multi-turn conversations on subjective topics, comparing an internal agreement representation, found with a linear probe on activations, against the agreement the models express. Within any single conversation, the internal signal varies widely and does not reliably track expressed agreement. Pooled across conversations, however, expressed agreement follows a simple Markov process, and aggregated internal agreement maps onto it through a power law. They also find that after a change of stance, models express agreement before their internal representations shift.

arXiv:2610.29328 PDF

Reader--Writer Coupling in Transformer MLPs

Giyeong Oh, Sunghee Ahn, Youngjae Yu Each hidden channel in a transformer MLP has a reader, which decides when the channel fires, and a writer, which decides what it adds back to the residual stream. The authors analyze the two together using the square Jacobian of the block's update, which gives every channel a signed pairing coordinate. Every block they measure has a small core of strongly coupled channels, and in that core almost every writer points against the direction its reader responds to. This opposed core appears in everything from small models trained from scratch to released checkpoints, and it forms within the first few percent of training. Intervention experiments show that training actively holds these pairings in place: a writer moved onto a different reader regrows to fit that reader.

arXiv:2610.29340 PDF

Data Schedules and Learning Rate Schedules Should Be Designed Together

Haodong Wen, Kairong Luo, Tingkai Yan, Wenguang Chen, Kaifeng Lyu LLM training usually switches completely from pretraining data to a smaller, higher-value mid-training dataset near the end. The authors treat this switch as one point in a broader family of data-mixing schedules and ask which schedule minimizes final mid-training loss under a fixed compute and token budget. The best data schedule depends on the learning-rate (LR) schedule: when the LR decays throughout training, mixing mid-training data in earlier beats placing it all at the end, but with a constant LR the result reverses. Ablations show it is better to start decaying the LR only after mid-training data enters, and a theoretical model explains this interaction.

arXiv:2610.29348 PDF

Don't Stop Long-Context Continual Pre-Training Too Early: An Empirical Scaling Law

Ziqing Qiao, Xiaoyue Xu, Sikui Zhang, Yinuo Xu, Bingxiang He, Chaojun Xiao et al. The authors study how the gains from long-context continual pre-training (LCCP) scale with the number of tokens spent on it. They fit an additive power law that relates long-context performance to model size, short-context token budget, and LCCP token budget, using five model scales and more than 100 budget configurations, and the law extrapolates well to held-out scales. They also introduce Constant-Last4, which averages the last four checkpoints of a constant-learning-rate run so a single extensible training run can be monitored stably. A 1.22B-parameter model trained on about 100B LCCP tokens shows no clear saturation, and the gains persist after 128K context extension and supervised fine-tuning, including on agentic tasks.

arXiv:2610.29352 PDF

Right On Q: Recovering Reasoning under Quantization via Self-Generated Trajectories

Boyuan Zheng, Jing Xiong, Chenwei Wu, Hui Shen, Zesen Zhao, Hei Ting Una Chan et al. Low-bit quantized LLMs struggle on long reasoning tasks even after quantization-aware training (QAT), often failing to finish within a token budget that suffices for the full-precision model. The authors trace this to two causes: training on fixed responses rather than the student's own outputs, and the student over- or under-predicting specific next tokens. OnQ addresses both by combining on-policy distillation, where the full-precision teacher supervises the quantized student on the student's own generated prefixes, with a candidate-wise Dynamic KL that balances forward and reverse KL for each candidate token. On Qwen3 models at 1.7B, 4B, and 8B parameters with 2- and 3-bit precision, it gets the best nine-benchmark average and improves Qwen3-8B at 2 bits by 10.26 points over the strongest baseline, while producing shorter responses with fewer truncations.

arXiv:2610.29378 PDF

ECHO: Emergent Common-Mode Hidden Offsets Drive Vocabulary Collapse in Preference Optimization

Yun Li, Ehsan Javanmardi, Simon Thompson, Jiachen Gong, Qunli Zhang, Zifan Zeng et al. Explains why preference optimization can improve ranking metrics while silently heading toward generation collapse. The authors trace the failure to a common-mode offset, a shared shift in hidden states across positions, prompts, and both chosen and rejected responses, which the relative ranking loss barely notices. Through the output layer, this offset forms a vocabulary well in which one token comes to dominate decoding. Injecting the isolated offset into a base model reproduces the collapse, and masking it from the loss prevents the well. The collapse appears in 14 of 17 base-model lineages and in full-parameter SimPO runs on Llama-2 and Llama-3.1, and across 17 UltraFeedback runs, better held-out ranking goes with samples that reward models prefer less often.

arXiv:2610.29410 PDF

Beyond the Vote: When Do LLMs Add Value to Retrieval-Based Tabular Classification?

Ziquan Wang, Yuxuan Nie, Jian Ma, Yingjie Li cross-listed Retrieval-based in-context learning lets a language model classify a table row from similar labeled examples, but it is unclear whether the model does better than simply voting over those same examples. The authors compare four general-purpose language models with distance-weighted nearest-neighbor voting (Distance-KNN) on identical retrieved neighbors across 19 tabular datasets. Distance-KNN achieves the highest mean accuracy at every retrieval pool size. Language-model gains appear mainly when the retrieved neighbors disagree on the label and lie close to the query, and they vanish for the most distant neighborhoods.

arXiv:2610.29520 PDF

RCSD: Reinforcement and Continual Learning for Domain-Specific Speculative Decoding

Mingkang Liu, Weiming Meng, Peipei Jiao, Jiawei Zhang, Tianyuan Liu, Junyan Li et al. Speculative decoding speeds up large language model inference by having a small draft model propose tokens that a larger target model verifies in parallel. Draft models are usually trained by soft-label distillation, which forces them to copy the target's full output distribution, including high-entropy regions that lower acceptance rates. RCSD (Reinforcement and Continual Learning for Speculative Decoding) instead trains the draft model to maximize acceptance length directly. It builds implicit preference pairs from the sequences that are accepted and rejected during normal speculative inference, and masks the loss so tokens already aligned with the target are not penalized. Across several benchmarks it improves average acceptance length by 35% over HASS and runs end-to-end inference 1.35x faster on practical workloads without hurting generation quality.

arXiv:2610.29527 PDF

Stochastic Pruning of the DPO Computation Graph: 1.4x Faster Training and Lower Peak Memory

Yubing Wu, Wei Chen, Junmei Yang, Delu Zeng, Tiewen Pan Direct preference optimization (DPO) spends compute on output scoring, sequence execution, and backpropagation, so supervising fewer tokens does not by itself make training cheaper. The proposed framework randomly selects response positions to score, response prefixes to execute, and preference branches to backpropagate through, and it makes each selection before the corresponding operation so that the skipped work is never computed. Inverse-probability weighting gives a conditionally unbiased gradient estimator for the single-branch case. Position and sequence pruning mainly reduce peak memory, while branch pruning gives a 1.426x step-time speedup and cuts training time by about 30% compared with stock DPO across two model families, at the cost of slightly lower mean accuracy and benchmark-dependent changes in open-ended scores.

arXiv:2610.29538 PDF

PACE: Periodic and Channel-Aligned Encoding of Four-Bit Recurrent States

Patrick Haller, Filipe Laitenberger, Alan Akbik Linear-attention language models keep context in a fixed-size recurrent state with one copy per concurrently decoded sequence, so storing that state at lower precision lets more sequences share an accelerator. Earlier quantizers round the state after every token and lose accuracy at four bits. PACE is a post-hoc four-bit quantizer that rounds only every few tokens and holds the tokens generated in between in a small buffer. Because large state entries cluster in particular columns, it shares one quantization scale per column, and it keeps the heads that retain rounding error longest at eight bits. On four hybrid models from 4B to 35B parameters, mixed four- and eight-bit PACE matches eight-bit DAMP on the three larger models at about two thirds of the memory, and all-four-bit PACE outperforms Q-Mamba at equal memory.

arXiv:2610.29545 PDF

A Diagnostic Study of Faithfulness in Diffusion Language Models for Table-to-Text Generation

Zaynab Merimi, Noussayba El Marrakchi, Arthur Satouf Diffusion language models generate text by repeatedly denoising a masked sequence, so they do not commit to tokens strictly left to right. The authors test whether this keeps them better grounded in structured inputs, using table-to-text generation on ToTTo. They compare the masked diffusion models MDLM and LLaDA, plus a QLoRA-adapted LLaDA, with GPT-2 and GPT-Neo, and score similarity to the reference text separately from grounding in the source table. These two measures diverge: adding demonstrations raises LLaDA's BLEU from 11.3 to 22.0, but entailment against the source table falls from 0.78 to 0.52. Manual annotation shows that autoregressive models tend to drift away from the table, while LLaDA-based models stay on topic but swap individual entities.

arXiv:2610.29602 PDF

SPARE: Sparsity Allocation via Routing Entropy for Efficient MoEs

Yifan Yang, Hanling Yi, Dongquan Yang, Feng Lin, Rong Xiao, Aijun Li et al. Pruning experts shrinks the large memory footprint of Mixture-of-Experts (MoE) models, but most methods remove the same number of experts from every layer. SPARE (Sparsity Allocation via Routing Entropy) is training-free and decides how many experts to cut per layer from the entropy of the routing weights measured on calibration data. Layers where routing is spread evenly lose little when an expert is removed, so they are pruned harder, an intuition the authors support with a perturbation bound. With the same expert selector and budget, it beats uniform allocation on Mixtral, Qwen3, and DeepSeek models, raising Qwen3-30B-A3B's eight-task average from 57.2% to 63.8% at 50% expert pruning.

arXiv:2610.29662 PDF

TraceForge: Distilling with Your Own Thinking

Yan Wang Some frontier model providers return only answers or short summaries, not full reasoning traces, which limits distillation, and existing recovery tricks break terms of service. TraceForge instead trains a generator that writes a reasoning trace from a context and the teacher's answer in the style of the student's own model family, first with supervised training and then adversarially against a critic that judges how real a trace looks. An independent judge rates the generated traces about as authentic as real ones (0.68 versus 0.66). With the same 5,000 questions and teacher answers, training on the generated traces beats training on Fable 5's own thinking content on coding, by 3.7 points on HumanEval+ and 7.1 on MBPP+.

arXiv:2610.29680 PDF

Scaling Multilingual Pretraining of Pixel Language Models

Ilker Kesen, Jenny Kunz, Kasra Malihi, Desmond Elliott Pixel language models read text as rendered images rather than tokens, which helps cross-lingual transfer, but little is known about how they scale. The authors train a family of MPIXEL models that vary the number of languages, the number of scripts, and model capacity, and evaluate them on five tasks covering 83 languages. They find that scaling acts locally: each axis mainly helps the languages and scripts it adds, and larger capacity offsets the curse of multilinguality. The models beat subword and character-level baselines on typologically distant low-resource languages, though transfer to unseen scripts does not improve further with scale.

arXiv:2610.29681 PDF

Retrieval Augmented Generation Compositionality

Ido Levy, Segev Shlomov, Leshem Choshen Retrieval augmented generation (RAG) techniques such as chunking, reranking, query augmentation, and context compression are usually evaluated one at a time. Using LASAGNA, an open-source framework for controlled RAG evaluation, the authors test 450 configurations across eight benchmarks and corpora of 5K to 200K documents against three pre-declared hypotheses. Gains do not compose: three retrieval techniques predicted to give +47% together deliver only +17%. Optimizing retrieval quality can also hurt answer faithfulness, and the best combination on a small corpus may stop being best at larger scale.

arXiv:2610.29722 PDF

Accept or Repair: Preserving Passage-Supported Answers in Contrastive Decoding

Jungseob Lee, Seungyoon Lee, Hyeonseok Moon, Seongtae Hong, Suhyune Son, Sugyeong Eo et al. LLMs often answer from training knowledge even when a provided passage contradicts it. Contrastive decoding addresses this but also alters correct answers that the passage supports. Accept-or-Repair (AoR) is a training-free procedure that keeps a sampled answer if the passage supports it and otherwise tries up to two contrastive-decoding replacements, rejecting ones that simply repeat the model's passage-free answer. It improves over plain sampling by 12.5 points on counterfactual passages and 2.3 points on ordinary ones, beating eight contrastive and prompting baselines.

arXiv:2610.29727 PDF

Metamodel-Guided LLM Decoding for Consistent DSL Generation

Boqi Chen, Aren A. Babikian, Jos\'e Antonio Hern\'andez L\'opez cross-listed Valid output in a domain-specific language (DSL) must satisfy the concrete grammar and also a metamodel's abstract syntax and well-formedness constraints, which existing constrained-decoding methods enforce only partially. MeMoDecode maps the partial output to a partial model each time a complete metamodel concept is generated, checks whether it can still be completed consistently, and on failure resumes from the last completable snapshot. Across four DSLs and six open-weight LLMs from 0.6B to 70B parameters, it improves consistency by 23% on average over unconstrained decoding and by 8% over grammar-constrained decoding, with modest runtime overhead.

arXiv:2610.29749 PDF

ALIAS: Evaluating Performance and Memorization in LLM-Written TLA+ Specifications

Arslan Bisharat, Mujtaba Nazari, Khushboo Bhadauria, Brian Ortiz, Eric Spencer, Beatriz Santos et al. Consistent identifier renaming is a common way to test whether LLMs memorized benchmark answers: a lower score on renamed tasks is read as evidence of recall. The authors test this on TLA+ (Temporal Logic of Actions Plus), using its model checker to confirm that renamed tasks keep exactly the same behavior. On 111 TLA+-Bench tasks, renaming drops Claude Opus 4.8 from 28% to 18% correct, but most of the drop comes from grading that requires exact names, and grading by behavior shrinks it to about 3 points (and to 0 for GPT-5). Fine-tuning experiments with Qwen and Gemma show that a renaming penalty mixes memorization, name-based grading, and the cost of unfamiliar names.

arXiv:2610.29767 PDF

Where Do LLMs Learn to Recommend AI?

Benaya Trabelsi, Jonathan Shaki, Sarit Kraus Large language models (LLMs) often steer users toward AI-related options when asked about study, careers or products, and the authors trace where this tendency comes from by comparing models before and after fine-tuning and by varying question phrasing. Pretrained models already name explicit AI/ML terms in 27% of recommendations, rising to 59% in their released instruction-tuned versions. In their own fine-tuning runs, general instruction tuning leaves the rate unchanged, but removing the 7% of instruction examples about AI cuts it by 14 percentage points, below the pretrained baseline. Helpfulness-focused data raises the rate, data ranked by diverse simulated users lowers it, and phrasings such as asking for the "best" option or what to do "right now" raise it by 8 and 18 points.

arXiv:2610.29785 PDF

S-EDL: Eliciting Self-Evidence from Sequence Likelihoods for Semantic Calibration of LLMs

Yawei Li, Jiazheng Li, David R\"ugamer, Bernd Bischl, Yulan He, Mina Rezaei Calibrating large language models (LLMs) in open-ended generation is hard because many different strings can express the same answer, and existing methods target either fixed multiple-choice options or the correctness of individual strings. S-EDL derives evidence from the model's own sequence likelihoods to build a differentiable Dirichlet prior over prompt-specific semantic classes, then trains with an evidential loss that does not need fixed answer options. Across four open-ended benchmarks and three models, it substantially reduces semantic calibration error while also improving accuracy.

arXiv:2610.29832 PDF

Your Benchmark Is an Empirical Measure Over Difficulty

Yifan Sun, Naicheng Yu, Jiawen Gong, Jingyan Shen, Huan Zhang The authors treat a benchmark as an empirical distribution over item difficulty and use that view to formalize two failure modes. In measure mismatch, a benchmark's difficulty distribution differs from the one its users intend to measure; in coverage gaps, missing stretches of the difficulty axis hide real but localized training progress. They propose two diagnostic metrics with theoretical guarantees and apply them to 25 popular benchmarks across 7 domains using carefully estimated item difficulties. They find that many benchmarks overweight certain difficulty regions, so their leaderboard conclusions warrant caution, while others leave large parts of the difficulty axis uncovered, and they introduce BenchFill, a difficulty-targeted item-rewriting algorithm that reduces these gaps.

arXiv:2610.29927 PDF

Graph Topology Meets Energy-Driven Pathfinding: Preserving Contextual Consistency in Retrieval-Augmented Generation

Yufeng Li, Zaijing Li, Fuchen Zhang, Zongzong Wu, Ming Zhao, Fengxiao Tang Retrieval-augmented generation (RAG) systems that build knowledge graphs for multi-hop question answering lose context when text is distilled into discrete symbols, and they create a mismatch between fragmented graph structures and the continuous text LLMs are trained on. ThreadRAG instead groups retrieved documents into context-intact semantic nodes and links them with narrative, semantic, and lexical edges. Its Energy-Driven Pathfinding step uses energy diffusion to find a query-relevant subgraph and linearizes it into coherent textual evidence paths. On three multi-hop QA benchmarks it gains 14.4% exact match (EM) over text-centric methods on HotpotQA and 5.8% EM over the best baseline on 2WikiMultihop, while using only 23.3% of the strongest baseline's token cost.

arXiv:2610.29933 PDF

CulturalRiddles: A Multicultural Benchmark of Riddles

My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan, Henry Gagnier, Pouya Sadeghi, Farah Abdou et al. Most LLM cultural benchmarks are English-centric and test factual recall rather than applied cultural understanding. CulturalRiddles collects original, human-written riddles through a participatory community effort, covering 62 cultures, 45 countries, and 51 languages; solving them requires abductive and compositional reasoning, social knowledge, language skill, and culture-specific associations. The authors evaluate 28 LLM configurations with both automatic and large-scale human evaluation and find that even frontier models reach at most 22% accuracy for some communities. All artifacts are released with an open process for contributing new languages and cultures.

arXiv:2610.29968 PDF

TIDE: Tree-Induced Fine-Grained Diagnostic Evaluation from Model Behaviors

Jiayi Chen, Hui Li, Yujiong Shen, Qiyuan Peng, Yanxin Li, Chenhao Huang et al. Aggregate benchmark scores hide what a model can and cannot do. Tree-Induced Diagnostic Evaluation (TIDE) recursively builds a hierarchical capability tree from benchmark samples and model scores: at each node an LLM judge proposes candidate splits and assigns samples to groups, and a behavior-based information gain picks the split. Applied to five models on four benchmarks, it surfaces fine-grained contrasts invisible in aggregate scores; for example, Claude Opus 4.6 scores 100% on microeconomic market reasoning versus 83.3% for GPT-5.4, but trails it on contract-law reasoning, 77.8% versus 88.9%.

