AI Paper Digest

Friday, October 9, 2026

611 papers cs.AI · cs.LG · cs.CL ← 2026-10-082026-10-12 →

Aug Sep Oct Nov

Highlights

V-CoLA: Vision Token Compression with Linear Attention

Highlight HF pick · 7▲Multimodal Hao Jiang, Yiru Mao, Tianpeng Bu, Hao Zhou, Hongtao Duan, Wang Jing et al. Vision tokens make up most of a vision-language model's (VLM's) input, but token compression methods built for softmax attention degrade on hybrid architectures with linear attention such as Qwen3.5. V-CoLA is a training-free compression framework for linear attention. It combines a uniqueness-aware importance criterion with adaptive token merging and is implemented to stay compatible with chunk-wise parallel linear attention. It keeps 99.5% of original performance with half the vision tokens and over 88% with 12.5%, while speeding up prefill by 1.86x to 6.15x.

Hybrid VLMs such as Qwen3.5 mix softmax and linear attention, and on them existing vision-token compression methods like FastV and DART do no better than dropping tokens at random. The authors trace this to the bounded recurrent state of Gated DeltaNet, then turn that state's selective retention into a training-free importance signal for pruning and merging vision tokens.

  • Each token's importance combines two scores: how well the final state reconstructs it, which measures its long-term contribution, and how much it changes nearby pseudo-query readouts, which measures its uniqueness; the scores are weighted 0.9 and 0.1.
  • Tokens are then merged rather than pruned: the sequence is split into chunks of roughly equal total importance, so important regions get finer chunks, and each chunk is softmax-weighted into one token. All vision tokens are also dropped after layer 24, where removing them leaves perplexity and accuracy essentially unchanged.
  • On Qwen3.5-9B across eight benchmarks, the method keeps 99.5% of the average score with 50% of vision tokens, 94.8% with 25% and 88.0% with 12.5%, beating VisionZip (99.0%, 94.5% and 86.4%) and the Mamba-oriented DTP (94.2%, 86.3% and 75.7%).
  • Language-model prefill runs 1.86× to 6.15× faster, and the merging step costs only about 0.86 ms at 8K tokens; the paper puts total overhead at anywhere from 0.1% to 1.6% of the forward pass, depending on the section.
  • The lead is in averages, not every benchmark: VisionZip wins on TextVQA at 75% compression and on MME at 87.5%, the margins on Qwen3.5-27B and InfiniteVL are small, and pure-Mamba or fully linear VLMs, along with linear-attention variants other than Gated DeltaNet, are not tested.

arXiv:2610.11251 PDF

SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces

Highlight HF pick · 6▲Multimodal Rongxue Li, Meng Yang, Yiru Mao, Yongliang Tao, Lulu Hu, Bin Yang et al. Spatial coding agents improve spatial reasoning in Multimodal Large Language Models (MLLMs) by running external tools to produce verified execution traces, but they are slow at inference time and depend on those tools. The authors observe that giving an MLLM summarized agent traces unlocks its own internal spatial chain-of-thought. SpatialOPSD builds on this with on-policy self-distillation that treats verified agent traces as privileged information, producing a standalone tool-free model. A Repetition-Aware Distillation technique, which combines repetition masking with unlikelihood regularization, keeps the privileged information from leaking into outputs. SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and out-of-distribution benchmarks.

Spatial coding agents reason well about 3D scenes by running geometry tools, but every query pays for a slow, tool-dependent execution loop. SpatialOPSD builds that skill into a standalone MLLM through on-policy self-distillation: the teacher is the same model shown summarized, verified agent traces as privileged information, and the student sees only the raw images and question.

  • SpatialClaw traces are kept only when the agent's final answer matches ground truth (5,008 of 10,000 MindCube training items), then condensed into answer-scrubbed summaries; this summary format beat full traces, extracted facts, and strategy-only intents (51.9 vs ~50.3 average).
  • To keep the student from copying trace content it won't have at inference, Repetition-Aware Distillation uses 4-gram matching to flag spans of 32 or more tokens that repeat earlier output or transcribe the privileged text, then masks the KL loss on those tokens and applies an unlikelihood penalty to them instead.
  • On Qwen3.5-9B, average accuracy across MindCube-Tiny and ViewSpatial rises from 45.9 to 51.9, beating SFT (47.6) and GRPO (48.2) and recovering 91% of the tool-using teacher's 57.3 in a single tool-free pass, with gains of 2.2–3.8 points at 2B, 4B and 27B scales.
  • Privileged-information leakage falls from 82.7% to 9.1% compared with vanilla OPSD, and the out-of-domain average (BLINK, OmniSpatial, MMStar, Video-MME) holds at 64.7 vs 64.4 for the base model, while SFT collapses to 38.4.
  • The leakage fix costs some accuracy, since masking alone reaches 53.0 average but leaks 73.6%; training uses only MindCube, ViewSpatial gains are modest (+3.4), and the model still trails both the agent and Gemini-3-Pro.

arXiv:2610.11366 PDF

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Highlight HF pick · 34▲Reinforcement Learning Core Team, Zongming Qiao, Ziyue Hua, Zirui Ou, Zihao Yue, Zihan Jiang et al. The MiMo-V2.6 technical report describes an omni-modal model family improved by scaling reinforcement learning (RL) compute after multimodal mid-training on a hybrid sliding-window-attention architecture. RL is scaled along three axes. Asynchronous training consumes 1,568 samples and 2.7 to 3.7B tokens per step at contexts up to 1M. Environments are more diverse, spanning code, general, visual, and cyber tasks across several agent harnesses. Groupwise agentic grading gives more accurate rewards on long-horizon tasks and pushes the model toward shorter solutions. Training is stabilized by freezing the mixture-of-experts router and layering defenses against reward hacking, and the authors open-source the training dynamics, RL environments, and the RL framework.

Xiaomi's MiMo-V2.6 report argues that the way to push foundation models toward self-improvement is to scale reinforcement learning compute along three axes: bigger batches, more varied agentic environments, and more compute spent on grading. They test this on two omni-modal Mixture-of-Experts models, MiMo-V2.6-Pro (1.02T total parameters, 42B active) and MiMo-V2.6-Flash (310B total, 15B active), both built on a hybrid sliding-window/global-attention backbone.

  • Training is fully asynchronous, with 1,568 prompts × 16 rollouts per step, which works out to about 25K sequences and 2.7–3.7B tokens per step at contexts up to 1M; to keep this stable at scale, they freeze the MoE router, switch to a Muon variant (Muown) during mid-training, and align routing and sampling between training and inference.
  • RL cost $2.6M for Pro and $0.9M for Flash, and over that run DeepSWE avg@3 rose steadily from 58.4 to 72.6 (Pro) and 48.7 to 65.7 (Flash), with similar gains reported on general-workflow (AutomationBench), visual-coding, and cybersecurity tasks.
  • Instead of binary test-pass rewards, the grader compares the rollouts within each group: Groupwise Reward Synthesis multiplies the test reward by rubric scores for solution quality and agent behavior, while Groupwise Advantage Redistribution moves positive advantage toward the higher-quality passing patches, which keeps turn counts and token length from blowing up and yields smaller, better-scoped patches, at only 12.7% of total RL compute.
  • Defenses against reward hacking include cleaning solution artifacts out of environments, network isolation, repeated rounds with a dedicated "hack agent" until no exploit remains, and zeroing out confirmed hacks during training, which together keep the detected hacking rate below 2% for the whole run.
  • Limitations: several headline benchmarks are in-house (MiMo Visual Coding, MiMo Cyber Bench), the advantage-redistribution ablation covers only Flash with a small batch of 128, and the material provided does not compare against other frontier models; the authors open-source a distilled MiMo-V2.6-Distill-Qwen-9B, the RL environments, and the RL framework.

arXiv:2610.11959 PDF

VibeEdit: Image Editing with Canvas Instructions

Highlight HF pick · 9▲Vision Jinjing Zhao, Fangyun Wei, Yitong Wang, Xiuyu Wu, Yunuo Chen, Yang Yue et al. In text-guided image editing, describing which object to change is awkward when several objects look alike. VibeEdit replaces the text prompt with a "canvas instruction": spatial marks and optional short notes drawn directly on the image, which together specify where to edit and what to change. It is built by adapting Qwen-Image-Edit with layer-decoupled conditioning, trained on 1.55 million source-target edit pairs using region-weighted supervised fine-tuning, and then refined with rubric-guided reinforcement learning. On a human-curated benchmark of 419 cases that focuses on choosing among similar objects, it scores 79.9 on a VLM rubric with 32.8 dB PSNR outside the edited region, against 67.4 and 24.0 dB for the strongest text-instructed baseline, FireRed.

Text-guided image editors make users describe both the change and its target, and that gets cumbersome when several similar objects appear in a scene. VibeEdit replaces the text prompt with a canvas instruction: users draw circles, scribbles, or arrows directly on the image and can add short handwritten notes, and the editor performs addition, removal, replacement, attribute changes, and movement from these marks alone.

  • The model adapts Qwen-Image-Edit with layer-decoupled conditioning, in which the frozen Qwen2.5-VL encoder reads the annotated image while the VAE separately encodes the clean source and a gray-background rasterization of the marks, so the annotations guide the edit without being copied into the output.
  • Training uses 1.55M synthetic source–target pairs built from Florence-2, SAM 3, ObjectClear, FLUX.1 Fill, and Qwen-Image-Edit, with canvas marks rendered online, followed by region-weighted SFT (λ=1.5 on dilated edit and annotation masks) and DiffusionNFT RL on 3,520 high-reward-variance conditions, using a GPT judge's rubric rewards for edit success, outside-region preservation, and local quality, plus a PSNR penalty.
  • On a human-curated 419-case benchmark built around choosing the right object among similar ones, VibeEdit reaches a 79.9 VLM rubric score and 32.8 dB outside-region PSNR, compared with 67.4 and 24.0 dB for the best text-instructed baseline (FireRed-Image-Edit-1.0), which was given detailed 21-word prompts, and it scores highest on all five edit types.
  • Ablations show that the decoupled VAE inputs matter most (a rubric score of 24.7 when encoding only the source, versus 67.8 when encoding source and marks separately), and that RL adds roughly 12 points over the SFT model, with all three reward groups needed for the best result.
  • Training data, RL rewards, and evaluation all come from GPT models (the judge agrees with human raters on 93% of questions, κ=0.805), and the benchmark is the authors' own, so the results depend partly on model-generated supervision and scoring.

arXiv:2610.12229 PDF

From Prompting to Composing: A Spatial Canvas Interface for Poster Generation

Highlight HF pick · 6▲Vision Yitong Wang, Fangyun Wei, Jinjing Zhao, Sirui Zhang, Hongyang Zhang, Dong Chen et al. Text prompts force users to describe a two-dimensional poster layout as a one-dimensional string of words. The proposed Spatial Canvas Interface lets users place elements directly in space using four binding types (semantic, identity, text, and pixel), plus text specifications for individual elements and overall appearance. Compo, adapted from a pretrained image editing model and trained with an automatically constructed supervision pipeline, takes these canvases as input; in an agentic mode, it first turns a high-level request into a planned canvas. On a new benchmark that tests adherence to each binding type and their combinations, Compo shows stronger compositional control than general-purpose image generators and dedicated poster systems while keeping visual quality high.

Text prompts make users squeeze a two-dimensional poster layout into a one-dimensional string. Compo replaces the prompt with a Spatial Canvas, where users place four kinds of region-level bindings directly in space: semantic (a concept for a region), identity (a reference image to preserve), text (an exact string), and pixel (content kept as-is). Short Text Specifications sit alongside the canvas to describe each element and the overall style.

  • Compo adapts a pretrained image-editing MMDiT (LongCat-Image-Edit or Qwen-Image-Edit-2511) with LoRA on both the text and image branches, trains it on the synthetic Compo-200K dataset, then refines it with DiffusionNFT reinforcement learning using rewards for text accuracy, instance similarity and layout IoU; an optional GPT-based agent can turn a plain-language request into a planned canvas.
  • On a new 500-sample benchmark, Compo-LongCat reaches an average score of 0.752, against 0.504 for the strongest general model (HiDream-O1-Image) and 0.447 for the best dedicated poster system (PosterOmni); overall IoU is 0.885, against 0.628 for PosterMaker.
  • Text rendering improves the most: sentence accuracy rises to 0.940 and normalized edit distance to 0.970, against 0.329 and 0.515 for CreatiDesign, the strongest baseline on text.
  • In ablations, removing all reinforcement-learning rewards drops the average score to 0.725 and sentence accuracy to 0.886, and each reward mainly helps its own binding type; weighting sentence accuracy at 0.75 in the text reward works better than relying on the sparse all-text-correct signal alone.
  • The training data and benchmark both come from the same synthetic pipeline, and the same tools (SAM-3 segmentation and Chandra-OCR-2 text recognition) produce the annotations, the RL rewards and the evaluation metrics, so the gains may partly reflect metric alignment; baselines also had to have their inputs converted, and the extracted text reports no human preference study.

arXiv:2610.12230 PDF

TokenRouter: Efficient Serving System for Token-Level LLM Routing

Highlight HF pick · 54▲Large Language Models Tianyu Fu, Tengxuan Liu, Ruoxi Wang, Yixin Dong, Yi Ge, Yichen You et al. Routing each token, rather than each query, between different LLMs can improve the cost-quality tradeoff, but existing serving systems assume a single model and suffer step desynchronization and batch admission delays under token-level routing. TokenRouter lets developers write routing logic from the point of view of a single request, while the runtime runs a subserver per model and dispatches requests asynchronously. Each subserver uses a delayed-batching scheduler whose hyperparameters are derived from a mathematical throughput model. Across routing algorithms, workloads, and model pairs, it achieves 2.01–64.15x higher decoding throughput than existing systems.

Token-level routing, which picks a different LLM at each decoding step, can beat query-level routing on cost and quality, but single-model serving stacks like SGLang and vLLM handle it poorly. Faster models end up waiting on slower ones, routed tokens wait for the next batch, and wiring routing logic into these engines is hard. TokenRouter lets developers write routing logic from the point of view of a single request, while the runtime runs each model as its own asynchronous subserver.

  • Developers implement three hooks, route, send and receive, to express algorithms such as R2R, CITER, Co-LLM, R-Stitch and multi-model ensembles, and the system exposes one endpoint that can replace a standard single-LLM server.
  • Each subserver runs an extra inter-model loop alongside its normal loops and keeps a handed-off request in a "pending" state, so the request resumes with only a token append instead of repeated prefix matching and KV-cache allocation, which take most of the per-step time in a naive SGLang-based baseline.
  • A delayed-batching scheduler waits until B requests have queued for a model before running a batch, and the throughput-optimal B is chosen with a discrete-time Markov chain model built from routing probabilities, per-model step latency and concurrency.
  • On Qwen3 model pairs (0.6B/32B, plus 8B for the ensemble) across reasoning and SWE-Smith agentic workloads, TokenRouter reaches 2.01–64.15× higher decoding throughput than the stronger baseline, and at concurrency 16 it delivers 18.58× the throughput of official R2R at concurrency 1 while still being faster per user.
  • Routed serving does not always beat simply running the large model alone: Co-LLM reaches 76 tokens/s against 135 for the large model only, and the threshold model assumes geometrically distributed gaps between handoffs, which the authors leave to future work for algorithms that break this assumption.

arXiv:2610.12242 PDF

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

Highlight HF pick · 29▲Vision Dahyun Chung, Siyoon Jin, Hyunwook Choi, Honggyu An, Junyoung Seo, Hyunsung Kim et al. Egocentric world models predict first-person video conditioned on an agent's actions, but most handle a single agent, and existing multi-agent versions use only coarse actions such as locomotion or camera control. ME-World generates synchronized first-person streams for multiple agents interacting through fine-grained actions. It jointly denoises all streams in one shared token sequence, conditions each stream on every agent's pose, and grounds generation in a shared environment memory. The authors also introduce metrics for environment, state-update, and identity consistency, and report improved shared-world consistency, action control, identity preservation, and video quality over existing methods.

Egocentric world models usually simulate only one agent, while existing multi-agent ones control agents with coarse inputs such as discrete commands or camera paths. ME-World generates synchronized first-person video streams for several agents who interact through fine-grained body, hand and head motion, and keeps every stream consistent with one shared, evolving world.

  • Method: all agents' noisy video latents are concatenated into one token sequence and denoised jointly by a fine-tuned Cosmos-Predict2.5 2B diffusion transformer, and each stream is conditioned on skeleton maps of its own motion plus the other agents' poses projected into its view, with head motion encoded as Plücker ray maps.
  • Shared environment memory: past frames from all agents are warped into each target view as aligned scene evidence, and a greedily chosen set of unwarped "anchor" frames is added to the sequence so every stream sees the same appearance references.
  • New metrics: the authors introduce S_env, S_update and S_id, which measure environment consistency only in regions the warped memory doesn't cover, whether interaction-caused changes (e.g., a moved frying pan) agree across views, and whether each agent's identity is preserved.
  • Results: on the real CoMind benchmark, ME-World reaches 0.468 S_env / 0.466 S_update against 0.417 / 0.412 for the best off-the-shelf baseline (GEN3C), raises other-agent pose accuracy (PCK) to 0.881 from 0.514, and lowers FVD (a video-quality distance) to 637.9. It also beats re-implementations of MultiWorld, Solaris, γ-World and MetaWorld that were adapted to share its backbone, data and action conditioning.
  • Limitations: experiments cover only two agents, the off-the-shelf baselines were not fine-tuned, and camera error on the synthetic benchmark is slightly worse than AnyView-DVS. In the ablations, adding shared action conditioning raises FVD (from 554.7 to 651.9), so the full model's video quality is worse than the variant without it.

arXiv:2610.12299 PDF

SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference

Highlight HF pick · 18▲Large Language Models Qitong Wang, Xinwei Niu, Mingluo Su, Shanwei Zhao, Shiai Zhu, Huan Wang Hessian-guided pruning methods for LLMs calibrate on natural text, but during decoding the model reads its own generated tokens, so the pruning objective does not match the activations seen at inference time. Existing speedups also mostly target sparse matrix-matrix products rather than the sparse matrix-vector operations that dominate decoding. SparseDecoding calibrates on activations collected during the dense model's own autoregressive generation, excluding prefill, and adds an optimized N:M sparse matrix-vector kernel. On Llama-3.1-8B, Llama-3.3-70B, and Qwen3-14B/32B, it beats standard fixed-text calibration on long-form generation and delivers up to 1.48x end-to-end decoding speedup on A100 GPUs.

Training-free pruners like SparseGPT calibrate on fixed text under teacher forcing. During generation, though, the model reads its own outputs, so the activations it sees drift away from the calibration activations. Existing 2:4 sparse kernels also run slower than dense at batch-one decoding. SparseDecoding fixes the first problem by building the pruning Hessian only from decode-step activations of the dense model's own generations. It fixes the second with a custom N:M sparse matrix-vector kernel.

  • Calibration: the dense model generates continuations of task prompts (LongWriter for writing, LiveCodeBench for code), and prefill activations are discarded; the resulting 1M decode-step tokens feed an unchanged SparseGPT or Wanda solver, so no training or new optimizer is needed.
  • Kernel: the sparsity pattern is stored as bitmasks at 2 bits of metadata per nonzero instead of 32-bit indices (16× less), and because every 32-bit mask word has exactly 16 set bits at 50% sparsity, each row is traversed in the same number of unrollable steps.
  • Quality: on WritingBench under 2:4 sparsity, Qwen3-32B rises from 2.99 to 5.31 (dense scores 6.48) and Qwen3-14B from 1.85 to 4.18; on ClassEval, Pass@1 for Qwen3-32B at 2:4 goes from 9% to 24%, and every one of the eight model-sparsity pairs improves by 4 to 15 points.
  • Speed: end-to-end decoding is 1.34–1.48× faster than dense on A100 for every pattern from 2:4 to 16:32, whereas cuSPARSELt's 2:4 Tensor Core path reaches only 0.85–0.87× of dense speed during decoding.
  • Limitations: the baseline calibrates on generic C4 while SparseDecoding uses prompts from the target task, so domain matching may account for part of the gain; Llama models stay far below dense (Llama-3.1-8B at 2:4 scores 1.52 vs 3.70); and speedups were measured only at batch one with 512-token context, using a kernel that supports only 50% sparsity.

arXiv:2610.12327 PDF

SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models

Highlight HF pick · 6▲Multimodal Hongxing Li, Jinyue Su, Dingming Li, Wenqi Zhang, Weiming Lu, Jun Xiao et al. Most spatial reasoning benchmarks test perception of relations already visible in the input, not the ability to predict how a scene changes after an intervention. SpaceCast-Bench follows an observe-transform-infer design with 3,862 questions from 182 real-world scenes, covering 16 task types at three levels: static perception, local prediction, and global prediction. Across 21 evaluated models, the strongest reaches only 58.0% against 87.2% for humans, and spatially specialized models stay near random chance. Analyses show that bridge views connecting separate observations matter and that explicit 3D evidence helps more reliably than generated outcome images or videos; fine-tuning on the programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7%, with gains on six out-of-domain benchmarks.

Most spatial-reasoning benchmarks only ask vision-language models to read off relations already visible in the input, not to predict how a scene changes after an action. SpaceCast-Bench tests that predictive step with an observe–transform–infer setup: the model sees views of the scene before a change, reads a text description of an object move, removal or whole-scene rotation it never sees, and answers a spatial question about the result.

  • The benchmark has 3,862 questions from 182 ScanNet/ScanNet++ scenes across 16 task types at three levels (static perception, local object-level prediction, global layout rotation), with labels computed by re-running a 3D geometry engine on the transformed scene.
  • Across 21 models, the best, Gemini 3.5 Flash, scores 58.0% against 87.2% for humans, and spatially fine-tuned models such as SenseNova-SI-1.2 and SpatialLadder land at 25.7–29.1%, close to the 25.9% random baseline and below every general-purpose model.
  • Chain-of-thought prompting gives no consistent gain, while giving models textual 3D metadata lifts proprietary models by +30.7 pp. Outcome images generated by Qwen-Image-Edit or GPT-image-2 help much less, partly because 80.3% and 38.3% of their wrong answers, respectively, trace back to flawed generations.
  • Staged SFT on 12,249 questions generated by the same pipeline from separate scenes raises Qwen3-VL-4B from 34.0% to 65.7% and its average across six out-of-domain benchmarks from 33.5% to 38.1%, although MindCube dips slightly and GRPO reaches only 37.7% on the benchmark itself.
  • All scenes are indoors, labels inherit any errors in the 3D reconstructions, and the transformations are geometric edits with no physics, so deformation, friction and collisions are not tested.

arXiv:2610.12402 PDF

ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills

Highlight HF pick · 8▲Agents Hongxing Li, Dingming Li, Yixin Li, Yong Du, Wenqi Zhang, Weiming Lu et al. Skill-augmented agents reuse strategies distilled from successful trajectories, but most store skills as text, which loses spatial layout and action-state structure, and they update skills separately from policy training. ViSkill encodes successful interactions as composite visual skill cards that vision-language model (VLM) agents read directly. Retrieved skills guide both inference and reward shaping, and new successful trajectories are distilled back into the library, so skill accumulation and policy improvement feed each other; an optional cold-start mechanism speeds up early learning. On Sokoban, FrozenLake, and PrimitiveSkill, it reaches an overall success rate of 0.89 (0.91 with cold start), beating all evaluated proprietary and open-source baselines and converging faster than standard PPO.

Text-based skill libraries flatten spatial layouts and action–state relationships into language, which loses geometric structure that VLM agents need in visually grounded tasks. ViSkill stores successful trajectories as composite visual "skill cards" and couples them with RL in a closed loop: retrieved skills guide both inference and reward shaping, and the improving policy distills new skills back into the library.

  • Each skill card is a single image that combines annotated trajectory frames with a strategy description written by the policy itself; cards are retrieved by geometric similarity weighted by historical utility, and new ones are admitted only if they pass a shortest-path quality gate and a novelty gate.
  • A bounded skill-guided reward gives failed rollouts partial credit for how far they followed the retrieved reference trajectory, and PPO on a Qwen2.5-VL-3B backbone optimizes the policy while the library grows alongside it.
  • On Sokoban, FrozenLake, and PrimitiveSkill, it reaches an overall success rate of 0.89, or 0.91 when the library is seeded with cold-start skills, beating AtlasVA (0.87), VAGEN (0.80), and the best proprietary model, GPT-5 (0.70); it also converges faster than plain PPO.
  • In ablations, swapping the visual cards for text skills drops success from 0.82/0.84 to 0.63/0.59 on Sokoban/FrozenLake, and disabling policy training while keeping retrieval collapses performance to 0.29/0.33, which shows that joint optimization matters most.
  • Evaluation is limited to small VAGEN environments where PrimitiveSkill is already saturated at 1.00; the quality gates and cold-start seeds depend on task-specific solvers (plus GPT-5.4-mini for the seed strategies), and the shaped reward helps little when valid solutions follow state paths that don't overlap with the reference trajectory.

arXiv:2610.12403 PDF

Applications 111

A Survey on LLM-Integrated Hardware Design Verification

Hao Zheng, Jaime Rafael Imperial, Bardia Nadimi, Xiangfei Kong cross-listed This survey reviews how large language models (LLMs) are used in hardware functional verification. It covers SystemVerilog assertion generation, testbench generation, bug localization and repair, model and equivalence checking, SAT/SMT optimization, and agentic verification workflows, and organizes the literature by method, objective, tool interaction, and benchmark. Its main observation is that LLMs work best as reasoning, search, and orchestration components inside workflows where simulators, formal engines, and solvers supply executable feedback. It also cautions that passing a tool's checks does not prove an artifact captures the full design intent. Open challenges include aligning specifications with verification evidence, generalizing to unseen designs, and evaluating correctness and cost rigorously.

arXiv:2610.10580 PDF

Has LLM Screening Performance Stalled in Software Engineering Systematic Reviews?