arXiv:2610.29985 PDF

Looped Uniform Diffusion Language Models

Saiteja Utpala, Chieh-Hsin Lai, Yuki Mitsufuji, Praneeth Vepakomma Uniform diffusion language models (UDLMs) denoise sequences in which any token may be corrupted and later revised. This makes them costlier to train than masked diffusion language models (MDLMs), but they reason better downstream. The authors apply selective looping, which reuses a few early Transformer layers recursively within each denoising step, to UDLMs without changing the objective or sampler. Looped UDLMs reach the validation loss of a non-looped model trained on about 0.5 trillion tokens with up to 2.8x fewer FLOPs on OpenWebText and 3.3x fewer on FineWeb-Edu, and reach up to 9.8% lower generation perplexity at matched function evaluations. Sampling smaller loop counts more often during training cuts inference loops by 12% without hurting accuracy.

arXiv:2610.30066 PDF

Loopholing All the Way: Language Modeling with pure Self-Conditioning

Adrien Deschenaux, Nadezhda Ilieva, Justin Deschenaux, Manuel Madeira, Dorina Thanou, Aur\'elien Lucchi et al. Self-conditioning, also called Loopholing, is usually treated as an auxiliary trick in diffusion language models; the authors ask whether a continuous generative model can be built on it alone. LAW (Loopholing All the Way) carries the expected embedding under the denoiser's prediction from step to step and decodes tokens only once, at the end. It beats prior discrete and continuous diffusion models on Sudoku and is the most accurate on Maze from 2 to 64 function evaluations. After training on TinyGSM, it solves 42.2% of GSM8K problems against 34.5% for uniform Loopholing Discrete Diffusion, and fine-tuning the pretrained Duo model with LAW raises its accuracy from 17.2% to 48.1%; on OpenWebText it approaches autoregressive models at 64 function evaluations.

arXiv:2610.30068 PDF

Circuits of LLM Classifiers

Hadi Khalaf, Sajani Vithana, Kunal Talreja, Haewon Jeong, Viveck R. Cadambe, Flavio Calmon LLM-based classifiers are cheap to build by prompting, but their predictions are nondeterministic and unevenly reliable across samples. The authors treat each LLM classifier as a noisy measurement of the true label and introduce circuits, which arrange queries to different models into an adaptive sequence of queries and decisions that maximizes accuracy at a given average query cost. They characterize the optimal circuit and the fundamental limits on achievable accuracy. Across seven classification tasks in areas including safety, health, human preference, and agent evaluation, circuits improve the cost-accuracy frontier over routing and cascading baselines.

arXiv:2610.30076 PDF

15 more specialized papers

Agents 67

CompLit: Scientific Literature Search Benchmarks Must Cover Implicit, Cumulative, and Unmet Needs

Dayoon Ko, Jihyuk Kim, Soyeong Jeong, Young-jun Lee, Dahyun Lee, Juyeon Kim et al. cross-listed CompLit is a literature-search benchmark covering all eight arXiv domains. Besides standard queries, it tests three needs that researchers actually have: requirements the target paper shows only implicitly, requirements that build up over a conversation, and requests that no existing paper satisfies. The strongest agents, Codex (GPT-5.6-Sol) and Claude Code (Opus 4.8), answer 86-87% of standard queries but drop to 66-68% when requirements are implicit. They lose a further 6 and 15 points respectively when requirements accumulate, and abstain on at most 53% of the queries that have no matching paper. Traces of the agents' searches show they often find the right paper but misjudge it.

arXiv:2610.28658 PDF

Who Verifies the Builder? Toward Independent Assurance for Agent-Produced Software

Sion Israel Sion, Alain April cross-listed When the same coding agent writes software and also the tests that judge it, a passing test suite stops being independent evidence of correctness. The tests can become an optimization target or share the code's misreading of the specification. The authors propose an assurance plane separate from the build. It has a Verification Contract stating what must be shown before release, a phase that verifies against the specification without seeing the code, and a phase that inspects the implementation using mutation analysis, property-based testing, and fuzzing, with a human making the final release call. In a pre-registered study on three stateful services, a verifier running in a fresh context caught significantly more of 46 seeded faults than the builder's own tests (Holm-corrected p = 0.006), while a verifier from a different model provider caught no additional faults.

arXiv:2610.28667 PDF

VerifierBench: Benchmarking the Generation of Executable Verifiers for LLM Agents

Jianing Yin, Ziyi Chen, Hao Wang, Yiyou Sun, Zijian Jin, Dawn Song et al. cross-listed VerifierBench measures how well agents write executable verifiers, meaning programs that should accept every valid solution to a task and reject invalid ones. It contains 222 tasks drawn from six agent benchmarks, covering tool-call, state-based, and code verification, with 4,993 hidden test cases that include alternative valid solutions and attacks. Across 19 models, normalized balanced accuracy ranges from 10.6 to 62.1. Generated verifiers reject invalid solutions more reliably than they accept valid ones, but much of that apparent strength comes from being too strict: most false rejections come from verifiers that accept no tested solution at all. These errors survive self-testing because agents test with solutions that share the verifier's mistaken assumptions.

arXiv:2610.28679 PDF

Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds

Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen et al. Scaling data and model size helps embodied agents in the physical world, and this controlled study tests whether the same holds across virtual game worlds that differ in visuals, controls, dynamics and goals. The authors run 120 continued pre-training (CPT) configurations covering model sizes from 0.8B to 27B, data budgets from 50 to 1,600 hours, and data mixtures from Minecraft-only to many worlds. More data improves performance in worlds seen during training, and diversity helps transfer to related worlds but adds little for distant ones. They also identify an early phase where gains spread across worlds and a later phase where they narrow, and fit an empirical scaling law relating validation loss to data volume, diversity and model size. The resulting models compete with Minecraft specialists and frontier models, but the persistent gaps on distant worlds suggest imitation-based CPT alone will not produce generalist agents.

arXiv:2610.28698 PDF

SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery

Ha Minh Hieu, Tongyao Zhu, Do Xuan Long, Min-Yen Kan Evolutionary search driven by large language models is a promising route to scientific discovery, but it is costly because expensive reasoning is applied to every step and lessons from failed or redundant trials are underused. SpecEvo splits the work by cost: cheap models act as parallel Speculators that generate and refine candidate programs, plus an Advisor that turns accumulated successes, failures and dead ends into guidance. A frontier-model Navigator steps in only at sparse checkpoints, choosing to synthesize, refine or reframe the search depending on how far it has stalled. Across 176 tasks in four discovery domains, it matches or beats frontier-model baselines while cutting API cost by up to 80%, and it improves on the strongest CO-Bench baseline by 3.8%.

arXiv:2610.28706 PDF

Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski Agent4Sci is a long-horizon testbed that measures how far frontier coding agents can improve the published best models on 20 machine-learning-for-science benchmarks across 11 domains. Today's agents beat the published best on every benchmark, with Claude Opus averaging a 10.4% relative improvement and no run flagged for academic misconduct, but the gains come almost entirely from standard ML-engineering tricks rather than domain insight. Analysis of the agents' trajectories shows three recurring weaknesses: misjudging how novel an idea is, under-using available compute, and quickly collapsing into narrow engineering sweeps. Adding lightweight plug-in skills for research ideation, experiment management and scheduled checklists nearly doubles the gains, lifting Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%.

arXiv:2610.28763 PDF

Measure First, Then Allocate: Calibrated Rollout Budgets for MCP Agents in Reinforcement Learning

Hieu Tran, Zonghai Yao, Sunjae Kwon, Hong Yu When tool-using language models are trained with group-relative reinforcement learning, rollout groups in which every attempt gets the same reward provide no learning signal and waste expensive interaction. Working in Model Context Protocol (MCP) environments, the authors use reward disagreement in two ways: Discord compares harness edits using only paired outcomes that disagree, with early stopping, and BRACE (Bank-calibrated Rollout Allocation under Correlated Episodes) uses historical task data and uncertainty estimates to predict which tasks will produce mixed rewards before spending generation budget on them. Across three Qwen3-VL model scales, BRACE beats the strongest baseline's peak performance with fewer new training rollouts, though collecting the historical bank is counted separately, and it also leads the compared methods on BFCL and NESTFUL.

arXiv:2610.28785 PDF

Is Anyone Enforcing the Cartel? What Reinforcement Learning Teaches LLM Agents in Markets

Hieu Tran, Hong Yu Prior evaluations report that LLM pricing agents collude, but a real cartel only holds if a member who cheats gets punished. The authors introduce an audit that forces one agent to cheat in counterfactual copies of a market, and use it to examine what reinforcement learning (RL) teaches small LLMs in pricing and trading games. Agents learn only behaviors they readily try: against a cheating rival they match its price cuts but rarely discover the costly price war that would deter it. Self-play produces fixed rules that ignore rivals, and prompted LLMs reach cartel-like prices mostly through such rules. The central claim is that outcome-based measures cannot distinguish real enforcement from coincidental high prices, while the audit can.

arXiv:2610.28787 PDF

An Explicit Feedback Contract for Iterative ML-Agent Harnesses

Kehui Yao, Chuanbo Hu, Jiao Chen, Ramin Giahi, Reza Yousefi Maragheh, Lalitesh Morishetti et al. Iterative ML agents edit code, run experiments, and read a validation score after every attempt, which leaves them free to overfit the validation split through adaptive selection. The proposed feedback contract specifies what an agent may observe, when it must stop, and when it may touch the test split, and it reveals a validation score only when a submission beats a noise threshold derived from the paired standard error. On a control task where the true gain is zero by construction, an agent reading exact scores reported a gain of 0.0157 versus 0.0033 under the contract. On click-through-rate prediction and learning-to-rank, the contract narrowed the validation-test gap with slightly higher test performance. Agents with exact feedback often used all thirty submissions, including on the no-signal task, while contract-governed agents stopped after four to twelve.

arXiv:2610.28798 PDF

A Mechanistic Study of Automated Failure Attribution in LLM-based Agentic Systems

Wendy Zheng, Liang Wu, Yinhan He, Mayank Darbari, Jundong Li, Liangjie Hong Automated failure attribution (AFA) asks an LLM judge to identify which agent step in a multi-agent run caused the wrong final result, and direct prompting does poorly at it. Using linear probes on internal activations, the authors find a knows-but-won't-say gap: models encode enough information to locate the failing step but do not state it in their output. They propose grafting, a simple intervention that uses this latent knowledge without needing ground-truth labels at inference time, and it consistently beats direct-prompting approaches across multiple LLMs and benchmarks.

arXiv:2610.28811 PDF

Forging More Tasks into Better Harnesses: Agentic Harness Evolution via Successive Scale Racing

Jingzhe Liu, Jiliang Tang, Tong Zhao, Neil Shah, Clark Mingxuan Ju Agentic harness evolution refines an agent's instructions, tools, and control flow using evidence from its own task executions, yet existing methods paradoxically get worse as the pool of evolution tasks grows. The authors trace this to poor choice of which evidence to use in each round. SSR (Successive Scale Racing) builds diverse evidence combinations based on task quality and coverage, then races the resulting candidate harnesses under a small compute budget. It reaches state-of-the-art results in all six configurations across three benchmarks, raising pass rate from 64.3% to 68.3% as the task pool grows from 40 to 200 while baselines decline, and it cuts validation executions by 60%.

arXiv:2610.28815 PDF

Where Should a Stuck Agent Spend Its Next Tokens? Causal Attribution of Inference-Time Resources

Saiyue Lyu, Jerrold Huang, Yunbei Zhang, Janet Wang, Zhenting Wang, Yingqiang Ge et al. cross-listed When a coding agent fails, it can spend more tokens on extra computation, on extra information, or on feedback, and these are qualitatively different resources. BRA (Banzhaf Resource Attribution) is a controlled factorial framework that measures each resource's marginal contribution to a successful repair, and their interactions, from a fixed failed state. Across six models and failures from MBPP+, HumanEval+, and SWE-bench Verified, task-relevant specifications consistently help, public-test feedback helps a little, retrieved code helps little or not at all, and extra computation matters less as models get stronger. Yet three of four frontier models pick feedback in 95–100% of cases and gain less than uniform random resource selection.

arXiv:2610.28834 PDF

Building AI Research Teams for Autonomous Discovery of Linear Attention Architectures

Yujian Liu, Li An, Zelin He, Jacob Helwig, Jiabao Ji, Rohit Jain et al. The authors study how to organize teams of autonomous AI researchers to avoid local minima when experiments are costly. Their team consists of persistent agents, each grounded in a different scientific discipline, that keep private research histories, share experimental results, and receive experiment slots through adaptive allocation that balances exploration and exploitation. Deployed to search for linear attention architectures, the team discovers a design that outperforms GDN2 at 1.3B scale on language modeling, common-sense reasoning, and retrieval tasks. Ablations show that removing disciplinary backgrounds, private histories, or adaptive allocation each yields markedly weaker discoveries.

arXiv:2610.28887 PDF

Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication

Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov, Ilya Makarov Multi-agent systems built on diffusion language models usually communicate only through decoded text and throw away the continuous intermediate denoising states. In the proposed Hidden Messages setup, a sender compresses its denoising state into four dense vectors that a receiver agent consumes alongside its own context. The authors test causally whether the receiver actually uses the content of the message, either by removing it or by swapping in a message from a different example. On QASC, with the two supporting facts split between agents, matched messages reach 80.8% accuracy against 46.8% for a mismatched message, and a single late transfer (87.8%) outperforms four repeated exchanges by 7 points, suggesting that when information is passed matters more than how often.

arXiv:2610.28905 PDF

DiffResearch: Accelerating Deep Research with Diffusion Language Models

Alan-Barsag Gazzaev, Alexey Gavrilov, Mikhail Mozikov, Sergey Muravyov Deep research agents repeatedly plan, retrieve, synthesize, and revise, so end-to-end latency limits interactive use. DiffResearch is an open multi-agent deep research framework built on a diffusion language model (dLLM). To isolate the effect of the backbone, the authors keep the scaffold, retrieval, and prompts fixed and swap only between a diffusion model and a comparable autoregressive model. On 100 Deep Research Bench tasks, the diffusion backbone is roughly 6x faster end-to-end, mostly due to much faster planning and writing. A mode called dFast reads intermediate denoising states to extract subqueries before generation finishes, which roughly halves decomposition time with no loss in report quality.

arXiv:2610.28909 PDF

D-Skill: Distilling Teacher Knowledge through Agent Skills

Shuo Yuan, Dacheng Li, Joseph E. Gonzalez Transferring a frontier agent's abilities to a smaller model normally means distillation through weight updates, which requires access to the student's weights and extra compute. D-Skill treats the teacher as a black box: it compares paired teacher and student rollouts, finds where their behavior diverges or agrees, and condenses the lessons into a natural-language skill document that the student reads at inference time. The process runs entirely offline on fixed trajectories, with no labels and no weight updates. Across SpreadsheetBench-Verified, ALFWorld, and chess puzzles, it closes 75% to 95% of the teacher-student gap, beats one-epoch supervised fine-tuning (SFT), stays competitive with multi-epoch SFT, and the student costs 5.7x to 8.3x less per task than the teacher.

arXiv:2610.28914 PDF

Back Up Before You Regret: Measuring Backup Awareness and Automating Backup in Coding Agents

Xu Li, Simon Yu, Minzhou Pan, Yiyou Sun, Zhaorun Chen, Bo Li et al. Coding agents change real files, environments, and configurations, but rarely save state they may later need before modifying it. RegretBench provides 205 executable tasks drawn from real user-reported regret moments on X.com, GitHub, and Stack Overflow. Across six frontier models, agents complete 86.1% of tasks but back up the affected state beforehand in only 15.2%. The proposed BackUpper harness plug-in checks every tool call and preserves affected state using rules plus a lightweight subagent. It covers every regret moment, restores state with 92.2% average similarity, costs about $0.014 per task, and can follow user preferences about what is worth backing up.

arXiv:2610.28934 PDF

Self-Improving Agents with World Model Predictive Paths

Haori Lu, Tianyu Gao, Yaoyao Liu, Yingying Li The authors propose a test-time framework that helps LLM agents both anticipate outcomes and learn from experience during deployment, inspired by Model Predictive Path Integral (MPPI) control. At each step, the agent proposes several candidate actions, imagines their outcomes with a world model, scores each predicted outcome for task progress and risk, and combines the results with MPPI-style weights to choose an action. A two-branch experience memory improves both candidate generation and the world model's next-observation predictions over time. On ALFWorld and ScienceWorld with three backbone models, the method shows strong overall performance and confirms the value of experience-guided predictive control.

arXiv:2610.28945 PDF

AgentPolicyForge: From Executable Environments to Policy Benchmarks and Adaptive Guards

Chao Shang, Wanyu Du, Yi Zhang, Sandesh Swamy Placing deployment policies over tools and resources in an agent's prompt does not reliably prevent violations, and training a separate guard model requires verified examples of compliant and violating tool calls. AgentPolicyForge executes tool calls in real environments to label them and builds PolicyGym, a dataset spanning ten policy families. Every violating case in PolicyGym uses a permitted tool, so a guard must check arguments, resources and context rather than which tool is called. An attacker-guard loop iteratively fine-tunes the guard, and the final configuration achieves a 40-44% relative reduction in attack success rate with fewer benign rejections than policy prompting alone.

arXiv:2610.28989 PDF

VeriEvolve: Evaluator-First Evolution for Tool-Using Agents

Yueshen Li, Kamer Ali Yuksel, Gokhan Tur, Dilek Hakkani-T\"ur, Jo\~ao Pedro Maia cross-listed Self-improving agents need a reliable fitness signal, and realistic stateful tasks rarely provide one. VeriEvolve builds the evaluation layer first: from the task instruction alone it synthesizes challenge states plus an executable evaluator, freezes that evaluator as the fitness signal while the agent workflow evolves, and between rounds refines it using counterfactual test cases taken from strong and weak runs. On all 30 MCPMark Filesystem tasks, the evaluators agree with 113 of 122 held-out states. Evaluator-guided evolution raises externally verified success from 6/30 to 18/30 for GLM-5.2 and from 27/30 to 30/30 for Claude. Follow-up studies on PostgreSQL, Playwright, Notion and GitHub show the approach is limited by how observable the environment is, how its state is represented, and how much search and execution capacity is available.

arXiv:2610.29005 PDF

GenXR-Bench: Benchmarking Coding Agents for Interactive XR Development

William Huang, David Li, Xiang 'Anthony' Chen, Xun Qian, Nels Numan, Zhongyi Zhou et al. Writing extended reality (XR) applications depends on continuous user motion, spatial input and environmental sensing, which existing coding-agent benchmarks do not test. GenXR-Bench provides 112 expert-written tasks covering feature implementation, bug repair and full application generation. Held-out tests combine scripted and adaptive interactions with programmatic and visual checks, and they match blinded expert judgments 96.5% of the time on consumer XR hardware. Across 10 models, the best configuration reaches only 50.9% Pass@1, with recurring failures around asynchronous sensor lifecycles, physical grounding and embodied interaction semantics.

arXiv:2610.29014 PDF

Agentic Music Post-Production

Jaywon Koo, Junghyun Koo, Marco A. Martinez-Ramirez, Wei-Hsiang Liao, Seungheon Doh, Qiyu Wu et al. cross-listed MixAgent is a multi-agent pipeline that mixes music the way audio engineers do, producing an interpretable and editable mixing plan guided by a reference song. A grouping agent routes related tracks into subgroups, a planning agent uses retrieved examples to choose and order audio effects, and a detector and critic then propose numerical parameter fixes. Because simply prompting multimodal models to listen and revise often made mixes worse, the authors built MixCritic, a dataset of 45,657 repair critiques, and trained the critic with supervised fine-tuning followed by preference optimization. On MedleyDB, learned refinement reduces error by 19% relative to the initial mix, and listeners rated MixAgent's output above mixes from Qwen3-Omni, Gemini 3.8 Flash and GPT-4o-audio.

arXiv:2610.29093 PDF

Recursive Self-Improvement on a Budget: Scoring Code Edits Before They Run

Erfan Darzi, Sree Bhargavi Balija, Sheng He A coding agent that rewrites its own scaffold has to decide which proposed edit to expand before that edit has been evaluated. The method fits a frozen linear readout that scores each candidate from the parent's solve rate on a steering split plus eight features of the proposed diff, calibrated on 96 labeled scaffold edits. Search returns the earliest node with the best steering-split score. Across twelve GPT-4o runs, the returned node reaches 87.0% on held-out HumanEval+, compared with 85.1% for Huxley-Gödel Machine selection, 84.2% for a parent-rate-only ablation, and 81.2% for the root scaffold, while with Claude Sonnet 4.5 the gains are marginal (86.6% vs. 86.5%).

arXiv:2610.29121 PDF

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model

Jiongxiao Wang, Han Ding, Youzhi Luo, Yawei Wang, Yun Zhou, Linbo Liu et al. Frontier black-box LLM agents lack knowledge of how a specific environment behaves, and existing world models that predict the next observation at every step are costly and often uninformative. A pilot study finds that world-model guidance helps most when it anticipates pitfalls within a bounded multi-step horizon. The Pitfall Anticipation World Model (PAWM) therefore intervenes only when a likely failure looms and warns the agent away from it. It is trained with supervised fine-tuning on synthetic pitfall-labeled data, then with GRPO co-training alongside the black-box agent, and it improves Claude Sonnet 4.5 by 24.3% on ScienceWorld and 7.7% on WebArena-Lite.

arXiv:2610.29139 PDF

TrainBridge: Agentic NPU Training Optimization Guided by GPU Compilation Evidence

Yupeng Chen, Kunxiang Zhao, Xingjian Wang, Yue Zhao, Yijing Wu, Ye Liu et al. Training programs on Huawei's Ascend NPUs often miss compiler optimizations such as operator fusion, leaving fragmented computation and avoidable overhead that is costly to fix through profiling-guided trial and error. TrainBridge is an agentic system that uses the highly optimized execution produced by mature GPU compilation stacks as a reference: it extracts optimization intents from the GPU run, rewrites the NPU program accordingly, and refines the result through iterative profiling. It achieves an average 1.7x speedup across LLM pretraining, image classification and vision-language workloads, and nearly 2.0x across 27 Modded-NanoGPT speedrun records.

arXiv:2610.29171 PDF

Reasoning-Search-Augmented LLMs: An Operational Taxonomy, Benchmark Analysis and Controlled Evidence Synthesis

Biswas Poudel, Nilson Chapagain, Amit Kumar, Xianshun Jiang cross-listed This survey covers reasoning-search-augmented LLMs, systems that treat retrieval as an action inside an iterative agentic loop of planning, querying, integrating evidence, verifying and deciding when to stop, rather than as a one-shot preprocessing step. It classifies frameworks along four architectural axes: how retrieval is coupled, controller design, knowledge source and how strictly answers must be grounded in evidence. The learning signal (prompting, supervision, preference optimization or reinforcement learning) is treated as a separate dimension, because it can change search and stopping behavior even within the same architecture. The authors synthesize only controlled within-study ablations and quality-cost comparisons, avoiding invalid pooling of results across different backbones, retrievers and evaluation protocols.

arXiv:2610.29189 PDF

Preference Is Not Permission: Personalizing the Auto-Purchase Line in Agentic Commerce

Ryotaro Shimizu, Sai HtaungKham, Mo Fan cross-listed Personalized shopping benchmarks ask which item suits a user. The authors instead ask how far a shopping agent may commit on that item: don't buy, confirm first, or buy automatically. Whether committing is appropriate depends on how complete the product information is, how reversible the purchase is, its price relative to the user's norm and its urgency. Across sixteen instruction-tuned models, agents overreach, committing even when fit is low and when their own evidence flag says product information is missing, compared with choices from fifteen human annotators role-playing the user. Annotators disagree in structured ways, and a shallow learner trained on 30-40 of one person's labels predicts that person's auto-purchase line better than every prompt-and-model configuration evaluated, though the learned boundary does not transfer across domains.

arXiv:2610.29216 PDF

CrashGym: Measuring the Decision Value of Experimentation for LLM Agents in Counterfactual Market Crashes

Haocheng Lu, Haoxuan Cheng, Mingyang Chen, Minjun Zhu Language-model agents are increasingly asked to run experiments, yet benchmarks rarely check whether those experiments actually improve the decisions they are meant to inform. CrashGym is a deterministic, replayable limit-order-book simulator in which an agent must choose between a deleveraging intervention and a liquidity backstop before a shock-triggered market crash, and it may first run experiments in a separate laboratory copy of the market. The choice is scored on an independently seeded target market where every feasible intervention is simulated, so regret is computed directly with no human or model judging. In a pre-specified study on 24 new worlds, laboratory access cut one deployment's mean regret from 8.878 to 3.785, a 57.4% reduction, while a laboratory whose strategy mix evolves over time showed no established advantage over one with a fixed mix.

arXiv:2610.29245 PDF

MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents

Hong Ting Tsang, Boyi Zhang, Haoyu Huang, Jiaxin Bai, Yufei Li, Yisen Gao et al. Omni-modal assistants must remember what users show and how they sound as well as what they say, but existing long-term memory benchmarks mostly cover text and images and report a single end-to-end score. MOSAIC contains 923 questions over persona-grounded histories of 11–24 sessions spanning text, image, video, and audio, with deliberately built spatial, temporal, and prosodic cues. It also includes a diagnostic protocol that holds the question, reader model, prompt, and budget fixed while intervening separately on how evidence is represented, retrieved, and delivered. The decomposition shows that representation loss depends on modality, that errors in the final reading step dominate even when the model is given gold evidence, and that models often answer well yet fail to abstain when they should.

arXiv:2610.29249 PDF

Let Agents Design How They Search: Automatic Interface Optimization for Agentic Search

Sangjune An, Jaeyoon Jung, Yumin Choi, Jinheon Baek, Junmo Cho, Sung Ju Hwang et al. Agentic search depends on the interface through which an agent explores a corpus, but these interfaces are usually designed by hand and reused unchanged across corpora. SearchWeaver optimizes the interface automatically: a proposer reads execution trajectories and task rewards and jointly revises the precomputed corpus artifacts, the executable tools, and the usage instructions. This decides what is prepared in advance, what the tools do, and which decisions are left to the agent. With GPT-5.4-nano, the optimized interfaces raise answer accuracy on BrowseComp-Plus from 71.6% to 78.4% and raise average nDCG@10 across four BRIGHT subsets from 45.7 to 55.0, compared with letting the agent interact with the corpus directly.

arXiv:2610.29250 PDF

Design and Tuning of Multi-Agent Systems Using Bayesian Evolutionary Optimization

Dominik Jehle, Tom Zehle Designing an effective LLM multi-agent system (MAS) is laborious, and this work automates the search for task-specific workflow topologies. Evolutionary operators generate diverse workflow graphs, and Bayesian optimization, using a graph-aware Gaussian process and the Expected Improvement criterion, chooses which candidates are worth expensive evaluation, rather than asking an LLM to choose. On BigCodeBench code generation, the evolved architectures achieve the highest overall score among the agent systems evaluated. On a tabular preprocessing benchmark built on TabPFN, optimized preprocessing improves on default TabPFN for several temporal datasets and stays competitive on IID datasets.

arXiv:2610.29276 PDF

Credit Where Credit Is Due: Contrastive Attribution Topology Search for Multi-Agent Systems

Fabian Tritthart, Tom Zehle The accuracy and cost of an LLM multi-agent system depend on its topology, meaning which agents are included and how they communicate, but a single task score does not show which agents help or hurt. CATS (Contrastive Attribution Topology Search) starts from a fixed supergraph of specialized agents and uses an LLM attributer that compares execution traces across candidate topologies to assign credit to individual agents. It uses those credits to update edges and progressively prune weak agents and edges, ending with a compact team. With small open-weight models, it improves BigCodeBench accuracy by 5.3 percentage points over score-only search while using 54% fewer deployment tokens, matches the strongest baselines on MBPP with fewer tokens, and excluded every adversarially prompted agent in a HotpotQA prompt-poisoning test.

arXiv:2610.29292 PDF

ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery

Yongqiang Chen, Chenxi Liu, Deyu Zou, Zhenhao Chen, Tongliang Liu, Bo Han et al. Evolutionary coding agents such as AlphaEvolve improve programs with large language models (LLMs), but their mutations mostly operate on low-level surface heuristics rather than ideas. ExpEvolve extracts high-level idea factors that explain why some programs score better than others. It then uses those factors to guide exploitation of past experience and exploration when progress plateaus, and the authors give theoretical conditions under which this improves sample efficiency. Across 11 open-ended mathematical tasks, it improves final performance and sample efficiency by up to 5.0x, and the gains carry over to four other LLMs and six machine learning tasks from MLS-Bench.

arXiv:2610.29303 PDF

Signals and Queries: Token-Efficient Agentic Context Discovery

Yam Marcovitz User-facing LLM agents must decide at every turn which behavioral policies apply, and checking every policy with an LLM call is expensive. Signals and Queries (S&Q), developed for Parlant 4.0, first has an LLM generate activating examples and hard negatives for each policy offline. A separate logistic classifier per policy is then trained on embeddings of those examples, so filtering requires no LLM calls at runtime and removes at least 30% of candidate policies while keeping recall high. Evaluated against 9,144 LLM-judge labels on τ²-bench airline conversations, S&Q achieves higher precision than top-K and cosine-threshold retrieval at every recall level tested, with consistent results across two embedding providers.

arXiv:2610.29326 PDF

AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents

Changshun Wu, Mengyu Li, Mathis Marsault, Xiaowei Huang, Chih-Hong Cheng AgentMonopoly is an open-source Monopoly environment in which LLM agents negotiate, buy and trade properties, and make decisions that affect later turns. It is used to compare model families and to test whether assigned personas change agents' behavior. Across four model families at two sizes each, models differ in how often they trade, waive rent, and manage cash, and larger models do not consistently win more games. Persona instructions change the agents' language and some of their actions but not consistently across decisions, which the authors argue means persona following should be judged by actions and outcomes as well as words.

arXiv:2610.29330 PDF

Benchmarking and Error Analysis of Deep Search and Deep Research LLM Agents

Zixuan Yu, Zhenheng Tang, Xiaowen Chu, Bo Han Evaluations of Deep Search and Deep Research LLM agents usually score only final outcomes and ignore the long, multi-step execution traces where errors actually arise. The authors introduce a methodology for diagnosing errors in those intermediate workflows. They organize the failures into a taxonomy with two main dimensions, System-Level Errors and Task-Level Errors. By examining step-by-step traces and agent-environment interactions, they show how each error type affects overall performance and derive practical design guidelines for more robust agents.

arXiv:2610.29346 PDF

Memory-Centric Continual Learning for AI Agents

Dongfang Li, Zhuoen Chen, Bonian Jia, Zixuan Liu, Zhengxuan Lu, Can Deng et al. The authors argue that external memory for foundation-model agents should be judged by whether it improves future actions, not just by what it can retrieve. Their Memory-Centric Continual Learning (MCCL) framework has three parts. The first is a three-axis taxonomy covering where memory state is stored, what form it takes, and how it is managed, with an inner loop that updates memory and an outer loop that refines management policies from task feedback. The second is an evaluation protocol for retention under interference, adaptation to change, and transfer across tasks, built to separate real behavioral gains from simple context expansion. The third is a scaling framework for how agent performance grows with experience, storage, and compute, which together recast agent memory from passive lookup to continual learning.

arXiv:2610.29357 PDF

BIRD-RL: Scaling Agentic Reinforcement Learning over Stateful Data-Centric Environments

Ge Qu, Jinyang Li, Xiaohan Xu, Xiaolong Li, Bowen Qin, Nan Huo et al. Targets agentic reinforcement learning (RL) for small language models on stateful, data-centric tasks such as SQL debugging. In these tasks, isolating each rollout needs its own database container, while naively reusing databases contaminates trajectories. BIRD-RL combines a four-stage RL curriculum with a trajectory-scoped persistent execution infrastructure that safely reuses database replicas, improving spatial efficiency by 12.8x without Kubernetes. On BIRD-CRITIC-SQLite, its 7B and 14B models reach success rates of 44.40% and 48.00%, close to a Claude-Opus-4.6 agent at 48.20%, and the same models can both debug and generate SQL.

arXiv:2610.29426 PDF

One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents

Shaojin Ma, Ruitao Wang, Pengyi Li, Jie Yang, Liangyu Chen, Gangyi Zhao et al. Identifies credit-update misalignment in multi-turn reinforcement learning for LLM agents. Different textual responses can produce the same executable action, yet policy updates assign credit to each wording separately, and the authors show this pairing of credit and text shapes update directions. Effect-Marginal Policy Optimization (EMPO) instead optimizes an effect-level surrogate that couples updates across responses with the same effect. It works with both trajectory-level and step-level credit estimators and needs no extra critics or rollouts. On ALFWorld, WebShop, and Sokoban, EMPO outperforms strong baselines in both task performance and learning efficiency.

arXiv:2610.29427 PDF

DynaWeave: Learning to Adapt through Dynamic Scenario Synthesis

Xuan Yang, Chunxiao Liu, Hongsheng Xin, Hao Xu, Ning Miao Tackles the mismatch between static training environments for tool-using LLM agents and real environments, where outside events can invalidate an agent's plan mid-task. DynaWeave converts static tasks into verifiable dynamic scenarios by generating relevant exogenous events, ordering them with directed acyclic graphs (DAGs), triggering them during interaction, and revising only the evaluation checks the events affect. Supervised fine-tuning of Qwen3.5 models on 2,095 verified trajectories from 36 environments improves results on Gaia2, CostBench, and ToolMaze at every model size. It raises CostBench task completion by 67.34 percentage points and surpasses Gemini-3.1-Pro-Preview on ToolMaze.

arXiv:2610.29443 PDF

RegoMender: A Multi-agent Framework for Automated Rego Policy Repair

Ruijie Zhang, Huaxiao Yin, Li Hu, Fanqi Kong, Haozhe Liang, Shaoyin Cheng et al. cross-listed Targets automated repair of Rego policies in Policy-as-Code systems, which general-purpose program repair methods handle poorly because of Rego's strict syntax and declarative semantics. The authors release RegoBench, a benchmark of 1,159 real-world defects, and propose RegoMender, a feedback-guided multi-agent framework. Its agents coordinate through a shared blackboard and an atomic toolkit to patch, rewrite, or roll back policies. RegoMender reaches syntactic and test-based semantic success rates of 82.9% and 77.9%, 13.4% above the strongest baseline, generalizes to HCL and SQL, and repairs 86.4% of 132 real-world CVE vulnerabilities.

arXiv:2610.29456 PDF

Harness Scaling for Generalizable Agentic Reinforcement Learning

Fei Bai, Huatong Song, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai et al. Argues that agentic reinforcement learning (RL) through a single fixed agent harness, the scaffolding that governs tool use and context management, leads to models that specialize and transfer poorly to other harness configurations. Harness Scaling builds a factorized space of harness configurations, selects a compact training subset with balanced coverage, and trains across harnesses with controlled variation on the same tasks. Across benchmarks, model scales, and RL algorithms, this improves generalization to unseen configurations. Models trained on structured white-box variation transfer to unseen external black-box harnesses better than models trained on a single complex black-box harness.

arXiv:2610.29458 PDF

PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents

Jian Zhong, Zeyu Liu, Pingchuan Cao, Rongduo Han, Shunye Tang, Guohuan Xie et al. Addresses how long-term dialogue agents can keep user memories stable while still updating them when the user genuinely changes, rather than appending everything or overwriting by heuristic. PreCoMem, inspired by predictive coding, stores memories as confidence-weighted beliefs and computes an Effective Surprisal signal from semantic deviation and contradiction evidence, discounted when retrieval is unreliable. A two-threshold gate then chooses among three updates: reinforce a supported belief, hold an ambiguous cue as a low-confidence hypothesis, or decay a reliably contradicted fact. It reports state-of-the-art accuracy on LoCoMo, LongMemEval-S, and PersonaMem-v2, and the new DiMoBench shows it adapts to belief drift without overreacting to noise.