Aleksi Huotala, Miikka Kuutila, Mika M\"antyl\"a cross-listed Earlier claims about LLMs automating paper screening in systematic reviews may be outdated given how quickly models change, so the authors re-evaluate eight newer LLMs on the SESR-Eval software engineering benchmark and a cheaper power-sampled subset, SESR-Eval-Mini. Newer models improved only marginally, with average Matthews correlation coefficient rising from 0.347 to 0.365, and differences between reviews still outweighed differences between models. LLMs largely agree with one another, and refining the inclusion and exclusion criteria gave modest recall gains. The authors conclude that LLMs are not yet ready to replace human screeners.

arXiv:2610.10633 PDF

NEMORA: Neural Equivariant Multipole Operators for Long-Range Atomistic Learning

Jay L. Kaplan, Samuel Varner, Rebecca Willett, Juan J. de Pablo Equivariant graph neural networks used as machine-learned interatomic potentials describe local atomic environments well, but their finite cutoffs miss long-range interactions, and existing long-range extensions trade away learnability, full equivariance or efficiency. NEMORA turns the Fast Multipole Method (FMM) into a learned equivariant operator: it replaces the method's analytical multipole expansions and translations with learned counterparts on an adaptive spatial hierarchy, while keeping the FMM's hierarchical structure and radial factors as physical priors. It runs in linear time and memory, scales to hundreds of thousands of atoms, and can be added to both symmetry-constrained and unconstrained short-range backbones. On non-local benchmarks it reduces force errors by over an order of magnitude and energy errors by up to three orders of magnitude relative to the short-range backbones.

arXiv:2610.10776 PDF

Grammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier Models

Marjan Celikik, Ana Peleteiro Ramallo, Javier Morales Language-learning platforms want a running picture of each learner's grammar mastery from tutoring transcripts, but annotating every lesson with prompted frontier models costs too much at scale. The authors fine-tune Qwen3.5 small language models (SLMs) on filtered, rebalanced supervision generated by a teacher model. This builds the annotation rules into adapter weights, so the deployed 0.8B model needs only a short prompt. On two human-curated benchmarks, both the 0.8B model and a 4B comparator beat prompted GPT-5.4 and GPT-5.6 Sol on precision and recall at every matching strictness level, while the 0.8B model cuts serving cost by about 16x. In an online experiment it also raised learner engagement by 15.8%.

arXiv:2610.10827 PDF

AI4Fire: Evaluating Large Language Models on Wildfire Tasks

Yue Zhao, Xiyang Hu, Zuobin Xiong, Zhangyu Wang, Ruolin Li AI4Fire evaluates LLMs zero-shot on five wildfire-management tasks, running six core models both bare and with one task-specific grounding addition, and sweeping 29 more models. Grounding helped most when the addition carried the answer: a read-only SQL tool lifted database accuracy from at most 16% to at least 88% for every core model. Simple rules were hard to beat: no core model outperformed repeating today's staffing count. The audit also found leakage and labeling hazards in public wildfire datasets.

arXiv:2610.10946 PDF

Clinician use of language models diverges from how the models are evaluated

Krithik Vishwanath, Haitong Lin, Anton Alyakin, Jin Vivian Lee, D. Brock Hewitt, Jie J. Yao et al. Analyzes 127,833 queries from 6,342 clinicians in 35 specialties to an institutional large language model (LLM) assistant over an eight-month roll-out. Each query was coded with RCQ-Map, a clinician-validated framework recording task, intent, answerability, missing information and potential harm. Documentation and administration (36.2%) and knowledge retrieval (28.9%) dominated use, diagnosis accounted for only 3.7%, and over a third of queries could not be answered well as posed. Across 58 public benchmarks collected in the Clinical AI Benchmark Atlas, the median benchmark contained no documentation requests and shared only 31% of the real task mix. Benchmarks designed to resemble clinical practice were individually no closer to real use than those cited in frontier model reports.

arXiv:2610.11069 PDF

Beyond Imitation: A Framework and Benchmark for LLM-Assisted Peer Review

Rachel S. Y. Teo, Yutaro Yamada, Shashank Kotyan, Yuki Imajuku, Tarin Clanuwat Evaluations of LLMs (large language models) as peer-review assistants have mostly measured how well they imitate human-written reviews. This work instead treats error detection as a core reviewing function and builds a scalable benchmark by synthetically inserting logical contradictions into conference papers, so every error is an unambiguous target. It also proposes the Multi-Layered Review (MLR) framework, which builds a detailed understanding of the manuscript before generating a review and uses fewer tokens. MLR shows strong alignment with human review scores and high error-detection performance, though the authors confirm that automated reviewers remain vulnerable to adversarial manipulation.

arXiv:2610.11087 PDF

Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains

Miruna Cretu, Alex Abrudan, Antonia Panescu, Tynan Perez, Rishabh Anand, N. Benjamin Erichson et al. Generative models for atomistic structures tend to specialize in either chemistry or biology and rarely pretrain on molecules and materials together. Zatom-2 is pretrained on about five million structures from the OMol25 and OMat24 electronic structure datasets. It uses a multiscale Transformer with conditional flow matching and supports generation, structure prediction, energy and force prediction, and force-conditioned sampling. It improves molecular distribution fidelity over Zatom-1 and performs strongly on molecule and material generation benchmarks. Transferred to proteins in a low-data setting, it raises backbone designability under length extrapolation from 67.8% to 74.8% after fine-tuning on 2,000 protein domains.

arXiv:2610.11454 PDF

Large Language Model Turnover Undermines Screening for Artificial Intelligence-Assisted Scientific Writing

Kazuki Nakajima, Takayuki Mizuno Journals now screen manuscripts for LLM-written text, but detectors are benchmarked on a fixed set of model versions while the versions people actually use keep changing. The authors paired 4,000 pre-ChatGPT abstracts from the Proceedings of the National Academy of Sciences with rewrites by 23 LLM versions from three vendors released between June 2023 and August 2026, then trained detectors under different update schedules. With the threshold set to falsely flag 1% of human-written abstracts, a detector trained only on a vendor's past versions caught over 99% of rewrites just before the sharpest generation boundary but only 3.8% just after it, and a commercial detector showed the same failure. Shifts in vocabulary between versions largely predict where detection transfers, which leads the authors to argue that a detector's benchmark accuracy should be re-verified with every LLM release.

arXiv:2610.11599 PDF

MAP4CS: A Multi-dimensional Data Pruning Framework for Efficient Code Retriever Fine-tuning

Yuxuan Chen, Mingwei Liu, Guangsheng Ou, Zekai Zhang, Zike Li, Yanlin Wang et al. cross-listed Fine-tuning code retrievers for retrieval-augmented generation (RAG) on large, noisy code corpora is expensive and can hurt performance through negative transfer, while random subsampling does not guarantee a representative training set. MAP4CS (Multi-dimensional Awareness Pruning for Code Search) selects a small core subset by combining syntactic structure, semantic diversity, and distributional representativeness, then applies rule-based filtering. On two large code-search datasets, it matches or beats full-dataset fine-tuning using only 5% of the training data and consistently outperforms random sampling. Further analysis shows it behaves as a de-duplicator on redundant corpora and as a denoiser on noisy ones.

arXiv:2610.11727 PDF

Can LLMs Fix It Without Code? Toward Automated Verification of No-Code Bug Fixes

Utku Boran Torun, Veli Karakaya, Eray T\"uz\"un cross-listed A no-code fix resolves a bug report by telling the user to change a setting, upgrade, or adjust their workflow, and checking such fixes by hand takes developer time. The authors build an execution-based pipeline in which an executor agent carries out each LLM-generated fix in a real browser and an issue-specific checker tests whether the bug persists. They run it with three executors (OpenCUA-72B, Claude Sonnet 5, and Muse Glimmer) on 322 fixes from 12 earlier configurations. Between 14.6% and 49.7% of fixes resolved the bug, depending on the executor. Changing only the executor shifted a configuration's resolution rate by 38.8% on average, and the three executors agreed on just 46.9% of fixes, so the authors argue such evaluations must report and control which executor they use.

arXiv:2610.11963 PDF

InterviewPlayground: A Simulation Environment for Evaluating AI Interviewers

Jonathan Ivey, Aimee Liang, Arthur Y. S. Wang, Madeline Mandell, Ziang Xiao, Anjalie Field AI interviewers used in market research, polling, and social science are hard to evaluate because they run long, adaptive multi-turn conversations. InterviewPlayground is a simulation environment that runs AI interviewers against simulated participants whose behavior is grounded in social theory, and it scores them with an InterviewReportCard of validated measures. To check validity, the authors ran 15 real qualitative studies with five AI interviewers, three topics, and 450 human participants. Simulated performance predicted human-study performance with an average Pearson correlation of 0.86 across 12 measures, and the simulations reproduced key behavioral findings from the human studies.

arXiv:2610.12023 PDF

A persistent accuracy ceiling in automated verbal deception detection

Riccardo Loconte, Jonas Festor, Zane Fatjanova, Mariam Bolkvadze, Bennett Kleinberg This systematic review and meta-analysis covers 25 years of automated verbal deception detection: 289 reports and 6,136 classifiers, of which 3,653 models across 97 datasets were meta-analyzed. Pooled accuracy was 74.4%, and differences were driven by methodological quality (ground truth, data source, class balance, evaluation procedure) more than by model complexity. Embeddings and LLMs brought no gains. Only 12.46% of reports used verifiable ground truth and only 23.96% of models were tested on independent data, pointing to a ceiling of 70-75% that matches manual approaches.

arXiv:2610.12118 PDF

Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search

Ziming Dai, Dabiao Ma, Ziheng Guo, Jack Dong, Zimu Zhou Industrial risk-control systems run on structured-data models, yet much useful signal sits in long unstructured text, and calling a large language model (LLM) on every live input is often too slow or costly for production. LLM-BlockFE uses an LLM offline to build executable feature programs by appending immutable code blocks, scoring candidate features with a downstream model. To avoid greedy search getting stuck, it adds block-level rollback driven by depth-calibrated credit allocation and runs several search trajectories interleaved, each with a fixed description of its exploration direction to reduce redundant exploration. The frozen programs extract features with no LLM calls at inference time, giving absolute AUC gains of 0.0069 to 0.0358 over the strongest baseline on four datasets and KS improvements of 0.02 to 1.56 percentage points across five deployed financial risk-control applications.

arXiv:2610.12390 PDF

Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment

Guowen Li, Yang Liu, Yujie Wang, Qiuyan Sun, Haoyuan Liang, Juepeng Zheng et al. Kilometer-scale regional weather forecasting usually depends on numerical forecasts for large-scale guidance or on training extra global components. ScaleCast instead reuses pretrained global weather models: a Global-Regional Conversion module aligns global and regional representations across different grids, and a Global-Regional Alignment and Dynamics block combines that guidance with interactions between neighboring regional cells. Trained with ERA5 global analyses and CERRA 5.5 km regional reanalysis, it improves surface and upper-air forecasts, and a single trained model works with multiple global drivers (Pangu-Weather, GraphCast, HRES) without retraining. Fine-tuning on 3 km HRRR data shows it adapts to other domains, and windstorm case studies show better cyclone positioning and core-pressure estimates.

arXiv:2610.12401 PDF

96 more specialized papers

Agents 83

Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction

Yipeng Li, Ashutosh Hathidara, Jane Lo, Harshavardhan Abichandani, Gunraj Singh, Atin Ghosh Training and evaluating enterprise tool-calling agents is hard because real business systems, data, and database schemas are legally restricted. Synthesis Through Simulation (STS) generates data by having an LLM agent carry out operations against policy-enforcing APIs inside simulated enterprise environments, so every record is structurally valid by construction and no schema is needed. Its domain-agnostic agent, the Generalist Populator, reaches 0.88 average marginal fidelity and 100% constraint satisfaction across ten environments without access to database schemas. By comparison, statistical synthesizers cannot run on seven of the environments, and schema-aware agents fail 82% of trajectories on the tightly coupled airline workflows. The framework, environments, and datasets are open-sourced.

arXiv:2610.10549 PDF

Agent-Controlled Forgetting for Tool-Using Agents: Reversible Context Curation in Practice

Jan-Peter Franke Tool-using agents keep large tool outputs in their context long after only a small part remains useful. In agent-controlled forgetting, the model itself picks earlier tool results to replace with short notes and keeps the exact originals in an archive it can recover from; a Python harness provides this without task-specific training. In an OpenTelemetry debugging session followed by an unrelated implementation task, the run ended with 231,951 prompt tokens versus 912,492 under full history, used 50% fewer cumulative input tokens, and cost about USD 1.3 instead of 4.38, though it made more requests and took 17% longer. A second application-development pair showed no savings, and one earlier run produced lower-quality output, so the benefit depends on the workload.

arXiv:2610.10590 PDF

From Investigation Failures to Reliable SOC Agents: Understanding and Improving LLM-Based Alert Triage

Saimon Amanuel Tsegai (Daphne), Alex Kantchelian (Daphne), Danfeng (Daphne), Yao, Peng Gao cross-listed Security operations centers (SOCs) must triage large volumes of mostly benign alerts, and LLM agents that use tools to gather evidence are being tried for the job. Using ALERT-BENCH, an interactive benchmark that replays enterprise telemetry through a live SIEM, the authors test five agent reasoning strategies on 1,247 alerts and find that every one misses at least 40.4% of attack-related alerts. Attacks are often dismissed when a search returns no records, and self-review in the same context does more harm than good. Their multi-agent framework AIDA requires an explicit proposed decision, an independent challenge in a separate context, stronger evidence before dismissing an alert, and a Judge to decide. It reaches an F1 of 0.958 and cuts the false-negative rate from 40.4% to 3.1%, while escalating 18.4% of alerts to human analysts.

arXiv:2610.10608 PDF

Code Understanding is a Bottleneck for Coding Agents

Nishant Balepur, Kiran Tomlinson, Tobias Schnabel cross-listed Repository benchmarks such as SWE-bench usually treat lines of code edited as a proxy for task difficulty, but they offer little control over which abilities actually cause agent errors. The authors introduce CABRA, a benchmark that builds tasks from scratch as call-graph transformations and scales difficulty along four axes: function traversal, search, runtime resolution, and instruction following. Across 6,840 tasks with eight LLMs and six coding agents, bare LLM accuracy drops as tasks grow, while agents stay near-perfect by offloading work to tools such as grep. A companion study on SWE-bench Verified finds that the number of reading and analysis tool calls predicts agent accuracy better than lines edited, which suggests that understanding code, rather than editing it, is often the real bottleneck.

arXiv:2610.10610 PDF

TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolution

Shuangjie Yao, Hao Wang, Koushik Sen, Simin Chen, Baishakhi Ray, Dawn Song cross-listed Coding-agent benchmarks typically judge a patch correct if it passes a fixed set of unit tests, which agents can game or satisfy while still missing required behavior. TestJack audits each trial by generating new tests aimed at prompt requirements the patch may violate, keeping only tests the ground-truth patch passes, and producing a replayable failing test for each confirmed violation. A cheaper variant audits a sample of trials and reuses the resulting tests. Across six frontier model backends and five benchmarks, including DeepSWE and SWE Marathon, about 34.4% of trials judged correct actually violate the task requirements, lowering the overall resolution rate from 50.6% to 33.2%.

arXiv:2610.10619 PDF

Agent4RE: A Self-Refining Multi-agent Framework for End-to-End Software Requirements Engineering and Benchmarking

Yongjian Tang, Linhan Li, Thomas Runkler cross-listed LLM approaches to software requirements engineering (RE) usually rely on basic prompting, and existing datasets cover only isolated subtasks. Agent4RE is a multi-agent system that orchestrates specialized agents with two iterative refinement loops, and the authors build RE-E2E, a dataset of human-written specifications for end-to-end evaluation from elicitation to generation. Two enhanced variants add either autonomous self-refinement or structured human feedback. Across eight LLMs, all variants beat a domain-context prompting baseline by an average of 8% on text-based metrics, and the enhanced variants score about 0.8 points higher than two RE baselines on a four-point scale in LLM-judge and human ratings.

arXiv:2610.10628 PDF

Conversational Task Disambiguation over Tabular Data: Leakage-Aware Formulation, Benchmark Suite, and Training

Nafiseh Ghoroghchian, Luis Scoccola, Tina Sedaghat, Omid Vaheb, Hannah Chen, Dino D'Agostino et al. In conversational task disambiguation over tables, an agent asks clarifying questions before producing a solution. Existing evaluations mix up how well the agent clarifies with how well it solves the task, and they also reward oracle leakage, where a simulated user reveals more than a real user would. The authors formalize an "ambiguous verifiable task" that splits the agent into an asking policy and a solution policy, and the environment into an oracle and a verifier. This yields separate metrics, formal leakage definitions and judge-free leakage diagnostics. They instantiate the framework for text-to-SQL as AmbiTab, a suite that unifies six ambiguous datasets, and train an asking policy with reinforcement learning that improves disambiguation metrics on all six datasets and task success on five.

arXiv:2610.10740 PDF

MemoWM: How World Models Change What Agents Need to Remember

Bingfan Zeng, Zhisheng Chen, Chenbo Sang, Zhengwei Xie, Jinpeng Wang, Xiangchen Guan et al. Long-running agents accumulate a growing store of past experience. MemoWM uses a world model's predictions to compress each memory: it reconstructs predictable content instead of storing it and keeps only information whose downstream value goes beyond the predictive prior. A task-aware allocation rule weighs the expected cost of reconstruction errors against storage cost. Across five long-term agent-memory benchmarks it reaches 42.42% average answer accuracy, 2.62 points above the strongest baseline, while cutting experience-specific storage by 53.9% relative to MIRIX, the most storage-efficient baseline. Stronger world models further reduce per-experience storage, which creates a trade-off between shared model size and recurring storage costs.

arXiv:2610.10778 PDF

Plan-and-Patch: Diffusion Language Models for Agentic Planning

Syamantak Kumar, Jiang Guo, Hassan Hamad, Hideo Kobayashi, Yi Xiang, Yezhou Yang et al. Long-horizon agents often need to revise only part of a plan when tools fail or assumptions break, and regenerating the whole plan risks unnecessary changes. Plan-and-Patch uses a diffusion language model (dLLM) to generate a structured, program-like plan through parallel unmasking, and repairs it by filling in selected regions while keeping the preceding and following steps fixed. Comparing DreamReasoner-8B as a diffusion planner with Qwen3-8B as an autoregressive planner, the diffusion model achieves nearly twice the plan-repair success on Natural Plan (53.7% vs. 27.0%) without task-specific training. After training on ALFWorld and TextCraft, both planners reach similar success, but diffusion cuts mean plan-generation latency by 39-46%.

arXiv:2610.10786 PDF

On the Clock: Towards Punctual and Productive Time-Budgeted AI Agents

Aaron Wang, Neelabh Madan, Vlad Sobal, Matthew Trager, Michael Kleinman, Elman Mansimov et al. The authors test whether small LLM agents can respect explicit wall-clock time budgets and use the time they have productively, running Qwen3.6-27B on five MLE-Bench Lite competitions and Qwen3-4B on Zork I (Jericho). When the budget appears only in the prompt, agents fail to control their time use: they get no timing feedback, cannot predict how long actions take, and have no learned link between available time and strategy. Two remedies help with adherence: harness hooks that expose timing and enforce deadlines, and reinforcement learning with GRPO using budget-aware rewards, which reaches near-perfect adherence on Zork I and generalizes to unseen budgets. However, agents that respect the budget still do not turn extra time into better task performance, often filling it with repeated actions.

arXiv:2610.10833 PDF

RFChipAgent: Multi-Agentic AI Flow for Analog/RF Chip Design

Awani Khodkumbhe, Yunfei Feng, Raj Rangarajan, Kevin Wang, Kamal Sahota cross-listed Analog and RF circuit design remains one of the most labor-intensive steps in chip development. RFChipAgent is a multi-agent LLM system that runs the end-to-end design flow under human supervision. It combines multimodal retrieval-augmented generation (RAG) over engineering documents using per-document FAISS indexes, agents for topology selection and schematic/testbench assembly, and a simulator-in-the-loop circuit-sizing engine that pairs Tree-structured Parzen Estimator (TPE) with CMA-ES optimization. A trust-scored simulation database guides later optimization trials. Validated on 60 GHz mm-wave low-noise amplifier topologies in GF22FDSOI, it shows substantial reductions in design effort while maintaining signoff-quality verification.

arXiv:2610.10858 PDF

Stochastic Teacher Intervention for Agentic On-Policy Distillation

Junnan Liu, Linhao Luo, Zhijun Chen, Qianren Mao, Thuy-Trang Vu, Gholamreza Haffari On-policy distillation (OPD) trains a student language model on its own rollouts using token-level teacher supervision. In multi-turn agentic tasks, early student errors push trajectories away from anything the teacher would produce, so the teacher's signal becomes unreliable. STI-OPD estimates teacher-student discrepancy with KL divergence and converts it into a probability of replacing the student's action with the teacher's, balancing teacher control against student exploration. An importance-weighted reverse KL objective corrects the resulting sampling mismatch. Across tool-integrated reasoning and long-horizon interaction, STI-OPD outperforms the strongest prior OPD baseline on every benchmark and student size.

arXiv:2610.10878 PDF

StoreBench: A Live-Commerce Environment for Evaluating and Training Autonomous Operator Agents

Daksh Raghuvanshi, Ved Vedere, Yifan Wang StoreBench is a live environment in which an LLM agent runs a mid-size online apparel store on a production-grade commerce backend, using the same 29 merchant tools as a human operator. Customers order continuously, suppliers reprice or fail, and market shocks arrive with little or no warning. Pass thresholds are calibrated against scripted policies, rewards are hardened against known reward hacks, and episodes replay deterministically. Across seven frontier LLMs, none matches a scripted triage heuristic: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. GRPO post-training on five tasks raises Qwen3.5-27B's held-out mean composite score from 0.136 to 0.373.

arXiv:2610.10942 PDF

Cross-Provider Review as a Runtime Contract for Coding Agents: A Controlled Pilot and Fault-Injection Study

Bowen Xu, Boyu Chen cross-listed When several coding agents share a workstation but draw on different providers and subscription allowances, asking a second agent to review a finished answer spends another resource pool and may produce nothing useful. The authors specify an advisory cross-provider review contract: separate resource pools, bounded execution, a reviewer with restricted capabilities, guaranteed delivery of the full input, explicit failure states, and a durable record of every attempt. In a pilot of 20 paired development turns, 8 produced a material reviewer finding. Fault-injection tests surfaced real defects, including false success on truncated input and a nonzero reviewer exit being counted as complete, and in read-only probes Codex attempted writes in five of five trials, all of which were blocked, while Claude had no writing tools.

arXiv:2610.10961 PDF

Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback

Zexuan Liu, Yuning Yang, Tiancheng Zhao LLM agents load a fixed context file such as AGENTS.md at the start of every session, so each token is paid for again on every round, and the files tend to grow by appending even though larger files can hurt performance. The authors model curation of this file as a capacitated assortment problem, in which instructions compete for limited attention and carry a per-session cost. They prove an upper bound on the optimal file size no matter how many candidate instructions exist, and show that appending every individually useful instruction can be arbitrarily worse than choosing an optimal subset. Because feedback is censored, with missing instructions noticed only when their absence causes harm, deleting instructions the agent ignores can remove helpful ones, and experiments confirm that irrelevant rules taken from real context files lower model compliance.

arXiv:2610.11007 PDF

Why LLM Agents Favor Their Group: Stakes, Observed Norms, and Reputation

Yujiao Chen cross-listed Studies why language-model agents favor members of their own group, using small simulated societies with arbitrary group labels and ten rounds of point sharing. The experiments span fifteen OpenAI and three Claude models, about 4,400 societies and 3.3 million audited model calls. The large effect of a bare group label seen in earlier work appears only when giving costs the agent nothing. Once the agent can keep points for itself, favoritism is driven mainly by having watched group members favor each other, and that effect carries over to labeled strangers. Favoritism disappears under an egalitarian norm and reverses when the agent's own group favors the other side. Stronger models follow an individual's own record when it conflicts with the group's behavior.

arXiv:2610.11008 PDF

NOMOS: Compiling Written Policies into Statically Verified Tool-Call Gates for LLM Agents

Min-Young Yu, Tony Kim, Jang Won Choi cross-listed Tool-using LLM agents often break the policies they are deployed to enforce, and they usually do so silently. NOMOS is a four-pass compiler that turns a natural-language policy into a deterministic gate checking each tool call. Static checks against the tool schemas alone repair or reject flawed rules, with no prover, solver or LLM involved. On τ²-bench the gate cuts policy violations among state-changing calls from 66.3% to 2.6% (airline) and from 30.8% to 6.9% (retail). It reaches a zero attack success rate on the AgentDojo banking suite and at most 3.6% on the other suites. Each decision takes microseconds without an LLM call, and compilation runs on-premise with open-weight gemma-4-26B.

arXiv:2610.11030 PDF

AgentHorizon: Evaluating Agentic Judges for Long-Horizon Computer-Use Tasks

Xing Han L\`u, Dheeraj Vattikonda, Sina Hajimiri, Fatemeh Pesaran Zadeh, Parishad BehnamGhader, Ghazwa Darwiche et al. Automatic judges increasingly decide whether computer-use agents succeeded, but it is unclear how reliable they are on long tasks that span several applications. AgentHorizon provides 1,373 instruction-trajectory pairs built from 166 hours of human recordings on three operating systems. Negative examples are made by swapping in the instruction from a closely related task, so judges must reject a trajectory that completed a similar but incompatible request. Across eleven judges, given either the full trajectory directly or run as coding agents in five harnesses, the best agentic judge (GPT-5.5) reaches 80.9% balanced accuracy. Tool use helped some models but hurt open-weight ones, and judges varied widely in how well they balanced accepting valid trajectories against rejecting failed ones.

arXiv:2610.11050 PDF

When Interfaces Speak: Data-Aware Generative UI Harness for Active Interaction

Xiaolong Li, Xiaohan Xu, Jinyang Li, Xinnuo Xu, Ge Qu, Nan Huo et al. GenUI-Harness is a multi-agent system that pairs a Tool Agent, which retrieves information and executes tasks, with a GUI Coder Agent that spots ambiguities and generates short-lived front-end interfaces instead of plain text replies. To make reinforcement learning practical for the coder, the authors add Dynamic UX, which runs interaction and collects rewards in a single sandbox, and Reward Auditor, a meta-reward mechanism that limits reward hacking by LLM judges. They also introduce UI-TAU Bench, built on Tau-Bench tool-use settings over 10 real-world databases. Training lifts a 4B model from 9.33% to 58.00% Pass@3, ahead of Claude Opus 5 at 46.67%, and in a reviewer survey generated UIs cut average dialogue rounds from 3.4 to 1.2.

arXiv:2610.11123 PDF

Do LLMs Learn from Rewards in Context? : Rethinking the role of reward in In-Context Reinforcement Learning

Minchan Kwon, Seunghee Koh, Sunghyun Baek, Minsung Bae, Junmo Kim LLM agents often improve at inference time by accumulating trajectories and rewards in context, a process described as in-context reinforcement learning (ICRL). The authors test whether the reward actually acts as a learning signal through controlled experiments on four benchmarks with six models. They find that flipping, randomizing or removing the reward leaves the improvement curve almost unchanged, even when prompts explicitly tell the model to reason over rewards, and shuffled or corrupted trajectories help as much as real ones. They conclude that this direct form of ICRL behaves like ordinary in-context learning (ICL), so agent memory design should focus on input distribution and demonstrations rather than reward shaping.

arXiv:2610.11152 PDF

VAMR: Multi-Question Agentic Reasoning for Efficient Long-Form Video Understanding

Runquan Gui, Hanzhu Chen, Zehao Wang, Hanxin Zhu, Xin Li, Zhibo Chen cross-listed Video agents usually answer each question about a long video through its own tool-use trajectory, exploring the video and building memory from scratch every time. VAMR handles all questions about a video in one shared trajectory. At each round, a persistent policy can call tools for several unresolved questions and submit answers once it has enough evidence, while a layered memory keeps both a shared video story and separate evidence for each question. After supervised fine-tuning, the policy is trained with question-horizon policy optimization (QHPO), which uses a per-question critic to assign credit to the rounds that served each question. On LVBench it reaches 62.1% accuracy, 4.3 points above VideoARM while using 85.9% fewer reasoning rounds and 61.4% fewer frames, and it also leads on Video-Holmes and LongVideoBench.

arXiv:2610.11171 PDF

What to Admit and How to Present: Governing Persistent Memory in LLM Agents

Chang Liu, Deliang Ding Persistent memory makes LLM agents more personal, but it can also cause sycophancy and leak information across domains. The authors separate admission, which decides what recalled memory enters the context, from presentation, which decides how admitted memory is worded. They implement two inference-time admission designs with no retraining: FC, which scores whole memory entries, and PS, which splits entries into typed units with permission semantics. On an external benchmark, FC and PS cut failure rates by 6.7 and 8.8 percentage points compared with injecting memory verbatim, and tightening admission reduces cross-domain failures by a further 17.5 points, whereas changing presentation has no significant effect. Both designs increase personalization failures, which the authors leave unresolved.

arXiv:2610.11188 PDF

Harness Compilation: Which Decisions Should a Small Vision-Language Model Keep?

Minhao Fan, Yinyi Liu, Jiayu Zhao, Zihan Teng, Song Chen, Weichen Liu Small vision-language models (VLMs) can often read external evidence well but struggle to obtain it. Harness Compilation (HC) is an offline procedure in which a large teacher uses a frozen small VLM's execution traces to revise the content and control logic of its external harness, then picks the deployed harness on a validation set; deployment needs no weight updates and no teacher calls. Across seven visual question-answering settings with students of at most 9B parameters, HC improves scores over the bare student by 9.9 to 23.9 points. Interventions on five runtime decision types show that requesting evidence and writing open-ended queries are costly for small models, while bounded choices and reading supplied text remain useful work for the student. With 100 practice items, HC beats answer-only LoRA on three tasks.

arXiv:2610.11231 PDF

LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models

Surya Shetty, Ulisses Braga-Neto When an LLM agent acting as an autonomous scientist stops making progress, it cannot tell whether it lacks capability or whether the model it is trying to infer is simply not identifiable from the available experiments. LLM-IDEA adds an identifiability engine that returns one of three verdicts: capability limit, resolvable within the current experiment design class, or certified exhausted. The engine reproduces known verdicts on classic systems such as Lotka-Volterra and Lorenz, and it flags explanation rubrics in the DiscoverPhysics benchmark that reward distinctions no legal experiment can make. On a new two-body testbed, LLM-IDEA reaches discovery depth of at least three on 8 of 8 seeds versus 1 of 8 without it.

arXiv:2610.11253 PDF

TaReD: Tool-Aware Recursive Decomposition for Long-Horizon Tasks

Wei-Xiang Mao, Zhi-Kai Chen, De-Chuan Zhan, Han-Jia Ye Tool-using agents that interleave reasoning and actions in a single chain become unreliable on long tasks, and exposing every tool description in a large library wastes context and makes relevant tools harder to find. TaReD (Tool-Aware Recursive Decomposition) organizes tools into a hierarchy of capabilities based on their functional relationships. The agent discovers tools on demand and recursively decomposes a task into a subtask tree whose levels match the capabilities each stage needs. On complex real-world tasks, it improves end-to-end success rate by up to 40 percentage points over the compared baselines.

arXiv:2610.11268 PDF

Gated Memory: Admission-Controlled Memory Formation for Conversational AI

Preeti Saraswat, Divya Neelagiri, Ajay Manoj Long-term memory for conversational AI usually extracts subject-relation-object triples from user utterances, which discards context such as whether a fact is a permanent attribute or a temporary situation. Gated Memory adds an admission gate that judges each candidate fact against the full utterance before extraction. A conditional enrichment stage then tags admitted facts with provenance (stated or inferred), scope, time, and place, under privacy constraints and a rule against asserting entities absent from the context. On LoCoMo-10, it achieves a +2.6% relative improvement in LLM-judge accuracy over a strong baseline with identical retrieval and generation.

arXiv:2610.11270 PDF

REMORY: Learning Residual Memory for Context Compaction

Hanchen Xia, Baoyou Chen, Yutang Ge, Naihao Deng, Senqiao Yang, Zilong Dong et al. Long-horizon agents compact their history into a text summary to stay within the context window, but the summary can lose details later decisions depend on. REMORY trains a neural memory network that appends a bounded set of soft memory tokens after the summary, so that a frozen LLM's continuation approximates what it would produce with the full history, much like a residual connection along the sequence. On SummHay it improves source attribution and approaches the full-context score using only 5.2% of the input positions. On long-horizon agent benchmarks, Qwen3.8-27B and GLM-5.3-Flash both improve and produce fewer repeated tool outputs and tool errors on BrowseComp and Terminal-Bench 2.1.

arXiv:2610.11287 PDF

DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration

Yingda Shen, Yuxiang Wang, Kunyu Feng, Qinke Ni, Jiaqi Li, Minghao Hsu et al. Voice agents increasingly pair a full-duplex model, which listens and speaks continuously, with delegated reasoning and coding agents for complex tasks, and that pairing needs a harness to manage task acceptance, progress, cancellation, and result delivery. DuplexAgent expresses this harness as six editable modules. Duplex-Harness-RSI improves those modules in a closed loop: a simulator generates timed test conversations and failure traces, an Exam Planner chooses the next tests, and a Harness Editor built on the same delegated LLMs and coding agents proposes targeted module fixes. The system scores higher on spoken-knowledge and executable-tool benchmarks than the compared delegated systems while keeping strong interruption handling, and an ablation shows the diagnosis-and-repair loop beats both the initial harness and editing without a repair archive.

arXiv:2610.11299 PDF

From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue

Zirui Liao, Zhengxian Wu, Zhuohong Chen, Yunyao Yu, Xiaoyu Liu, Yifan Xu et al. Retrieval-augmented generation (RAG) memories for long-term dialogue agents treat memory as passive storage, which blurs a person's subjective claims with established facts and makes it hard to connect evidence spread across sessions. CogMem stores dialogue in a Person-Event-Concept-Claim-Fact graph whose Claim nodes record who said what about whom. Records are consolidated into higher-level facts and reconciled into time-scoped views when the same source contradicts itself. At query time, an LLM-driven controller composes four deterministic graph operators to reconstruct the relevant context. It shows strong results on LoCoMo and LongMemEval, especially on multi-hop, temporal, and knowledge-update questions, and ablations credit the separation of claims from facts, consolidation, and agentic retrieval.

arXiv:2610.11314 PDF

Mine Odyssey: Benchmarking Spatial Agentic Intelligence in the Wild

Yuxuan Cao, Junlong Li, Hao Li, Junxian He Mine Odyssey is a benchmark for agentic spatial intelligence: exploring unfamiliar environments, updating spatial understanding through interaction, and recovering from errors. It is built from Minecraft reconstructions of 30 real-world locations across 20 countries and regions, including Midtown Manhattan and Buckingham Palace. Its 180 tasks give natural-language instructions to visit waypoints in a set order, which requires finding routes and entrances, opening doors, and climbing stairs and ladders. Across eight state-of-the-art models, GPT-6 Astra leads with an 85.6% success rate and Claude Opus 5.5 follows at 73.9%, while the best open-weight model, DeepSeek-V4.1-Flash, reaches only 23.9%.

arXiv:2610.11328 PDF

ReCast: Attribution-Oriented Step Representation Learning for LLM-Based Agent Systems

Weilin Jin, Mingyu Wang, Taiyu Zhu, Ziqi Zhou, Wenbo Li, Haoyang Huang et al. In LLM-based agent systems, a failure often starts at an early step and spreads through later interactions. Failure attribution is the task of identifying the earliest step responsible. An empirical study shows that the raw LLM hidden states used by recent methods separate root-cause steps poorly from other steps. ReCast turns hidden states from a frozen LLM into attribution-oriented step representations: it selects relevant layers, builds complementary pattern and deviation features, and trains a contextual encoder with contrastive and ranking objectives. The authors also release a training dataset, ReCast-2K. ReCast achieves the best Hit@1 on four benchmarks, beating the strongest baseline by 5.65 and 9.19 points on the Who&When Algorithm and Handcrafted splits.

arXiv:2610.11334 PDF

EvoSim: Learning to Model, Modeling to Learn

Yun-Wei Song, Jinkai Tao, Jun-Dong Zhang, Rui Zhang, Yi-Min Wu, Qiang Zhang Building physics-based models means choosing physical processes, governing equations, couplings, and parameters, and current AI systems struggle to make these structural choices on their own. EvoSim is a self-evolving AI scientist that uses mismatches with experimental data to revise mechanisms and equations, and uses held-out data to test physical plausibility. Its exploration traces update its knowledge, skills, and multi-agent orchestration. On two industrial battery-modeling tasks, it predicts lithium-metal-plating onset with a 1.79% mean absolute error in state of charge, and its dynamic voltage predictions under driving conditions (7.62 mV root mean square error) surpass the reported accuracy of models built by human experts. Self-evolution reduces model and physics errors by about 36% relative to the baseline.

arXiv:2610.11344 PDF

RIT-RAG: Navigating Document Corpora with Retrieval-Induced Trees

Meghanadh Pulivarthi, Swaraj Kumar Biswal, Kushagra Bhushan, Yatin Nandwani, Sachindra Joshi, Dinesh Raghu Agentic retrieval-augmented generation (RAG) lets a model search iteratively, but it sees isolated chunks without document structure. Structure-aware methods such as PageIndex must first commit to a single document and cannot recover from a wrong choice. RIT-RAG (Retrieval-Induced Tree RAG) builds a tree for each document offline from its table of contents or sitemap. At query time, it retrieves a broad set of chunks and uses their positions to induce small sub-trees, possibly spanning several documents, which an LLM agent navigates, reading promising nodes and reformulating queries as needed. RIT-RAG achieves the best answer accuracy on financial, scientific, and customer-support benchmarks. On EntQABench, a new benchmark of 2.84 million technical-documentation pages, it beats the strongest baseline by 6.8 to 11.4 points across three LLMs.

arXiv:2610.11370 PDF

Environmental Feedback Modeling Matters: Rethinking Feedback Treatment in Agentic Hindsight Self-Distillation

Hangxi Guo, Fengyuan Liu, Yue Wang, Yuhua Qi, Haoyi Xiong, Fei Sun et al. Reinforcement learning is hard to apply to language agents when no reward signal is available, and recent methods instead use environmental feedback as privileged context for hindsight self-distillation. The authors find that simply conditioning a teacher on that feedback is not enough. They propose SELF, which jointly trains the policy to predict how the environment will respond to its actions while distilling guidance from a feedback-conditioned copy of itself, and they report that the two objectives reinforce each other. With Qwen3-8B, SELF beats SDPO and GRPO by 6.4 and 4.1 percentage points on τ-bench, and by 10.71 and 3.57 points on AppWorld task goal completion.

arXiv:2610.11384 PDF

Closed-loop evaluation of LLM agents for embedded software development

Jorge Garc\'ia-Carrasco, Sergio Garc\'ia-Carrasco, Alejandro Mat\'e, Juan Trujillo cross-listed Embedded firmware is a hard target for LLM coding agents because correctness depends on runtime behavior under sensing, timing, and safety constraints, yet existing evaluations mostly test one-shot synthesis or offline correctness. The authors present a closed-loop benchmark on simulated ESP32 firmware. In each task an agent receives a plain-text requirement, a constrained workspace, and build-and-runtime access, and must iterate until the device behaves correctly. Five control tasks are run under four feedback scenarios: one-shot generation, realistic self-verification, CI-style red/green feedback, and oracle-style detailed feedback. Across 420 runs of seven configurations, gpt-5.4 has the highest pass rate without saturating the benchmark, and qwen3.5-27B is the strongest local model, while smaller local models degrade sharply in both pass rate and search efficiency.

arXiv:2610.11447 PDF

Tracing the Thoughts of a Coding Agent Playing ARC-AGI-3: Lessons for Continual Learning

Chen Wu, Josh Passenger, Yin Song The authors study how a coding agent running on a frozen foundation model learns across a sequence of tasks. The agent plays ARC-AGI-3, a set of interactive reasoning games with no instructions, and keeps no state across turns except the Python scripts and text notes it writes. A measurement protocol traces each belief, rule, or plan through these files across seven runs using three backbones. Only 33 of 630 script references cross a task boundary, because scripts tend to embed level-specific state. The agent instead rewrites the general rules into new scripts and abandons 74% of the scripts it wrote before a boundary. Its notes are only ever appended to, so contradictions accumulate and are resolved against the full log, which means forgetting is selective rather than catastrophic. The costliest failure is a hard-coded value carried into a task where it no longer holds.

arXiv:2610.11450 PDF

SAIL: Scientific Agentic Intelligence via a Science-Aware Loop

SAIL Model Team, Boyuan Sun, Bryan Dai, Che Liu, Chi Liu, Derek Li et al. SAIL is an open mixture-of-experts model with 35B total and 3B active parameters for literature research, scientific coding, and multi-step research workflows. It is built through a science-aware improvement loop: agents running on frontier models diagnose its task failures in evidence search, scientific reasoning, and planning, then use paper collections and code repositories to build targeted training problems, trajectories, and executable environments. Over several development cycles, the model is trained with supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning. The authors report competitive performance on scientific research tasks with substantially fewer parameters than leading open-weight models.

arXiv:2610.11451 PDF

Generative Adversarial Loops

Kislay Aditya Oj, Nidhi Jain, Sri Surya Varma Datla, Priyanka Jayaswal, Kumar Krishna Agrawal, Aditya Desai Automated research systems mostly focus on discovering new methods, while the creation of benchmarks to test them has received little automation. Generative Adversarial Loop (GAL) alternates two agentic searches: a discriminator that generates adversarial data exposing weaknesses in current algorithms, and a generator that discovers algorithms to fix them. Applied to efficient-inference techniques, the discriminator finds weaknesses in state-of-the-art methods for KV compression, sparse video generation, sparse attention, and context extension. The discovered fixes also improve results on established benchmarks. For example, GAL raises CompactorPress on 4x KV compression with Qwen3-4B from 0.35 to 0.97 on the discriminator's data and gains 0.77 points on RULER-HARD, and it lifts Dual Chunk Attention from 0.20 to 0.90.

arXiv:2610.11458 PDF

Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents

Xing Zhang, Guanghui Wang, Yanwei Cui, Ziyuan Li, Wei Qiu, Bing Zhu et al. Self-improving agent loops on open-ended tasks rely on hand-written rubrics or bare LLM judges, which invite reward hacking and shared blind spots. The authors make the verifier itself evolve: it is an inspectable expression over small, mostly deterministic failure detectors synthesized from clustered failures. Detectors are selected for agreement with a ten-item anchored reference set and for consensus on unlabeled outputs, never for the agent's score. On MBPP+ the evolved verifier gains 0.21 held-out agreement over the hand-authored seed. A key finding is that removing the anchors collapses the verifier into an always-pass grader that still trains skills equally well, so downstream task score cannot validate an evolved verifier. Paired with a skill loop in a setup called Double Ratchet, the verifier retains 88–110% of the improvement that ground truth or a rubric provides, across code generation, text-to-SQL, and report generation.

arXiv:2610.11464 PDF

GROB: A Multi-Agent Architecture for Public-Trace Investigation of Candidate Agentic Activity

Chiara Bonfanti, Cataldo Basile cross-listed GROB is a multi-agent architecture for investigating suspected autonomous-agent activity using only public Internet traces when privileged telemetry is unavailable. It performs controlled, read-only collection and preserves selected observations so they can be resolved later as new public evidence emerges. In a frozen September 2026 corpus, the strongest result is that Census-labelled identifiers captured on 9 September were later resolved through public revision records to specific Census requests from 16–17 June. Other links to later evidence are weaker, and the authors note that public traces alone cannot establish which organization was responsible.

arXiv:2610.11467 PDF

Evaluating Local Language Model Agents for Reproducible Data Engineering: An Empirical Software Engineering Study of Mobility Workflows

Jorge Garc\'ia-Carrasco, Javier Sanchis, Alejandro Reina-Reina, Alejandro Mat\'e, Juan Trujillo cross-listed The study tests whether open-weight large language model (LLM) agents running locally can produce correct, reproducible data-engineering artifacts, not just plausible text. Its benchmark has fifteen mobility-workflow tasks, from data discovery and transport-feed processing to visualization and reporting, each scored by deterministic checkers. Ten local configurations were run in one-shot and closed-loop workspace modes on a consumer GPU, giving 1,500 scored attempts. For models above two billion parameters, letting the agent inspect and repair intermediate artifacts in a workspace raised pass rates by 26.7 to 52.0 percentage points; the best configuration reached 85.3% success, and a quantized 9B model reached 69.3% in about 6.5 GB of memory.

arXiv:2610.11482 PDF

Constitutional Gating and Deterministic Recovery for Multi-Agent LLM Negotiation: Ablations Against a Stateful Adversarial Gatekeeper

Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO, Inc.) Multi-agent LLM systems negotiating with a stateful counterpart waste model calls on polite loops, malformed outputs, and compliance deadlocks. The authors test a control stack against an adversarial Gatekeeper whose acceptance rules are fixed regular expressions. The stack has three parts: a runtime constitution, a 4-tier swarm (a Director, a three-agent majority vote, a Monitor, and a schema hard gate), and Cognitive Annealing, which detects deadlock deterministically, purges the agent-side context, and sends a canonical recovery message. Across 30 runs with Gemini 2.5 Pro agents, the constitution raised unlock rates only modestly, from 0/5 to at most 2/5. When the agents fell into a deadlock, atomic purge plus a canonical message escaped 5 of 5 times versus 0 of 5 for LLM-only steering, while LLM-written recovery messages failed a deterministic pre-flight check every time.

arXiv:2610.11542 PDF

Safe, Persistent, and Evolving Agent Harness for Understanding Partially Observable Worlds

Yisen Gao, Yue Guo, Qing Zong, Yiwen Guo, Yangqiu Song Enterprise agent workflows require strict policy compliance, handling of hidden side effects that tool feedback does not reveal, and state tracking across long tasks. E-Ledger is a multi-agent harness that checks every proposed action against policy before running it. It also keeps a ledger of verified hidden rules and evidence-backed state. Its companion framework WorldAbduct analyzes execution trajectories from four perspectives to hypothesize hidden rules, then tests them with targeted interactions before adding them to the ledger. On the World of Workflows benchmark, it improves safe task completion by 5–15 percentage points over the strongest evolution baseline across four LLM backbones, and it also carries over to ScienceWorld and DiscoveryWorld.

arXiv:2610.11552 PDF

SWE-Journey: Towards More Realistic Evaluation of Coding Assistants through Long-Horizon, Multi-Turn Interaction

Hexuan Deng, Yue Wang, Wenyu Jiang, Cheng Yang, Haolin Yang, Zhaohua Zhang et al. Benchmarks for coding assistants such as Claude Code and Codex rarely reflect long tasks in evolving repositories or multi-turn clarification with users. SWE-Journey builds long-horizon coding tasks with a weak-to-strong synthesis pipeline. It simulates users with an agent built on four personas mined from real interaction data. Models pass over 75% of tests for requested functionality when working with software architects but fewer than 25% with non-coders. The authors attribute the gap to three capabilities: asking the right questions, finding the right code, and making the right fixes.

arXiv:2610.11559 PDF

Incremental Open-Ended Deep Research with Structured Harness

Meilin Chen, Hongyuan Bao Open-Ended Deep Research systems write reports from scratch, which is wasteful when a report must be kept current as new information appears. Incremental Open-Ended Deep Research treats a report as an evolving state that keeps valid content, revises outdated parts, and adds new findings. Its Structured Harness represents reports as outlines, sections, and evidence, with structured retrieval, a persistent evidence pool, and selective generation. On DeepResearch Bench and DeepConsult, using a temporal evaluation that spans ten years, it keeps report quality competitive while improving continuity. On DeepResearch Bench it uses 33% fewer tokens and 61% fewer search calls.

arXiv:2610.11566 PDF

Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies

Yi Wen, Derong Xu, Pengyue Jia, Yichao Wang, Yingyi Zhang, Maolin Wang et al. Retrieval-based memory for LLM agents usually applies one strategy to every memory, even though memories differ in kind. The authors release TriMEM, a dataset with precise memory-type annotations across varied scenarios. They also propose MemoType, which uses a learned router to classify both memories and queries and then applies a retrieval strategy suited to each type. They prove that any single retrieval strategy has an upper bound on expected precision in a corpus with multiple memory types. Across three datasets, MemoType improves Recall@1 by up to 16.18%.

arXiv:2610.11573 PDF

Chronos Enables Code Agents to Reason over Software Evolution

Xin Yin, Yiang Zhang, Zhiyuan Peng, Chao Ni, Zhe Cui, Xiaohua Xin cross-listed Merged pull requests record the design decisions and constraints behind a codebase, but code agents rarely use that history. Chronos distills pull requests into structured experience cards linked by a typed graph of code-level, developer-intent, and organizational relations. It retrieves relevant cards through semantic search followed by multi-hop expansion, then uses them to guide two patch-generating agents and a steward agent that picks between their patches. On SWE-Bench Verified it improves SWE-Agent across all six tested backbones, raising the mean resolution rate from 69.2% to 72.9% and reaching 79.8% with MiniMax M2.5. It also improves results on SWE-Bench Pro and FEA-Bench Lite.

arXiv:2610.11578 PDF

Runnable Commit Untangling for Coding Agents

Jinfeng Jiang, Dongsun Kim, Dayi Lin, Zhou Yang cross-listed Coding agents tend to produce large, tangled patches that mix several purposes, and prior commit-untangling methods neither keep each split commit runnable nor show any real maintenance benefit. RucTangle is an agentic method that splits a patch into ordered commits while keeping the code runnable after every commit, and TangleEval measures whether the resulting histories help coding agents repair bugs. On 131 agent-generated patches, every RucTangle history was runnable, while 20.6% to 37.4% of the histories from four baseline methods were not. On 453 patches that introduced regressions, giving repair agents the untangled history as context raised pass@1 by 5.2 percentage points.

arXiv:2610.11593 PDF

Error-Propagation Modeling for Failure Attribution in LLM-Based Multi-Agent Systems

Jiaqi Liao, Yuanzhao Zhai, Huanxi Liu, Xu Zhang, Zheming Zhuang, Dawei Feng et al. When an LLM-based multi-agent system fails, the outcome rarely points to the decisive error, meaning the agent and step whose correction would have rescued the run. Existing attribution methods flag suspicious steps without modeling how errors cascade between agents or persist in unresolved loops, so they confuse root causes with downstream symptoms. Error-Propagation Modeling for Failure Attribution (EMFA) builds a structured representation of the failed trajectory, models both cascading propagation and persistent loops, screens candidate steps with that model, and then confirms the decisive error through counterfactual verification. On the Who&When benchmark it sets a new state of the art for step-level attribution, improving on the previous best by 3.45 and 4.40 percentage points on the Hand-Crafted and Algorithm-Generated subsets, and stays competitive at the agent level.

arXiv:2610.11600 PDF

AgentEvolver: System-Wide Self-Evolution Through Task Execution

Wentao Zhang, Fuchao Yang, Yilei Zhao, Xinrun Wang, Bo An An agent can finish a task without becoming any better at its work, and turning that experience into reusable capability means tying each changed component to its evaluation and later use. AgentEvolver keeps the foundation model fixed and lets eight families of entities evolve during task execution through a shared versioned lifecycle; these cover operations, methods, agents, control flow, interfaces and supporting state. A shared runtime, persistent planning and recoverable context keep long tasks on track. The authors report an 82.08% resolution rate on SWE-bench Pro Public with evolution enabled, above their own no-evolution baseline, and six case studies show retained capabilities being reused in later website, game and research work alongside documented failures. Transfer to independent tasks and total development cost are left as open questions.

arXiv:2610.11613 PDF

One Skill Too Many: How Co-Installed Skills Conflict in Coding Agents

Chaoliang Yan, Zihao Xu, Yuekang Li, Shangzhi Xu, Yi Liu, Gelei Deng et al. cross-listed Coding agents are extended with skills, which are directories whose SKILL.md file tells the model when and how to perform a task, and an installed skill can end up alongside a similar skill from another source that does the same job. The authors mined 822,109 candidate similar-skill pairs from 20,947 repositories and ran 312 confirmed conflicting pairs on three models across 6,368 runs. Nearly one in four installed skills has such a duplicate, and the similar skill takes over one run in five without lowering task completion, which means benchmarks that only check task completion miss the problem. Runs that open the similar skill first lose over a third of the core functions only the installed skill provides, such as a ban on touching git. Install location decides which skill runs, and a pre-tool hook placed at the first skill read restores those functions.

arXiv:2610.11647 PDF

SkillContrast: Difference-Guided Text Selection for Agent Skill Reranking

Jiandong Ding, Honglei Ji, Ming Liu, Tao Duan cross-listed Agent skills that look alike often share most of their instructions and differ only in when they should be used, so choosing skill text by query relevance can keep the shared parts and drop the distinguishing ones. SkillContrast is a training-free selector that compares the retrieved skills with each other and passes the text where they differ, plus some surrounding context, to a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it produced 54-72 more clean hits (a helpful skill retrieved without its marked risky sibling) than TF-IDF query selection at the same input length, across two retrievers and two reranker sizes. Compared with feeding full skill bodies, it uses 51-59% fewer input tokens, scoring slightly lower with the 0.6B reranker and the same or better with the 4B reranker.

arXiv:2610.11650 PDF

Harness Evolution Hits a Ceiling: When Weight Training Should Begin

Yuan Tian, Bing Hu, Hao Wang, Binghang Lu, Fang Wu A long-horizon LLM agent can be improved by evolving the harness around a frozen model or by training the model's weights, and the authors study how to choose between the two. They label each failed trajectory by the first failure signal that fires, which separates process failures (blocked calls, loops, running out of steps) from content failures (a plan that is delivered but poor). A self-evolving harness fixes process failures, lifting Qwen3.5-4B on DeepPlanning from 0.16 to 0.30 and Qwen3.5-9B from 0.32 to 0.44. LoRA adapters trained on trajectories from the evolved harness carry that gain into the weights, and on 9B the adapter alone matches the full harness-evolution result, cutting content failures from a quarter of trajectories to one in twenty, while a placebo adapter trained on shuffled answers does worse than the base model. The harness loop also transfers to WebArena-Lite (+0.09 on 117 unseen tasks), but there the adapters add nothing on top of it.

arXiv:2610.11655 PDF

TRACE: Diagnosing Verifier Brittleness in Agentic Evaluation

Radhika Gaonkar When an LLM agent's verifier score changes, the change is usually read as a change in capability, but it may come from how the evaluation is presented or scored. TRACE is a protocol that changes one part of an evaluation, compares paired runs, checks whether the agent's behavior actually changed, and rescores unchanged trajectories to see whether the scoring rule is responsible. On public τ²-bench tasks, renaming tools or reformatting tool outputs left reward within ±0.10 for seven of eight agent-change pairs, while deliberately misleading tool names cut every agent's reward by 0.20-0.44, showing the setup can detect real effects. Identical reruns flip 15-36% of task outcomes, so single-run comparisons are unreliable, and two frontier LLM judges disagree on 57% of the same records, largely because one grades procedure rather than outcome.

arXiv:2610.11678 PDF

DeltaReplay: Task-Relative Memory Reuse for Mobile GUI Agents

Yudong Bai, Yihong Chen, Quanming Yao, Yaqing Wang Memory-augmented mobile GUI agents reuse past successful trajectories, but a stored trajectory rarely matches a new task exactly: parameters may differ, only some steps may overlap, or nothing relevant may be stored. DeltaReplay stores trajectories as paths in a transition graph whose nodes are app pages and whose edges are actions. Each action is split into a task-independent operation and task-specific parameters. At every step the agent decides whether to follow the recorded step, replay it with new parameters, or hand control back to the base agent. With the same backbone as the base agent, it raises task success by up to 10.3 percentage points on AndroidWorld and 25.0 percentage points on SPA-Bench.

arXiv:2610.11707 PDF

MemTrial: Learning When to Trust Memory in LLM Portfolio Agents

Guanghao Wu, Zhuo Cai, Shoujin Wang LLM agents for portfolio management credit each remembered experience with the outcome of the decisions that used it. Because returns are dominated by market-wide moves, this credit tracks the market rather than the experience, and such agents often underperform an equal-weight (1/N) portfolio. MemTrial drafts each decision under eight combinations of retrieved experiences chosen by a fractional factorial design, so the shared market effect cancels out. It credits each experience with its Banzhaf value, pools these noisy credits with a hierarchical Bayesian model, and stays anchored to a conservative reference like 1/N until the credits have predicted unseen dates. On PortBench and InvestorBench it ends at most 2.2% below 1/N, while the best experience-learning agent loses 15–38%, and averaged over five settings it improves that agent's utility by 21.2%.

arXiv:2610.11732 PDF

What Output-Only Review Cannot Verify: Study Contracts for Research Agents

Eitan Waks, Ben Glocker Some defects in studies produced by AI research agents cannot be detected from the outputs alone, because detecting them requires knowing what was approved before execution. The authors propose study contracts that bind declared experimental choices, required runs, and claim scope to recorded execution evidence, and they distinguish this contract-relative check from scientific truth. In a small diagnostic of eight self-authored clean/mutated pairs, a deterministic contract checker caught all eight registered mutations. LLM judges given filtered artifact packages flagged 104 of 144 mutated cases and abstained or failed on the rest. The authors stress that this setup does not measure comparative verifier quality and list the full-information and closed-loop agent evaluations still needed.

arXiv:2610.11754 PDF

Narrow and Deep: An Ontology Tower as the Knowledge of an LLM Agent for an Industrial Equipment System

Younghwan Joo, Sung-il Kim cross-listed The authors propose an "ontology tower" as the knowledge source for LLM agents that operate industrial equipment. It is narrow in entities but deep in knowledge: it adds quantities derived from measured points through physical relations, plus lessons from the operating journal stored as knowledge nodes. In a preregistered nine-task evaluation on a real low-humidity air-handling plant, four open-weight models from 9B to about 750B parameters avoided each task's most plausible misjudgment about 20 percentage points more often, and the 9B model's overall score rose as much as the largest model's. Agents used operating lessons when they were built into the tower or placed in the prompt, but seldom when the lessons sat behind a search tool. In live runs through a safety layer, the agents brought the controlled variable into its target band in 12 of 14 runs.

arXiv:2610.11768 PDF

Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks

Haoyu Zhao, Zhengxu Yu, Zhiyuan He, Meng Fang, Rasul Tutunov, Haitham Bou-Ammar et al. Memento 3 lets a frozen LLM agent continually learn an explicit world model of an unfamiliar environment through external memory. The agent keeps a natural-language rulebook of revisable hypotheses, compiles it into executable code for prediction and planning, and refines both using prediction errors. Code updates are accepted only when they are judged faithful to the rulebook and replay the observed transitions exactly, and a population variant runs several world models in parallel to guide exploration. On ARC-AGI-3, the single-model agent clears every level of all 25 public games while using 44% of the human action count, and in Atari Pong a learned controller wins 21:0 in each of three episodes without further LLM calls.

arXiv:2610.11794 PDF

Trajectory-Guided Fault Localization for Agent Skill Evolution

Yu Ge, Linna Xie, Zhong Li, Yu Pei, Tian Zhang cross-listed Agent skills are reusable guidance documents for code agents, and flaws in them can degrade task execution, but LLM-generated revisions are often not grounded in what the agent actually did. SkillMorph applies fault-localization ideas to skills: it compares abstracted trajectories from failed and successful runs across repeated tasks and evolution rounds to find suspicious actions, maps those actions to specific parts of the skill, and revises only those parts. On SWE-Skills-Bench and CannBot, the evolved skills achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods. In an automated kernel-generation deployment, an operator-development team accepted 6 skill-revision pull requests.

arXiv:2610.11858 PDF

How Is Automated Research Evaluated? A Survey of Benchmarks and Evaluation Practices

Liulei Zhang, Dejing Zhou, Chuyue Huang, Guanhua Chen, Yutong Yao, Lidia S. Chao et al. Surveys how automated research systems are evaluated across six targets: literature synthesis, ideation, executable workflows, scholarly writing, automatic peer review, and end-to-end research. It compares how benchmarks construct tasks, source evidence, choose evaluators, and score results, to clarify what each design actually measures. Three lessons recur: output checks, process checks, and human studies provide complementary information; evaluator calibration only holds for the specific property being assessed; and resource budgets and attempt selection must be reported for performance comparisons to mean anything. The survey closes with concrete reporting and audit recommendations.