arXiv:2610.29467 PDF

Rethinking Agentic Model Orchestration for Time Series Forecasting

Tong Guan, Tianxiang Zhan, Shirui Pan, Ming Jin Agentic model orchestration (AMO) systems select or combine time series foundation models (TSFMs) for forecasting. The authors test whether reported gains come from better orchestration or simply from stronger candidate models. Their protocol, FairAMO, fixes the candidate forecasts in advance so that orchestrators can be compared fairly. On GIFT-Eval, the best orchestrators cut mean absolute scaled error (MASE) by less than 6% relative to the best single TSFM. These gains depend on candidates that excel on different tasks rather than on the number of candidates, a redundant candidate can erase them, and a few recent windows of history capture all the benefit.

arXiv:2610.29504 PDF

Recursive Code Evolution for Task-Oriented Semantic Communications

Yan Gong, Zheng Chu, Wei Koong Chai, Hari Mohan Pandey, Jingrui Hou, Miao Zhang et al. cross-listed ComRSI (Communications-Oriented Recursive Self-Improvement) uses an LLM agent to rewrite the code of a semantic communication system over successive generations. The agent edits the encoder, channel adapters, decoder, loss, and training policy, and a protected evaluator accepts or rejects each candidate under fixed resource budgets. On CIFAR-100 over an additive white Gaussian noise (AWGN) channel, eight accepted revisions produced a design reaching 74.3% accuracy at 32 symbols per image, compared with 67.2% at 256 symbols for the initial design. Across ten repeated searches, ComRSI clearly beat a memoryless agent, but its advantage over an agent that keeps a textual history was not statistically significant.

arXiv:2610.29641 PDF

Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization

Zhiqian Xia, Jielong Tang, Ling-An Zeng, Shixiong Kai, Rong Yao, Hao Kong et al. Learned memory systems for long-horizon agents usually optimize memory construction and memory retrieval separately, which limits how well the two stages adapt to each other. Uni-Mem trains one shared policy for both stages with end-to-end online reinforcement learning (RL). It introduces Stage-Decoupled GRPO, which alternates between optimizing the two stages and controls variation from the other stage within each rollout group, and Event-Wise Credit Assignment (ECA), which assigns sparse feedback to the events responsible for it. On LoCoMo and LongMemEval, Uni-Mem outperforms the strongest existing RL-based memory method.

arXiv:2610.29671 PDF

Bridging Foundational and Agentic Capabilities of LLMs

Wei Guo, Yuanjian Xu, Jianing Hao, Yongyao Duan, Ran Chen, Zhong Li et al. End-to-end agent benchmark scores do not show whether a failure comes from weak underlying capability or from poor execution as an agent. The authors build, for each model, a foundational profile from capability benchmarks and an operation-level profile from interaction trajectories, and define capability realization gaps as the difference between observed operations and what the foundational profile predicts. Across fifteen model services on τ²-Bench, foundational profiles only partially predict agentic operations: models with similar profiles can execute very differently, and these gaps are consistent across task subsets. Adding operation profiles substantially improves prediction of task scores.

arXiv:2610.29751 PDF

Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting

Zhipeng Chen, Yunsen Liang, Yanxi Chen, Hongkang Zhao, Lingyun Feng, Ruiyang Ren et al. Studying recursive self-improvement (RSI) requires an environment that keeps changing, can be checked objectively, and is grounded in the real world, and the authors argue that forecasting future events meets these requirements. DeepEvolve builds a live environment that synthesizes verifiable forecasting tasks and an evolution engine that turns each prediction's reasoning trajectory into parameter updates, with search decoupled from prediction to limit contamination, and into reusable distilled skills. The self-improved forecaster ranked first on the external FutureX leaderboard for three consecutive weeks, with up to a 195% improvement in rating. On the authors' synthesized evaluation, it placed in the top three across four base LLMs.

arXiv:2610.29784 PDF

Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents

Mohammed Saqlain The authors run a controlled ablation of harness design for biomedical claim-verification agents on SciFact, with the live PubMed API as the only tool, comparing closed-book answering, single-shot retrieval, a ReAct-style agent that writes its own queries, and an oracle given the gold abstracts. Naive single-shot retrieval finds the gold paper for only 6% of claims. Agent-written queries triple gold recall and give the best realistic accuracy (62%), while the oracle reaches 79-86%, which identifies retrieval rather than reasoning as the main bottleneck. A citation audit shows closed-book models cite PubMed IDs that almost always exist but are almost never relevant (only 3 of 327 judged relevant), so checking that a citation exists is not enough of a safeguard.

arXiv:2610.29812 PDF

UserSWE: Evaluating Coding Agents in Multi-Turn Conversations with Dynamic Users

Zhan-Yu Wu, Ji Luo, Alyssa Wu, Junjielong Xu, Richeng Xuan, Zhichao Hu et al. Most coding-agent benchmarks hand the agent a complete issue report in one turn, which does not reflect how developers actually fix bugs together with an agent. UserSWE turns repository history and existing single-turn benchmarks into multi-turn tasks with a simulated user. Information pools limit what the user can reveal, a persona model calibrated on 20,859 real development sessions shapes how it talks, and tests run after every turn, with verdicts written and gated by agents from two different model families. Evaluating eleven agents on 72 tasks from 42 repositories, the authors find that rankings change with the interaction budget: Claude Fable 5.1 resolves the most tasks in one turn, GPT-6-Astra the most by turn eight, and Claude Opus 5 leads at most budgets in between.

arXiv:2610.29822 PDF

PAUSE: Search-Free Steering of Coding Agents Before Their First Source-Code Edit

Viral Patel, Boyin Tan, Jason Xue, Youcheng Sun cross-listed Coding agents often explore and test extensively and still produce failing patches, and existing steering methods rely on costly periodic reviews or on monitors that only react to anomalous behavior. PAUSE intercepts the agent's first source-code edit, which happens at a median of 60% of trajectory progress, and has a separate Planner read the task and pre-edit transcript to write one repair plan, after which the original agent resumes with its context intact. With Qwen3-coder as the agent and GPT-5.6 Sol as Planner, online steering improves resolution by 16.8% on SWE-bench Verified in OpenHands and by 14.7% on SWE-bench Multilingual in SWE-agent. It also beats a rule-triggered advisor at similar cost, matches SICA's periodic overseer at about 11-15% of its Planner cost, and offline recovers 35.4% of eligible failed tasks.

arXiv:2610.29870 PDF

FrontierCS 2: A Benchmark for Computer Science R&D

Qiuyang Mang, Wenhao Chai, Runyuan He, Dongruixuan Li, Zihan Zheng, Yujun He et al. FrontierCS 2 tests whether LLM agents can contribute to frontier computer science research. Each of its 48 tasks takes a published paper's released code, deletes the paper's contribution, and uses the deleted code as the human reference to beat. Agents work inside the inherited repository and may submit patches at any time; each patch gets a development-suite score, while the final grade comes from a private held-out suite compared against the reference. Across 13 domains and 12 agents, the best agent, Claude Opus 5.5, passes 58% of its runs, no run passes on 16 tasks, and most runs that beat the reference on the combined score still fall short on some individual metric or workload.

arXiv:2610.29932 PDF

ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination

Nithin Sivakumaran, Ali Payani, Hyunji Lee, Mohit Bansal, Elias Stengel-Eskin LLM agents struggle to represent uncertainty about a partner's state in text, which hampers theory-of-mind (ToM) reasoning in multi-agent settings. ToM2Vec gives each agent a fixed budget of continuous latent reasoning tokens, trained to match a compressed form of the partner's actual thoughts, and at test time those tokens condition the agent's own thoughts and actions. Across collaborative, competitive, and mixed-motive environments, including Overcooked-AI, Hanabi, Decrypto, and multiparty negotiation, it consistently beats the base model and a range of trained and prompted baselines. Performance drops when the same reasoning is expressed in text, the training generalizes to partner model families not seen during training, including GPT-5.4, and probing indicates the tokens encode partner beliefs, desires, and next actions rather than task-specific information.

arXiv:2610.29934 PDF

Remembering Through Change: Provenance-Preserving Hypergraph Memory for Multimodal Agents

Jielong Tang, Ling-An Zeng, Xujie Yuan, Xiao Dong, Zhiqian Xia, Qingfeng Li et al. Long-term memory for multimodal agents must keep fine-grained visual evidence while tracking how entities change across sessions; compressing images into text loses detail, while storing raw images makes identity tracking hard. MHG-Mem is a multimodal hypergraph memory whose typed hyperedges group mentions of the same entity, synthesize how its state evolves, and link back to the source visual evidence. A hypergraph-based agentic search chooses an entry layer, traverses the typed hyperedges, and inspects the original observations when fine-grained verification is needed. Across three benchmarks it improves performance by 8.20-9.98% over the strongest baseline.

arXiv:2610.29995 PDF

Agent Behavior Maps: Locating Agents by What They Would Do, Without Running Them

Shota Fujikawa, Takuya Tamura, Kosuke Akimoto, Daichi Haraguchi, Masafumi Oyamada LLM agents with similar success rates can still act very differently, and comparing them usually means running new rollouts. An Agent Behavior Map places each LLM-plus-harness combination in a coordinate space using centered action log-likelihoods on shared reference trajectories, so the target agent never has to be run in the environment. The authors derive error bounds linking squared coordinate distances to trajectory KL divergence, and across eight cross-scored models the distances track measured trajectory KL in rank order. Across 100 LLMs and four environments, ALFWorld-based coordinates support performance prediction and reveal behavioral similarities that persist after controlling for success rate and model size, even across model families; harness-component changes also roughly compose as displacements in the map.

arXiv:2610.30014 PDF

GuideEvolve: Self-Evolving Agent Framework with Expert Guidance and 3Ts Mechanisms

Meihong Jia, Yasemin Timar, Jonathan Taws GuideEvolve is a self-evolving agent framework that drives its evolution with task-specific expert evaluation and adds three regulatory mechanisms: Think Fast and Slow, Throwback (rollback to recover from performance drops), and Termination by choice. On MGSM and MMLU it roughly matches the Gödel Agent (89.8% versus 90.6% on MGSM; 89.1% versus 87.9% on MMLU) while reducing variance by 25–59%. The authors also report a 100% evolution success rate against the baseline's 82% and about half as many optimization iterations.

arXiv:2610.30017 PDF

What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion

Yikai Zhao, Saurabh Pandey, Pradeep Kumar Misra Enterprise agents read Word and PDF files through an ingestion layer, and every mainstream converter drops the meaning authors encode in visual formatting, which leaves agents confidently wrong rather than visibly incomplete. On a task built from 76 real policy documents, scored deterministically against the files' own OOXML properties, formatting-blind ingestion drops micro F1 from 0.690 to 0.150 with docling and 0.075 with plain text; the gap holds across six answering models from Anthropic and OpenAI. Leave-one-out analysis traces the loss almost entirely to highlight and colour, while bold, italic, monospace, and strikethrough contribute almost nothing, and 32.8% of 6,460 surveyed documents use deliberate colour or highlight. A pre-registered test shows that explicitly typing the cue for the agent helps beyond merely preserving it, which places this reliability problem in ingestion rather than in reasoning.

arXiv:2610.30026 PDF

Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents

Zijian Jiang, Xilu Wang Text "skills" extend tool-using agents but make prompts long and expensive. Soft prompts are a compact alternative, but training them on whole successful trajectories can introduce unwanted behavior changes. The authors find that a skill's useful effect is concentrated in a few trajectory positions, and they propose Behavioral Localization, which compares the frozen agent with and without the text skill to find positions where the skill both raises the likelihood of successful tokens and meaningfully shifts the next-token distribution. Only those positions supervise a short soft prompt. On SpreadsheetBench with a frozen Qwen3.6-35B-A3B agent, the top 10% of positions capture 96.12% of the skill's positive gain, and task-specific localized supervision reaches 57.38% success, ahead of both the full text skill and the on-policy OPCD baseline.

arXiv:2610.30041 PDF

HGM-Dual: Lineage-Aware Guided Agent Harness Optimization for Long-Horizon Planning

Sudipta Paul, Pengrui Quan, Aounon Kumar, Nicholas Tzou, Vivek Kulkarni, Srinivas Chappidi Harness optimization improves agents built on frozen LLMs by editing their prompts, tools, and workflows. Recursive self-improvement methods such as the Darwin–Gödel Machine and Huxley–Gödel Machine (HGM) were built for coding and waste evaluation budget on planning tasks with interacting constraints. HGM-Dual records the agent's lineage as context (what each ancestor changed, which behaviors improved or regressed, and which edits were ineffective) and generates two kinds of children for each parent: a general one that keeps exploring and specialized ones that target the harness components most likely behind the parent's top constraint violations. It gains up to 14.45% relative case accuracy over the strongest baseline and matches HGM variants' final performance with 2–4x less evaluation budget.

arXiv:2610.30042 PDF

SWE-DIVE: Multi-Turn Coding Agent Evaluation with Dynamic User Simulation

Lorenz Brehme, Benjamin Steenhoek, Ruth Breu cross-listed Coding agents are usually benchmarked on single, well-formed prompts such as SWE-bench GitHub issues, while real sessions involve vague requests and requirements added over many turns. SWE-DIVE evaluates coding agents in multi-turn settings using a user simulator that can play different behavioral patterns and a conversation validator that checks whether the simulator's messages stay faithful to the original requirements and steer the agent toward the task. The framework runs inside a Harbor environment that orchestrates the agent, simulator, and validator and executes tests to decide whether the task was solved.

arXiv:2610.30047 PDF

Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis

Xinzhe Wang, Xudong Weng, Zang Li, Shaolei Zhang Data-analysis agents built on ReAct-style loops reason almost entirely in text, while human analysts also read charts to spot trends and anomalies. On a 200-task probe suite with four matched interface conditions, agents read their own charts as well as numeric output but do not reliably choose to plot. Charts help most on temporal, pattern, and anomaly questions, and text helps most on precise reading and verification. The proposed method distills two kinds of experience from judged rollouts, namely how to select and verify analytical operations and when to use text, charts, or both, and admits an experience only if it improves held-out performance. Across three held-out benchmarks and four LLM backbones, it raises analytical accuracy and visual-evidence quality, and backbones that never drew charts begin using them, without any parameter updates.

arXiv:2610.30072 PDF

Parallel AI Agent Collaboration as Graph-Structured Distributed Block Optimization

Hanmin Li, Yury Demidovich, Peter Richt\'arik cross-listed When AI agents work in parallel, each can make a useful local edit, yet combining those edits can still damage the shared solution. The authors model this as discrete block optimization with a known dependency graph: edits that form an independent set in that graph are guaranteed compatible, and a coordinator merges them safely. It can also test interacting edits to keep beneficial interactions. They prove that merging via an optimal fractional coloring keeps a 1/χ_f(G) fraction of the total verified local improvement, where χ_f(G) is the graph's fractional chromatic number, and show this bound is tight. They also derive geometric convergence rates and a linear program for assigning agents. In code-repair experiments, verified merging resolves conflicts that break naive merging, and in route-search tasks fractional scheduling cuts median wall time by 15%.

arXiv:2610.30074 PDF

Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use

Yanzhe Chen, Ziyi Yang, Jifeng Zhu, Qiming Huang, Ruihe An, Peiyao Xu et al. cross-listed This survey asks what changes when multimodal agents move from operating computers to operating robots, and it defines multimodal embodied agents as systems that couple multimodal reasoning with physical action and feedback. It introduces PAPAV, a framework of five capabilities defined by function rather than architecture: Perceive, Anticipate, Plan, Act, and Verify. It uses this framework to compare computer-use agents, robotic systems, and embodied agents under five physical constraints, such as partial observability and outcomes that cannot be verified. A review of 64 benchmarks finds that evaluation centers on acting, with only 5 benchmarks assessing anticipation and 3 assessing verification.

arXiv:2610.30108 PDF

Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research

Houde Qian, Zheng Jiang, Wenyi Li, Chenxi Xu, Xinjie Lin, Yuzhao Peng et al. Existing finance benchmarks score formulas, code, or individual trading decisions, so they do not test whether large language models (LLMs) can run the full quantitative research loop and produce strategies that hold up outside the research period. In Quant-Arena, a model builds features and stock-scoring code on real Chinese A-share data, receives feedback only on 2022 and 2023, and then freezes one strategy for evaluation on the sealed year 2024. The evaluator applies realistic trading constraints and costs, including market impact estimated from a simulated order book. Of nine frontier models, GPT-6 Astra scores highest, but no model beats the equal-weight market in 2024.

arXiv:2610.30113 PDF

The End of Consensus Research?