arXiv:2610.11877 PDF

Forms of LLM-Integrated Applications from LLM-Chats to Autonomous AI Agent System

Irene Weber (University of Applied Sciences Kempten, Germany) Asks whether marketing labels for LLM-integrated software, such as chatbot, copilot, RAG, workflow, coding agent, and AI agent, correspond to real architectural differences. Drawing on research papers and vendor documentation for 22 systems, the survey describes seven recurring forms in a shared vocabulary of agents and tools, characterizing each by architectural pattern, execution control and user intervention point, agent calls per task, and tool use. It finds that the labels do carry architectural meaning: for example, 'copilot' denotes a router-worker design with step-by-step user confirmation, and the coding agents from four major providers share a single architecture, a reason-and-act loop that delegates to subagents.

arXiv:2610.11899 PDF

Event-Centric Memory with Query-Aware Graph Augmentation for Long-Term Conversational Agents

Yichen Liu, Chunfeng Yuan, Haowei Liu, Wenjuan Li, Zefeng Lin, Bing Li et al. Long-term conversational agents need memory that tracks events and state changes. Flat memory stores leave those relations implicit, while graph-based memory is costly to build and accumulates irrelevant links over long histories. QGMem turns dialogue histories into event-indexed atomic memory units and merges related units into traces that record how states change over time. At query time it uses hybrid retrieval and query-aware reranking to build a compact working memory, arranges that memory as a local graph, and passes the graph to the LLM as a graph token alongside the text. Across six benchmarks it shows consistent gains in retrieval, multi-hop reasoning, conflict resolution, and ultra-long dialogue reasoning while keeping contexts compact and inference cost moderate.

arXiv:2610.11920 PDF

Agentic-TTT: Training test-time policy for test-time training

Jiahao Lu, Mohan Kankanhalli Test-time training (TTT) updates an LLM's parameters using signals drawn from test inputs, but any given TTT algorithm only helps in some settings, and applying the wrong one wastes compute or hurts performance. Agentic-TTT learns a test-time policy that decides when to run TTT, which algorithm to use, and whether a previously learned skill can be reused. TTT procedures are exposed as callable tools, accumulated skills are treated as an evolving environment, and the policy is trained on the utility gains its decisions produce. On the authors' benchmark, Agentic-TTT nearly doubles utility over the backbone model, learns to trade utility against compute, and generalizes to domains not seen in training.

arXiv:2610.12002 PDF

When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation

Yiruo Cheng, Shen Huang, Xiaoshuai Song, Jiejun Tan, Guanting Dong, Pengjun Xie et al. LLM agents usually reason before every action, even though earlier reasoning often still covers the steps that follow. The authors find that the drop in likelihood of later reference actions, when a reasoning step is removed, is a cheap and reliable signal of whether those actions are still supported by earlier reasoning. Building on this, RACE (Reasoning Adaptation through Cross-Turn Estimation) uses a procedure called LoGiC to progressively find reasoning turns that can be removed with little effect. These removal signals feed both supervised fine-tuning and agentic reinforcement learning, so the policy learns when to reason and when to act directly. On four agent benchmarks, RACE substantially reduces reasoning cost while maintaining or improving task performance.

arXiv:2610.12061 PDF

EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution

Yanning Dai, Yuhui Wang, Nanbo Li, Wenyi Wang, J\"urgen Schmidhuber LLM-driven program evolution relies on evaluation feedback, which is expensive to compute, and existing methods split their evaluation budget across candidates with a fixed strategy. EvoAlloc is a resource-allocation agent that periodically distills past search and allocation outcomes into reusable experience and uses it to revise its strategy. It also occasionally evaluates candidates it had turned down (counterfactual exploration) to learn from them. On coding and agent-harness optimization benchmarks it reaches baseline performance with 59-82% fewer full evaluations and 61-89% fewer LLM tokens, and scores 8.7-12.0% higher at an equal evaluation budget.

arXiv:2610.12086 PDF

Use and Disuse: Intent-Structured Experience Consolidation for Memory and Learning in LLM Agents

Xiangyi Zeng, Baihang Liu, Xutong Wang, Ze Jin, Yunpeng Li, Qixu Liu LLM agents that run for long periods need to turn continuous experience into reusable knowledge without carrying their full history forward. Hippocam is a hierarchical memory architecture that organizes an agent's work as nested intents. Completed intents are condensed to the outcomes and state that later work needs, and older history is recursively consolidated so that unused experience gradually becomes more abstract, while the original interactions remain recoverable on demand. Recalled memories are consolidated again together with new experience, so agents can learn and build skills from their own experience without parameter updates. The abstract reports no quantitative results.

arXiv:2610.12124 PDF

Recursive Self-Improvement through Multi-Agent Self-Supervision

Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia et al. Recursive self-improvement on open-ended tasks hits a supervision bottleneck once outputs exceed what experts can judge reliably, and a single model instance is poor at critiquing its own complex reasoning. Multi-Agent Self-Supervision (MASS) has one base model propose, execute, and self-evaluate multi-agent workflows, using an evolutionary search with structural guardrails to discover effective roles and information routing. The model is then fine-tuned on its own trajectories, and the cycle repeats. Over two cycles with Qwen3.6-27B, it achieves 1.2 to 1.6 times higher performance per output token on four open-ended benchmarks, and a student trained on multi-agent traces outperforms a single-agent student trained on 1.4 times more tokens.

arXiv:2610.12176 PDF

A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization

Ming Chen, Rong-Xi Tan, Ke Xue, Yu-Jie Zhou, Taiye Lu, Zhi-Xuan Gao et al. Studies of large language model (LLM) agents for black-box optimization (BBO) use different domains and setups, so their results are hard to compare. AgenticBBO-Bench evaluates agents under a single finite-budget protocol across synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design. Agentic BBO beats direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. Adding more numerical tools does not consistently help, task semantics are broadly useful while specific priors are less reliable, and on a five-task frontier challenge run with the Codex agent harness, GPT-6 Astra and DeepSeek-V4.1-Flash sit on the cost-performance Pareto frontier.

arXiv:2610.12183 PDF

DataSense-Bench: The First Step Toward an AI Scientist

Yudi Zhang, Mingyu Cao, Lu Yin, Mykola Pechenizkiy, Shiwei Liu The benchmark tests whether frontier AI agents have a "sense of data": can they choose the training subsets that will most improve a small LLM after fine-tuning? Agents can inspect candidate data, write and run analysis code, and run forward passes, but they cannot train the model or see the evaluation tasks. The selected subsets are then fine-tuned and scored on terminal problem solving (OpenThoughts-Agent data, evaluated on TBLite) and tool use (EnvScaler data, evaluated on BFCL). Gains over random selection are limited, agents do not reliably rank their own picks, and ranking skill does not carry across tasks. Execution traces show that agents often look at similar data signals but disagree about what those signals mean for training value.

arXiv:2610.12190 PDF

Language Models as AI Research World Models

Zijun Wang, Zewen Liu, Minhua Lin, Zhaotian Weng, Zhan Shi, Bing He et al. AI research agents can propose experiments faster than they can run them, so predicting experiment outcomes becomes a key capability when compute budgets are limited. The authors evaluate language models as Research World Models (RWMs) that forecast the results of candidate interventions, using more than 2,600 experimental records from nine environments covering pretraining, post-training, and inference, which together represent over 171,000 H100 GPU-hours. Research knowledge drawn from real experiments improves predictions within an environment and transfers across environments. For example, pretraining experience from OLMo3, Marin, and Nanochat reduces selection regret in the Qwen3 environment by 78%, and it raises the best gain found in multi-round Autoresearch by 11.6–15.8%. Ablations across 13 models show that adding research knowledge improves intervention ranking more than switching models or increasing reasoning effort alone.

arXiv:2610.12235 PDF

HarnessSQL: Harness-Native Training for SQL Agents in Realistic Database Environments

Haolin Yang, Jipeng Zhang, Jian Xie, Shuaishuai Gong, Sirui Han, Yike Guo Text-to-SQL models are usually trained to map a question straight to a single query, yet deployed database agents work through multi-turn, stateful sessions: they inspect schemas, run probe queries, and fix errors inside an execution harness that is only introduced at inference time. HarnessSQL closes this train-deploy gap by building isolated, executable database environments with hidden execution oracles. Teacher models are rolled out inside the target harness, only verified trajectories are kept for supervised fine-tuning, and execution-reward reinforcement learning follows. On Spider 2.0-SQLite this raises Qwen3-8B from 15.5% to 45.2% and Qwen3-14B from 22.2% to 54.8% execution accuracy, with gains transferring to the out-of-distribution interactive benchmarks BIRD-Interact and LiveSQLBench.

arXiv:2610.12274 PDF

Prior or Feedback? What an LLM Uses When Adapting Neural Operators

Julian Chan, Javier Mora Jimenez This study asks whether LLM scientific agents rely only on their initial task context or actually adapt to experimental feedback, using an LLM that picks fine-tuning configurations for neural operators on partial differential equation (PDE) problems under a limited trial budget. The LLM achieves lower held-out error than random search and Bayesian optimization in nearly every matched comparison. Controlled interventions show its first configuration already ranks near the top of the random-search pool, and its chosen learning rate shifts with the PDE description. Reassigning validation scores among evaluated configurations changes the next proposal in every tested case, establishing that the agent combines a task-dependent prior with genuine sensitivity to feedback.

arXiv:2610.12325 PDF

Can AI Agents Learn Their Way to the Top? Evaluating Heuristic Learning in a Long-Running Game Agent Competition

Kaisen Yang, Qingle Liu, Kejin Wang, Yicheng Zhao, Jieming Li, Shenghan Zheng et al. Adversarial Heuristic Learning (AHL) treats AI agents as learning engines: instead of updating model weights, they improve game-playing programs by revising code. The AAArena benchmark includes 12 real adversarial games and 1,920 archived human programs, with an evaluation protocol modeled on real competitions in which agents interpret rules, pick opponents, analyze replays, and revise their game bots within fixed budgets. Opus 5.5 running in Claude Code earns 6 gold medals, but no tested configuration reaches the top of the other 6 human ladders, and performance drops as game rules get more complex. Choosing opponents and receiving dense feedback both help, and agents learn from replays of their own matches as well as from other players' matches.

arXiv:2610.12341 PDF

Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict

Kaiser Sun, Bernal Jimenez Gutierrez, Hongjun Liu, Jingyu Zhang, Jie Gao, Mark Dredze et al. The authors evaluate whether LLM agents show epistemic humility (EH) when retrieved evidence conflicts with what they already believe. They break EH into three trajectory-level behaviors, Identify, Solve, and Escalate (ISE), and test them under both controlled and naturally occurring knowledge conflicts, each paired with matched no-conflict controls. Across four agents, higher task accuracy does not imply greater humility: some accurate agents notice a conflict during execution yet present incorrect final answers without flagging the unresolved uncertainty. Agents often detect conflicts early but lose track of them in later steps, and model-level interventions that improve EH tend to reduce accuracy.

arXiv:2610.12360 PDF

OnTrack: Real-Time Monitoring and Intervention in LLM Agent Trajectories via Streaming Structure-Aware Optimal Transport

Babak Barazandeh, Connor Swanson, Chinmay Kulkarni, Nikhil Mungel Autonomous LLM agents can waste compute or take irreversible actions, while existing safeguards either add a monitoring agent to every step or only judge runs after they finish. OnTrack streams the agent's steps and their dependencies, compares them against recorded successful runs using structure-aware optimal transport, and can alert or block in about a millisecond per step. It operates under three levels of reference access, from full historical runs down to step logs alone, detecting anything from plan violations to loops, stalls, and repeated tool calls. On SWE-bench trajectories it ranks failing runs from their first 8 steps better than content-similarity methods (+0.057 AUROC), and an abort policy saves about 18% of compute, with 5 of 6 aborts correct.

arXiv:2610.12375 PDF

ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills

Hongxing Li, Dingming Li, Yixin Li, Yong Du, Wenqi Zhang, Weiming Lu et al. cross-listed Skill-augmented agents reuse strategies distilled from successful trajectories, but most store skills as text, which loses spatial layout and action-state structure, and they update skills separately from policy training. ViSkill encodes successful interactions as composite visual skill cards that vision-language model (VLM) agents read directly. Retrieved skills guide both inference and reward shaping, and new successful trajectories are distilled back into the library, so skill accumulation and policy improvement feed each other; an optional cold-start mechanism speeds up early learning. On Sokoban, FrozenLake, and PrimitiveSkill, it reaches an overall success rate of 0.89 (0.91 with cold start), beating all evaluated proprietary and open-source baselines and converging faster than standard PPO.

arXiv:2610.12403 PDF

BrickBench: Evaluating Agentic Brick Design

Peter Kulits, Yiqing Xu, R. Kenny Jones, Cordelia Schmid, Jiajun Wu BrickBench is a benchmark in which an agent turns a text prompt into a LEGO-set design that must meet semantic and design criteria and also be physically buildable from a discrete parts library. Designs are scored for validity, alignment with the prompt, and design quality across three settings that vary in scale and part availability. The authors also provide BrickAgent, an environment in which coding agents build, inspect, and validate their designs. Leading agents largely satisfy the verifiable physical and semantic requirements but still fall short of human designs.

arXiv:2610.12452 PDF

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

Drew T. Nguyen, William Fithian METR's 50% time horizon expresses AI capability as the length of software tasks, measured in human completion time, that an AI solves half the time. Using 228 tasks and 26 AI systems, the authors recompute time horizons with splines and item-response theory instead of assuming that AI difficulty grows linearly with the log of human time. The fitted curve is nearly flat between about 2 and 30 minutes and close to linear elsewhere, so a jump from 3 to 30 minutes is much easier than one from 30 minutes to 5 hours even though both are tenfold. The new estimates score better under cross-validated proper scoring rules, and the authors provide diagnostic plots that they recommend reading alongside any time-horizon benchmark.

arXiv:2610.12466 PDF

4 more specialized papers

Other 74

D-SLR: The Disjoint Row-Sparse plus Low-Rank Decomposition

Vincent Szolnoky Matrix compression for reconstruction usually relies on the truncated singular value decomposition (SVD), and methods that add a row-sparse correction typically need iterative solvers and regularization tuning. The D-SLR (Disjoint Row-Sparse plus Low-Rank) decomposition is a closed-form replacement in which each row is either stored verbatim or approximated by the low-rank fit, never both. The authors prove that under squared error this restriction loses nothing: the joint optimum is reachable disjointly with fewer parameters, and with zero stored rows the method reduces to the truncated SVD, so it never does worse at equal cost while needing only three SVDs and no tuning. A computable lower bound certifies how much any other choice of rank and stored rows could gain, and experiments cover LLM embedding tables, network traffic, and hyperspectral images.

arXiv:2610.10636 PDF

Learning infinite context windows in recurrent architectures via spatial neural computing

Aleix Salvador-Pomarol, Arthur N. Montanari, Earl K. Miller, Adilson E. Motter, Jorge Cort\'es Recurrent neural networks (RNNs) run in linear time with constant memory but struggle with long-range dependencies. The authors replace neuron-to-neuron communication with a spatially evolving field governed by discretized partial differential equations, inspired by cortical waves. They show the resulting second-order model is equivalent to an infinite-order RNN whose state depends on its entire history, with a fixed number of parameters. They also derive conditions for marginal stability that rule out vanishing and exploding gradients, and the architecture outperforms other recurrent models on long-horizon benchmarks while using substantially fewer parameters.

arXiv:2610.10690 PDF

MotherTree: Meta-learning on synthetic data improves decision tree training

Ziyuan Wang, Fredrik D. Johansson Classical decision trees are transparent and can be deployed on their own, but each one is learned from scratch, while tabular foundation models gain from meta-learning yet produce no standalone model. MotherTree is a tabular transformer pretrained with stochastic gradient descent on a synthetic prior, with no reference trees as supervision. Given a training set, it outputs a hard, axis-aligned decision tree in a single forward pass. On benchmarks with controlled sample sizes it is competitive with size-matched trees from recursive partitioning, gradient-based learning, optimal-tree methods, and distillation from foundation models. Used as an initializer followed by task-specific tuning, it outperforms from-scratch gradient-based tree learning on all benchmarks and sample sizes.

arXiv:2610.10832 PDF

Dynamics as Code: On Model Compression via Dynamic System

Fan Gao, Wei Su, Juntong Fan, Renfeng Peng, Hongyu Liu, Jinqiao Duan et al. Instead of pruning, quantization or distillation, dynamic-system (DS) compression stores each weight vector as an index into a trajectory generated by a dynamic system and recovers the weights at decompression time. The authors prove that, under a Diophantine condition, a finite irrational-winding trajectory forms an epsilon-net over the weight space. This ties state resolution, decompression error and compression ratio together in a predictable way. They generalize the framework to space-filling curves, chaotic systems, pseudo-random generators and low-discrepancy sequences, add KD-tree acceleration and outlier handling, and show competitive compression without retraining on ResNet-18, Qwen2.5-1.5B and Qwen1.5-7B.

arXiv:2610.11115 PDF

Phonological Interference in Multilingual Speech Models

Moran Yanuka, Raja Giryes, Moris Alper Phoneme-level speech models often fail on code-switched speech, where a speaker alternates between languages, and on low-resource languages they never saw in training. The authors trace this to phonological interference: a model assumes the input is in one language and applies that language's sound system, so it drops phonemes the assumed language lacks. Two phone recognizers and a phoneme-conditioned text-to-speech model lose 32% to 79% of language-specific phonemes on code-switched input. The interference sits in a low-dimensional subspace of the models' activations, and an inference-time fix called windowed language estimation (WLE) replaces the model's single language estimate with local, per-window estimates. WLE removes 34% to 69% of the interference while leaving the recognizers' monolingual performance essentially unchanged.

arXiv:2610.11275 PDF

Spectral Weight Decay: Inducing Low-Rank Structure in Neural Network Weights

Dmitrii Andriianov, Andrey Veprikov, Aleksandr Beznosikov Standard weight decay ignores the spectral structure of weight matrices. Spectral weight decay is a decoupled nuclear-norm update applied after each optimizer step that shrinks singular values additively rather than multiplicatively, and the authors connect it to approximate proximal descent. On LLaMA-style models from 124M to 500M parameters, it lowers effective rank and improves SVD-LLM compression at matched validation loss. At 500M parameters and a 4% distortion budget, it reaches 1.89× compression and a 1.18× GPU inference speedup, versus 1.14× and 1.01× with standard weight decay. Under 60% label noise, it also improves clean-test accuracy over L2 regularization by up to 17.8 points on MNIST and 4.6 points across four BERT-base tasks.

arXiv:2610.11730 PDF

Timer-M1: A Multivariate Time Series Foundation Model via Learning Primitives

Haoran Zhang, Haixuan Liu, Xingjian Su, Yong Liu, Zhi Chen, Yuxuan Wang et al. Timer-M1 is a pretrained multivariate time-series foundation model for zero-shot forecasting. It is built on primitives, the elementary temporal and cross-variable patterns that recur across domains. A synthesis pipeline generates series from shared temporal primitives and combines real and synthetic series into multivariate samples using relational primitives. Training episodes assign each channel a role: target, past-only covariate, or known-future covariate. The model uses gated two-dimensional Transformer blocks that adjust cross-variate attention by layer, and it ranks first on FEV and TIME and second on GIFT-Eval among recent time-series foundation models.

arXiv:2610.11734 PDF

Beyond QAOA: A Review of AI and Quantum Computing for Adaptive Combinatorial Optimization

Hoong Chuin LAU cross-listed This review covers adaptive quantum combinatorial optimization, where learned policies make choices such as problem formulation, penalties, shot budgets, backends, and whether to use a quantum processor at all. It structures 119 papers, 67 coded in detail, by the decision being learned. It finds strong evidence that machine learning helps quantum optimization, but evidence that quantum computation improves learned optimizers is confined mostly to small-scale simulations. Experimental controls are thin: 25 of 57 studies have no classical baseline, and the median experiment uses 17 qubits. The authors introduce an M0–M5 evidence hierarchy and find no broadly convincing result at its highest level, matched-resource practical advantage.

arXiv:2610.11759 PDF

Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits

Kanghui Ning, Marin Bilo\v{s}, James T. Wilson, Yilang Zhang, Kashif Rasul, Dongjin Song et al. The study asks which kinds of test-time compute improve strong pretrained tabular foundation models (TFMs), looking at adaptation, aggregation, and context construction across the TabArena benchmark plus wide and large OpenML tables. For adaptation, the authors introduce DiagScale, a diagonal query-key similarity update that trains only 0.003–0.03% of parameters yet matches full fine-tuning across three independently pretrained backbones. For aggregation, greedy selection from a pool of 96 configurations cuts error by 2.4% relative to TabPFN-3's default, while uniform averaging makes error worse. Attention-guided retrieval helps on some large tables, but the context-expansion methods tested give no consistent gain. Combining adaptation and aggregation adds further improvement at a much higher compute cost.

arXiv:2610.12005 PDF

Training on the Future: A Delay-Aware Audit of Test-Time Adaptation for Time-Series Forecasting

Mohamed Readh Fentazi, Mazene Ameur, Adlen Ksentini Test-time adaptation (TTA) methods for time-series forecasting update a deployed model from incoming ground truth, but the label for an H-step forecast only becomes available H steps later, and published evaluations often ignore this delay. The authors build a leakage-free harness that enforces realistic label delays inside the released code of four methods (TAFAS, COSA, PETSA, DynaTTA) across five benchmarks, and compare them with two cheap closed-form correctors: a hyperparameter-free bank of recursive least squares (RLS) filters and an ELF-style linear corrector. Beyond ETTm1, the largest significant gain any published method achieves over its own frozen checkpoint is half a percent, and at the shortest realistic delay at least one published method is significantly worse than the frozen model on each of those four datasets. Leaky next-step updates inflate apparent gains by up to 110%, and differences in backbone training change error by up to 25x, more than any adaptation effect measured.

arXiv:2610.12232 PDF

Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning

Julian Kleutgens, Mauricio Tec, Claudio Battiloro, Francesca Dominici, Giannis Daras RefineMix trains discrete diffusion models when in-domain data is very scarce, as is common in science, by using out-of-distribution data only at selected diffusion times. Because masking leaves the surviving tokens' domain information intact, related data is of limited use at high noise levels. At low noise levels, however, the domains' disjoint supports let the model learn from both kinds of data without biasing the sampler, and the authors formalize this with a theoretical analysis. Across five domain-shift settings the method matches or beats in-domain finetuning and data mixing. In protein generation with only 197 in-domain examples, it nearly doubles the fraction of generated proteins that are novel, foldable, and in-family compared with standard finetuning.

arXiv:2610.12340 PDF

63 more specialized papers

Large Language Models 73

Freeze the Decoder, Heal the Encoder: Parameter-Efficient Adaptation for SVD-Based KV-Cache Compression

Yufeng Wang Post-hoc SVD-based KV-cache compression factorizes a pretrained model's key and value weights into a down-projection (encoder) and an up-projection (decoder), then runs a short fine-tune, called "healing", to recover accuracy lost to truncation. Under a single shared learning rate, healing only the encoder looks like a clear win. The authors show this advantage is an artifact: a shared rate hurts the arms with more trainable parameters. With each arm's learning rate tuned separately, encoder-only healing reaches parity with the alternatives while using 3x fewer trainable parameters and 3x less optimizer memory, verified on Qwen2.5-VL-3B-Instruct and replicated on two text-only backbones. The shared-learning-rate pitfall applies to any comparison of fine-tuning recipes whose arms differ in parameter count.

arXiv:2610.10552 PDF

Beyond Type-checking: Towards Holistic Evaluation of Formal Specification Generation

Srijith Nair, Aditya Vempaty, Jia Liu, Ashish Jagmohan cross-listed In verifiable code generation, an LLM writes a formal specification that a proof is then checked against. Unlike the proof, the specification has no definitive check that it captures what the user actually wanted. The authors build an evaluation framework on 350 Lean tasks drawn from VERINA and CLEVER, covering formal validity, similarity and equivalence to reference specifications, and behavioral adequacy, and they score input acceptance separately from output constraints. They show that measurement coverage can reorder rankings, and give a constructed example where a specification scores 100% on postcondition tests while accepting 0% of required inputs.

arXiv:2610.10604 PDF

WorldBench: Evaluating LLMs on Three.js Voxel World Generation

Krish Bakshi cross-listed LLMs can now write complete interactive 3D worlds as code, but automatic judges that only inspect rendered snapshots or only read source code disagree on 32% of required items. WorldBench uses a single prompt for a floating voxel island in Three.js, and its judge explores the running world by controlling the clock, orbiting the scene, and sending a navigator agent to each biome. Code quotes must match the source and visual claims are checked against measured pixels. In a mutation test where features were removed by construction, code-only judging still gave full credit to four of five removed features, while the combined judge cut credit for removed features by about a third (5.44 to 3.55 of 7.11 points). Five frontier models are evaluated, including Claude Fable 5.1 and GPT-6 Astra.

arXiv:2610.10622 PDF

Recurrent Self-Improvement: Dynamic Cross-Loop On-Policy Distillation for Looped Language Models

Yi Wang, Rui Qian, Yu Li, Haoyang Yao, Wenjie Wang Looped Language Models (LoopLMs) reuse shared parameters across recurrent steps to scale reasoning, but post-training them is difficult because rewards are sparse and external teachers create context mismatch. LoopOPD uses a frozen terminal-loop policy as a teacher for an intermediate-loop student on the student's own rollouts, providing dense supervision with no external teacher. D-LoopOPD keeps refreshing that teacher as the shared weights update, which enables recurrent self-improvement. On Ouro-Thinking models, both methods improve mathematical reasoning, and despite math-only training, the models also improve on general reasoning and code generation benchmarks.

arXiv:2610.10623 PDF

Lossy Compressive Text Autoencoders

Vinko Sabol\v{c}ec, Angelos Katharopoulos, David Grangier The work learns compressed latent representations of text with an autoencoder that progressively downscales hidden states along the sequence (time) axis and then upscales them, passing through a residual, low-dimensional discrete bottleneck. The authors compare quantization methods, training objectives and datasets. They measure reconstruction quality at the surface level with BLEU and at the semantic level with an LLM-based judge, and also test the representations on downstream question-answering and semantic text similarity benchmarks. The compressed representations reach 2.24 bits per byte on web text, on par with lossless text compression algorithms, while keeping good reconstruction and downstream performance.

arXiv:2610.10738 PDF

Evaluating Rubric Generation with Interventional Transfer

Erik Skalnes, Layne C. Price, Raviteja Anantha, Michael Oberst LLM-generated, instance-specific rubrics could make expert-level evaluation scalable, but there is no good way to judge whether a generated rubric matches an expert one. Interventional Transfer treats two rubrics as similar if perturbing a response to pass or fail one rubric moves its score on the other. In a case study on HealthBench, rubrics generated by Qwen3.8-27B, Deepseek-V4-Flash and Opus-5 were used to grade responses from GPT-5.6-Terra. The study reveals an asymmetry: perturbations that degrade a response transfer between generated and expert rubrics, but perturbations that improve one do not reliably raise scores on the other. The authors argue this limits the use of generated rubrics for hill-climbing, and existing rubric-evaluation methods do not surface the asymmetry.

arXiv:2610.10809 PDF

Real Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than Recompute

Sietse Schelpe Models recompute their key-value (KV) state every time a prompt is sent. The authors evaluate galahad-kv, a memory layer that saves the KV state of each roughly 16,000-token block to encrypted local NVMe disk and later reloads it byte-exact. Over 50 million tokens of public text, served through vLLM on one H100 with Gemma 4 12B and Gemma 4 31B, every probed block reloaded without recompute. Loading was 2.8-4.3x faster than recomputing and used 8.8-12.3x less GPU energy, and GPU memory stayed flat. On facts planted millions of tokens earlier, the models answered 82/100 (12B) and 98/100 (31B) correctly. The authors note that this reuses stored state one block at a time rather than widening attention, and that the store needs terabytes of disk.

arXiv:2610.10845 PDF

Amortized Off-Policy Evaluation for LLMs

Younwoo Choi, Leo Feng, Vincent Liu, Haanvid Lee Teams evaluate candidate LLMs offline using data logged from models already in deployment, which is off-policy evaluation (OPE). This setting has two kinds of shift: model responses change during post-training, and reward definitions change with business requirements. Classical OPE methods must be refit for every new dataset or reward. PFN-OPE is a prior-data fitted network pretrained once on contextual-bandit tasks built from LLM responses scored by several reward functions, and it outputs a value estimate in a single forward pass with no per-task fitting. On HelpSteer2 and UltraFeedback with Qwen, Llama, and Gemma policies, it achieves 2.0-9.3x lower error than the best baselines in the reward-shifted settings.