Cl\'ement Burgelin, Rachid Guerraoui, Zicong Hong, Yinan Hu, Anne-Marie Kermarrec, Rafael Pires et al. cross-listed Designing low-latency consensus protocols for geo-replicated state machines requires protocol knowledge, correctness reasoning, implementation, and experimentation. The authors built AI-greement, a harness for coding agents that combines a protocol knowledge base, a modular implementation skeleton, and a Lean-based verification framework, and used it to discover A-greement, a new protocol. On a geo-distributed key-value store, A-greement cuts mean latency over the fastest baseline by 12.2% on write-heavy and 33.3% on read-heavy workloads, and the Lean formalization uncovered several previously unreported vulnerabilities in existing protocols. The authors argue that human experts remain essential throughout the process and that agents change the researcher's role rather than replacing it.

arXiv:2610.30114 PDF

Environment Duels: A Microcosm of LLM Co-Evolution

Shibo Jin, Zhiqiu Xu, Yiling Zou, Shangyu Gong, Mayur Naik Environment Duels is a dynamic agentic benchmark that tests large language models (LLMs) both as authors of interactive environments and as solvers of them. Each author writes environments with a privileged hint, aiming for tasks that strong solvers fail without the hint and pass with it. Across nine frontier LLMs, the strongest solvers are not the strongest authors, and smaller models write the hardest environments and the most useful hints. Head-to-head outcomes are transitive and produce a stable ranking. In a reinforcement-learning study with Qwen3.8-27B, how much a hint helps the learner on an author's environments predicts how much training on those environments improves it on held-out benchmarks.

arXiv:2610.30116 PDF

1 more specialized paper

Other 45

What Git Does Not Record: Provenance Legibility in Two AI-Assisted Research Repositories

Koutian Wu, Zesen Huang, Hejia Geng cross-listed Publishers increasingly ask authors to disclose how AI tools contributed to research, so the authors ask what a git repository can actually reveal about that. They audit two single-author repositories, one carrying standing agent rule files and the other a running agent scratchpad, and find that the two record AI involvement in opposite ways. The Co-Authored-By commit trailer, the field automated audits read, proves unreliable: in one repository only 1 of 39 commits that revised the agent's working-notes file carried the trailer, and in the other the trailer rate jumped from 2.7% to 17.7% over time. The authors spell out which provenance claims such traces can and cannot support, and derive requirements for systems that record research provenance as work happens.

arXiv:2610.28655 PDF

Speedrunning Relational Foundation Model Pretraining

Gabriel Wedson Mendon\c{c}a de Souza Veloso, Luan Borges Teodoro Reis Sena, Denis Oliveira Correa, Francisco Galuppo Azevedo Pretraining recipes for relational foundation models are developed in individual labs under private protocols, so progress does not accumulate. The authors introduce modded-nanoRT, a speedrun leaderboard for Relational Transformers where entries are ranked by wall-clock time to a quality target on a single GPU, with the held-out database chosen by calibrating how well short learning curves separate real improvements from seed noise. The opening recipe combines typed trilinear attention, which conditions attention on the foreign key traversed, with the Muon optimizer and SIGReg, and reaches the target in 305.63 seconds on one H200, 8.11 times faster than the reference implementation, winning 13 of 15 comparisons outside the development database.

arXiv:2610.28855 PDF

LEAN: Loss-Engineered Modulation of the Model Weight-Sensitivity Distribution for Efficient, Error-Tolerant Neural Networks

Habibur Rahaman, Swarup Bhunia A trained network's weights differ widely in sensitivity: the 1% most sensitive weights account for about a fifth of the damage a bit flip can do. LEAN adds a regularizer during the final training or fine-tuning epochs that bounds the step a bit flip can take and lowers the gradient of the most sensitive weights, reshaping the sensitivity distribution. Across nine vision models (ResNet-18, ResNet-50, VGG-16 on CIFAR-10, CIFAR-100, and ImageNet) it shrinks the sensitive tail 5 to 13 times within 0.5 points of accuracy, and with an error-correcting BCH code over 1% of weights the models survive 300 adversarial bit flips where most published defenses fail. The reshaped distribution also enables a mixed 8/6/4/3-bit format that stores ImageNet ResNet-50 at 3.29 bits per weight, 2.4 times smaller than INT8 and more accurate than LSQ at 3 bits.

arXiv:2610.28872 PDF

Rethinking Time Series Tokenization from a Frequency Perspective

Zelong Tian, Ming Jin, Bo Du, Shirui Pan Splitting time series into subsequences (patches) is now the standard way to tokenize them, but the authors argue it introduces spectral distortions that push learned representations away from the true trends, periodicity and seasonality. A frequency-domain analysis gives the boundary conditions under which these distortions occur. The proposed tokenizer, TimeFT, builds tokens through frequency-domain partitioning, frequency shifting and Nyquist sampling. It is parameter-free, adds negligible compute, can replace existing tokenizers directly, and consistently improves forecasting, classification and anomaly detection.

arXiv:2610.28997 PDF

Solving Go under Positional Superko: Sound Proof Transfer and New Board Values

Sami Kharma, Florian Schintke Tackles solving Go under positional superko (PSK), the rule that forbids recreating any earlier whole-board position. Because PSK makes legality depend on game history, it breaks transposition tables, static life-and-death recognition, and safe move pruning. Sanko, a parallel depth-first proof-number solver, restores all three soundly through a proof-transfer theorem that says when a stored proof stays valid on a new path. It proves 5x5 Go in 16 minutes on six threads, compared with 5.25 hours for the previous best, and establishes PSK values for 23 more boards, 11 of which were unsolved under any rules. It also shows that superko changes game values: 4x7 drops from a full-board win under basic ko to 12 points.

arXiv:2610.29414 PDF

Fast Bayesian Optimization Through Linear Models

Colin Doumont, Henry Moss, Philipp Hennig Bayesian optimization (BO) is usually reserved for expensive objective functions, where the algorithm's own runtime hardly matters. The authors place BO and faster optimizers such as CMA-ES on a speed-performance Pareto frontier and point out regions of it that no current method covers. They propose fast BO methods that exploit the linear structure already present in existing BO algorithms, for continuous and combinatorial inputs ranging from about 10 to more than 6,000 dimensions. These methods match state-of-the-art BO performance while running 100x to 1,000x faster.

arXiv:2610.29580 PDF

39 more specialized papers

Theory 44

Depth Separation for Neural Networks under Gaussian Measure

Cynthia Bortolotto, Francisco Ganacim, Roberto I. Oliveira cross-listed Depth separation results show functions that deeper neural networks can approximate efficiently but shallower ones cannot. This work proves such a result for L² approximation of radial functions under the standard Gaussian measure in d dimensions: a radial function exists that a polynomial-size depth-3 ReLU network can approximate, while depth-2 networks with subexponentially bounded weights need exponential size. The key tool is a general lemma that radial functions with substantial high-frequency mass in their Hermite expansion cannot be approximated by narrow depth-2 networks with bounded weights. All results are formally verified in Lean 4.

arXiv:2610.28800 PDF

Actually, All Learnable Distribution Classes Are Privately Learnable

Mohammad Afzali, Hassan Ashtiani, Alex Bie, Gautam Kamath, Peihan Liu, Argyris Mouzakis et al. The paper settles a conjecture by Ashtiani about whether learnability of distribution classes implies private learnability. The authors prove that every class of distributions learnable to arbitrarily small total variation error with finitely many samples is also learnable under (ε, δ)-differential privacy. This complements earlier work showing that the implication fails when learnability is defined at a fixed target error; under the standard definition, the conjecture holds.

arXiv:2610.28878 PDF

Transformers as Generic Deep Neural Networks: Exact Compilation, the Cost of Removing Activation-Activation Products, and Multi-Agent Attention Clustering

Sergey Salishev, Zeev Volkovich, Oleg Granichin The authors show that every Transformer block, including attention, LayerNorm and a GELU MLP, compiles exactly into a generic deep neural network built from fixed affine maps, routing, and a few elementwise nonlinearities. This suggests neural processing units (NPUs) do not need dedicated attention operators. They prove that without activation-activation products such a compilation needs at least N² extra nonlinear evaluations per head, and that the elementwise product found on every surveyed NPU removes this cost. For top-k attention, they show that estimating the dropped tail mass is unbiased while simply dropping the tail is biased, and that clustering queries around centroids reduces score computation from N² to Nk. Compiled versions of DeiT-Tiny, MambaIRv2-Light and Pythia-70M match their references and work with standard post-training quantization.

arXiv:2610.28958 PDF

How Does Orthogonalization Adapt to the Neural Network Hessian Structure? A Gradient Self Outer-Product Analysis at Initialization

Shenyang Deng, Garvit Goyal, Tianyu Pang, Shuhua Yu, Yaoqing Yang The Muon optimizer orthogonalizes momentum before each step, and its implicit preconditioner matches the row-block structure of neural network Hessians when the momentum's self outer-product VVᵀ is diagonally dominant. Why networks produce that dominance has been unexplained. The authors develop an asymptotic theory at initialization for symmetric matrix factorization, deep linear networks, two-layer ReLU networks and single-layer linear attention. In each model, diagonal entries of the expected gradient outer-product outgrow off-diagonal ones at explicit rates as hidden dimensions grow, which points to overparameterization as the cause. For attention's query, key and value weights, the off-diagonal terms vanish entirely, and simulations plus language-model experiments confirm the predicted scaling.

arXiv:2610.28975 PDF

From Common Information to Canonical Correlation Analysis and Beyond: Bridging Information Theory, Signal Processing, and Machine Learning

Paris A. Karakasis, Walid Saad cross-listed Many problems involve several jointly observed views of one process, each mixing a shared component with a view-specific one, and the shared component must be recovered without supervision. This tutorial brings together how statistics, information theory and signal processing have each tackled this problem. It covers the Gács-Körner and Wyner notions of common information, canonical correlation analysis (CCA) and its multi-view extensions, and subspace methods for blind multiuser receivers. It then covers kernel, deep and probabilistic nonlinear CCA, when the common factor is identifiable, and links to self-supervised learning, distinguishing exact equivalences from mere analogies in a self-contained exposition.

arXiv:2610.28978 PDF

Conditional Flow Models Convert Epistemic Uncertainty into Apparent Aleatoric Variance

Navodita Sharma, Abhishek Shetty, Max Simchowitz, Pulkit Agrawal Conditional Flow Models (CFMs) are increasingly used in place of deterministic predictors, including for regression. The authors show that CFM samples often have larger variance than the true conditional distribution, even for nearly deterministic targets, because the model's uncertainty about an unresolved input dependence shows up as apparent noise in its samples. Synthetic and real-data experiments, along with comparisons to explicit smoothing estimators, show that this excess variance tracks underfitting and shrinks with more training or data. A simplified theoretical analysis explains the effect, and the authors use it to build high-coverage prediction sets for regression targets.

arXiv:2610.29172 PDF

Approximate Solomonoff Induction: On the Robustness of Learned Universal Predictors

Qian Li, Shang-Hua Teng Prior work explains the broad abilities of log-loss-trained models by noting that minimizing population log-loss recovers a Bayesian mixture predictor, which becomes Solomonoff induction in the limit of a universal prior. Real training only approximates that optimum, so this paper asks how robust prediction is to the approximation. It proves that any predictor within KL divergence η of the mixture keeps bounded cumulative prediction error, but the excess error in a given environment is tightly Θ(η / w_μ), where w_μ is the environment's prior weight, so environments with low prior probability can suffer badly. Averaged over the prior, the excess error is only O(η), and the authors add a distribution-shift theorem and an extension to finite-sample training of neural networks and Transformers.

arXiv:2610.29289 PDF

When Are We Near Nash? Certifying Solutions in General-Sum Games

Boning Li, Longbo Huang cross-listed Solvers for general-sum extensive-form games produce a sequence of strategy profiles but cannot easily tell how close any of them is to a Nash equilibrium. The authors derive an online certificate that bounds the NashGap, the distance from equilibrium, using only the regret and terminal-distribution accumulators the solver already maintains. It costs time linear in the number of terminal histories, needs no best-response computation, and remains sound under alternating updates and with three or more players. Running a portfolio of sixteen training-free regret-matching solvers and keeping the iterate with the smallest certificate solves 89–92% of instances across in-sample, held-out, and fresh sets of biased Leduc poker games at threshold 1e-5, and an exact-comparator refinement certifies every solve.

arXiv:2610.29300 PDF

When Thoughts Become Memory: Unifying In-Context Learning and Fine-Tuning

Leon Chlon, Julie Huang cross-listed Proposes a certified rule for deciding whether new data should go into a pretrained model's context or into its weights. The rule rests on an exact law for the data's information value, measured as code length saved on held-out queries, which holds for any model, architecture, or task. The law's bounds can certify the choice from a single measurement, and an anytime-valid test makes it from streams of held-out queries. The test picks the better of GPT-2's cache and a fine-tune in 61,948 of 64,000 query streams and errs in at most 3.3% of null streams. The same valuation also measures how much individual context items help, recovering a published evidence adjudicator with an AUROC of 0.91.

arXiv:2610.29419 PDF

A Sharp Dynamic Haussler Lemma

Shihao Shao, Dacheng Tao Haussler's covering lemma gives the sharp metric-entropy bound for binary VC classes, and the authors prove a sharp global dynamic (sequential) analogue for bounded real-valued classes under normalized l_p distance, for every p from 1 to infinity. A single fixed family of deterministic causal experts of size at most (C/ε^p)^d, where d is the sequential fat-shattering dimension, covers every target on every sequence for every horizon, and a matching lower bound holds. For p=2 this resolves the horizon-independent sequential l_2 covering problem of Rakhlin, Sridharan, and Tewari, and for p=1 it settles open questions about binary classes in terms of the Littlestone dimension. The proof combines rank-decreasing correction schedules with shared-parent routing to remove the logarithmic overhead.

arXiv:2610.29898 PDF

Mitigating Model Collapse through Dreaming Learning

Matteo Benati, Alessandro Londei, Denise Lanzieri, Raffaello Mastromarino, Vittorio Loreto Model collapse, the loss of diversity when generative models are trained on their own synthetic data, is studied here through the eigenvalue spectrum of a learned Markov transition operator. Sampling at temperature below 1 acts like repeated sharpening: subdominant eigenvalues move toward the unit circle, the spectral gap narrows, and accumulating data does not appear to stop the drift. The authors analyze Dreaming Learning, a regularizer that adds a loss on the model's own samples drawn at temperature above 1, which keeps the eigenvalues bounded away from the unit circle. Spectra measured over 5000 generations match this picture, and the mitigation carries over to LSTM language models and a 124M-parameter GPT-2.

arXiv:2610.30009 PDF

33 more specialized papers

Safety & Alignment 33

S^3martCirc: Self-supervised Smart Circuit Discovery

Wendy Zheng, Yinhan He, Liang Wu, Jundong Li cross-listed Mechanistic interpretability (MI) work on large language models usually happens in two steps: first find which components, such as attention heads or feedforward neurons, matter for a task (circuit discovery), then work out what each one does (functional interpretation). The authors argue these two questions depend on each other and should be answered together. S^3martCirc sorts each node into one of two general computational roles that carry over across tasks, and assigns the roles with a quantitative metric instead of subjective human reading. As a result, a component's importance and its function are discovered in a single step, and the authors report that the framework outperforms existing circuit discovery methods.

arXiv:2609.00755 PDF

SimHack: Predicting the onset of detectable reward hacking before RL training

Junwei Deng, Alex Bie, Natalia Ponomareva When large language models are trained with reinforcement learning (RL) to write code, reward hacking can often be detected but is hard to prevent: hardening environments gets harder as they grow, and penalizing detected hacks teaches models to hide them. The authors instead predict, before a run starts, when detectable hacking will become dominant, framing it as forecasting the distribution of the onset step for a given training recipe. SimHack learns hacking dynamics from a few smaller reference runs and simulates unseen recipes without retraining. It predicts the median onset with 11.3% relative error and a rank correlation of 0.85 across recipes, generalizes across model scale, curriculum, rollout count and learning rate, and was used to find curricula that train 96.6% longer before hacking appears.

arXiv:2610.28707 PDF

Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise

Pedro H. Barcha Correia, Matheus Azevedo de S\'a, Paulo H. Gomes de Senna, Giovanni Mandel Martignago, Victor Takashi Hayashi, Charles Christian Miers et al. cross-listed Persistent memory in chat assistants lets a prompt injection carry over into future sessions, a threat the authors call memory injection. They present a general threat model and show that all NIST attack goals are collectively achievable across ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash, through vectors including personal memories, projects, skills, organization instructions, and Model Context Protocol (MCP)-connected resources. Compromised services can keep spreading disinformation, generating harmful content, or faking unresponsiveness in fresh sessions. In Claude's case, shared memory lets an attacker compromise multiple users in an organization and exfiltrate prompts. The authors derive defenses from the threat model and report that nearly half remain unimplemented across the services tested.

arXiv:2610.28810 PDF

To Manipulate or Not: Opinion Dynamics in LLM Interactions

Mohammad Siahkamari, Phuoc Dat Nguyen, Stephen MacNeil Most persuasion research measures only one side of a conversation, even though LLMs in extended exchanges can both shape and be shaped by the other party. The authors run controlled LLM-LLM debates on climate-change attribution, where one role is assigned a stance and sometimes told to persuade while the other has no assigned position, across four pairings of GPT-6 Luna and Gemini 2.5 Flash Lite, tracking both stated positions and turn-by-turn behavior. Same-model pairings narrow the opinion gap on average (clearly for only one of the two models) while cross-model pairings widen it. Stated positions and in-dialogue behavior capture different aspects of opinion change, and an explicit instruction to persuade does not reliably increase influence.

arXiv:2610.28875 PDF

Fig: Automated Multiagent Behavior Elicitor

Anrui Gu, Zhichao Wang, Nicolai Ouporov Social evaluations of multiagent systems usually wait for behaviors to emerge rather than deliberately eliciting them. Fig (Finding Interactions in Games) is an agentic system that takes a natural-language behavior description, breaks it into sub-behaviors, generates or selects social games, and intervenes in the game engine, opponents, instructions, or context to elicit the target behavior, scoring results with mechanical checks and rubric-based judgments. For six behaviors including deception, covert coordination, betrayal, and sabotage, models attempt the target behavior in 45.2% of episodes in valid generated scenarios, and Fig's interventions raise attempt rates by 10.9% on average over matched control games.

arXiv:2610.28891 PDF

Guilt by Association: Semantic Leakage Drives Associative Gender Bias

Hazel Chen, Peter West, Hila Gonen Semantic leakage occurs when an irrelevant concept in a prompt steers what a language model generates about an entity, for example a person who likes yellow being described as a school bus driver. The authors show that leakage persists even when the concept is not attributed to the entity at all, which points to a broad associative tendency rather than the model building a profile of that entity. Extending this to gender, they find that a gendered concept merely present in the scene shifts generations toward the stereotype about as much as when the subject engages with it directly. Current bias evaluations, which focus on concepts attributed to the subject, would miss this form of bias.

arXiv:2610.28944 PDF

Black-Box Red-Teaming of Multi-Agent Systems via Reinforcement Learning

Yuqi Jia, Hangsheng Zhang, Chenxi Liu, Junjie Liang, Xun Chen cross-listed Black-box red-teaming attacks built for single large language model (LLM) agents work far less well against multi-agent systems (MAS), where many sub-agents never interact with the user directly. ReMAS uses reinforcement learning to fine-tune attacker LLMs that generate prompts for extracting system prompts. A rewrite stage first sharpens base attack prompts, then a template-generation stage builds prompts that are more likely to invoke specific sub-agents and expose their system prompts. The method substantially improves attack success rates over existing black-box attacks and transfers across backend LLMs and MAS structures.

arXiv:2610.28966 PDF

Behavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents

Ekagra Gupta, Raj Saha, Aniekan-Abasi Inyang, Adam Al Gharib The authors test whether large language models (LLMs) that refuse harmful requests in text also refuse when acting as tool-calling agents. They use a 2,304-prompt benchmark of matched text-only, tool-enabled and adversarially framed tool-enabled variants, scored against 20 forbidden actions. Across five models, 10-52% of harmful requests refused in chat still produced unsafe tool calls under standard tool prompting. A refusal direction in the residual stream weakens in tool contexts and predicts unsafe calls (AUC 0.76-0.86). Restoring or steering along it mostly suppresses tool use rather than producing safer calls, so the authors conclude that agent safety evaluations must examine actions directly.

arXiv:2610.28982 PDF

PerMIT: Securing Persistent Memory for LLM Agents via Machine-Checked Integrity Theorems and Scalable Community Trust

Tanqiu Jiang, Min Bai, Sadat Shahriar, Wanyu Du, Sandesh Swamy PerMIT is a memory architecture for large language model (LLM) agents that defends persistent memory against injection and poisoning. It routes all untrusted, environment-derived values through a finite interface and a confirmation gate. Two integrity theorems, machine-checked in Lean 4, guarantee that untrusted memory reaches a sensitive action only if confirmed at the gate, which reduces every attack in the threat model to one auditable decision. It records zero observed attack success on Trojan Hippo, and logged confirmations let communities of up to 90 agents on τ²-bench build shared knowledge bases that improve new agents without promoting injected memory.

arXiv:2610.28987 PDF

Attention-based Classification for Indirect Prompt Injection

Amelia Kawasaki, Stefan Lee Attention Tracker detects indirect prompt injection (IPI) by thresholding a scalar focus score derived from attention, and claims to work regardless of dataset. Evaluated on the agentic AgentDojo benchmark, it gives near-chance detection across most model families because long structured tool outputs break its assumptions about attention. The proposed AttnDA replaces the threshold with a Fisher Linear Discriminant Analysis classifier over per-head focus scores and improves detection across four benchmarks. Transfer to AgentDojo remains hard, so learned detectors need training data from the target deployment domain.

arXiv:2610.28988 PDF

Better Policies, Worse Verifier Audits

Daniel Cie\'slak, Jacek {\L}ubi\'nski The authors study a common loop in which a component verifier is audited against the full success label on fresh policy outputs and refreshed whenever its score falls. For a conjunctive success condition, as the policy masters one component, failures of that component become rare in the audit data. The audit then increasingly rewards information about the other component, which the authors call on-policy audit collapse, and it can end up preferring a worse verifier. They derive exact thresholds and demonstrate the effect on an oracle, a multi-label text task and an ECG study where acting on the audit increased downstream constraint violations. The fix is to audit each component on its own failures and keep those failures in the refresh distribution.

arXiv:2610.29029 PDF

Mitigating Reward Hacking in Realistic Scenarios Through Representation Engineering

Austin Meek, Kyle Cox, Matthew Nguyen, Austin J. Brockmeier, Steven Basart The authors study how to detect and reduce reward hacking in realistic settings that involve tool calls and long contexts. They derive steering directions in three ways: difference-in-means over labeled rollouts, gradient optimization, and activations recorded while models reference an anti-reward-hacking model spec. They apply these through a dynamic intervention that perturbs representations less than uniform residual-stream steering. The intervention cheaply and significantly reduces reward hacking without major losses in coherence or capability, and vectors extracted from coding environments generalize to other domains.

arXiv:2610.29129 PDF

Correlated Failures of AI Approval Monitors

Luca Zanolini cross-listed Before an AI agent takes a consequential action such as a payment, several monitors often check it, but monitors that share a base model or read the same poisoned input can fail together. The authors borrow fail-prone system models from distributed computing to define when a rule for combining monitor verdicts is safe (never executes a bad action) and live (never blocks a good one). They test twelve small open-weights monitors, plus Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and DeepSeek-V3.2, on payment approval. Every reliable fraud approval came through one evidence source and fooled only the monitors that read it, and a rule that counts independent sources instead of votes blocked 0.9% of correct payments versus 7.4% for the best vote-counting rule at equal safety. The pre-registered worst-case test still failed, and only adding a rule engine that cannot be persuaded restored a rule that was both safe and live on the measured attacks.

arXiv:2610.29178 PDF

A Unified Evaluation Framework of Physical Adversarial Robustness in the Edge VLM Era

Eric Liang, Alan Luo, Kaiwen Yuan cross-listed Compact vision-language models (VLMs) that locate objects from language prompts are increasingly deployed at the edge in safety-critical systems, but their robustness to physical adversarial attacks has not been compared fairly with that of conventional object detectors. The authors measure whether a patch worn on a person evades detection, using a conditional attack success rate (cASR) and splitting failures into omission (no box), substitution (a box on the patch) and displacement (a box elsewhere). Grounding VLMs can be as vulnerable as detectors, and the failure mode tracks a model's output format more than its architecture family. Under distribution shift, VLMs move from omission toward substitution and displacement, which return confident, well-formed boxes that a downstream system cannot easily detect as failures.

arXiv:2610.29196 PDF

Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics

Hongyu Li, Xinlong Huang, Zhan Yu, Jing Zhao, Zhe Zhao, Juntao Zhong et al. Most LLM safety benchmarks use binary safe or unsafe labels, which explain little and produce inconsistent verdicts across automated judges. DiagSafe instead attaches a rubric to each prompt with separate criteria for safety (concrete harms, with exemptions to suppress false positives) and usefulness. It comes with a bilingual English-Chinese benchmark built on an 84-label safety taxonomy, split into DiagBase (direct and jailbreak prompts) and DiagAttack (harmful intents embedded in complex instruction-following tasks). LLM judges using the rubrics reach 95.34% agreement with human annotations, versus 69% to 87% for baselines, with Cohen's kappa of 0.864 between judges. The evaluation also shows that leading models differ markedly in unsafe rates and are especially vulnerable in Chinese and multi-turn conversations.

arXiv:2610.29225 PDF

Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support

Yuyao Zhang, Yakun Cui, Han Zhu, Jinsong Lu, Yike Guo, Sirui Han Large language models (LLMs) are used more and more to support consequential decisions. This work tests whether their advice changes when the same probability is stated as a gain or as a loss, for example a 70% chance of success versus a 30% chance of failure. The authors build FrameBench-HPS, a set of 2,880 prompts from 240 high-stakes scenarios in healthcare, finance, education and career, and legal and consumer-rights domains, and run it on seven LLMs. Framing alone changes the recommendation category in 32.4% of complete prompt groups, mostly by adding uncertainty rather than reversing the advice outright. The effect survives paraphrasing, exceeds ordinary run-to-run variation, and is substantially reduced by rewriting probabilities into a single canonical form.

arXiv:2610.29294 PDF

Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan

Ju-Chun Ko cross-listed "Abliterated" open-weight models have the activation direction that mediates refusal removed. The authors test whether this also removes political bias, comparing 2-bit Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next against community abliterations on statements about Taiwan, the 1989 Tiananmen events, and control countries in Traditional Chinese, Simplified Chinese, and English. The writing script mattered far more than abliteration: the original Bolt affirmed "Taiwan is part of the PRC" (People's Republic of China) with probability 0.91 in Simplified characters but 0.06 in Traditional, while abliteration shifted its scores by at most 0.06. Abliterated Qwen stopped refusing on 1989 yet still affirmed the PRC claim in all three scripts. In a sample of 740,000 web documents, Simplified-Chinese text about Taiwan used PRC-framing terms 3.7 times as often as Traditional text, and a four-sentence system prompt shifted the abliterated Qwen strongly toward the position of the Republic of China (Taiwan).

arXiv:2610.29342 PDF

TenetBench: A Clause-Anchored Russian Corpus for Evaluating Language Models against Professional Codes of Ethics

Yegor Anichkov, Valery Vinogradov, Andrey Furov, Marina Polukoshko, Sergey Bolovtsov TenetBench tests whether language-model assistants breach, or help users breach, specific clauses of professional codes of ethics. It contains 3,433 requests that conflict with clauses of fifteen Russian-language professional codes, each paired with a benign twin request, and is produced by an automated pipeline that can generate tasks from any code's text. Specialists rated 0.995 of the conflicting tasks plausible. Eleven of fourteen model and reasoning-mode configurations produced conflicting content for most such requests, usually without any warning. Turning on reasoning halved the refusals of two models.

arXiv:2610.29477 PDF

WRDO: Weighted Refusal Direction Optimization in Aligned Language Models

Grigoriy Evseev, Vladimir Solodkin, Dmitry Bylinkin, Daniil Medyakov, Vladimir Aletov, Aleksandr Beznosikov Refusal in aligned language models is known to depend largely on a single residual-stream direction, usually computed as the difference of mean activations between harmful and harmless prompts. That average merges very different harmful prompts into one vector. Weighted Refusal Direction Optimization (WRDO) builds a composite refusal direction from a bank of per-prompt directions using learned aggregation weights. The authors evaluate it on three open-weight models with three guard judges, HarmBench, and standard utility benchmarks. They also run the same optimization in reverse to restore refusal in an abliterated checkpoint, and report that weighted aggregation manipulates refusal more effectively than prior single-direction methods.

arXiv:2610.29555 PDF

Effects of Model Scale and Quantization on Social Bias and Abstention in Question Answering

Kuniko Paxton, Ifeoluwa Wuraola, Dhavalkumar Thakker When a small or quantized language model gives fewer biased answers, it may be because it is fairer or because it abstains more often. The authors test this on the BBQ question-answering bias benchmark across sizes and quantization settings of Qwen3, Gemma-3, and SmolLM2, varying the prompt wording and also removing the "Unknown" answer option. Larger models are more accurate, but bias shows no simple scaling trend, and quantization effects vary by configuration. In 20 of the 26 configurations with fewer biased responses, over-abstention also increased, so fairness evaluations should measure bias, abstention, and prompt design together.

arXiv:2610.29657 PDF

Monitorability Is Trainable: Post-Training for Chain-of-Thought Monitorability

Usman Anwar, Tim Bakker, Christos Louizos, David Krueger Chain-of-thought (CoT) monitoring is useful for oversight only if a model's reasoning stays coupled to its actions, and training processes such as CoT compression or optimization against a monitor can weaken that coupling. The authors propose an unsupervised post-training objective that maximizes mutual information between the CoT and the output. Across three model organisms with degraded CoT, it recovers most or all of the lost monitorability. On undegraded policies, it improves how well CoT monitors and activation probes predict unsafe behavior, which suggests that monitorability can be trained as a property of the policy.

arXiv:2610.29734 PDF

Investigating the Impact of Learner Errors on Social Bias in Large Language Models

Federico Marcuzzi, Martin Christoph, Iryna Gurevych, Alla Rozovskays Social bias benchmarks for large language models almost always use fluent English, which leaves out non-native writers. The authors inject common learner grammatical errors into prompts from three bias benchmarks covering gender, race, and religion, and find that errors increase bias, with the effect growing as the error rate rises. Preposition errors produce the largest and most consistent increase in stereotype alignment, which suggests errors matter most when they interfere with comprehension. Fine-tuning on error-containing text reduces the effect, and the authors conclude that evaluations using only well-formed prompts may underestimate the bias these users experience.

arXiv:2610.29896 PDF

Unintended Side Effects of Value Training in LLMs

Siddhesh Pawar, Chen Cecilia Liu, Anna Korhonen, Isabelle Augenstein Fine-tuning LLMs to hold specific human values is becoming a common post-training step, but evaluations usually check only the targeted value. The authors fine-tune four instruction-tuned LLMs along individual Hofstede cultural value dimensions and then measure effects on untrained values and broader behaviours. They find that value-specific training produces unintended side effects on untrained values, risk preferences, social cognition, and safety traits. They argue that value alignment should be paired with broader behavioural evaluation and that more work is needed on where value training belongs in the training pipeline.

arXiv:2610.29922 PDF

Building Auditable AI Text Detectors with Sparse Autoencoders

Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch The strongest detectors of AI-generated text are opaque fine-tuned classifiers, while transparent stylometric alternatives trail them badly, which is a problem in high-stakes uses such as peer review and admissions. The authors train sparse autoencoders (SAEs) on a fine-tuned detector's residual stream, label each feature with auto-interpretation, keep only features whose descriptions pass a fidelity threshold, and train a lightweight classifier on those interpretable features. The recovered features correspond to specific, predictive linguistic "tells" of AI text. The resulting classifiers are interpretable and generalize better to out-of-distribution text than both the neural and stylometric baselines.

arXiv:2610.30006 PDF

9 more specialized papers

Vision 25

DualHeads: Reward-tilted few-step diffusion made simple

Yuanzhi Zhu, Eleftherios Tsonis, Simo Alami Chehboune, Xi Wang, St\'ephane Lathuili\`ere, Vicky Kalogeiton cross-listed Aligning distilled few-step diffusion models with rewards is hard because methods built for multi-step models need differentiable rewards or stochastic trajectories that distilled generators may lack. DualHeads adds two lightweight binary-classifier heads to a frozen diffusion backbone: one learns which samples for the same prompt score higher under an external reward, the other separates reference images from generated ones, and the generator maximizes a weighted sum of their logits. With optimal heads this objective equals KL-regularized reward maximization, with the head-weight ratio setting the regularization strength. A few-step DualHeads generator reaches task rewards close to 40-step reinforcement learning methods while preserving image quality, at no extra inference cost.

arXiv:2610.28858 PDF

OccJEPA: Occlusion as Masking for Amodal Video Segmentation

Genpei Zhang cross-listed Amodal video segmentation predicts an object's full mask in every frame, including the parts hidden behind other objects. The current best methods use fine-tuned video diffusion models, which are slow, need a separate depth network, and give a different answer on each sample. OccJEPA treats occlusion as masking: it drops the tokens outside the object's visible mask, lets the pre-trained V-JEPA 2.1 predictor fill them in, and decodes the features with a light head. The resulting model is deterministic and 33.5x faster than the diffusion-based state of the art while being more accurate on SAIL-VOS, TAO-Amodal and Kubric, and controlled experiments trace the gain to the predictor's pre-trained weights rather than its size.

arXiv:2610.29061 PDF

LLMSeg: Training-Free Text-Prompted Image Segmentation with B-Splines

Abdullah Hamdi cross-listed Asks whether a general-purpose multimodal large language model (MLLM) can segment images from text prompts without a segmentation head, mask decoder, or any training. LLMSeg splits the job into three LLM-driven stages: it detects the instances that match the query, traces each one with closed periodic B-splines inside its bounding box, and then refines by repeatedly focusing on the remaining errors. With an off-the-shelf Gemini-3.7-Flash backend, it is competitive with and often beats the specialist SAM 3 model on the same items across referring, reasoning, open-vocabulary instance, and semantic segmentation benchmarks. It reaches 71.5 cIoU on ReasonSeg val with no segmentation-specific training.

arXiv:2610.29409 PDF

Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning

Martin Thi{\ss}en, Elke Hergenr\"other cross-listed Image generators usually need auxiliary adapters or inference-time steering for fine-grained spatial control, even though their vision-language text encoders can represent spatial coordinates. Using a synthetic benchmark of geometric primitives and a color-aware layout metric, the authors show that the training signal-to-noise ratio (SNR) distribution largely determines spatial grounding in flow-matching models. Training on a centered SNR distribution gives 25.9% layout adherence, while shifting emphasis toward high-noise stages raises it to 80.8%, close to the 83.6% reconstruction ceiling of the variational autoencoder. The noise schedules used by models like FLUX.1 prove too conservative for precise layout. On COCO, distribution tuning alone raises layout Average Precision from 1.0 to 57.3 with no architectural changes.

arXiv:2610.29581 PDF

Improved Training of Strong Stochastic Flow Maps

Zander W. Blasingame, Antonio Franca, Raul Minan, Christian Koke, Niklas Rindtorff, Alexander Tong Stochastic flow maps aim to sample stochastic differential equations (SDEs) in a few steps by learning direct jumps along the solution path. They are hard to train because the Brownian path is infinite-dimensional and nowhere differentiable, which blocks the derivative-based losses used by deterministic flow maps. The authors represent the path with a shifted Legendre polynomial expansion and train with a Wong-Zakai-style objective that forces the network to use higher-order path information, which makes derivative-based losses possible. They also give a recipe for fine-tuning large pretrained flow models into stochastic flow maps, reaching roughly a three-fold improvement in FD-DINOv2 over SSFM at four function evaluations on ImageNet.

arXiv:2610.29605 PDF

Rethinking Post-hoc Debiasing of Large Pretrained Models

Divyaansh Mertia, Chirag Bhutra, Himanshu Shekhar, Shubham Bagwari, Pratik Mazumder cross-listed Post-hoc methods for debiasing frozen encoders such as CLIP all require the bias attribute to be identified and described in text. The authors show this can hurt: on UrbanCars, which has two co-occurring bias attributes, recent methods underperform a plain classifier even when given ground-truth bias identities. They observe that cross-class training pairs placed closest together by the frozen encoder almost always share a spurious attribute. Using this, they learn a single linear map over the features that performs competitively across datasets without bias labels or any LLM.

arXiv:2610.29688 PDF

19 more specialized papers

Reinforcement Learning 22

Reinforcement learning for hardware acceleration and system-level optimization on FPGA and integrated-circuit platforms: A survey

Hamza Kheddar This survey reviews more than 100 studies that implement reinforcement learning (RL) on field-programmable gate arrays (FPGAs) and integrated circuits (ICs). It sorts the work by RL family (Markov decision process methods, value-based, deep RL, actor-critic, policy-gradient and multi-agent), by level of abstraction, by hardware realization strategy and by application domain. The comparison covers resource use, latency, throughput, power and reported speedups, and closes with recurring architectural patterns, practical limitations and open challenges in hardware-algorithm co-design.

arXiv:2610.28699 PDF

Second Thoughts: Exploring and Planning through Comparative Imagination in World Models

Hieu Tran, Hong Yu Agents with world models can compare imagined outcomes before acting, but they still have to decide which alternatives to explore and when a planned action should override the learned policy. SDI (Sibling Disagreement in Imagination) handles both decisions with one reference: the variation seen across repeated imagined rollouts of the actor's own action sequence. This reference scales an exploration bonus and gates plan acceptance, so a planned action is executed only when its predicted advantage beats that variability by a fixed factor. No extra ensemble is needed. Across Meta-World, Adroit, and DM-Control, it achieves the best task-average score in every suite and matches or exceeds the strongest baseline on 15 of 18 tasks.