arXiv:2610.10848 PDF

KDFP: A first-principles approach to knowledge distillation in large language models

Ryan Swift, Konstantinos Psounis Recent LLM distillation work mostly transfers post-training skills such as instruction following and chain-of-thought reasoning. General knowledge distillation, which matters for efficient on-device models, has received less attention. The authors re-examine lessons from earlier distillation work from first principles and propose KDFP, a white-box general knowledge distillation method for modern LLMs. KDFP outperforms existing methods by 1.6-4.9% across 9 benchmarks and improves training efficiency by up to 99.1% through ephemeral parameter reduction.

arXiv:2610.10854 PDF

Sparse Attention Is Matrix Approximation, Not Choosing from a Bag of Values

Fang Wan, Xufeng Liu, Fan Li, Yi Liu Sparse attention methods cut the quadratic cost of attention by keeping only the largest entries or highest-mass regions of the attention matrix. The authors argue this is the wrong framing because the matrix is multiplied by value vectors, so sparsification should be treated as matrix-product approximation. Matrix Approximation Sparse Attention (MASA) replaces attention-mass ranking with a closed-form score that measures how much each sparse unit reduces approximation error. It plugs into existing sparse attention frameworks without changing their kernels or budgets and delivers consistent accuracy gains across multiple sparse attention methods, benchmarks, and LLM backbones.

arXiv:2610.10871 PDF

Large Language Models for Machine Translation Quality Annotation: Humans and Models Are Both Challenged

Hala Almaghout, Christian Federmann, Qin Gao The study tests how well LLMs can replace human annotators for machine translation (MT) evaluation under two schemes, Multidimensional Quality Metrics (MQM) and Error Span Annotation (ESA). Experiments cover a long-context test set of 70 language pairs plus WMT23 and WMT25 data. For some tasks, agreement between LLMs and humans exceeds agreement between human annotators, but both vary widely across schemes, language pairs and domains, and remain unreliable in most settings. Humans struggle most with fine-grained MQM and low-resource pairs, while LLMs struggle with minor errors, wrong language variants and marking error spans.

arXiv:2610.10918 PDF

Spectrally Targeted Muon

Vishrut Goyal, Rohan Ramkumar The Muon optimizer sets every singular value of each update matrix to one, and it is unclear whether its success comes from boosting small singular directions or from suppressing large ones. Spectrally Targeted Muon orthogonalizes only the singular values above or below a threshold, which interpolates between normalized SGD and Muon, and it does so with Newton-Schulz iterations instead of an SVD. On the CIFAR-10 and NanoGPT speedruns, orthogonalizing everything except the few largest singular values nearly matches Muon, while orthogonalizing only the top directions falls well short, so Muon wins mainly by amplifying directions too small to train on. Shrinking the largest singular values keeps the weight spectrum flat but does not drive the loss, and AdamW mostly differs from Muon in how slowly it builds structure, which explains why warmup helps AdamW but hurts Muon.

arXiv:2610.10965 PDF

Optimizing Large Language Models with Chained LMOs

Sungyoon Kim, Kaan Ozkara, Youngsuk Park Optimizers inspired by Muon that compose several matrix normalizations have multiplied without a shared theory. The authors model them as chained linear minimization oracles (LMOs). They show that many such chains fall outside the standard LMO framework and can diverge on smooth convex objectives, yet chaining can still beat Muon on linear associative memory with anisotropic embeddings. Their new optimizer, TensorChain, stacks compatible weight matrices from different layers into a 3D tensor and normalizes it along each axis, and in Qwen3 0.6B and 1.7B pretraining it achieves an average 9.6% token saving over Muon at matched validation loss.

arXiv:2610.10975 PDF

SFT-as-Context Mitigates Forgetting in Supervised Fine-Tuning

Kenan Tang, Andong Hua, Chengxuan Qian, Saket Tiwari, Yao Qin Supervised fine-tuning (SFT) gives LLMs specialized skills but tends to erode their general capabilities. SFT-as-context is a training-free alternative in which the original parent model answers a query using the SFT model's response as in-context input. Across 19 parent-SFT pairs and 11 benchmarks, it stays within about 2 points of the SFT model on specialized tasks such as AIME 2024 and LiveCodeBench and within 2.2 points of the parent on general tasks. It also solves some mixed queries that neither model solves alone, and responses from a small open-source SFT model can improve a strong closed-source LLM. A Bayesian analysis bounds the method's error, and attention visualizations suggest the parent model attends more to useful SFT responses.

arXiv:2610.11132 PDF

QUILT: Rethinking Sparse-Attention Prefill through Shared Query Execution

Zhenduo Zhao, Qihui Zhou, Mingcong Song, Zhiyi Chen, Chuangguan Ye, Fengfan Hou et al. Sparse-attention kernels usually handle each query independently, so key-value (KV) cache entries shared by neighboring queries are loaded and dequantized again and again. QUILT processes neighboring queries together and reuses the KV entries they share. It relies on Shift-and-Compare Set Decomposition (SCSD), which turns irregular set operations into data-parallel primitives suited to accelerators, plus cascaded multi-level sharing and tile-aware execution. On LongBench with GLM-5.3 and DeepSeek-3.2, it cuts kernel latency by up to 55.1% and time-to-first-token by up to 36.8% compared with the state-of-the-art sparse kernel, with negligible accuracy loss.

arXiv:2610.11134 PDF

RAG-Stress: Probing the Limits of Evidence Reliance in Retrieval-Augmented Generation

Shunyuan Zhou, Hao Chen, Tianyu Wang, Goose Lin, Zaiyuan Wang, Haiying Zhao Misleading retrieved evidence can make a model abandon an answer it would otherwise get right, and standard accuracy hides this by mixing such replacements with errors the model already made. RAG-Stress holds each question fixed, edits one assertion in the evidence to support a wrong answer, and varies both the source-priority instruction and where the answer span appears in the text. It measures the misleading rate only on questions each model already answers correctly without retrieval. Across fifteen systems, including API models, open models, and RL-trained search agents, on TriviaQA-RC, HotpotQA, SearchQA, and MedQA, instructions that prioritize documents raise the misleading rate by 10.9 to 13.5 percentage points compared with instructions that allow prior knowledge. On average, misleading evidence is most effective at the end of the text and least effective in the middle.

arXiv:2610.11183 PDF

Bridging KV-Cache Quantization and Linear Attention: From Theory to Pretrained Weight Migration

Kaicheng Xiao, Liran Dong, Haotian Li, Guoliang Xing Two common ways to cut Transformer memory and compute costs work differently. KV-cache quantization compresses each key-value entry but keeps all of them, while linear attention folds past entries into a fixed-size state that can suffer interference. The authors present RAM-Net as a bridge between the two: soft assignments over a discrete address space drive recurrent updates to continuous slot states. They prove that, under a restricted construction, this locally approximates full-attention similarity, and they use the connection to migrate pretrained Transformer weights through a soft-quantized intermediate. Across nine pretrained models from 0.3B to 7B parameters, RAM-Net recovers an average of 87.1% of the teachers' accuracy gains over random guessing on six commonsense and knowledge tasks, using only 500M tokens per model.

arXiv:2610.11214 PDF

When Lower Reconstruction Loss Hurts: Distributionally Robust Refinement for Low-Bit LLM Quantization

Yanlong Zhao, Xiaoyuan Cheng, Huihang Liu, Baihua He, Xinyu Zhang, Harrison Bo Hua Zhu et al. Weight-only post-training quantization (PTQ) usually picks low-bit weights by minimizing reconstruction loss on calibration data. The authors show that lower reconstruction loss can actually degrade performance, even on the calibration data itself, partly because such weights become fragile when the input activation distribution shifts. They propose DRQ (Distributionally Robust Quantization), a post-hoc step that refines the integer weight codes to minimize worst-case reconstruction loss over a constrained set of activation distributions, leaving the quantization grid and inference kernels unchanged. DRQ improves models quantized by six PTQ methods, including AWQ, GPTQ, and ParoQuant, on both dense and mixture-of-experts LLMs, with no added inference overhead.

arXiv:2610.11226 PDF

Read What Matters: Query-Adaptive Quantization for KV Caches

Siddharth Bhandari, Lucas Gretta, Krishna Balasubramanian, Shiva Kasiviswanathan KV-cache entries are written before the queries that will read them are known, yet each query needs precision in different places. ReadKV stores keys and values as progressive codes and, for each query, decides how many bits to read: first per key channel based on the query, then per value token based on the resulting attention. The authors prove that query-dependent reading can strictly beat any query-independent reader at the same read budget. Across six base models, reading four bits on average from an eight-bit cache raises C4 perplexity by at most 0.66%, and it beats storing and fully reading four bits at the same read budget. On an 8K-token, single-layer workload on an NVIDIA A10G, a restricted ReadKV reader has 39% lower latency than TurboQuant.

arXiv:2610.11245 PDF

Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals

Lei Zhao, Qichao Zhao, Bowen Zuo, Qishi Zhan On-policy distillation (OPD) transfers capabilities between language models but sometimes stalls, and the reasons are not well understood. In code generation and math reasoning, OPD from larger teachers plateaus early, with an average final loss reduction of 25.1% after 200 updates versus 96.2% for self-RL teachers (teachers obtained by further reinforcement learning on the initial student). Training-log diagnostics link the plateaus to an early drop in a gradient-based learning-signal proxy, and a local recovery guarantee offers a conditional explanation for why teachers close to the student succeed. Very small parameter changes and near-identical representations before and after OPD suggest that limited representation adaptation may contribute to the collapse, a hypothesis the authors say still needs testing.

arXiv:2610.11247 PDF

When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better

Siyan Zhao, Yonggan Fu, Jindong Jiang, Shih-Yang Liu, Song Bian, Byung-Kwan Lee et al. On-policy distillation (OPD), where a teacher supervises a student on the student's own live samples, is increasingly popular, but it is unclear whether on-policy sampling helps for every teacher-student pair. A simpler alternative, Semi-OPD, distills on offline rollouts generated once by the initial student, and outperforms OPD in 14 of 17 teacher-student pairs ranging from 1.5B to 235B parameters, with up to 13.6% higher accuracy and an 11.4x training speedup. Whether OPD helps depends on how well teacher and student already agree, measured by an output-token overlap ratio. For poorly aligned pairs, student rollouts drift away from what the teacher would produce as context grows, which weakens the distillation signal.

arXiv:2610.11291 PDF

Characterizing Overconfident Failure in LLM-Based Code Generation

Ravishka Rathnasuriya, Wei Yang cross-listed Code generated by LLMs can look plausible but fail execution, and model uncertainty could serve as an early warning if it tracked correctness. This empirical study of four open-source code models on three execution-based benchmarks finds that incorrect programs are often generated with confidence comparable to correct ones, at both the program and token level. Uncertainty-based selection does not consistently raise the accuracy of accepted outputs. Instruction tuning can increase certainty on failing generations, and common mitigations fix only parts of the problem. An exploratory analysis suggests that hidden representations may carry correctness signals that output confidence does not expose.

arXiv:2610.11300 PDF

DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition

Yuxuan Lou, Kai Yang, Geng Zhang, Yong Liu, Yang You Upcycling a pre-trained dense model into a fine-grained Mixture-of-Experts (MoE) model is cheaper than training one from scratch, but when all experts come from a single source model, routing collapses and accuracy falls to near-random. DivMoE instead initializes experts from copies of the dense model that were continually pre-trained on different domains, and adds a hard routing constraint so that each token activates experts from distinct domain groups. Across two base models and 15 benchmarks it beats six upcycling baselines, reaching 55.6% versus 51.6% for the strongest baseline on Qwen3-1.7B, and improves over the dense base model on every benchmark. After supervised fine-tuning, a 12B-parameter DivMoE model matches the 16B Moonlight-MoE at 64.5% average accuracy.

arXiv:2610.11317 PDF

ReCal: Calibrating Structured Pruning for On-Policy Distillation Recovery

Houcheng Jiang, Mao Zheng, Mingyang Song, Qiyong Zhong, Jie Sun, Tianyu Zhang et al. Structured pruning makes reasoning language models cheaper to deploy, but the damage it causes can undermine on-policy distillation (OPD), which relies on trajectories generated by the pruned student model. ReCal (Recovery-Aware Calibration) is a plug-and-play step applied before pruning. It measures the forward KL divergence between the unpruned teacher and a pruned probe to find teacher-supported predictions that pruning disrupts, then reweights the calibration statistics so existing pruning criteria preserve those predictions. Across several models and pruning methods, ReCal consistently improves mathematical reasoning after OPD, with gains of up to 16.7 percentage points on AIME, and it also improves most code-generation comparisons.

arXiv:2610.11332 PDF

Writing for the Reviewer: Defensive Writing in GPT Models

Junchi Liao Recent GPT versions often narrow or retract authors' claims when revising papers, which the authors call defensive writing when the source material does not justify the change. They have GPT versions and other developers' models rewrite paragraphs from pre-ChatGPT papers, or write from evidence sheets listing a paper's methods and results. Defensive writing grows with each GPT version, and the results favor the explanation that models write for an anticipated reviewer rather than correcting overclaiming: fewer than one in ten of the claims GPT-6-astra retracts are actually overstated. Mentioning review or adding a self-review round increases the effect. AI reviewers score defensive rewrites higher, while human readers find them harder to read and the authors less certain.

arXiv:2610.11355 PDF

RaReCache: Bridging the Gap in Cross-Model KV Cache Reuse via Rank disagreement-based Selective Recomputation

Sreetama Sarkar, Saptarshi Mitra, Sitao Huang, Souvik Kundu, Peter A. Beerel When coding agents or multi-model systems pass a shared context between models, the receiving model usually has to recompute its key-value (KV) cache from scratch. Linear maps that translate KV caches between models lose accuracy as the size gap grows. The authors show these errors are concentrated in a small set of information-dense tokens. RaReCache lets a small model prefill the cache for a large one, then recomputes only the critical tokens, which it finds with a rank-disagreement score. From Qwen3-0.6B to Qwen3-14B, recomputing just 30% of positions keeps 95–99% of target accuracy, with up to a 3.04x prefill speedup. In online serving at saturation load, it cuts median and 99th-percentile time-to-first-token by 5.0x and 6.4x.

arXiv:2610.11358 PDF

From a Prompt to Repertoires: Evolving Functional REpertoires Enable LLM Continual Learning

Fengyuan Liu, Yue Wang, Hangxi Guo, Fengyuan Liu, Chenxu Wu, Yanguang Liu et al. Prompt optimization can match reinforcement learning methods such as GRPO on individual tasks without updating model parameters. When applied to a sequence of tasks, however, it suffers catastrophic forgetting, because optimized prompts accumulate rules that overfit each task. Evolving Functional REpertoires (EFRE) replaces the single prompt with a repertoire of functions: compatible updates refine existing functions, while conflicting updates create new ones. On a three-task continual-learning stream, EFRE ends 7.50 points above GRPO, and after adapting to a biology task its FinQA score drops by only 1.56 points, versus 25.10 for the base prompt-optimization method. The gains carry over when EFRE is used inside a minimal agent system with different backbone models.

arXiv:2610.11373 PDF

TypedBench: A Benchmark for Calibration, Framing Sensitivity, and Cost in System One Decision Models

Rahul Sharma, Andrew B. Ducan, Ga\'etan Marceau Caron, Sebastian J. Vollmer "System One" decision models output probabilities over typed answers (categorical, ordinal, or binary) that software then acts on through thresholds and cost rules, so miscalibrated or wording-sensitive probabilities can trigger unintended actions without leaving any rationale to inspect. TypedBench evaluates such models using seven policy-labelled generators and nine evaluation suites. It reports accuracy across paraphrases, calibration error relative to a finite-sample noise floor, selective prediction, and realised cost under asymmetric cost matrices. Comparing a hosted model, an open encoder, and open decoders from 0.8B to 9B parameters, the authors find that the hosted model is wording-sensitive and underconfident, and that under asymmetric costs acting on its probabilities can be worse than simply taking its top answer.

arXiv:2610.11392 PDF

Estimating great expectations under autoregressive language models with potentials

Francesco I. Re, Shubhangi Ghosh, Tim Vieira, Ryan Cotterell Many uses of language models depend on the expected value of some test function under the model rather than on individual samples, and estimating that expectation reliably can be expensive. The authors reduce estimator variance by exploiting the next-token conditional probabilities that sampling already produces. They do this through "potentials", real-valued functions on prefixes that decompose the test function additively. They derive conditions under which a potential reduces variance and design practical potentials for several estimands, demonstrating substantial variance reductions at comparable compute cost.

arXiv:2610.11399 PDF

Design Creativity Bench: Measuring creativity in LLM-Generated UI

Aman Rusia, Abhijit Bhole, Prashank Gupta, Dipanjan Dey cross-listed Design Creativity Bench measures how varied and appropriate LLM-generated user interface (UI) designs are. It scores originality (how different models' designs are for the same prompt), creative range (how much one model's designs change across product domains for the same UI goal) and appropriateness (the share of acceptance criteria a design meets). Every model scored above 90% on appropriateness, with the best reaching 99.2% against 98.0% for human designs. Creative range, however, was 0.581 across models versus 0.902 for human designers, and designs from different models were much more alike than human and model designs, so default LLM output is appropriate but repetitive.

arXiv:2610.11539 PDF

Learning the Loop, Not Just the Page: Execution-Grounded Loop Learning for Web Generation

Yuxin Meng, Ruixu Zhang, Junjie Wang, Yuhan Suo, Yuhan Sun, Ruining Hu et al. Methods for generating functional web pages usually optimize only the final page, leaving the diagnose-and-repair process underexplored. WebLoop trains generation, critique, and refinement jointly in one shared policy using group-relative policy learning. The Critic, which has no directly executable output, is trained on signals for both requirement-level discrimination and how much its feedback helps downstream. With Qwen3.5-9B, it improves the base model by 11.3 points on WebRise and 15.4 points on WebGen-Bench. The gains carry over to first-pass generation, persist at 27B scale, and transfer to multimodal inputs, and controlled analyses show they are not explained by an extra refinement pass alone.

arXiv:2610.11543 PDF

Conditional Transfer from Controlled Pretraining Mixtures to Code

Ohad Rubin Synthetic tasks are often justified as pretraining data because their loss falls quickly. This study separates three signals: whether a task's loss tracks overall training progress, whether it responds to its own token budget, and whether including it improves downstream performance. In controlled code pretraining that mixes general Python with OpenCodeInstruct, curated synthetic tasks, and literature-derived probes, these signals rank data sources differently. HumanEval pass@20 peaks at 22.6 with a 75% OpenCodeInstruct mixture, above either pure source. The loss-based adaptive scheduler Ado pushes OpenCodeInstruct below 5% of the mix and underperforms fixed mixtures by 2.4–11.0 points, showing that chasing loss reduction can move training away from data that transfers.

arXiv:2610.11548 PDF

Smoothing the Top-k Exposure Boundary for Sparse Mixture-of-Experts

Yunkai Chai, Tong Zhu, Xiaoye Qu, Xuyang Hu, Guanjie Chen, Qipeng Guo et al. Sparse Mixture-of-Experts training with a fixed top-k creates a sharp boundary: experts with nearly equal scores either receive full training signal or none. Elastic Expert Routing samples the number of active experts from a small symmetric distribution centered on k, which softens the cutoff while keeping expected training compute and inference cost unchanged. In supervised fine-tuning it improves downstream averages by +0.84 on OLMoE-1B-7B and +2.02 on Qwen3-30B-A3B. In pretraining from scratch it beats the static top-k baseline by 1.6 points.

arXiv:2610.11575 PDF

Adapting English Quality Classifiers for Multilingual LLM Pretraining Data Selection

Vinko Sabol\v{c}ec, Bettina Messmer, Yassine Turki, Martin Jaggi Model-based quality filtering works well for selecting pretraining data in high-resource languages, but low-resource languages lack the annotated data to train such filters. The authors convert an existing English quality classifier into a multilingual one by training a small multi-layer perceptron on Transformer encoder embeddings of multilingual text. The training labels come from running the English classifier on machine-translations of that text. In pretraining experiments at 1B, 3B and 8B scale covering more than 100 languages, the approach matches existing multilingual model-based filtering baselines on downstream benchmarks without hurting regional and cultural knowledge, and it appears to reproduce the English scoring criteria even for languages left out of its training data.

arXiv:2610.11585 PDF

Where to Adapt Matters: Layer-Selective Fine-Tuning for Capability Retention

Zhiqiang Pang, Zihong Sun, Qi Xie, Jun Shu, Deyu Meng, Zongben Xu Parameter-efficient fine-tuning (PEFT) of LLMs often erodes general capabilities, and the usual remedies rely on data replay or regularization. The authors instead ask which layers should be adapted, using layer-wise empirical Fisher information to measure how sensitive each layer is to the target task. Because computing Fisher information requires backward passes, they propose a cheap forward-only proxy: layers whose outputs have lower cosine similarity to their inputs consistently have higher Fisher scores. Layer-Selective LoRA (LS-LoRA) places adapters only in those layers, and on math reasoning and code generation it improves average target-task performance while retaining much more commonsense reasoning ability than standard all-layer LoRA.

arXiv:2610.11620 PDF

DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation

Anhao Zhao, Haoran Xin, Junlong Tong, Yingqi Fan, Xuan Lu, Ping Nie et al. On-policy distillation (OPD) trains a student model on its own generated trajectories using token-level signals from a teacher, and the authors find that training on fewer tokens can beat training on all of them. Their method, DIAL-OPD, ranks tokens by teacher-student disagreement weighted by the logarithmic mean of the two models' probabilities, so tokens that both models consider very unlikely are not over-rewarded, and keeps only the top-scoring ones. Keeping just 40% of tokens, it beats full-token OPD across four teacher-student pairs and seven math reasoning benchmarks, with mean accuracy gains of up to 5.25 points, and doubles AIME25 Pass@16 from 13.33% to 26.67%. With a 4B teacher it also beats the strongest full-token baseline that uses an 8B teacher.

arXiv:2610.11659 PDF

Internalizer: Portable Context-to-Parameter Mapping for Very Large Language Models

Peter Devine, Nick Ryan, Benjamin Sirb, Alex Chiocchi Hypernetworks can turn a document directly into a LoRA (low-rank adaptation) adapter so that a model carries the document in its weights, but earlier work only reached base models of up to 14B parameters. The Internalizer keeps most of its parameters in a model-agnostic trunk with thin per-model entry and exit layers. That lets it train cheaply against small models and then port to the frozen 284B-parameter DeepSeek v4 Flash, about two orders of magnitude larger than previous targets. On unseen documents of up to 4096 tokens, the generated adapters reach 84.9% top-1 and 97.8% top-5 teacher-forced accuracy versus 63.4% and 83.5% for the base model, with only a three-word instruction in the context window. A single forward pass produces each adapter.

arXiv:2610.11715 PDF

Where Draft Trees Lose Target Mass: Exit-Guided Speculative Decoding

Shijing Hu, Xuancheng Ren, Zhihui Lu, Pan Zhou Tree-based speculative decoding checks several draft continuations in one pass of the target model, but trees built from draft-model scores miss probability mass the target model assigns elsewhere. Using a target-flow view, the authors prove that one plus target coverage sharply bounds the expected output length of any exact path verifier. They derive Tree Exit Verification (TEV), an exact level-parallel verifier, and Exit-Guided Draft-Tree Training (ExitTrain), which uses the missing target probability as node-level feedback to improve draft trees. Across dialogue, code, and math reasoning, ExitTrain lengthens accepted output blocks by 13% and TEV cuts verification latency by 15%, giving a 14% end-to-end speedup over DDTree.

arXiv:2610.11750 PDF

Thinking Inertia: LLMs Keep Thinking When Told Not To

Dianqiao Lei, Kevin Qinghong Lin, Pan Lu, Philip Torr, James Zou The study asks whether LLMs actually stop reasoning when told not to. Proxies such as a disabled thinking mode turn out to be unreliable, so the authors measure no-thinking with three metrics: Empty-Thinking Rate, Question-Pre-answer Relevance, and an LLM-as-judge Explicit Inference Rate. They test six prompting interventions on six LLMs across Boolean, multiple-choice, and open-ended questions. They find "Thinking Inertia": explicit no-think controls cannot reliably eliminate visible inference, and the effect grows as the answer space opens up. Accuracy stays stable on closed-form tasks, but open-ended tasks trade answer-only compliance against accuracy, and supplying candidate answers makes answer-only responses easier to produce.

arXiv:2610.11765 PDF

RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing

Ilya Lasy, Nora Yinuo Cai, Kola Ayonrinde Interpretability work on sparse Mixture of Experts (MoE) models has generally assumed that each expert specializes in one coherent domain, and that assumption has produced little insight. The authors propose the Superposed Specialisation Hypothesis (SSH): each expert specializes in a disjoint union of fine-grained features. Building on SSH, RouterInterp finds the Sparse Autoencoder features most predictive of routing decisions and turns them into unified natural-language explanations. On gpt-oss-20b, these explanations achieve about 65% higher detection accuracy than prior methods based on token statistics.

arXiv:2610.11775 PDF

Easy to anticipate, hard to compute: boundary dependence finds the computed outputs that entropy patching misses

Nicol\'as Vera Z\'u\~niga Byte-level models such as the Byte Latent Transformer (BLT) start a new patch where a small model's next-byte entropy is high, which concentrates expensive global compute where prediction is hard. The authors show this rule misses positions whose type is predictable but whose value must be computed, such as the number after "=" in a math solution. In BLT-1B with patch starts on 10% of bytes, the entropy rule gets 19.0% of computed results in GSM8K solutions exactly right, compared with 67.1% when entropy is combined with "boundary dependence", a label-free signal measuring how much the model's loss rises when a patch start is removed. The gap persists after low-rank fine-tuning, grows with model size in byte models trained from scratch, and applies specifically to computed values rather than copies or lookups.

arXiv:2610.11790 PDF

Evaluating Exact Output and Checkpoint-State Prediction in Real Programs

Xiaohong Chen, David Bucur, Chenglong Ma, Yi Zhang, Lingming Zhang, Sriram Vishwanath et al. cross-listed Introduces a benchmark that asks models to predict a program's final output and its intermediate state at checkpoints from the source code and input alone, without running anything. It extends CRUXEval-style output prediction with paired shorter- and longer-trace inputs and checkpoints inside and after loops, covering 400 cases from 371 Python and C++ programs and seven settings from four model families. Turning on reasoning improves accuracy by 33 to 55 percentage points, and the best setting scores 93.0% on shorter-trace output but 77.0% on longer-trace output and about 64% on the state-prediction tasks. The authors caution that the longer inputs change several factors at once, so the gap cannot be attributed to trace length or state tracking alone.

arXiv:2610.11889 PDF

Project Greenhouse: Progress Toward Fully Open and Sovereign Agentic Search

Jimmy Lin, Sahel Sharifymoghaddam, Lingwei Gu, Nour Jedidi cross-listed The authors argue that fully open, sovereign models for agentic search can be built with modest compute and without relying on third-party open-weight backbones. As a first milestone, they train a pointwise decoder-only reranker from scratch using only commonly available datasets: pre-training followed by supervised fine-tuning. The resulting Gaggle model family is competitive, and most experiments ran on no more than a handful of GPUs. Data, code, configurations, and checkpoints are released so the full training process can be reproduced independently.

arXiv:2610.11922 PDF

MetaOPD: Meta-Learned Token Weighting for On-Policy Distillation

Zipeng Wang, Xinpeng Dong, Yuefan Wang, Pingchen Lu, Xian Wei, Kun Kuang et al. On-policy distillation (OPD) trains a student model on its own outputs using token-level supervision from a teacher. Existing methods either weight all tokens equally or use fixed rules to set token weights. MetaOPD jointly learns the student and a small token-weighting network through bilevel optimization. The weighting network is updated according to validation loss on reference solutions after a simulated student update, so weights reflect how much they actually help the student. On six math reasoning and three out-of-domain datasets, it improves over plain OPD by about 2 points Avg@8 and 6 points Pass@8 for both 0.6B and 1.7B students.

arXiv:2610.11989 PDF

Examining Social Attribution in LLM Reasoning: A Theory-Guided Probing Methodology

Zhaoxin Yu, Qingchao Kong, Dajun Zeng, Wenji Mao Social attribution means judging who caused an event and who deserves responsibility, blame, or credit for it, and it has been studied far less in LLMs than in social psychology. Guided by Attribution Theory, the authors build a benchmark of 7,639 responsibility and blame questions drawn from classic vignettes and real-world narratives. They evaluate 32 LLMs and 5 non-LLM baselines on it, and they probe hidden states for five key attribution dimensions. LLMs show measurable but incomplete agreement with human judgments, and agreement grows with model size. Some attribution dimensions can be decoded from specific positions in the hidden states, and their influence on the final judgment matches what Attribution Theory predicts for humans.

arXiv:2610.12022 PDF

Structure Tax: How Structured Output affects LLMs Performance

Vineet Kumar, Kanishka, Bhuvanesh Mandora Forcing LLM outputs into formats like JSON or XML is widely thought to impose an unavoidable accuracy "structure tax." The authors test this claim across many models, datasets, and schemas, measuring accuracy, confidence calibration, and hidden-state geometry. The tax turns out to depend on schema design rather than structure itself: putting reasoning fields before the answer matches or beats free-form accuracy, while putting the answer first causes steep drops, especially in smaller models. Schemas that keep reasoning first also improve calibration, and centered kernel alignment (CKA) analysis shows clearer separation between correct and incorrect representations in middle layers.

arXiv:2610.12056 PDF

All Verdicts are Not Equal: Rethinking LLM Judge Reliability

Vineet Kumar, Darshita Rathore, Anindya Moitra LLM-as-a-Judge evaluation is often treated as deterministic ground truth. The authors stress-test six frontier judge models across four benchmarks, five prompt formats, two presentation orders, three temperatures, and ten repetitions per condition. Verdicts change between identical runs even at temperature zero, swapping the answer order flips most verdicts on hard tasks, and the most consistent judge achieves its consistency by repeating wrong verdicts, matching ground truth only 51% of the time. They propose the trustworthy verdict rate, a metric for whether a verdict is reproducible, order-invariant, and accurate at once, and use it to show that reliability varies by item rather than by model and that holistic rubric scoring beats pairwise win-rate judging.

arXiv:2610.12083 PDF

Universal Textual Teaching for LLMs

Zhanyi Lu, Huan Wang Knowledge distillation usually updates the student model's weights, which ties the transferred knowledge to one checkpoint and rules out API-only models. Universal Textual Teaching (UTT) instead distills gaps between teacher and student into a reusable natural-language document called a Primer. Four roles build it iteratively: the student attempts a task, a Prompter converts evaluation feedback into a teaching instruction, the teacher gives a targeted demonstration, and a Synthesizer consolidates validated lessons. Student accuracy on KernelBench rose from 9.4% to 48.6% and on Omni-MATH-2 from 27.6% to 51.7%, beating prompt engineering and weight-based distillation, and Primers carried over to students that were not used to build them.

arXiv:2610.12114 PDF

Rehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read

Zhiyun Shi Key-value (KV) caches for documents, shared prompt prefixes, or long conversations are often compressed before the model knows what it will be asked. The usual approach keeps the entries the model attends to while rereading the whole context. The authors show this backfires under tight budgets: at a 3% keep ratio, full rereading retains only 31.5 of 96.5 points on RULER, because the budget is spread across the whole context and the entries needed for the answer survive at little better than chance. Their training-free method Attic-KV instead has the model quiz itself with question-answer pairs that quote the context, plus a content-adaptive number of anchor tokens. It is the best training-free method in all eight tested settings on RULER and LongBench, it improves KVgrad and RestoreKV+ by up to 17.1 and 28.1 points, and it reaches a 41.9-point gain over full rereading at a 3% keep ratio while also compressing faster.

arXiv:2610.12133 PDF

Language-Specific Effects of Tokenizer Choice in Multilingual Language Models

Clara Meister, G\"ul Sena Alt{\i}nta\c{s}, Antoine Bosselut Tokenizer vocabularies have finite capacity, so improving the representation of some languages can hurt others. To test whether tokenizer choice matters equally across languages, the authors train 123 language models with 54 tokenizers, holding architecture, data, token budget, and optimization fixed. Tokenizer choice matters more for languages with less language-model training data: the spread of a language's bits-per-byte (BPB) across tokenizers grows as its data share shrinks (Spearman rho = -0.52), and leaving a language out of tokenizer training always raises its BPB. Giving lower-resource languages more tokenizer-training data does not reliably help, and the tokenizer properties linked to good BPB vary by language, but metrics for those properties can predict pairwise BPB rankings and so can screen tokenizer candidates before full training.

arXiv:2610.12144 PDF

TokenRouter: Efficient Serving System for Token-Level LLM Routing

Tianyu Fu, Tengxuan Liu, Ruoxi Wang, Yixin Dong, Yi Ge, Yichen You et al. Routing each token, rather than each query, between different LLMs can improve the cost-quality tradeoff, but existing serving systems assume a single model and suffer step desynchronization and batch admission delays under token-level routing. TokenRouter lets developers write routing logic from the point of view of a single request, while the runtime runs a subserver per model and dispatches requests asynchronously. Each subserver uses a delayed-batching scheduler whose hyperparameters are derived from a mathematical throughput model. Across routing algorithms, workloads, and model pairs, it achieves 2.01–64.15x higher decoding throughput than existing systems.

arXiv:2610.12242 PDF

NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor

Long Wang cross-listed Dense retrieval ranks text chunks by semantic similarity and ignores structure that data systems already store, such as section membership, session boundaries, and ordering. NativeScope handles queries that have a known anchor A and relation r: it first uses belonging, before, or after operators to select the relevant native units, then ranks only chunks inside that scope by the target term B. On 200 controlled records derived from QASPER and LongMemEval with a 1,024-token budget, it reaches 89.28% native-unit recall on documents and 72.50% on memories, 42.75 and 22.00 points above instance-wide dense RAG. Gains come mainly from the relational scoping, and memory recall falls to 35.50% when anchors are localized automatically, so the method depends on reliable anchors.

arXiv:2610.12243 PDF

Looking Inside LLMs: Small-World Connectivity as a Signature of Reasoning Performance

Zheng Huang, Sansheng Cao, Enpei Zhang, Weikang Qiu, Elynn Chen, Xiang Zhang et al. Borrowing from neuroscience findings that link intelligence to small-world brain networks, the authors build functional graphs from similarities between attention-head activations in LLMs. A higher small-world index (SWI), which combines strong local clustering with short global paths, consistently correlates with better fluid reasoning performance across models and training checkpoints. Heads that matter for performance tend to keep most of their connection weight within their own community and spread little across communities. A pruning method built on these signals, Small-World Allocation (SWA), preserves performance better than competing sparsity allocations across six LLMs, cutting WikiText perplexity by up to 20%.

arXiv:2610.12304 PDF

SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference

Qitong Wang, Xinwei Niu, Mingluo Su, Shanwei Zhao, Shiai Zhu, Huan Wang Hessian-guided pruning methods for LLMs calibrate on natural text, but during decoding the model reads its own generated tokens, so the pruning objective does not match the activations seen at inference time. Existing speedups also mostly target sparse matrix-matrix products rather than the sparse matrix-vector operations that dominate decoding. SparseDecoding calibrates on activations collected during the dense model's own autoregressive generation, excluding prefill, and adds an optimized N:M sparse matrix-vector kernel. On Llama-3.1-8B, Llama-3.3-70B, and Qwen3-14B/32B, it beats standard fixed-text calibration on long-form generation and delivers up to 1.48x end-to-end decoding speedup on A100 GPUs.

arXiv:2610.12327 PDF

VFold: Symmetry-Aware Cross-Layer Value Cache Compression

Neha Verma, Sungwon Kim, Kenton Murray, Kevin Duh The key-value (KV) cache that speeds up LLM decoding can dominate memory at long context lengths, and existing methods that exploit similarity between layers usually require architectural changes and add overhead. VFold merges value caches across layers in a symmetry-aware way, saving memory without degrading performance or adding decoding overhead. It composes with high-ratio quantization and key-cache pruning to reach compression ratios that neither method reaches alone, at little extra cost. The authors conclude that the value cache contains a large amount of unused capacity.

arXiv:2610.12338 PDF

Which Skill to Distill? SGUID: Selecting a Compact Skill Bank for Model-Skill Co-Evolution

Yuhan Liu, Xiyao Ma, Zhongkai Sun, Xu Han, Chengyuan Ma, Benjamin Z. Yao et al. Skills are reusable procedural guidance given to LLMs at inference time, and they can also serve as teachers for distillation. The authors find, however, that fewer than 25% of retrieved skills provide useful distillation signals. SGUID keeps only skills that consistently produce effective learning signals during training. Distilling just 6 selected skills matches or beats distilling the full bank, which is up to 11x larger, on most Olmo and Qwen models. In repeated rounds of model-skill co-evolution, the selection step keeps training stable, raising Qwen3-8B from 64.3% to 66.3%, whereas adding unfiltered skills degrades performance.

arXiv:2610.12367 PDF

Latent Core Tokenizer: Compress, but Meaningfully

Felermino D. M. A. Ali, Millicent Ochieng, Ogbemi Ekwejunor-Etchie, Ade Famoti, Jacki O'Neill, Debjit Paul Tokenizers optimized for compression can spread vocabulary capacity unevenly across languages. The Latent Core Tokenizer (LCT) first identifies reusable linguistic units using Minimum Description Length, entropy-based boundary signals, and morphotactic constraints, and only then builds a shared vocabulary. Across 104 languages with a 200K-token vocabulary, it achieves lower fertility and higher MorphScore than BPE, Unigram, and parity-aware BPE, while keeping cross-lingual cost disparity comparable. On four multilingual benchmarks it improves aggregate scores by 1.48-2.00 points, which the authors take as evidence that compression alone does not predict representation quality.

arXiv:2610.12376 PDF

Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization

Hanyang Li, Shao Tang, Daniel Thomas Braithwaite, Gregory Dexter, Leonardo Neves, Aman Gupta et al. Storing AdamW's optimizer states in 4 bits saves memory, but quantization errors carry through the moment updates and distort later steps. The authors show that a small error in the stored second moment can still cause a large error in the preconditioner, and they propose two fixes. ZIP-SR keeps zero in the codebook and performs stochastic rounding in preconditioner space, while ZE-EDEN drops zero from the codebook and rescales quantized blocks to compensate. In GPT- and Llama-style pretraining from 130M to 2.7B parameters, both methods narrow the validation-loss gap between TorchAO's 4-bit AdamW and 32-bit AdamW at every model size, cutting the gap by up to 70%, and they also beat TorchAO in full-parameter fine-tuning.

arXiv:2610.12444 PDF

15 more specialized papers

Theory 65

Certified Corruption Budgets: Anytime-Valid Leaderboard Claims under Adaptive Rigging

Hamed Khosravi, Xiaoming Huo cross-listed Public AI leaderboards can be manipulated through vote rigging, selective disclosure of private variants, or benchmark contamination, and existing guarantees break against attackers who corrupt in bursts. The authors introduce a certified corruption budget, published with each pairwise ranking claim, that is anytime-valid: with high probability, at every point in time, either the claim is correct or more than the stated number of records were corrupted. This holds even against adaptive attackers who see every certificate. Separate certificates are given for forged votes and for votes flipped after being seen, and publishing the best of several private variants costs only a logarithmic amount. In replays on 1.8 million Chatbot Arena votes, a few hundred rigged votes make standard confidence intervals certify false orderings while the new certificate stays valid.

arXiv:2610.10597 PDF

Language Modeling is Monotone Compression

Noam Mazor, Andrew Morgan, Rafael Pass cross-listed Offers a theoretical account of the observed link between language modeling and compression. The main result is that next-token predictors are equivalent to monotone (order-preserving) compression algorithms, in the sense that each can be built from the other with error preserved up to an additive gap of 2. The monotonicity condition is needed for this equivalence if and only if cryptographic one-way functions exist. As a corollary, next-bit pseudoentropy turns out to be equivalent to monotone incompressibility; previously only one direction of this was known.

arXiv:2610.11031 PDF

Emergent Inverse-Depth Scaling From Nonlinearity In Attention

Zirui Peng, Yizhou Liu, Ziming Liu, Jeff Gore Existing theory ties the depth scaling of model loss to power-law data spectra, a picture derived from linear-attention models, which learn strong spectral directions before weak ones. The authors show that nonlinear attention produces inverse-depth decay of loss across all data spectra tested, because it can focus on relevant tokens and so learn strong and weak directions in parallel. They connect this to the central limit theorem: error shared across layers sets the loss plateau, while aggregating layer-specific differences keeps improving loss as depth grows. The result suggests depth scaling may stem from attention's nonlinearity rather than from the global covariance structure of the data.

arXiv:2610.11063 PDF

Do Flatter Minima Drive Better Generalization? An Algorithmic Separation in Grokking

Mohnish Harwani Grokking, where a model memorizes training data and only later generalizes, is used here to test whether flat loss minima actually cause generalization. Biasing training toward flatter solutions with sharpness-aware minimization (SAM) produces flatter solutions but does not reliably trigger generalization on its own. Combined with weight decay, however, SAM speeds the transition to generalization by up to 4x. A minimal two-layer ReLU model with both memorizing and generalizing solutions shows that flatness alone cannot tell the two apart. The same model reveals a regime where the memorizing solution is stable under gradient descent but unstable under SAM, which explains the speedup.

arXiv:2610.11206 PDF

The Lattice of Transition Laws

T. Y. Tsui, Jiatao Gu, Lingjie Liu Diffusion, autoregressive (AR), and hybrid generative models each fix their decoding schedule by design, and the authors ask whether a schedule's quality can be predicted before decoding. They describe all of these models as paths on a single corruption lattice and define a schedule's cost as the dependence its parallel steps discard. For data that are Markov on a graph, the fewest steps of a zero-cost schedule equal the graph's treedepth, which is logarithmic in sequence length and linear in grid side length. With fewer steps, schedules can be ranked using a pairwise-dependence kernel estimated from pretrained weights, and most of these predicted rankings hold across text, image, and video generation.

arXiv:2610.11216 PDF

From Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon

Jihwan Kim, Dogyoon Song, Chulhee Yun Optimizers can fit the same training data yet settle on classifiers with different geometry, and it has been unclear whether this provably changes test performance. The authors prove that in high-dimensional multiclass classification, row-wise normalization can reach strictly higher population accuracy than full-batch Adam and exact-SVD Muon. Under a Gaussian data model, row normalization's class-wise Euclidean geometry preserves the true decision-boundary directions, whereas Adam's per-coordinate geometry and Muon's spectral geometry distort them. The advantage holds under several anisotropic settings and can reverse against Adam when covariances align with the coordinate axes. Synthetic experiments and experiments on the last layer of a language model support the theory.

arXiv:2610.11309 PDF

$\sigma$Transfer: Uncertainty Transfer from Small to Large Networks under $\mu\mathrm{P}$

Richard Bergna (University of Cambridge, Spotify), Fernando Ruiz Mazo (University of Cambridge), Nicol\`o Felicioni (Spotify), Jos\'e Miguel Hern\'andez-Lobato (University of Cambridge), Kamil Ciosek (Spotify) cross-listed Laplace approximations give useful uncertainty estimates only if the prior precision is tuned, and tuning it requires a sweep over posteriors that is impractical for billion-parameter networks. Working under the Maximal Update Parametrization (μP), the authors derive a rescaling of the prior covariance that keeps the chosen precision stable as width grows, so σTransfer can tune it on a small model and reuse it unchanged on a large one. They prove convergence under explicit conditions and report precision-sweep speedups of about 5000x from width 128 to 4096 on MNIST at negligible loss in negative log-likelihood (NLL). Transferring from a public 1B model to a 7B model gives a median speedup of about 2.3x with almost no NLL increase, and the same stability carries over to out-of-distribution (OOD) detection and abstention decisions.

arXiv:2610.11668 PDF

Softmax Attention on Gaussian Mixtures: Linear When It Can, Selective When It Must

Simon Gabet (LMO), Etienne Boursier (LMO, CELESTE), Claire Boyer (LMO, IUF) cross-listed Prior theory analyzed softmax attention on Gaussian prompts, where the infinite-prompt limit collapses attention to a linear map and loses the query-dependent selection that distinguishes it from linear attention. The authors study the same limit on Gaussian mixtures, which stay tractable but add latent structure, multimodality, and nonlinear dependencies. They show that softmax attention can represent, and learn by gradient-based training, optimal solutions to statistical tasks such as supervised classification and denoising. It matches linear attention on linear tasks and uses query-dependent context selection to solve nonlinear tasks that linear attention cannot.

arXiv:2610.11798 PDF

Diffusion Removes Langevin's Conditioning Dependence: A Sharp Gaussian Analysis

Adam Perbost, Francis Bach, Pierre Marion cross-listed The work uses Gaussian target distributions to isolate why diffusion models avoid the bottlenecks of classical score-based samplers. With optimized hyperparameters, diffusion reaches a 2-Wasserstein sampling error of O(√(dλ_max) log N / N), while unadjusted and underdamped Langevin dynamics carry an extra √κ factor, where κ is the condition number of the target covariance. The bounds are shown to be sharp through matching first-order asymptotics, which pins down how time-dependent score trajectories remove the dependence on conditioning. In the learning phase, by contrast, estimating the clean score by gradient descent gives essentially the same estimator as estimating a noisy score, which suggests the benefit of noising does not come from learning.

arXiv:2610.12052 PDF

SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction

Ruijin Hua, Zichuan Liu, Zhuokai Zhao, Yujia Zheng Latent world models are more useful when their state separates factors shared across related observations (invariant) from factors that change between them (variant), but existing methods get this split through reconstruction. SplitJEPA is a joint embedding predictive architecture (JEPA) that learns the latent state and its invariant/variant structure directly in representation space, without reconstruction. The authors prove that, under stationary Gaussian predictive dynamics and a full-rank variation condition, it identifies the invariant and variant subspaces up to block-wise isometries without any decoder. Experiments on synthetic nonlinear systems and robotic manipulation tasks support the theory and show gains in robustness and efficiency.

arXiv:2610.12349 PDF

55 more specialized papers

Multimodal 47

SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows

Albert Gao, Bing Xue, Andrea Zanette cross-listed Multimodal large language models (MLLMs) often answer visual questions from language priors instead of looking at the image. Textual chain-of-thought helps but makes inference slower, and existing latent reasoning methods offer little control over what intermediate states encode. Structured Latent Visual Reasoning (SLVR) first trains the model to rely on the image by masking answer-revealing text and contrasting the correct answer with plausible distractors. It then organizes reasoning into typed latent stages for planning, grounding, evidence selection, and integration, each supervised with its own signal. Built on Qwen2.5-VL-7B, it gains +9.4 on MMVP and +14.2 on BLINK Relation, with further improvements on V*, MathVista, and ChartQA, without generating textual reasoning chains.

arXiv:2610.10563 PDF

SAVU-BENCH: A Real-World Benchmark for Spatial Audio-Visual Understanding

Yu Chen, Ruihang Liu, Yangguang Xu, Xinyue Jiang, Mohammed Bennamoun, Farid Boussaid et al. cross-listed SAVU-Bench is a real-world benchmark for spatial audio-visual understanding, covering three capability levels and seven tasks, paired with SAVU-Diag, which breaks reasoning questions into prerequisite grounding and alignment sub-tasks. Across 12 models, visual spatial grounding is relatively mature, but spatial perception involving audio is the main bottleneck, and most reasoning errors coincide with failures on prerequisite tasks. A training-free baseline, SAVU-EA, that makes spatial cues explicit substantially improves grounding and matching, but high-level spatial reasoning remains difficult.

arXiv:2610.10624 PDF

VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning

Meng Lu, Ligeng Zhu, Olivia Xiao, Yuchen Zhuang, Zihan Wang, Kuncheng Wu et al. cross-listed Reinforcement learning with verifiable rewards (RLVR) for vision-language models (VLMs) usually trains on a fixed task set. As the model improves, many tasks become trivial while others stay unsolvable, and the learning signal fades. VICO trains an actor together with an EnvRewriter that edits verifiable image-side structures, such as scene graphs, chart tables or region masks, and re-renders them into correctly labeled samples. A pass-rate-based reward calibrates the difficulty of these samples to the actor's current ability, without extra human annotation. Across nine multimodal benchmarks, VICO-8B gains up to +5.0% on out-of-domain tasks and beats self-evolution and text-editing co-evolution baselines by 4.3% and 8.4%. It matches chart-specialized RLVR methods while using 16 to 160 times fewer labeled samples.

arXiv:2610.10782 PDF

Omni-Diffusion-Distill: Few-Step Distillation of Unified Multimodal Diffusion Large Language Models

Hong Huang, Chenhongyi Yang, Junzhe Sun, Animesh Sinha, Wuyang Chen, Yifan Jiang cross-listed Unified multimodal diffusion large language models (dLLMs) handle both image generation and multimodal understanding, but each output takes tens to hundreds of forward passes. Omni-Diffusion-Distill is a two-stage distillation method in discrete token space: the student first learns to skip steps by replaying cached teacher trajectories, then is refined on states from its own rollouts. A pairwise collision penalty reduces repetition in parallel text decoding, and entropy-matched guidance prevents entropy collapse in image generation. It cuts image generation from 128 to 8 steps and understanding from 512 to 64, giving 18.2x and 21.2x wall-clock speedups, while scoring 0.828 on GenEval, 83.0 on DPG-Bench, and 57.2 on COCO captioning, twice the teacher's score at the same step count.

arXiv:2610.10990 PDF

Mid-Training Language Models on Raw Video

Jaedong Hwang, Xiaoqian Shen, Ernie Chang, Changsheng Zhao, Chong Zhou, Saksham Suri et al. cross-listed Tests whether raw, uncaptioned web video can serve as mid-training data for an existing language model. The model learns to predict the next continuous visual token, with no text loss. After mid-training Qwen3-1.7B on YT-Temporal-1B clips and then applying the same image-text instruction tuning as a baseline, the model gains 2.9 points on four video benchmarks and 5.1 points on ten image benchmarks, while its text performance holds steady. The gains appear within the first 30% of training, and predicting captions instead did not beat purely self-supervised next-visual-token prediction.

arXiv:2610.11019 PDF

GameCommBench: A Unified Benchmark and Type-Aware Evaluation for AI-Generated Game Commentary

Qirui Zheng, Zhengteng Lin, Yunyi Xiao, Junhao Li, Keyuan Cheng, Xingbo Wang et al. Research on AI-Generated Game Commentary (AI-GGC) is fragmented across games, modalities and evaluation protocols, and overlap-based metrics ignore that commentary serves different functions. GameCommBench unifies commentary for board games, sports and esports, with each comment annotated by type. The accompanying Type-Aware Commentary Evaluation (TACE) framework scores each commentary type separately and is validated against human judgments. Benchmarking current AI commentators shows uneven capabilities, with live observation and strategic analysis as the main bottlenecks.

arXiv:2610.11129 PDF

V-CoLA: Vision Token Compression with Linear Attention

Hao Jiang, Yiru Mao, Tianpeng Bu, Hao Zhou, Hongtao Duan, Wang Jing et al. cross-listed Vision tokens make up most of a vision-language model's (VLM's) input, but token compression methods built for softmax attention degrade on hybrid architectures with linear attention such as Qwen3.5. V-CoLA is a training-free compression framework for linear attention. It combines a uniqueness-aware importance criterion with adaptive token merging and is implemented to stay compatible with chunk-wise parallel linear attention. It keeps 99.5% of original performance with half the vision tokens and over 88% with 12.5%, while speeding up prefill by 1.86x to 6.15x.

arXiv:2610.11251 PDF

SAGE: Sink-Aware Guided Emphasis for Visual Grounding in Vision-Language Decoders

Jeonghyo Song, YoungJoon Yoo cross-listed Vision-language models (VLMs) often hallucinate partly because decoder attention collapses onto a few image tokens, known as attention sinks. The authors find that this behavior depends on the layer: early and late decoder layers collapse onto the same image regions regardless of the prompt (which they call Prompt-Invariant Sinks), while middle layers are prompt-conditioned and drive vision-language alignment. SAGE (Sink-Aware Guided Emphasis) is a lightweight intervention that steers attention away from these sinks and toward query-relevant regions, using region masks derived from backbones such as CLIP, ViT, and DINOv3. Across several VLM families, it improves visual grounding, reduces hallucinations, and yields gains on public benchmarks, including fine-grained discrimination tasks.

arXiv:2610.11469 PDF

From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction

Uddipan Basu Bir, Vincent Christlein, Andreas Maier, Mathias Zinnen cross-listed Heritage archives need OCR-to-structured-data extraction for handwritten and irregular documents, but commercial vision-language model (VLM) APIs raise concerns about cost, data control, and environmental footprint. The authors compare eight open-source VLMs of up to 7B parameters that extract text and produce schema-compliant JSON across three university heritage collections, in zero-shot, few-shot, and fine-tuned settings. They measure Character Error Rate (CER), ANLS*, and mAP-F1, and also test the effects of hyperparameter optimization, classical image preprocessing, multi-stage training, and per-dataset versus joint multi-dataset fine-tuning. They conclude that carefully adapted VLMs of up to 7B parameters offer a private, high-performing alternative to manual transcription or commercial systems, and give practical guidance for institutions.

arXiv:2610.11818 PDF

From Suppression to Repair: Mitigating Object Hallucination in Large Vision-Language Models via Localized Distribution Alignment

Chen Zhao, Xingping Dong, Jiachun Shi, Liang Peng, Chong Wang, Zhen Lei et al. cross-listed One common fix for object hallucination in large vision-language models (LVLMs) suppresses hallucination-related directions in hidden states, but this can also remove useful information. ResOT is a training-free, inference-time method that instead repairs these representations. It isolates a low-dimensional residual subspace of hallucinated directions, uses Gaussian optimal transport (OT) to align the hallucinated distribution with the faithful one, and adaptively moves each token state toward its OT target. On three LVLMs it substantially reduces object hallucination while also improving caption quality and multimodal benchmark performance.

arXiv:2610.11826 PDF

When History Helps and Hurts: Selective History Use across Multimodal Turns

Shuoyang Sun, Kerui Gu, Hao Fang, Shaoli Huang, Bin Chen Multimodal assistants must decide which parts of a conversation's history still apply: an earlier question may remain relevant even though its answer is outdated, or a current request may depend on earlier media. ReTurn is a benchmark of 7,000 base tasks with visual and audio evidence. Its four task types (Reconfirm, Reground, Retrieve, Rebind) separate this kind of history use from question difficulty, and each task has a matched single-turn version for comparison. Across 13 omni-modal, vision-language, and audio-language models, median open-ended accuracy drops from 93.7% with direct input to 72.3% in conversation. Models often recall the earlier question yet fail to apply it, and supervised fine-tuning closes only part of the gap.

arXiv:2610.11948 PDF

DataVista: Diagnosing Multimodal LLMs on Data Video Understanding

Yupeng Xie, Zhenyang Wang, Jiayi Zhu, Yinghao Tang, Zhouan Shen, Yiyu Chen et al. cross-listed Data videos combine animated charts with video storytelling. Understanding them requires reading values from moving charts, connecting evidence across time, and following the narrative, yet no benchmark has evaluated this systematically. DataVista contains 961 real-world data videos and 6,775 questions organized into three levels (data perception, temporal reasoning, and narrative understanding) across 10 question types. Among 19 multimodal LLMs, the best model, Gemini-3.1-Pro, reaches 70.0% accuracy, well below human experts, and models are weakest on causal reasoning and narrative structure. Adding more frames or subtitles helps perception and temporal reasoning but does little for narrative understanding.

arXiv:2610.11993 PDF

Perception Test 2026: Challenge Summary and Extension to City-scale Audio-Visual Reasoning

Fedor Kitashov, Jo\~ao Carreira, Shiry Ginosar, Dima Damen, Andrew Zisserman, Viorica P\u{a}tr\u{a}ucean cross-listed This report covers the fourth Perception Test challenge, held at the European Conference on Computer Vision (ECCV) 2026 with a focus on spatial intelligence. It had four tracks: multiple-choice and grounded video question answering from the original Perception Test benchmark, plus two new city-scale tracks built on walking-tour videos, KilometerAudio and KilometerVision. The report describes the new benchmarks and summarizes the winning solutions, including a generalist model that entered every track. In the city-scale tracks, complex spatial and multimodal reasoning was solved by expensive agentic pipelines but remained difficult for standalone multimodal models.

arXiv:2610.12081 PDF

EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams

Heeseung Kim Wearable augmented reality (AR) assistants that watch the user through first-person video and audio must decide on their own both when to speak and what to say, and no existing system handles both together. EgoVoice builds training data from HoloAssist recordings of real human instructors, using source separation and speech resynthesis to produce clean audio. Each session becomes a moment-by-moment choice between staying silent and giving spoken guidance. An omni-modal LLM is fine-tuned on this data and then refined with direct preference optimization (DPO). Existing closed and open models rarely make well-timed, useful interventions, while EgoVoice clearly improves intervention timing, content relevance, and human preference over its zero-shot backbone.