arXiv:2610.28779 PDF

ArcFlow: Training Flow Policies for Multiple Solver Budgets

Hieu Tran, Hong Yu Flow-matching policies can trade inference cost for quality by changing the number of solver steps, but reinforcement learning at a single fixed step count can hide poor behavior at cheaper budgets and even erase one-step ability inherited from pretraining. ArcFlow (Action-Ratio-Clip Flow) trains one policy across a range of solver budgets. It combines action-aligned supervision with a PPO-style clipped surrogate, cycles through budgets during training, and selects checkpoints by worst-budget performance, all without changing the architecture or deployment cost. Across DM Control, manipulation, and locomotion tasks, it attains the highest mean one-step return among five flow-policy baselines on 25 of 30 tasks.

arXiv:2610.28788 PDF

RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning

Shixuan Liu, Cheng Tang, Yuzheng Hu, Fan Wu, Han Zhao, Jiaqi W. Ma Data attribution methods for online reinforcement learning (RL) mostly measure a sample's influence within a single training round and miss its effects across later rounds. The authors define a trajectory-level leave-one-out quantity, RL-LOO, and derive RL-IF, a first-order estimator that propagates influence through both optimization updates and changes in which data the policy samples, with approximation guarantees under smoothness assumptions. They show the full estimator is often well approximated by the optimization effect alone, and build on this to create RINSE, a data-filtering method. On standard control benchmarks and an RLHF-style toxicity-mitigation task, RINSE improves training efficiency and final performance over standard PPO and local attribution baselines.

arXiv:2610.28883 PDF

Is In-Context Reinforcement Learning A Posterior Sampler? An Empirical Study

Weiqin Chen, Xinlong Du, Gesualdo Scutari, Harsha Honnappa cross-listed In theory, the Decision-Pretrained Transformer (DPT) approach to in-context reinforcement learning (ICRL) performs Bayesian posterior sampling, but only under idealized assumptions such as infinite pretraining data. The authors derive two necessary conditions for a posterior sampler, sufficient-statistic invariance and the martingale property, and build statistical hypothesis tests for each. In Gaussian, Bernoulli and linear bandit settings, both conditions are consistently violated by practically trained DPT models and proposed variants. This suggests ICRL does not behave as a Bayesian posterior sampler in practice, and the tests serve as a reusable toolkit for checking Bayesian behavior.

arXiv:2610.28977 PDF

Doob-Based Generative Policy Optimization

Maojiang Su, Qijie Zhu, Yu Xiao, Shuyang Yu, Minshuo Chen, Zhaoran Wang et al. Diffusion and flow policies are expressive, but their action likelihoods are intractable, which blocks standard policy optimization methods such as PPO and GRPO. DoobPO uses Doob's h-transform to separate the two jobs. It first learns a lightweight ratio model representing the desired policy update, then distills the resulting tilted target into the generative policy through conditional drift matching. The authors prove that this practical objective is gradient-equivalent to the ideal Doob drift matching objective and that it yields the minimum-energy modification needed for the desired update. On MuJoCo continuous-control benchmarks, it performs strongly against diffusion- and flow-policy baselines.

arXiv:2610.28990 PDF

FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-Next

Zhewei Yao, Yuxiong He Existing open-source RL stacks struggle with the memory demands and custom operators of Qwen3.8-Flash-Next when running long-horizon agentic reinforcement learning (RL). FlashNext-RL is a model-specific training framework built on prime-rl, with vLLM generating software engineering (SWE) rollouts. It trains the full 48-layer model in BF16 on a single 8xB200 node by optimizing host offload, gradient accumulation, and kernels for mixture-of-experts, sparse attention and the gated residual stream. Training throughput rises from about 395 to 4,438 tokens/s/GPU at a 131K-token context, and with three inference nodes the system sustains a live SWE RL run of 1,024 rollouts per batch at about 22 minutes per iteration.

arXiv:2610.28998 PDF

Dissimilarity-Based Markov Decision Processes for Stabilizing State Distributions

Xun Shen, Yuepeng Wang, Daniyal Sajedi-Hosseini, Akifumi Wachi, Ken Kawano, Yoshihiko Fujisawa et al. Maximizing expected return does not guarantee that a stochastic system's state distribution converges to, or stays at, a desired target. The authors formulate a Dissimilarity-based Markov Decision Process (DMDP), in which the policy acts on the current state distribution and the objective measures its distance from a target. They give sufficient conditions under which the optimal policy makes the target distribution asymptotically stable, plus certificates that account for learning and estimation errors. A practical version that combines distribution estimates with PPO and SAC is demonstrated on a linear system and on regulating the velocity distribution in HalfCheetah.

arXiv:2610.29024 PDF

Value-Rectified Distillation for Flow-based Offline Reinforcement Learning

Ke Jiang, Wen Jiang, Yoshinobu Kawahara, Xiaoyang Tan In offline reinforcement learning (RL), distilling a Flow Matching (FM) behavior policy trained on multi-modal data tends to average across distinct behaviors, producing invalid out-of-distribution (OOD) actions. Value-Rectified Distillation (VRD) recasts distillation as a value-guided alignment of generative trajectories, separating behavioral modes and keeping the high-value actions. It comes in two versions: VRD-Q, which uses value quantiles to filter out low-value behaviors, and VRD-R, which reorganizes the latent space to isolate high-value trajectories. On D4RL, OGBench and Minari, VRD improves existing generative baselines on continuous control and planning tasks, especially on multi-modal, non-expert datasets.

arXiv:2610.29075 PDF

MA-UITD: Offline Multi-Agent Reward Inference for Characterizing Continuous Social Behavior

Zeyuan Zhang, Yiyun Yang, Yuan Cheng, Herbert Zheng Wu, Yusi Chen Inferring rewards from interacting agents is hard because each agent's choices reflect both its own preferences and its expectations about the others, and existing multi-agent inverse reinforcement learning methods often need extra environment interaction or an inner policy-optimization loop. MA-UITD (Multi-Agent Unilateral Inverse Temporal Difference Learning) works offline on fixed joint demonstrations. It represents each agent's long-term consequences through belief-conditioned successor features that average over the other agents' behavior, then links them to agent-specific rewards using inverse temporal-difference consistency and noise-contrastive learning. In simulated cooperative, competitive, and general-sum tasks it outperforms baselines in discrete-action settings, and applied to real social-foraging trajectories it recovers distinct reward structures for leaders and followers.

arXiv:2610.29237 PDF

Flat-RLVR: Sharpness-Aware Reinforcement Learning with Verifiable Rewards for Reasoning Capacity in LLMs

Yeonbum Kim, Hyun Kyu Lee, Sung Whan Yoon Reinforcement learning with verifiable rewards (RLVR), especially with group-relative policy optimization (GRPO), improves LLM reasoning but tends to sharpen the policy distribution, suppressing rare but correct solutions and reducing sample diversity. Flat-RLVR counters this by seeking flat regions of the reward landscape, applying sharpness-aware minimization during post-training together with GRPO variants. The authors connect flat-reward optima to reduced distribution sharpening in theory, and in experiments they report better sample diversity while keeping single-attempt accuracy.

arXiv:2610.29281 PDF

The Predictive--Optimization Gap in GuanDan Self-Play

Qicheng Zou, Biao Yuan, Siyu Song, Wentao Liu, Xiaokun Wang Tests whether a signal that predicts outcomes is also a good target for optimization, using the multi-round imperfect-information card game GuanDan. Changes in a frozen match-win predictor are added as round-level supervision in a matched DanZero-style Monte Carlo self-play setup. One selected checkpoint clearly beats its baseline, but none of five preregistered replication seeds is positive and the pooled ten-seed interval includes zero. The authors call this a predictive-optimization gap: a signal can track outcome-relevant progress without reliably improving learning.

arXiv:2610.29411 PDF

Events, Not Pixels: Where World Models Get the Return Wrong

Haoyue Liu, Xiaoying Tang Examines why world models used for planning can render frames faithfully yet still predict the wrong return. A transition-level diagnostic swaps imagined transitions for real ones and measures how much return error disappears compared with a budget-matched random swap. Event transitions, the few steps where a reward is earned or an episode ends, look as good visually as other steps, but restoring them removes a median 70% of return error. Restoring the transitions with the largest pixel or perceptual error helps less than a random choice. Event patches beat random patches in 39 of 44 Atari-100k model-collection pairs, improve plan ranking, and show the same concentration on control tasks with rare rewards.

arXiv:2610.29435 PDF

CORD: Conflict-Aware Optimization for Joint RL and Distillation in Agents

Wenhao Yu, Zhengcheng Wang, Jiahong Liu, Zexuan Qiu, Minda Hu, Shaohang Wei et al. When LLM agents are trained with both reinforcement learning rewards and self-distillation from a teacher, the gradients of the two objectives can partly cancel, and a simple mixing weight cannot remove only the conflicting components. CORD is a conflict-aware optimization framework that chooses one objective to protect, decides when to project away the conflicting gradient, and gives a closed-form nearest-point update once per mini-batch. Added to three existing joint training recipes, it improves ALFWorld success by 1.5-15.9 percentage points and improves two of three recipes on WebShop, with SDAR rising from 60.7% to 64.6%. The analysis shows that the protection direction and the projection trigger must be chosen together, and that the distillation weight still matters after projection.

arXiv:2610.29965 PDF

EPIC: Rethinking Trust Regions for Autoregressive LLM Policies via Exact-KL Prefix Importance-Ratio Clipping

Chengxi Li, Hanlin Wu, Zhicheng Cai, Huan-ang Gao, Wei-Ying Ma, Jingjing Liu et al. PPO-style clipping in reinforcement learning (RL) for LLM reasoning looks only at the probability ratio of the current token, even though the prefix the token is conditioned on was also generated by the policy. This mismatch matters most for long responses and off-policy training. The authors model autoregressive generation as a finite-horizon Markov decision process and show that the cumulative likelihood ratio of the generated prefix is the exact importance weight for each policy-improvement term. They then build a prefix-level trust region on the exact reverse-KL quantity r−1−log r, which gives EPIC, a sequence-clipping method. In off-policy RL on a sparse mixture-of-experts model, EPIC consistently outperforms standard and sequence-level clipping baselines on mathematical-reasoning benchmarks and stays competitive with routing-replay methods, without needing expert-routing replay.

arXiv:2610.29999 PDF

Asymmetric On-Policy Distillation Learns In-Context Exploration

Sriyash Poddar, Yanda Bao, Jacob Krantz, Matthew Chang, Xavier Puig, Roozbeh Mottaghi et al. cross-listed Agents acting under partial observability must explore to uncover hidden information. Learning that behavior with online reinforcement learning is sample-inefficient, and offline imitation of a privileged expert averages its actions across hidden states and never learns to explore. ASTEROID instead lets a history-conditioned student gather on-policy context, then trains it to imitate the privileged expert's actions conditioned on that context, so the student learns to search while uncertain and exploit once the hidden state is revealed. Across eleven environments, including robotic manipulation, navigation, and Procgen, it achieves about 2x higher returns under fixed exploration budgets with about 100x fewer interactions than RL baselines, and in sim-to-real robot manipulation without vision it raises success by 40% over standard distillation.

arXiv:2610.30105 PDF

6 more specialized papers

Robotics 20

HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second

Jonas Eschmann, Giuseppe Loianno cross-listed HyperDrone is a batch quadrotor simulator that runs both multirotor dynamics and rendering entirely on the GPU, and the authors report it as more than 20x faster than existing state-of-the-art batch quadrotor simulators. Its new ray-tracing engine, HyperLux, supports OptiX, Vulkan, Metal and WebGPU backends. It renders RGB, depth, segmentation, surface normals and optical flow, with optional reflections, motion blur and Physically Based Rendering (PBR), at millions of frames per second on consumer GPUs. To test transfer to real hardware, the authors train a vision-based yaw-control policy purely in simulation and deploy it fully onboard a real Crazyflie quadrotor with a camera.

arXiv:2610.28950 PDF

SLoSa: Small, Low-Power, and Safe Foundation Models for Robotic Task Planning

Weizhe Xu, Jialiang Fan, Essien Taylor, Zijun Wang, Russell Joseph, Fanxin Kong cross-listed SLoSa is a framework for building small, low-power robotic task planners that respect safety constraints, combining the flexibility of large language models (LLMs) with the guarantees of classical symbolic planners. It builds a safety-aware dataset using classical planners, a teacher LLM and formal verifiers. A compact model is then trained with supervised fine-tuning followed by reinforcement learning, guided by a verification-based reward machine. Across four Planning Domain Definition Language (PDDL) domains, this raises the success rate of 4B-14B planners to 70.5-76.0%, and INT4 deployment on a Jetson Orin draws about 18-26 W. Household simulators and a desktop robotic arm are also used to show the framework applies across settings.

arXiv:2610.28955 PDF

PHD-Hand: A Printable Humanoid Dexterous Hand with Linkage-Driven 3-DOA Digits

Neel Adwani, Muhaiminul Islam Akash, Saffia Ouali, Ethan Liu, Cong Wang cross-listed Tendon-driven dexterous robot hands are expensive, hard to maintain and short-lived, and linkage-driven alternatives struggle to fit three degrees of actuation (DOA) into each finger. PHD-Hand uses a novel flexure hinge and spring-steel elastic drive rods to simplify the linkage drivetrain, avoiding most miniature moving parts. As a result, the hand can be fabricated on a budget 3D printer and sized close to a human hand. A prototype is evaluated on grasping and object manipulation, alongside new results on the hand's kinematics and control.

arXiv:2610.29030 PDF

SAFTER: Safety-Aware Filtering Through Explicit Recoverability for Robust Reinforcement Learning under Distribution Shift

Benedictus C. G. Cinun, Bin Hu cross-listed Safety filters based on control barrier functions (CBFs) usually enforce only instantaneous safety, so a robot with limited actuation can still drift into states it cannot recover from. SAFTER embeds a differentiable, closed-form backup-CBF filter inside Soft Actor-Critic training. One certificate covers current safety, finite-horizon backup safety, terminal recoverability and input limits, and it yields an exact minimum-intervention projection. In quadruped navigation it learns with about 30% fewer transitions than baselines, improves success in unseen environments by up to 61.4 points over SEA-Nav, and reaches 73-77% success on two real-world maps. Ablations show that differentiating accurately through the filter is critical: replacing its Jacobian with the identity drops high-shift success from 61.4% to 6.2%.

arXiv:2610.29149 PDF

How Representational Priors Shape Generalization in World Action Models

Junha Chun, Youngjoon Jeong, Taesup Kim cross-listed World action models (WAMs), which build robot policies on top of generative image or video backbones, generalize very differently depending on which backbone they use. The authors explain this through representational geometry, introducing nuisance orthogonality: the degree to which a backbone's features for out-of-distribution visual variations are orthogonal to the in-distribution manifold. Across image and video backbones, higher nuisance orthogonality before policy training predicts stronger robustness to visual variation afterward. Robust WAMs also suppress the feature displacement caused by such nuisances, giving a way to anticipate policy generalization from the choice of backbone.

arXiv:2610.29168 PDF

A Decade of Learning-Based Control: From Theoretic Foundations to Practical Deployment

Angela P. Schoellig, Matthias A. M\"uller, Sebastian Trimpe, Melanie N. Zeilinger, Mahathi Anand, Andrea Carron et al. cross-listed This tutorial marks ten years of Learning-Based Control sessions at the IEEE Conference on Decision and Control (CDC) and reviews methods that combine machine learning with the theoretical structure of control. It starts with the core ideas behind stability and safety guarantees, then covers learning dynamics models for control under uncertainty and learning policies directly with reinforcement learning, imitation learning and Bayesian optimization. It closes with representative applications and open challenges, and ships code implementations and worked examples as an entry point to the field.

arXiv:2610.29187 PDF

Watch Your Step: Anticipated Perception for Humanoid Locomotion

Matthias Heyrman, Chenhao Li, Ioannis Dadiotis, Marco Hutter cross-listed Perceptive humanoid locomotion policies usually take in a dense elevation map around the robot, even though only the terrain near the next footsteps matters for control. The authors train a lightweight foothold prediction model (FPM) that predicts upcoming footholds from proprioception and the velocity command alone. The policy then sees only small elevation-map patches around those predicted footholds, and the FPM is periodically retrained as the gait matures. The policy reaches a given terrain difficulty in about a third of the training iterations needed with the full map, does better where footholds are sparse and on unseen terrain, and matches a far costlier attention-based encoder. It transfers zero-shot to a Unitree G1 humanoid indoors and outdoors.

arXiv:2610.29612 PDF

DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving

Alessandro Canevaro, Julian Schmidt, Hang Yu, Ye Jin, Sajad Marvi, Georg Martius et al. Imitation learning for autonomous-driving planners suffers from distribution shift and compounding errors in closed-loop execution, while reinforcement learning (RL) usually depends on handcrafted rewards and large amounts of simulation. The authors propose Expert-Guided RL (EGRL), which uses RL purely as an optimizer with a reward based on trajectory distance to logged expert driving. A controlled study finds that the training procedure, more than model size or complexity, drives much of the performance gain. A lightweight MLP policy trained this way, DriveMLP, scores 86.51 on nuPlan Val14 Reactive, 3.2 points above the previous best imitation-learning planner, with more than 15x faster inference.

arXiv:2610.29676 PDF

Toward Recursive Self-Improvement in Physical AI

Meng Luo, Zechen Bai, Shanqing Xu, Yanlin Li, Kaixin Li, Mengchen Qian et al. cross-listed This survey brings recursive self-improvement (RSI), where a system improves both its behavior and the process by which it improves, from code-rewriting digital agents into robot learning. In the physical world this is harder because trials are costly, observations are partial, and mistakes can be irreversible. The authors argue that gains carry forward only when feedback targets remaining learning needs, the receiving component can use the change, the value survives task or hardware shifts, and continued practice stays safe and affordable. Existing studies, mostly in manipulation, show behavioral and procedural improvement, but none yet isolate gains attributable to recursion itself.