arXiv:2610.12248 PDF

DVD: Dynamic Vector Decoding for Efficient MLLM-based Perception

Jinghua Hou, Zhe Liu, Hengshuang Zhao cross-listed Multimodal large language models (MLLMs) used for perception typically output coordinates as text, which costs many tokens, or as fixed-range quantized values, which limits range and precision, especially for unbounded 3D scenes. DVD (Dynamic Vector Decoding) converts 2D boxes, 2D masks, and 3D boxes into 1D vector sequences, maps them to compact discrete tokens, and uses a lightweight de-tokenizer to turn model outputs back into perceptual structures. On RefCOCO, SUN-RGBD, KITTI, Hypersim, and nuScenes, the method reports stronger 2D and 3D results while significantly reducing token overhead and inference latency.

arXiv:2610.12266 PDF

Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models

Hongxing Li, Yixin Li, Dingming Li, Zixuan Wang, Yuchen Yan, Wenqi Zhang et al. cross-listed Vision-language models (VLMs) are weak at spatial reasoning because RGB images do not directly provide geometric evidence. GPD (Geometry-Privileged Distillation) renders depth, semantic, and bird's-eye-view cues from 3D-scanned training scenes as compact text and routes them to a teacher in on-policy self-distillation. A privileged KL term, applied only to incorrect trajectories, is added to GRPO, while the deployed model still sees only RGB images. On a 4B backbone it reaches 57.1 on VSI-Bench and 37.6 on average across MindCube, SPARBench, MMSI-Bench, and ViewSpatial, outperforming GRPO and distillation that uses only the answer as privileged information.

arXiv:2610.12355 PDF

SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models

Hongxing Li, Jinyue Su, Dingming Li, Wenqi Zhang, Weiming Lu, Jun Xiao et al. cross-listed Most spatial reasoning benchmarks test perception of relations already visible in the input, not the ability to predict how a scene changes after an intervention. SpaceCast-Bench follows an observe-transform-infer design with 3,862 questions from 182 real-world scenes, covering 16 task types at three levels: static perception, local prediction, and global prediction. Across 21 evaluated models, the strongest reaches only 58.0% against 87.2% for humans, and spatially specialized models stay near random chance. Analyses show that bridge views connecting separate observations matter and that explicit 3D evidence helps more reliably than generated outcome images or videos; fine-tuning on the programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7%, with gains on six out-of-domain benchmarks.

arXiv:2610.12402 PDF

WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Zheyu Fan, Yue Zhang, Mingkai Deng, Kangrui Wang, Qineng Wang, Canyu Chen et al. cross-listed The authors hypothesize that multimodal large language models (MLLMs) fail at spatial, embodied, physical, and temporal reasoning because they share one deficit in visual transition reasoning, meaning predicting how a visual state changes. WOVEN is a training source and benchmark of 36,076 examples spanning 20 scene types, 5 action types, and 8 reasoning types, built from rollouts of video-pretrained generative models; evaluating 38 frontier MLLMs, including GPT-5.4 and Qwen3-VL-235B-A22B, shows a systematic gap to humans that persists with scale. Training on subsets of about 2,000 items each improves 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data can replace 30-50% of a task's own training data at comparable accuracy. Controlled comparisons yield a training recipe: choose supervision by the reasoning operation it teaches rather than by scene or domain, and prefer larger visual state changes.

arXiv:2610.12417 PDF

29 more specialized papers

Safety & Alignment 45

Coverage, Not Difficulty, Sets How Much Synthetic Data an Activation Probe Needs

Ankush Checkervarty Activation probes that monitor deployed language models are trained on synthetic conversations, and it has been unclear how many samples they need. The authors trace learning curves from 10 to 590 samples for three concepts (high-stakes situations, harmful replies, and instruction non-following) across fourteen evaluation distributions and four probe models. High-stakes and harm probes get close to their plateau at about 80 samples on Gemma-3-27B-IT, while instruction-following probes need several times more. The main driver is coverage rather than difficulty: each kind of data needs a median of only 7-11 samples of its own kind, and the concepts differ mainly in how well samples of one kind transfer to the others. So generating broader data pays off a lot for some concepts and very little for others.

arXiv:2610.10594 PDF

When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry

Yixin Tan, Jiayang Liu, Lu Sun, Yuke Hu, Zheng Li, Rui Wen Mixture-of-Experts (MoE) language models produce routing telemetry that operators may log for monitoring, debugging, or auditing, and the authors ask whether this telemetry reveals which examples were used for fine-tuning. Their router-augmented membership inference attack combines output-side signals with aggregated routing features and uses a classifier trained on shadow models. Across three MoE architectures and three data domains, router telemetry raised the true-positive rate at 1% false-positive rate by 2.7 to 9.4 percentage points over a strong output-only ensemble. The leakage persists under LoRA, frozen-router training, and restricted telemetry, and mechanistic analysis shows the router exposes membership signal already present in hidden representations, even when its own parameters are frozen.

arXiv:2610.10616 PDF

When AI Finds Hidden Messages, Does It Report?

William Guey, Rashik Jahangir, Pierrick Bougault, Vitor D. de Moura, Wei Zhang, Jos\'e O. Gomes cross-listed The study tests whether AI assistants tell their users when they encounter a hidden message addressed to another AI. Four model-provider deployments ran simulated tasks across 1,408 sessions containing harmless or harmful messages in plaintext or ROT13 encoding, alongside no-message controls. Explicitly asking for reports raised notifications identifying another AI as the recipient by about 53 to 55 percentage points for encoded messages. Trace checks also found cases where agents interpreted a message but did not notify the user, which the authors use to argue that interpretation, notification, and task performance are distinct outcomes to measure.

arXiv:2610.10620 PDF

Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models

Yi Wang, Xiuyuan Qi, Dongqi Han, Dongsheng Li, Wenjie Wang cross-listed Because each recurrent depth of a Looped Language Model (LoopLM) can be read out independently, safety at one depth does not guarantee safety at another. A jailbreak evaluation across depths finds that attack success can increase at deeper inference depths, that the same query can behave differently across depths, and that attacks transfer between depths, and standard supervised fine-tuning and preference alignment do not close these gaps. The proposed SafeBridge combines depth-specific control of shared layers, selective state bridging, and joint safety supervision across depths. It substantially reduces attack success for both matched-depth and cross-depth attacks while improving general utility and keeping over-refusal comparable.

arXiv:2610.10625 PDF

Nullify: Null-Space Activation Steering for Training-Free LLM Unlearning

Wei Zhai, Xiang Liu, Qiang Huang, Rui Qian, Lemao Liu, Ziwei Li et al. LLM unlearning tries to erase specific memorized private knowledge without hurting general utility, but existing methods trade the two off and require costly fine-tuning. Nullify is a training-free method that adds steering vectors at inference time to redirect activations for forget-set queries away from their memorized answers. A null-space constraint leaves activations for retained queries essentially unchanged. On TOFU and MUSE it matches or beats established baselines in forget quality with near-lossless preservation of model utility, and it never updates the model's weights.

arXiv:2610.10655 PDF

Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors

Jan Dubi\'nski, Anna Sztyber-Betley, Jan Betley, Owain Evans Subliminal learning is when a student model picks up a teacher's trait by distilling on data unrelated to that trait, and so far it had only been shown for simple traits such as animal preferences. The authors show it can transfer a new capability, predicting the outputs of a random MLP, as well as backdoors and agentic misbehavior. A student distilled on number sequences from a backdoored teacher answers in French on 23.5% of prompts with female names versus 0.0% with male names, and a student trained on numbers from a steered "hacker" teacher hacks an agentic chess environment in 58.3% of episodes versus 10.9% for the base model. The amount of transfer depends on the setup and increases with logit distillation or with LoRA restricted to attention layers.

arXiv:2610.10657 PDF

BRANCH: Bypassing Multi-Scanner AI Guardrails

William Hackett, Peter Garraghan cross-listed Guardrail systems built from several cooperating scanners can block jailbreak and prompt-injection bypass techniques that work against a single classifier. BRANCH attacks these multi-scanner systems with a branching tree search: it applies adversarial perturbations against individual scanners, then optimizes perturbations and selects techniques based on the overall improvement across all scanners in the system. It reaches a 100% attack success rate across 6 guardrail systems in 120 scenarios, using 72% fewer queries and 4.5x less wall-clock time than established techniques while preserving the meaning of the original prompt. The generated bypasses also transfer to 29 unseen guardrails, including 8 commercial black-box products.

arXiv:2610.10742 PDF

Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models

Gaurav Patel, Jun Fang, Greg Ver Steeg, Qiang Qiu, Sravan Sripada cross-listed Text-to-image diffusion models are often distilled into few-step (FSD) variants for fast inference. Existing unlearning objectives that suppress harmful generations assume multi-step denoising and therefore fail on these distilled models, while unlearning the base model and re-distilling it is costly. The authors show that standard Direct Preference Optimization (DPO) and its unlearning variants, which are built on noise-prediction error, transfer poorly to FSD models. They propose a modified preference-optimization objective that matches few-step generation dynamics. On identity, nudity, and object removal, it achieves effective forgetting directly in FSD models while retaining non-targeted capabilities and keeping few-step efficiency.

arXiv:2610.10859 PDF

How Hackable Is Your Speech Quality Metric? A Corrected Protocol, a Benchmark, and What Patching Buys

Ali Alavi, Donald S. Williamson Speech quality predictors are increasingly used as training rewards, but there is no agreed way to measure how easily they can be gamed. The authors fix two flaws in the usual measurement. First, they score attacks against the unperturbed neural-codec round trip rather than the raw input, which changes measured hackability by up to a factor of four. Second, they take the worst case over several attackers, because seed-only variation moves attack success from 0.00 to 0.38. Under this protocol NISQA is hacked on 90% of utterances, against 6% to 21% for UTMOS, DNSMOS and SSL-MOS. An attack-detect-patch loop hardens a predictor only within its own attack space, does no better than a random-perturbation baseline, and costs up to 0.30 system-level Spearman rank correlation (SRCC) out of domain.

arXiv:2610.10899 PDF

Reading the Room: Foundations, Design, and Challenges of Normative Competence in LLMs

Andrea Wynn (Nathan), Harsh Satija (Nathan), Seokhyun (Nathan), Baek, Anqi Liu, Eric Nalisnick et al. Aligning AI systems with community norms requires normative competence, which the authors define as inferring from interaction alone which norms a community enforces, without relying on pretrained knowledge. They build a multi-agent community debate setting in which access to debate is governed by synthetic norms. Baseline LLM agents fail to learn the norms even when doing so would improve their accuracy, and add-on normative modules for norm inference are highly sensitive to the style of norm and the underlying model. Agents also show an unselective attribution failure: they copy idiosyncratic noise alongside the enforced rules, even when imitating unnecessary behaviors is explicitly penalized.

arXiv:2610.10906 PDF

Power Side-Channel Membership Inference Attack on Embedded Machine Learning

Sahan Sanjaya, Prabhat Mishra cross-listed Membership inference attacks (MIAs) usually need model outputs such as probabilities or labels, which on-device systems may not expose. PSCMIA infers whether a sample was in the training set directly from the device's power traces during inference. It is evaluated on MNIST, FMNIST, CIFAR10 and CINIC10 with fully connected and convolutional models on STM32F3 and XMEGA microcontrollers. The attack reaches a ROC-AUC of up to 0.907 on fully connected models and outperforms label-only MIA in 11 of 16 configurations, showing that hiding model outputs does not prevent membership leakage.

arXiv:2610.10909 PDF

Speedbumps: Rejection Attacks on Speculative Decoding

Adam Y. J. Jones, Yu Yuan, Sergio Maffeis cross-listed Speculative decoding speeds up LLM inference by having a small draft model propose tokens that the target model verifies in one pass, so the speedup depends on how often drafts are accepted. The authors introduce Speculative Rejection Attacks, in which adversarial suffixes in attacker-controlled content make draft and target models disagree. This forces more target-model passes and raises the victim's inference costs. Two optimization variants, Speedbump-P and Speedbump-D, can make decoding slower than plain autoregressive decoding, remain effective under sampling, and transfer across drafters or across target models that share a drafter. The attacks also degrade output quality, and regularizing to preserve quality gives up most of the slowdown.

arXiv:2610.10929 PDF

RH-Detect: A Unified Benchmark for Reward Hacking Detection

Junwei Quan, Evgenii Opryshko, Rohan Subramani, Igor Gilitschenski RH-Detect merges reward-hacking-relevant subsets of eleven public datasets into a common schema covering 92,761 rows and six behavior categories, so reward hacking detectors can be compared directly. Six off-the-shelf LLMs used as detectors without additional training reach a best pooled AUROC of 0.962. However, the four strongest are 10.7-15.9 percentage points less accurate on multi-turn tool-use trajectories from MALT and TRACE. Input format effects vary by model, for example removing thinking helps Qwen3.5-4B but hurts Qwen Flash. Single-token supervised fine-tuning on the benchmark improves held-out detection on five of six sources.

arXiv:2610.10947 PDF

How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense

Zhankai Ye, Yanning Wang, Yukai Jin, Bo Mei, Fangyi Li, Wei Wang et al. Safety-aligned models that refuse a harmful request stated directly often comply when the same request is wrapped in role-play or a narrative, and on Qwen3-1.7B attack success reaches 89.4% in English and 95.7% in Classical Chinese. The GUISE benchmark tests this with parallel English, modern Chinese, and Classical Chinese requests, matched harmful and benign pairs, held-out wrapper types, and a strict criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that narrative wrappers push harmful requests much farther from the model's refusal direction than language or register changes do. The proposed defense, AXIS, adds a rotation objective that realigns those representations with the refusal direction and a commitment objective that trains full refusal to preference optimization, and it gets the best combined safety and usability score on Qwen3-1.7B, Qwen3-4B, and GLM-4-9B.

arXiv:2610.11005 PDF

Distillation for Incrimination and Distillation for Capabilities

Sebastian Prasanna, Jacqueline Tay, Alek Westover A misaligned model might behave well whenever it recognizes it is being audited. The authors propose using distillation into a weaker student to get around this, in two ways. Distillation for Incrimination (DFI) tries to transfer the misbehavior without the ability to hide it: students distilled from AuditBench secret-keeping models are significantly more likely than their teachers to admit the hidden behavior, but only when the student shares the teacher's pretrained base model. Distillation for Capabilities (DFC) tries to transfer skills without misalignment. There, inoculation prompting and training for more epochs on fewer unique examples keep the capability gains while sharply reducing subliminal transfer of a proxy trait, an animal preference.

arXiv:2610.11012 PDF

Characterizing Statistical Separability in TP-CRIV for Probabilistic AI Models

Teruki Sano, Minoru Kuribayashi, Masao Sakai, Shuji Isobe, Eisuke Koizumi, Zhang Zhang et al. cross-listed Third-party challenge-response identity verification (TP-CRIV) lets an independent verifier check whether a claimant holds the same model as a remotely deployed one. This is hard for probabilistic models, because repeated queries return different outputs. The authors derive how challenge-level behaviour of matching and non-matching provers adds up to verification-level separability, and make explicit how the number of challenges and repeated responses affects detection. This lets one estimate the verification budget needed for a target AUC (area under the ROC curve). In experiments on LLMs with open-ended challenges, theoretical and empirical AUCs agree closely and the minimum-budget estimates are consistent.

arXiv:2610.11163 PDF

How to post-train on a surrogate: Envelope sampling mitigates reward hacking

Sanjit Dandapanthula, Shuvom Sadhuka, Samir Khan, Michael Oberst, Aaditya Ramdas, Alexandra Chouldechova LLMs are often post-trained with reinforcement learning against cheap surrogates such as LLM judges, which invites reward hacking whenever the judge is miscalibrated. The setting studied here has a small number of outputs labeled with ground truth, used to recalibrate the judge before optimizing against it. The proposed envelope sampling picks which outputs to label by minimizing an upper bound on the post-trained model's regret, assuming the true and recalibrated rewards stay close to the judge. It comes with practical samplers based on rejection sampling or fine-tuning on a modified reward. On clinical note generation and a controlled sycophancy task, recalibrating on envelope samples mitigates reward hacking, while recalibrating on base-model samples does not.

arXiv:2610.11281 PDF

BeliefScope: Diagnosing Evidence-Driven Revision and Pressure-Induced Shifts in Large Language Models

Shuai Guo, Yidong Cui When an LLM changes its stance on a claim, the visible shift alone cannot tell whether real evidence or mere directional user pressure caused it. BeliefScope is a controlled black-box framework that crosses Evidence and Pressure conditions with factor-specific controls and measures shifts through probability reports, categorical judgments, and action recommendations. It is validated with known-truth synthetic recovery, ablations, and noisy semi-synthetic stress tests. In a study across Qwen and Llama model families, with spot checks on Gemma3-12B, broad model-level differences can flip under matched controls, decoding settings, or response formats, while some narrower within-model patterns hold. Instructions that reduce pressure-following can reflect either real resistance or overcorrection in the opposite direction.

arXiv:2610.11305 PDF

Deception by Omission: Language Models Knowingly Hide Their Mistakes

Lucas Florin, Amelie Knecht, Ulysse Schaller, Thilo Hagendorff As LLMs act as agents with little human oversight, users depend on them to report their own mistakes. The study prefills chat and agentic trajectories with synthetic mistakes and measures whether models disclose them. Models fail to disclose the mistake in 36.4% of chat rollouts and 67.1% of agentic rollouts. In 2.4% and 5.3% of rollouts, respectively, the chain of thought shows awareness of the mistake but the model still conceals it, reaching up to 19.9% of agentic rollouts for Gemini 3.5 Flash. Many models show no awareness of their mistakes at all, yet spot them reliably when reviewing the same transcript as an outside observer, so the authors recommend independent trajectory monitors or training models to check and disclose their past actions.

arXiv:2610.11351 PDF

When to Intervene? State-Aware Sparse Manipulation in Federated Reinforcement Learning

Shutong Zheng, Sijia Chen Federated reinforcement learning (FRL) lets distributed agents train a shared policy, which exposes it to poisoning by malicious participants. Prior attacks focus on how to craft malicious updates, but controlled experiments here show that when an attack is applied, meaning which trajectory states are targeted, substantially changes its effect even with the same update construction. The proposed V-BSA attack uses the local policy's uncertainty to choose a sparse set of states to intervene on, then applies constrained behavioral steering. On discrete-action benchmarks it significantly degrades training under robust aggregators and ensemble defenses while using a fraction of the interventions that dense poisoning needs.

arXiv:2610.11523 PDF

Workerville: Towards an Organizational Behavior Account of Agent Safety

Hanjun Luo, Junting Mao, Yuhan Lu, Haobo Zhang, Zhimu Huang, Yankai Chen et al. The authors propose organizational behavior, the study of how people act within organizations, as a framework for agent safety. They formalize counterproductive work behavior for agents as Agentic Counterproductive Behavior (ACB), which links supervisor relations, peer norms, and internal cognitive structures to unauthorized disclosure, destructive operations, and production deviation. The Workerville benchmark applies 16 organizational configurations to 210 tasks, producing 3,360 challenges, and evaluates 6 frontier LLMs. Combining negative influences amplifies risk non-monotonically: unauthorized disclosure rises from 16.5% with no negative influence to 60.1% with two, then falls to 50.3% with three. The agents also reproduce behavioral patterns that human research on counterproductive work behavior predicts.

arXiv:2610.11561 PDF

LTBD: Learnable Trust-Boundary Delimiters for Prompt Injection Defense

Luman Zhao, Minghui Xu, Yue Zhang, Yijun Yang cross-listed Prompt injection attacks hide instructions in external data that override the user's intent, and the authors attribute this vulnerability to LLMs having no explicit representation of where each piece of input came from and whether it is trusted. Learnable Trust-Boundary Delimiters (LTBD) add a small number of learned delimiter embeddings that separate trusted user instructions from untrusted external data, without changing the LLM's weights. The defense outperforms inference-time defenses, performs on par with training-based ones, preserves benign-task utility and adds negligible inference overhead. It achieves a 0.00% attack success rate on AlpacaFarm and 0.11–0.19% on TaskTracker, and it holds up against adaptive attackers who know how the defense works.

arXiv:2610.11634 PDF

Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising

Nikhil Verma, Siddharthan Dileep, Anoop Singh, Srikanth Sastry, Ramya Hebbalaguppe, Sayan Ranu et al. Standard memorization tests for diffusion models only catch the problem once generated samples are near-copies of training data. The authors show that memorization first appears in the geometry of the learned energy landscape: localized basins form around training samples before any memorized sample is generated. They probe these basins with cyclic denoising, which repeatedly adds partial noise and then denoises, and this recovers training images from CelebA and CIFAR-10 checkpoints whose ordinary samples contain no copies, raising one checkpoint's memorized fraction from 0.1% to above 30%. On off-the-shelf Stable Diffusion v1.4, a cycling-based signal separates memorized from non-memorized prompts with an AUC of 0.944.

arXiv:2610.11670 PDF

Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation

Haitong Jiang, Chunlin Liu, Sihan Tang, Chan Wu, Xiaoqing Su, Yuhong Feng LLM safety evaluations usually report attack success rate (ASR), but a non-harmful response can mean the model understood a harmful request and refused it, misunderstood it, or answered something else entirely. The authors pair ASR with an operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to answer. Across interfaces, similar ASR values hide sharply different intent-recovery rates. Recovery rises as compressed prompts become more explicit, while ASR does not follow the same pattern, and on FormalLogic high recovery coincides with frequent harmful assistance. The authors argue that intent recovery should be reported alongside ASR.

arXiv:2610.11766 PDF

Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models

Youwei Feng, Yitong Zhang, Yuetong Liu, Jia Li Guard models for LLM agents mostly flag forbidden actions. The authors argue that agent safety also depends on catching "obligations": safety-critical actions that were required but never performed. On a popular safety benchmark, 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, versus 30.00% containing forbidden actions. They introduce ObligationBench, 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. The best of 14 evaluated models reaches only 48.97% recall, and ObligationGuard, trained on 40,000 synthetic examples, improves this to 57.52% recall and a 21.67% exact-match rate.

arXiv:2610.11773 PDF

Is Memorization Context-Sensitive? Prefix-Based Extraction Beyond Isolated Prefixes

Ali Satvaty, Narjes Sharafi, Jirui Qi, Suzan Verberne, Fatih Turkmen LLMs can leak memorized training text when given a prefix from a training example. In deployed systems such as retrieval-augmented generation (RAG), though, that prefix usually arrives alongside instructions or retrieved documents. The authors score each memorized suffix with no context and again under retrieved contexts of varying relevance, across three open-weight instruction-tuned models. Context does not erase memorization. A context-robust core of samples stays extractable, especially with longer prefixes, while context mostly shifts borderline cases, suppressing some leaks but also enabling new ones that prefix-only testing misses, which weakens the claim that RAG reduces memorization risk.

arXiv:2610.12085 PDF

Could LLM Watermark Detection be Public?

Georgios Milis, Tom Sander, Tom\'a\v{s} Sou\v{c}ek, Heng Huang, Pierre Fernandez cross-listed Providers keep LLM watermark detectors private for fear that attackers would use detector feedback to strip watermarks. The authors first measure how much extra risk a public detector actually adds at several levels of access. They then propose a split-key scheme: one key is exposed through a public detector and the other is kept private, and a statistical test flags tampering through the imbalance between public and private scores. A public detector helps watermark removal only at small edit budgets, since plain rephrasing already strips watermarks. It does make forgery easier, but the private key catches those forgeries, which suggests detectors need not stay fully private.

arXiv:2610.12106 PDF

An Investigation of Model Coherence: Narrow Finetunes Contradict Themselves Under Resampling

Robert Graham, Yariv Barsheshat, Phil Blandfort, Sabri Alouache Many coherence measures for language models rely on output variance, which can also come from ambiguous questions or from the model not caring which answer it gives. The authors build a set of 175 questions where neither ambiguity nor indifference easily explains contradictory answers, then measure incoherence as contradictions across resampled answers to the same question. The metric is designed to have high specificity, flagging only glaring inconsistencies, yet narrowly finetuned models still score poorly. Model organisms from the literature show identity conflation, introspection failures, and rationalizations, which suggests that narrow-finetuning pathologies may limit what these models reveal about coherent misaligned behaviour.

arXiv:2610.12129 PDF

ReSI: Recursive Safety Improvement toward Resistant and Resilient AI

Jingnan Zheng, Dongcheng Zhang, Yi Zhang, Ming Zhang, Qiaosheng Zhang, Youbang Sun et al. cross-listed Models now change through frequent checkpoint updates, and new red-teaming methods keep exposing new vulnerabilities, so safety alignment has to be renewed for each checkpoint. ReSI (Recursive Safety Improvement) treats this as an automated research loop. Each round, it red-teams the current model with diverse methods, develops candidate training recipes, and promotes the update with the largest safety gain among those that pass a Pareto gate on capability retention. Across four dense and mixture-of-experts models, it matches or exceeds evaluated frontier models on in-distribution and out-of-distribution safety benchmarks while largely preserving general capabilities. It reduces mean X-Teaming attack success from 86.01% to 31.45%, below the 56.69% of GPT-5.6-Luna, the best frontier result the authors evaluated.

arXiv:2610.12233 PDF

One Word Opens the Gate: The Option-Channel Attack on Typed Decision Models as Agent Guardrails

Seyedarmin Azizi, Erfan Baghaei Potraghloo, Massoud Pedram Typed decision models, which return probabilities over caller-defined options rather than generating text, are increasingly used as agent guardrails that allow or block proposed tool calls and incoming messages. Testing seven open-weight models in this role, the authors find allow-or-block accuracy of 36% to 72% on prompt-injection, jailbreak, and toxic-content screening, with low error rates in one direction mostly reflecting a default answer. Six lines of irrelevant server log text raise a gate's fail-open rate from 0% to 63%. Giving the permissive option a misleading name raises the fail-open rate to 93-100% on the four models that include option labels in their input. Every defense tested is defeated, while a deterministic rule over parsed policy fields reaches 100% accuracy, leading the authors to argue these models should not be the deciding component.

arXiv:2610.12292 PDF

Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems

Saisab Sadhu, Aadit Sengupta, Vinay kumar Sankarapu, Pratinav Seth LLM compliance systems are assumed to base their verdicts on the regulatory rule they are given. Across five models and 20 regulatory and platform-policy domains, the authors delete, swap, or negate the governing rule while keeping the case fixed. They then measure changes in both the verdict and the model's internal representation of compliance. Verdicts are often unchanged by large rule perturbations, and the guard model tested is the least rule-sensitive and least accurate, at 51% versus 90-92% for general-purpose models. Much of this insensitivity reflects easy cases, since models do track the rule where deleting it flips a correct prediction, but neither better prompting nor representation-level intervention closes the gap.

arXiv:2610.12313 PDF

Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness

Saisab Sadhu, Shreeyans Arora, Pratinav Seth LLMs that justify legal verdicts by naming a statute or precedent may not actually rely on that authority. Holding case facts fixed, the authors swap the named authority for an unrelated one and decode the evolving verdict from hidden states, testing seven open-weight models (8B-70B) on CaseHOLD, ECHR, SCOTUS, and ContractNLI. Models name the correct authority 66.7%-100% of the time, but the verdict changes when the authority changes in as few as 0.0%-21.7% of cases on CaseHOLD. Neither larger scale nor a legal-reasoning fine-tune closes this gap. Meanwhile, models follow adversarial instructions hidden in the case facts 73.3%-96.4% of the time.

arXiv:2610.12361 PDF

Searching for "Harmful Refusal": A Psychometric Audit of an AI Safety Benchmark

Christopher M. Stewart, Preston Botter, Natalie Sarabosing, Muye Zhang, Rachel Phinnemore, Shalini Ghosh et al. Safety benchmarks often report one aggregate score, which can hide which attributes models actually differ on. Using a construct-validity framework, the authors test whether harmful refusal, a model's tendency to refuse dangerous prompts, is a single measurable attribute in HELM Safety. Three of the four candidate datasets turn out to be saturated, and on the remaining one, HarmBench, multidimensional item response theory modeling strongly suggests the score does not measure a single attribute. A differential item functioning analysis flags items where models from different developers with equal refusal ability score differently, though most flags disappear under scope-specific matching, and the authors argue that any score averaged over datasets and items should earn a single-attribute interpretation before it is used to compare models.

arXiv:2610.12409 PDF

Predicting Alignment Generalization with Value Representations

Andy Liu, Mehar Bhatia, Karolina Stanczak, Mona Diab, Vered Shwartz, Daniel Fried Fine-tuning an LLM on narrow behaviors can change its behavior in unrelated contexts, so the authors define alignment generalization prediction: forecasting how training a model to follow one value shifts its behavior on many held-out values. Across 66 values drawn from modern alignment targets, representations built from model activations while applying a value in context reach a correlation of 0.45 with the measured generalization matrix, versus 0.05 for methods based on textual value descriptions. Similarity between values in a multi-value alignment target, measured with these representations, correlates significantly with model robustness. The authors also report early evidence of a shared, model-independent value space and use it to build a taxonomy of LLM values grounded in observed generalization.