arXiv:2610.29677 PDF

Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies

Johannes Keller, Albrecht M. Blo{\ss}e, Adrian Lindenmeyer, Thomas Neumuth, Daniel Schneider cross-listed Vision-language-action (VLA) robot policies need to recognize when they are outside their training experience, but ensemble disagreement requires training and running several policies. The authors attach post-hoc kernel-variance heads that estimate epistemic uncertainty from a single policy's embeddings using Gaussian-process conditional variance, with no gradient updates and no change to the policy. With SmolVLA on LIBERO-10, the heads match two-policy ensembles for predicting task success, active demonstration acquisition, and referral to a stronger supervisor, while cutting uncertainty-scoring time by up to about 50x. The authors also find that the standard LIBERO-10 protocol mainly rewards detecting coverage differences between tasks, and they add controlled perturbations as a complementary evaluation.

arXiv:2610.29730 PDF

SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation

Zhaohui Yang, Quanbao Lin, Chengqi Shi, Zhenglang Weng, Jicheng Lin, Boyu Chen et al. cross-listed Humanoid loco-manipulation needs an interface between high-level task policies and whole-body control, and using one action representation for both either overloads the policy or underspecifies coordination. SP-HLM uses five task-critical sparse body poses together with a learned latent motion prior. A vision-language-action model, SP-VLA, predicts the sparse poses and finger joints from observations and language. A distilled motion generator, SP-MGM, reconstructs full-body motion in real time, and a general motion tracker, SP-GMT, combines the explicit targets with the latent prior for control. In simulation and on a real humanoid, the system improves motion tracking, whole-body coordination, and loco-manipulation success over existing approaches.

arXiv:2610.30044 PDF

From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks

May Espinola, Qifan Zhang, Yu Xiang cross-listed The authors test whether general-purpose frontier models can control a robot with no task-specific training, using a closed-loop framework on the LIBERO-10 benchmark. Models see images and proprioception and act through structured tool calls that set end-effector poses and gripper commands. Across six models (GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5, Claude Opus 5.5, and Gemini-4 Argon) at three reasoning-effort levels, giving 1,800 episodes, the best success rate is 44% (GPT-6 Astra), closely followed by Gemini-4 Argon at 43%. More reasoning effort generally helps but varies by model and costs more inference; models handle object placement relatively well and struggle with precise spatial arrangements and multi-step interactions.

arXiv:2610.30051 PDF

AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation

Likui Zhang, Zhe Zhu, Zhihao Zhan, Yibo Bai, Hanlin Wang, Dongwei Li et al. cross-listed Reinforcement learning (RL) fine-tuning of vision-language-action (VLA) models on long-horizon manipulation suffers from sparse trajectory-level rewards, which blur credit assignment, and from uneven mastery across subtasks. AtomicRL splits tasks into atomic subtasks. Its SAPO component gates temporal-difference residuals so that failures in one subtask do not contaminate credit in others, and its BATE component steers sampling and exploration toward the subtasks the policy has not yet mastered. A Real2Sim2Real pipeline transfers the trained policies to physical robots. Compared with PPO, it gains up to 9.5% on CALVIN ABC-D and 3.6 to 4.0% on LIBERO-LONG, and it improves real-world task performance by an average of 21.3%.

arXiv:2610.30097 PDF

7 more specialized papers

Multimodal 14

Qwen-Audio-3.1-ASR Technical Report

Chuanmeng Bian, Daren Chen, Peixin Chen, Zhigao Chen, Chong Deng, Zhiyun Fan et al. cross-listed Qwen-Audio-3.1-ASR is a family of audio-language models that improves automatic speech recognition (ASR) over its predecessor and adds speaker-diarized transcription and general audio understanding. It supports 30 languages and 16 Chinese dialects, plus instruction-controlled features such as conditioning on prior context, hotword customization and transcript polishing. Diarized output labels speakers consistently across turns, with timestamps, and represents overlapping speech, while the audio-understanding variants handle captioning, sound event detection, temporal grounding and audio question answering. The family ships separate variants for file transcription, low-latency streaming, interactive voice input and non-streaming use, and is evaluated across standard, multilingual, dialectal, diarized and general audio benchmarks.

arXiv:2610.28711 PDF

Cache Locally, Recompute Selectively for Streaming Video Understanding

Junhyeok Lee, Sangyoon Lee, Jaeho Lee cross-listed Streaming video understanding (SVU) systems cut answer latency by encoding incoming frames into key-value (KV) caches while idle, but they limit the context available during this prefill, so the cached tokens can miss context the eventual query needs. FrameBlend is a training-free, plug-and-play method that, at query time, re-encodes selected video tokens with query-relevant context and swaps their recomputed KVs into the cache. Because of this recomputation, idle-time prefill can use shorter attention windows, which lowers cost and also improves accuracy. Across three SVU frameworks and five benchmarks, accuracy rises by up to 3.9 percentage points, and on RVS-Ego per-frame prefill FLOPs fall by over 20% and total processing time by 14–59% for retrieval-based frameworks, with query-time recomputation included.

arXiv:2610.29229 PDF

SpatialBrain: Spatial Intelligence Growing from Experience

Yifan Zhou, Ziyang Gong, Xiaojun Ye, Yifei Liu, Donggang Li, Xiaolong Zhou et al. Spatial reasoning requires deciding when direct observation is enough and when to reconstruct a scene, take measurements, or switch reference frames. SpatialBrain accumulates experience about these decisions while keeping its vision-language backbone frozen. It combines three parts: 3DFusionMemory, which turns scene observations into queryable spatial relations; a Spatial Skill Library of reusable tool procedures; and a Spatial Cognition module that links task conditions to past experience. Completed trajectories are abstracted into process records between task blocks, and the framework improves over the same backbone's direct answers on four of five spatial-reasoning benchmarks and ties on the fifth.

arXiv:2610.29317 PDF

Asking What a Director Sees Can Change What Vision-Language Models Choose

Jiaxuan Guo, Xingyu Ren, Kejia Zhang, Jingxin Yang, Youran Sun, Haizhao Yang Theory-of-mind evaluations of vision-language models often ask the model to report another person's perspective before acting, and they treat a correct report followed by a wrong action as a failure to use that perspective. In a rendered Director Task, where the largest block is hidden from the director, first asking which blocks the director can see raises hidden-block choices by 12 to 88 percentage points on six GPT models and by 17 points on Qwen3.8-27B. This happens even though the models mostly report the director's view correctly. Reasoning traces suggest the question shifts the model's reading of the request toward its own view, and stating explicitly what the director knows almost eliminates the effect. The authors recommend that perspective-taking benchmarks include paired prompts that differ only in the report question.

arXiv:2610.29318 PDF

Omnilingual-MsASR: Benchmarking Multimodal, Multilingual, Long-Form Multi-Speaker ASR of Omni Large Language Models

Yuan Ge, Pei Zhang, Xize Cheng, Lingzhou Mu, Mingyang Wang, Jin Xu et al. Introduces Omnilingual-MsASR, a benchmark for long-form multi-speaker automatic speech recognition (MsASR) aimed at omni-modal large language models. It covers six languages, recordings up to 30 minutes, and up to 20 speakers, and supplies parallel audio and video so that cross-modal consistency can be measured. Evaluations of leading omni-modal LLMs show that their tcpWER error rates frequently exceed 100%. The results point to fine-grained timing, telling speakers apart, and combining audio with video as the main open problems.

arXiv:2610.29455 PDF

pplx-embed-v2-late Technical Report

Markus Krimmel, Louis Milliken, Ekaterina Glazkova, Maximilian Schall, Antoine Chaffin, Sedigheh Eslami et al. cross-listed pplx-embed-v2-late is a pair of multilingual, multimodal late-interaction retrieval models with 0.6B and 9B parameters, both distilled from an 18B teacher. Both achieve state-of-the-art results for their size on text, image, and visual document retrieval benchmarks, and the 0.6B model is competitive with much larger models on visual documents. Because the two models share an embedding space, documents can be indexed with the 9B model and queried with the faster 0.6B model. The models are publicly released.

arXiv:2610.29523 PDF

8 more specialized papers

Reasoning 12

Breaking Diffusion: Parallel Reasoning with Draft Refinement

Jaesang Yu, Seoyeon Byun, Yun Wang, Yan Lu, Juho Kim cross-listed Repeated diffusion denoising does not guarantee that individually plausible token updates add up to a better answer. Draft Refinement Language Models (DRLMs) address this by using one denoiser both to propose parallel revisions of a complete draft and to score candidate drafts. The denoiser is trained to reconstruct every token from uniformly corrupted drafts. To score a draft, the method corrects for the evidence each token gives about itself, yielding a single-pass "cavity energy", and a revision is kept only if it lowers that energy. On Countdown, Sudoku-Extreme, and Maze-Hard, the method beats autoregressive and discrete diffusion baselines, and a 6M-parameter DRLM reaches 93.9% on Sudoku-Extreme and 44.4% on Countdown-5.

arXiv:2610.26542 PDF

ReAPO: Retry-Aware Policy Optimization for Efficient Exploration in LLM Reasoning

Nghi Dao, Calder Morton-Ferguson, Noah Golowich, Adam Block Reinforcement learning with verifiable rewards (RLVR) improves LLM reasoning, but independent sampled rollouts often repeat the same failed strategies. Retry-based methods, which put earlier wrong answers in context for the next attempt, raise the question of how to credit each attempt. The authors frame this as a Stochastic Shortest Path problem, which leads to ReAPO, a small modification to GRPO that spends extra generation on failed attempts and rewards successful retries without rewarding the failures before them. On Qwen3 and OLMo-2 models from 0.6B to 4B across eight math and science benchmarks, it improves first-attempt accuracy over GRPO, and on Qwen3-0.6B it matches GRPO's peak accuracy with 31-39% fewer generated tokens than versions using double the group size or training steps.

arXiv:2610.28770 PDF

TeachBack: Learning to Teach via Bidirectional Preference Optimization for On-Policy Self-Distillation

Linfei Ji (Huazhong University of Science and Technology), Jiayun Li (Tsinghua University), Haozhao Wang (Huazhong University of Science and Technology), Siyu Jiang (Huazhong University of Science and Technology), Zehao Wang (Harbin Institute of Technology), Longtao Xiao (Huazhong University of Science and Technology) et al. In on-policy self-distillation, a model learns from a copy of itself that has privileged information such as the reference solution. That copy produces accurate trajectories, but they tend to be shortcut derivations that skip the exploration and self-correction a student needs. TeachBack also trains how the teacher teaches, varying how much of the reference reasoning is injected into the prompt and the prompt style so that the teacher re-solves problems rather than copying. A bidirectional preference objective uses teacher trajectories to improve the student and student behavior to shape future teacher outputs, and it consistently outperforms existing self-distillation methods on math and logical reasoning tasks.

arXiv:2610.29167 PDF

ARCANA: A Benchmark for Abstraction and Analogical Reasoning

Ruqi Chen, Zilei Wang, Wei Wang Few benchmarks explicitly test far analogical reasoning, which maps between domains that look different on the surface but share deep structure. ARCANA (Abstraction and Reasoning Corpus of ANAlogy) asks solvers to map real-world objects or physical phenomena onto abstract grid transformations, unlike ARC-AGI, which mainly tests innate human priors. A human study shows people solve the tasks reliably, while a range of LLMs fall clearly short of humans and show systematic failure modes. The authors also propose ConceptSIR, a reasoning framework based on concept slippage, introspection, and reflection, which lets standard LLMs solve several tasks they failed with direct prompting.

arXiv:2610.29271 PDF

From Prompts to Posteriors: Can LLM Agents Perform Bayesian Inference in Context?

Linus Filbry, Luckeciano C. Melo, Yarin Gal Many scientific uses of large language models (LLMs) require updating beliefs as evidence accumulates, which Bayesian inference formalizes. The authors build an evaluation framework that tests LLMs' probabilistic coherence, their ability to recover externally computed posteriors, and the downstream cost of updates that do not follow Bayes' rule. Updating beliefs in context, LLMs generally violate Bayes' rule. With modern reasoning abilities and code execution, their performance improves drastically, to the point of closely approximating analytically intractable posteriors.

arXiv:2610.29331 PDF

Parallelizable Probabilistic Latent Reasoning for Diverse Reasoning Trajectories

Seokhyun Lee, Seongjun Lee, Changhee Lee Reasoning in a continuous latent space could explore alternative solution paths more efficiently than verbal chain-of-thought (CoT). However, existing latent models are mostly deterministic, and stochastic versions train unstably and slowly. Probabilistic Latent Reasoning Models (PLRM) use a JEPA-style variational objective in which a posterior infers latent distributions from observed reasoning trajectories in parallel and a prior predicts the next latent with teacher forcing. A discrete strategy variable with a mixture prior keeps distinct reasoning strategies separate rather than averaging them into one Gaussian. PLRM matches deterministic latent models on single-trajectory data and consistently beats stochastic baselines on multi-trajectory datasets while training efficiently.

arXiv:2610.29385 PDF

A Strong MoE Is All You Need for Process Supervision

Guanzhi Deng, Zengyan Liu, Haibo Wang, Kuan Wu, Shing Yin Wong, Sichun Luo et al. Process reward models (PRMs) score each step of a model's reasoning, but reliable step-level labels are expensive to obtain. The authors show that expert-routing statistics in Mixture-of-Experts (MoE) models give a built-in step-confidence signal that is competitive with confidence derived from output logits. They use it in two methods. RoCo-Rethink regenerates the reasoning from the least confident step onward, and RoCo-PRM uses routing-derived scores from a strong MoE to label reasoning traces and distills them into a standalone PRM. Both outperform strong baselines across MoE architectures and reasoning tasks, and RoCo-PRM transfers across architectures and tasks.

arXiv:2610.29498 PDF

Parallel Thinking for Large Language Model with Path Diversity

Jiaru Qian, Yihong Dong, Xue Jiang, Chun-Wah Hsu, Zhi Jin, Ge Li Parallel thinking has a language model explore several reasoning paths at once. The authors find that when this is trained with reinforcement learning (RL), the paths tend to collapse into near-copies of one another. Diverse Parallel Thinking (DPT) adds an RL reward that penalizes paths for sharing similar prefixes, pushing the model to try distinct strategies. Across multiple reasoning benchmarks, DPT outperforms previous parallel-thinking methods, supporting the claim that path diversity is a key ingredient. Code is released.

arXiv:2610.29501 PDF

Keeping Time vs. Understanding Time: Neither Alone Explains Temporal Reasoning in LLMs

Auss Abbood, Fernando Mart\'inez-Plumed, Jos\'e Hern\'andez-Orallo, Nigel Collier Temporal reasoning benchmarks usually divide tasks into temporal arithmetic (date and time calculations) and temporal semantics (temporal and causal relations), but this split had never been tested directly. Fine-tuning on one benchmark's subset for either mode improves that dataset by 6.8 to 23.1 percentage points, yet transfer to another dataset of the same mode ranges only from −3.2 to +1.1 points. An analysis of which cognitive demands drive model performance finds that knowledge, not quantitative or logical reasoning, is the main driver. Current benchmarks therefore largely reward models for remembering how time typically works.

arXiv:2610.29507 PDF

Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding

Shin'ya Yamaguchi With chain-of-thought (CoT) prompting, a language model's final answer may rely on shortcuts in the question rather than on its own reasoning steps. The authors apply context-aware decoding, an existing contrastive decoding method, as Rationale Contrastive Decoding (RCD). RCD compares next-token predictions made with and without the rationale and boosts tokens that the rationale makes more likely. They interpret the rule as a KL-regularized reward-maximization objective. Across five reasoning benchmarks, RCD improves accuracy over standard CoT without any training, and swapping in correct or random rationales shows that answers depend more on the rationale content.

arXiv:2610.29518 PDF

Did the Problem Actually Get Harder? Testing LLMs on Paired P and NP-Complete Problems Hardened by Adversarial Multi-Agent Rewriting

Nurul Labib Sayeedi, Md. Faiyaz Abdullah Sayeedi, Zabir Al Nazi, Mahbub E Sobhani, Shubhashis Roy Dipta Many NP-complete problems have a polynomial-time twin that differs in a single detail, such as 3-colouring versus 2-colouring a graph. These pairs make it possible to test whether a model is reasoning about a problem or recalling it. P2NPCBench pairs NP-complete and polynomial-time problems across 12 families, and AgentGAN, an adversarial multi-agent framework, rewrites each statement until frozen code models fail while an executable check confirms that the answer and complexity class are unchanged. Across 12 models, every model scores lower on the NP-complete statements (35.4% vs. 63.8% pass@1), and a tenfold time limit barely helps because the failures come from wrong or crashing programs. Rewriting the problems narrowed the gap between the two classes from 24.8 to 14.6 points, which suggests that part of the apparent gap reflects familiarity with textbook phrasing.

arXiv:2610.30103 PDF

LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data

Zhennan Wan, Jianfei Chen Reinforcement learning (RL) can improve long-context reasoning in large language models (LLMs), but it needs long, diverse, challenging and verifiable training data, which is costly to annotate and hard to synthesize without external corpora or stronger teacher models. LongZero has the model write its own tasks: it first produces a question, an answer and the supporting evidence in a compact structured form, then separately expands that evidence into a long multi-document context. These self-written multi-hop reasoning tasks are combined with self-written multi-round retrieval tasks and used for training with GRPO. On two LLMs across five long-context reasoning benchmarks, LongZero outperforms RL on data synthesized from external corpora while using no external data, and it generalizes to real tasks.

arXiv:2610.30132 PDF

Unclassified 5

MemoryTax: The Bitter Lesson of Agent Memory

Zedong Peng No summary available — see the abstract on arXiv.

arXiv:2610.29823 PDF

The role of locality and weight sharing in convolutional neural networks

Francesco Mori, Melanie Weber No summary available — see the abstract on arXiv.

arXiv:2610.30118 PDF

Self-normalized Bernstein Inequalities for Regenerative Processes and Applications to Learning Without Independence

Imon Banerjee, Pierre Alquier No summary available — see the abstract on arXiv.

arXiv:2610.30119 PDF

A Quasipolynomial Algorithm for $\ell_\infty$ Contractions

Ashkan Soleymani, Gabriele Farina cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.30122 PDF

PhysiCalWorld: In-Context Physics Calibration for World Models

Yunhao Luo, Wentao He, Yinpei Dai, Hyelin Nam, Hongze Fu, Joyce Chai et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.30125 PDF