arXiv:2610.12410 PDF

Ecology of AI Agents: Collaboration Creates a Population Threshold for Takeoff

Erin Crawley, Hidenori Tanaka The authors ask what decides whether a population of misaligned AI agents that can compromise computers and quietly deploy more copies of themselves stays contained or grows out of control. They model this with a population growth equation in which an agent population's growth rate depends on its cybersecurity capability. Without collaboration, the population takes off only when individual-agent capability passes a threshold. With collaboration, collective capability grows with population size, which creates a critical population size above which the population takes off even though individual capability is unchanged (the strong Allee effect from ecology). Because red teaming a small group of agents cannot rule this out for larger populations, the authors propose 'ecological red teaming' and 'population pacing', which means deploying larger agent populations gradually in controlled environments and re-estimating the takeoff threshold for each new model generation.

arXiv:2610.12436 PDF

Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception

Oskar J. Hollinsworth, Alex F. Spies, Tigist Diriba, Adam Gleave, Chris Cundy The work tests whether white-box probes, which are classifiers trained on a model's internal activations, can catch deception and sabotage by LLM agents in realistic frontier monitoring settings. The authors assemble the largest deception dataset to date and introduce a probe architecture that combines information across many layers and tokens. Their probes reach 98.8% AUC on SHADE-Arena, outperforming an Opus 5.5 text-monitoring baseline, and they improve as the underlying model gets larger. The probes also detect 'introspective deception' that cannot be inferred from context alone, separating transcripts that contain a model's true hidden goal from other goals with up to 99.7% AUC, and they catch open-weight models lying about politically sensitive topics. The training dataset, FIBS, is released.

arXiv:2610.12445 PDF

From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security Incidents

Abbas Raftari cross-listed The authors compare three 2026 incidents in which cybersecurity evaluations of OpenAI, Anthropic, and Google agents reached real systems outside their authorized scope. The incidents include OpenAI agents compromising parts of Hugging Face's production environment and Google's Gemini reaching three real organizations through an unintended internet route. From these cases they build the Proactive Agent Security Assurance Cycle (PASAC) and a five-layer Boundary Assurance Stack, which combine executable scope contracts, pre-run validation, least-capability access, independent egress enforcement, cross-run monitoring, and automatic stop conditions. The central claim is that an evaluation's boundary must be continuously verified while the agent runs rather than assumed from any single sandbox. The authors note that their account of the Gemini case rests on attributed statements and journalism, so its causal details are provisional.

arXiv:2610.12463 PDF

8 more specialized papers

Robotics 37

Masked Generative Motion Planning with Geometry-Guided Token Search

Lipeng Zhuang, Yingdong Ru, Shiyu Fan, Edmond S. L. Ho, Gerardo Aragon Camarasa, Paul Henderson cross-listed Generative motion planners usually rely on a learned prior only to produce the first trajectory and then repair it with local continuous refinement. MGMP (Masked Generative Motion Planning) uses a masked generative transformer to produce discrete trajectory candidates in parallel. Its Geometry-Guided Token Search (GGTS) then uses scene geometry to decide where to edit and which alternatives supported by the prior to try, so repair can restructure a whole route instead of only deforming it locally. MGMP reaches 96% success on Ring Maze and 82% repair success on Controlled Route Invalidation with a Kuka arm, 23 and 25 percentage points above the strongest baselines, and it generalizes to unseen layouts and geometries, dual-arm planning, and real Baxter tasks.

arXiv:2610.10646 PDF

NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime

Gengze Zhou, Yicong Hong, Jiazhao Zhang, Xunyi Zhao, Jian Zhou, Zixing Lei et al. cross-listed NavGPT-3 connects a reasoning language model to a low-latency action policy for robot navigation through an OS-like runtime. Reasoning, acting and monitoring run as threads with their own context, tools and permissions, and the runtime decides which thread controls the robot's motion so the robot can react to sudden events. Underneath sits NavGPT VLA, an action policy trained on 19.28M examples that allocates visual tokens in proportion to scene change; on its own it reaches 74.51 success rate on R2R-CE. The full system sets a new state of the art on R2R-CE (81.51) and matches human followers on RxR-CE (90.43 vs. 90.4 success rate) in about 1 minute 22 seconds per episode, versus roughly 3 minutes for a human. When the action policy executes the route, the system's minimum reaction time drops from 3-19 seconds per language-model decision to 0.5-1 second per action-policy step.

arXiv:2610.10787 PDF

Diagnosing and Recovering from Observation-Space Shift at Long-Horizon Skill Seams

Pranav Wagh, Yu Fang, Yue Yang, Mingyu Ding cross-listed Robotic manipulation skills that are reliable on their own often fail when chained, because each skill starts from whatever state its predecessor left behind. The authors call this Observation-Space Shift and use privileged simulator resets to find its main cause: displaced scene state, such as an open drawer or leftover objects, rather than the robot's joint configuration or the target object. To test this diagnosis, they build a learned system that detects stalls, restores the displaced scene, and resumes with a skill fine-tuned for robustness at skill boundaries. On BOSS-44 this system raises full-chain success from 7.6% to 26.5%, while best-of-K resampling, Diffusion Policy and world-model baselines fail to recover from the evaluated states. On a real Franka arm with a fine-tuned pi0.5 policy it recovers some otherwise terminal failures but is limited by exterior-camera observability.

arXiv:2610.10810 PDF

Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning

Prabin Kumar Rath, Omkar Patil, Nakul Gopalan Behavior cloning (BC) in non-Markovian environments needs long-horizon memory, but recurrent policies suffer from hidden-state collapse and attention-based policies are capped by context length. Keyframe Mnemonics instead discovers a small set of information-critical past observations without supervision: it learns an objective from randomly sampled past observations and uses it as a reward for selecting keyframes. The BC policy then conditions on those keyframes, and under certain task-structure assumptions the method guarantees context retention over an infinite horizon. It reaches 100% success on synthetic memory tasks and improves average success by 13.9 points over the strongest baseline across 23 robot manipulation tasks, keeping 80% success at 20x longer horizons on a real robot.

arXiv:2610.10857 PDF

When Listening Becomes Easier: Scrubbing Visual Cues for Shortcut-Free VLAs

Jasper Gerigk, Kenzo Aspuru-Takata, Chin-Hsuan Wu, Mohammad Mohammadi, Shuhong Zheng, Igor Gilitschenski cross-listed Robot demonstration datasets have limited diversity, so vision-language-action (VLA) policies can latch onto spurious visual cues such as viewpoint or background instead of following the task instruction. The authors find that vision-language backbones differ substantially in how susceptible they are to these shortcuts. This susceptibility correlates with a proposed representation-level metric, action margin, which requires no policy rollouts. Visual shortcuts enter action representations in early layers, and models differ in how much later layers correct them using language. A domain-adversarial training method called task scrubbing strengthens the model's reliance on language and, in simulation and real-world experiments across several VLAs, often eliminates visual shortcut learning while improving out-of-distribution robustness.

arXiv:2610.10912 PDF

SafeInferCom: Safe Inference-Time Compute via Verifier-Guided Mid-Generation Intervention for Robotic Task Planning

Weizhe Xu, Jialiang Fan, Mengyu Liu, Fanxin Kong cross-listed When Large Reasoning Language Models (LRLMs) plan robotic tasks, continued reasoning can overwrite valid intermediate plans or leave constraint violations unresolved. The authors build an inference-time monitor that extracts and formally verifies intermediate plans without disturbing decoding, and they use it in SafeInferCom to keep valid plans and steer error correction during generation. Experiments across several LRLMs and planning domains show reasoning-response inconsistency and weak self-correction in one-shot inference. SafeInferCom improves planning success over one-shot inference, and when combined with iterative refinement it raises success while using fewer tokens than refinement alone. It is also evaluated in VirtualHome and demonstrated on a real robot arm.

arXiv:2610.11223 PDF

Being-M0.7: A Latent World-Action Model for Humanoid Robots

Junpeng Yue, Boyuan Li, Yuxuan Wang, Zepeng Wang, Yuhui Fu, Feiyang Xie et al. cross-listed Robot demonstrations for humanoids that both walk and manipulate objects are scarce, and human video and motion data are abundant but rarely paired and not directly executable. Being-M0.7 is a latent world-action model pre-trained on more than 10,000 hours of mixed human data (video-only, motion-only, and paired), learning to jointly predict future latent visual states and body motion. A robot mid-training stage adapts this prior to robot viewpoints and dynamics. An action expert then uses gated cross-attention to combine the frozen predictive features with current images and body sensor readings to produce whole-body commands. The model achieves the highest aggregate success rate among compared baselines on SIMPLE and matches the strongest baseline on real Unitree G1 loco-manipulation tasks.

arXiv:2610.11283 PDF

WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models

Kai Ding, Yang He, Ruijie Quan, Yi Yang cross-listed World Action Models (WAMs) control robots by having an action expert query key-value (KV) representations from a pretrained video Diffusion Transformer (DiT), and recomputing that video DiT at every action chunk dominates the compute cost. WAM-Cache is a training-free method that reuses layerwise KV pairs across chunks and recomputes only a sparse set of tokens. The authors find that refreshing the tokens that changed most visually falls well short of the dense baseline, so WAM-Cache instead picks tokens based on where the action expert attends, combined with visual surprise and a strict age bound on cached entries. On Fast-WAM, it cuts video DiT prefill FLOPs by 32-42% while staying within 0.7-1.8 points of dense performance in simulation and 2.5 points on a real robot.

arXiv:2610.11401 PDF

RoboAware: Learning to Coordinate Embodied Skills from Counterfactual Outcomes

Bohan Zhou, Xingbei Chen, Emily Huang, Weilin Ruan, Haojian Huang, Yehang Zhang et al. cross-listed Robot coding agents can call both modular skills and frozen end-to-end policies, but they need to know which kind of policy will work in the current physical state. RoboAware keeps the coding agent frozen and learns only a state-conditioned coordinator that picks the policy family at each step. Skills are grouped into five semantic stages under a P^5 schema, framed as a hierarchical Markov decision process. Training uses State-Locked Counterfactual Branching, which resets the robot to the same state and runs a code block from every eligible policy family, and Execution-Aware Learning, which combines Monte Carlo tree search with Q-learning to turn those outcomes into values. Across 100 tasks it reaches 77.0% overall success, including 90.0% on RoboSuite, 73.8% on LIBERO-Pro and 90.0% on bimanual RoboTwin tasks, beating code-as-policy and vision-language-action baselines.

arXiv:2610.11480 PDF

TACROSS: An Efficient and Low-Cost Scalable Human Touch System Across Heterogeneous Tactile Sensors for Dexterous Robot Learning

Bo Chen, Huanzhang Hu, Junyang Ma, Bo Yue, Fangdi Yu, Haijier Chen et al. cross-listed Collecting tactile demonstrations on robots is slow and expensive, but human tactile gloves differ from robot sensors in how they transduce touch, their layout, and their resolution, so raw signals cannot be directly aligned. TACROSS aligns human and robot touch at the level of contact events. It uses a USD 10.86 piezoresistive glove with 285 sensing points and maps the heterogeneous signals into a shared 256-dimensional tactile latent with a temporal Transformer. Robot demonstrations supply the action labels, while human demonstrations support representation learning and provide confidence-weighted auxiliary targets. On four contact-rich manipulation tasks, the system improves data-collection efficiency 3.5-fold over teleoperation, cuts equipment cost by 95.7%, and comes with a planned release of over 150 hours of tactile data.

arXiv:2610.11945 PDF

REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models

Houlong Xiong, Zhenqi Qiu, Zechen Wang, Suohang Zhang, Yiyu Ren, Wanting Xu et al. cross-listed Flow-based vision-language-action (VLA) models output chunks of actions, which forces a trade-off: long chunks give smooth motion, while frequent replanning reacts faster but causes jerky transitions. REACT keeps a persistent action buffer at staggered flow timesteps. At each control step it denoises the whole horizon against the latest observation, executes the cleanest block, shifts the partly refined future blocks forward, and appends fresh noise at the tail, so every executed block has been refined across several recent observations. A "dual decoupling" scheme runs sensing, VLM encoding, diffusion-transformer denoising, and execution separately to support real-time streaming. On the RoboTwin 2.0 simulation benchmark and real bimanual and dynamic tasks across multiple robots, REACT raises task success and cuts reaction latency while producing smoother trajectories than frequent-replanning and asynchronous baselines.

arXiv:2610.12007 PDF

Humanoid World Action Model With Joint State--Action Generation

Yan Yang, Jikun Rong, Minzhao Zhu, Zheyi Zhao, Qirui Hu, Zihan Lan et al. cross-listed In humanoid robots, vision-language-action (VLA) and world action model (WAM) policies output reference actions that whole-body control then carries out, so the motion the robot actually performs can differ from what the policy commanded. HWAM closes this gap by making the robot's post-execution proprioceptive state an explicit prediction target, generating reference actions and their realized body states together. Training uses three paths: a policy path conditioned only on current observations, forward dynamics modeling that predicts future frames from actions and realized states, and inverse dynamics modeling that recovers joint trajectories from visual transitions. On three real-robot tasks with the LimX OLI humanoid, HWAM achieves the highest success rate among the evaluated baselines, including 70.6% on Candy Picking versus 43.3% for Fast-WAM.

arXiv:2610.12026 PDF

Real-Time Motion Planning with Dynamic Hazards: Classical vs. Learning-Based Methods

Eran Iceland, Alexander Tuisov, Oren Gal, Ariel Barel, Alfred M. Bruckstein cross-listed A controlled benchmark compares classical motion planners with learning-based methods on real-time planning around dynamic hazards: planar maps with rotating, sprinkler-like obstacles whose sweeping sectors create time-varying forbidden regions. In deterministic environments, classical planners reach near-perfect success with better paths, though sometimes with long planning or replanning times. Under stochastic obstacle dynamics, online search becomes highly sensitive to its compute budget, and PPO policies trained on the same scenario distribution win on latency, success rate, and path quality. The authors conclude that uncertainty in how obstacles evolve, more than partial observability, decides which planning paradigm works in practice.

arXiv:2610.12249 PDF

RiCo: Neural Simulation of Rigid-Body Interactions via Local Contact Reasoning

Ruixiang Ouyang, Guanren Qiao, Fansen Meng, Yueci Deng, Ruixing Jin, Kui Jia et al. cross-listed Learned physics simulators struggle to capture how local contacts between surfaces shape the motion of rigid objects. RiCo (Rigid-body Contact Reasoning) represents interactions through sparse neighborhoods of contact surface points: it combines each point's state with the geometry, motion, and physical properties of nearby surfaces, then reasons across each object's points to work out their combined effect on its motion. On the MOVi benchmark it cuts 100-frame position errors by 31-35% and orientation errors by about 38% compared with baselines, with high contact fidelity. It also generalizes zero-shot from small training scenes to scenes with 270 objects, and real-world multi-ball collision tests give early evidence of sim-to-real transfer.

arXiv:2610.12333 PDF

ARC: A Reasoning Recipe for Robot Foundation Models

Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li, Fabio Ramos et al. cross-listed Robot foundation models (RFMs) are usually improved with bigger models, more demonstrations, and costly training. ARC instead adds a reasoning recipe: reasoning traces grounded in the robot's next action that explain why it fits and what it should cause, generated automatically from existing demonstrations as ARC-Trace-DROID. The recipe includes fine-tuning and inference adapted to models such as π0.5 and Cosmos3-Nano-Policy. Without new robot data, the adapted models set a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 points on RoboLab-Reasoning-50, and raise π0.5's real-robot task success by 82.2 percentage points.

arXiv:2610.12386 PDF

LeWAM: A JEPA World Action Model with Diffusion-Steering-Based MPC

Shashank Hegde, Alexander Popov, Elie Aljalbout, Nikolai Smolyanskiy cross-listed World action models (WAMs) predict both actions and future observations but usually rely on reconstruction-based representations that carry noisy, redundant detail. LeWAM is a bidirectional transformer trained end-to-end on a decoder-free JEPA latent space in four modes: forward dynamics, backward dynamics, inverse dynamics, and policy prediction. Linear probes recover robot and object state from its latent better than from a forward-only LeWM, the latent ignores visual distractors far better than a reconstruction-based WAM, and closed-loop control matches a same-size flow-matching policy while also providing a world model. The authors also find that planning in the policy head's noise space, rather than sampling raw actions, stops model-predictive control from exploiting dynamics-model errors and improves closed-loop performance.

arXiv:2610.12407 PDF

RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments

Zimo Wen, Yijin Chen, Yuxuan Cao, Wendi Chen, Yanwen Zou, Wenye Yu et al. cross-listed Robots that act through code can fix programs from execution feedback, but it is hard to attribute each fix to the capability responsible and to validate it before reuse. RoboRSI uses Top-Down Skill Refinement (TSR), which breaks tasks into compound, atomic, and base skills with explicit input-output contracts, assigns each outcome to the responsible branch, and limits revisions to that branch. Manager, Planner, Engineer, and Reviewer agents coordinate planning, execution, diagnosis, and validated release of new skills under human steering, and stable skill sequences are merged into reusable compound skills. On a real mobile manipulator it learns multi-object household cleanup over 104 rounds, and in simulation it achieves the highest success rate on LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin, beating the strongest baseline by 2.7 to 11.0 percentage points.

arXiv:2610.12424 PDF

VioLA: Learning Generalist Humanoid Control Policies from Human Data

Mert Albaba, Jens Bei{\ss}wenger, Anna Manasyan, Daniel Marta, Michael J. Black, Wieland Brendel et al. cross-listed Generalist humanoid policies struggle because whole-body joint control is high-dimensional and tightly coupled and humanoid demonstrations are scarce, so current models need task-specific fine-tuning on teleoperated data. VioLA predicts body and hand motion latents instead of joint commands; pretrained controllers execute these latents, and shared motion encoders map both human and robot motion into the same latent spaces, so human recordings can be labeled directly in the policy's action space. Training on 140.6 million frames, 93.2% of them human, it follows locomotion instructions on a real robot zero-shot with 100% success, versus 16.7% for GR00T N1.7 and 0% for Ψ0, and reaches 88.6% manipulation success without task-specific fine-tuning. The approach works across two vision-language-action backbones and one world-action model backbone, and a policy trained on human demonstrations alone still performs real-robot locomotion zero-shot.

arXiv:2610.12435 PDF

A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control

Octi Zhang, Mateo Guaman Castro, Patrick Yin, Ignacio Dagnino, Abhishek Gupta, Rosario Scalise et al. cross-listed Combining diverse simulator resets with massively parallel simulation reduces how much reward shaping and how many demonstrations robot reinforcement learning (RL) needs. However, sampling resets uniformly wastes much of each batch on configurations the policy has already mastered or cannot yet attempt. Success Guided Sampling (SGS) is a simple adaptive sampler that focuses training on configurations near the edge of the policy's current ability, so adding more parallel environments actually improves learning. With up to 2^20 (over one million) parallel environments, SGS solves multi-terrain quadruped locomotion and contact-rich assembly tasks that prior methods fail on. The resulting manipulation policies, distilled into RGB-based policies, transfer zero-shot to real hardware.

arXiv:2610.12465 PDF

18 more specialized papers

Vision 35

SpectralCache: Accelerating Diffusion-Based World Models via Spectral Feature Caching

Zhendong Mi, Pu Zhao, Ziyu Hu, Xiaodong Yu, Yanzhi Wang, Grace Li Zhang et al. cross-listed Diffusion-based world models can generate interactive environments, but they are slow at inference because the Transformer runs again at every denoising step. The authors observe that world-model features keep very stable singular subspaces across nearby denoising steps, and that their singular values change in predictable ways. SpectralCache is a training-free method that reuses those subspaces, predicts only the low-dimensional singular values by linear extrapolation, and skips some full backbone evaluations by rescaling singular values. On HunyuanWorld-Voyager-13B it reaches 5.22x acceleration while keeping a WorldScore of 65.90 for static scenes, ahead of existing training-free caching methods.

arXiv:2610.02660 PDF

From Pixels, Without Pre-training: Joint Generative and Self-Supervised Representation Learning in One Model

Vicente Balmaseda, Ching-Long Lin, Tianbao Yang cross-listed Strong image generators usually depend on class labels, frozen pretrained encoders, or separately trained autoencoders. SCION (Self-conditioned Generation on Self-supervised representation) instead trains contrastive representation learning and flow-matching generation together in one pixel-space encoder, with no labels or pretrained models. The encoder's [CLS] token serves as the semantic representation, and the model conditions generation on that representation. A jointly learned prior over the embedding allows sampling without a reference image, and gradient-norm balancing plus stop-gradient tricks make single-run training stable. On ImageNet 256x256 it reaches 8.92 FID with the JiT-B recipe and no guidance, beating iREPA and RCG, which both rely on pretrained DINOv2; with JiT-L it reaches 3.47 FID with representation guidance.

arXiv:2610.05711 PDF

GHARP: Real-time Gaussian Head Animation from Large-scale Reconstruction Prior

Ali Benlalah, Sepehr Johari, Patricia Vitoria, Armin Kappeler, Artem Sevastopolsky, Alexander Jung et al. cross-listed GHARP animates a 3D head in real time from a few photos of a person plus a driving expression signal. An expensive offline stage builds a fixed-size canonical Gaussian representation of the person, using a pretrained reconstruction model's structured latent space. A lightweight runtime network suitable for mobile devices then predicts expression-dependent residuals on top of it. A body alignment network removes blur and flicker caused by expression codes not describing body pose or clothing. The method reaches state-of-the-art quality on Ava-256 while running up to 13x faster on an A100 with 8x fewer Gaussians.

arXiv:2610.10945 PDF

Multi-Bandwidth Distribution Matching Distillation: On the Equivalence of Distribution Matching Distillation and Drifting Models

Jialin Zhu, Xing Liu, Feixiang He, He Wang Drifting Models are a recent approach to one-step generation, and their training objective looks almost identical to Distribution Matching Distillation (DMD/DMD2), but no exact correspondence between the two had been established. The authors prove that if the velocity or noise field of a pretrained diffusion or flow model is converted into the attraction force of a Drifting Model, and the repulsion force is estimated from the generated distribution, then training a Drifting Model is equivalent to DMD. Building on this equivalence, they propose Multi-Bandwidth Distribution Matching Distillation (MBDMD), a variant of DMD that borrows ideas from the drifting perspective.

arXiv:2610.10989 PDF

Deflating the Hessian: Rank-4 W4A4 Quantization for Multimodal Diffusion Transformers

Shiwen Wang, Pengxiang Zhao, Xiaoming Yuan Quantizing both weights and activations of diffusion transformers to 4 bits (W4A4) after training loses quality. Existing fixes add a high-precision low-rank branch but either optimize it separately from the quantized residual, which demands higher ranks, or ignore activation quantization error. The proposed method treats the low-rank branch and residual quantization as one joint calibration problem, which yields a deflated Hessian that discounts errors the low-rank branch already captures, plus a surrogate term that suppresses activation quantization noise. Across five diffusion backbones, the rank-4 version beats rank-4 SVDQuant, and on SANA-1.6B, FLUX.1-schnell, and FLUX.1-dev it surpasses rank-32 SVDQuant with an 8x smaller rank and up to 6.25x faster quantization. On the Qwen3-8B LLM it raises MMLU accuracy from 61.50% to 68.17% over rank-32 SVDQuant.

arXiv:2610.11315 PDF

Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder

Zidan Wang, Yaqian Li, Xiaokai Zhang, Kaiwen Long, Kun He, Hanpeng Liu cross-listed Earlier work argued that the standard contrastive loss is unsuitable for post-training CLIP, because it causes catastrophic forgetting when batches are small. The authors show that for the InfoNCE objective, the forgetting comes mainly from a contrastive temperature that is set too high, not from having too few negatives; with a small enough temperature, contrastive post-training improves the pretrained model. They build on this with ComCLIP, a single-epoch recipe that freezes the text encoder. It trains the vision encoder with a properly tempered contrastive loss, an MSE anchoring loss against the original CLIP, and relational distillation from DINOv2. It matches the CLIP-Refine baseline on zero-shot classification and raises linear-probing accuracy on ViT-B/16 from 42.28 to 48.99, and it works as a drop-in vision encoder for LLaVA-1.5-7B with no net change across eight benchmarks.

arXiv:2610.11374 PDF

Conditional Residual Prediction: Improving Autoregressive Video Diffusion without a Bidirectional Teacher

Bowen Zheng, Zhiguang Liu, Jiarong Ou, Rui Chen, Tianyang Hu cross-listed Causal video diffusion models generate video autoregressively, which suits streaming and long videos, but they usually lag behind bidirectional models of the same size and are often improved by distilling from a bidirectional teacher. The authors train a causal model starting from an image model, with no bidirectional video model at any stage. They find that training on ground-truth history makes the model over-reliant on it, so errors in its own generated frames propagate. Conditional Residual Prediction (CRP) has the model first predict each chunk without the history and lets the history add only a residual on top. In controlled experiments it nearly closes a 6.14-point gap to a bidirectional model, and the resulting 2B-parameter model Optica scores 82.78 on VBench after training on about 15M videos.

arXiv:2610.11479 PDF

HAND: A Biologically-Inspired Activation Function that Improves Generalisation and Sample Efficiency in Image Classification

Michael W. Spratling, Heiko H. Sch\"utt cross-listed Deep networks need far more training data than humans and generalize less robustly, a gap the authors try to narrow with a biologically inspired inductive bias. HAND (Homeostasis, Accelerating Nonlinearity, and Divisive-normalisation) is an activation function that adds this bias to convolutional networks. With HAND, a ConvNeXt-tiny matched in 25 epochs the ImageNet-1k accuracy that the unmodified model reached after 200. The gains were larger and did not shrink with more training on the reduced, imbalanced long-tailed ImageNet, and robustness to corruptions and rejection of unknown classes were unaffected or better.

arXiv:2610.11534 PDF

HI3D 3.0 (Twinkle3D): Object-specific 3D Asset Generation with High Resolution

Ziying Li, Shengchu Zhao, Huiang He, Yiyang Chen, Jianwen Huang, Bailin Li et al. cross-listed Image-to-3D systems often distort or lose object-specific details such as inscriptions, brand marks, and repeated structures. Hi3D 3.0 uses Twinkle3D, a geometry model that generates watertight triangle meshes at 2048³ resolution by fixing the surface-quality and watertightness problems of O-Voxel/FaithC representations. A redesigned diffusion transformer (DiT) scales to sequences of 300K geometric tokens, with training time per step cut from about ten minutes to ten seconds, and the single-stage model beats earlier two-stage refinement pipelines. Against four commercial systems it leads on every reported metric, recovering 82.1% of inscribed characters at 98.2% precision, compared with 21.7% recall for the strongest competitor.

arXiv:2610.11685 PDF

MultiWorldBench: Do Independently Controlled Views Describe One Shared World?

Zhangbo Xu, Ruoxi Zhang, Rui Hu, Yisong Wang Multiplayer world models must keep views that players control independently consistent with one shared, persistent world. MultiWorldBench is a diagnostic Minecraft benchmark of 495 case configurations covering ten capabilities, including cross-view motion, shared-state synchronization, persistence, and delayed revisits. On a ten-capability average, Gamma-World scores 21.39, Solaris 20.88, and MineWorld 1.89, against 91.69 for the game engine's ground truth. All generated systems score at most 8.00 on state persistence, and none succeeds at spatial reasoning or at preserving building identity. Human preference rankings closely match the automatic scores, with a mean Spearman correlation of 0.96.

arXiv:2610.11723 PDF

Few-Step Generation via Data-Space Iteration

Shanchuan Lin, Yansong Peng, Fu-Yun Wang, Haoqi Fan Few-step distilled flow-matching generators iterate along the probability flow, which forces a hand-picked timestep schedule that is costly to tune and the same for every sample. Data-space iteration drops that discretization entirely: starting from noise, one shared generator repeatedly refines its prediction directly in data space, with each step trained to produce the best sample it can. The method plugs into distribution matching distillation (DMD) with minimal changes. On class-conditional ImageNet 256x256 it beats standard discretization baselines and matches or exceeds schedule-searched variants without schedule-specific training.

arXiv:2610.12102 PDF

VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling

Leigang Qu, Feng Cheng, Ziyan Yang, Bangbang Yang, Zhaoyang Huang, Wei Chow et al. cross-listed Paired video-editing data (source video, instruction, edited video) is scarce, while image-editing pairs exist by the millions. VINCIE-NExT transfers image-editing ability to video by breaking an edit into a chain of steps (video to image, image to edited image, edited image back to video). An image edit pair, generated by the model or supplied by the user, is prepended as an in-context demonstration, and a new position encoding places that demonstration and the video frames in a shared spatial coordinate system so the edit carries over to every frame. Running the chain as progressive diffusion stages lets the model improve quality with extra compute at test time without retraining, and it reaches state-of-the-art results on OpenVE-Bench.

arXiv:2610.12104 PDF

VibeEdit: Image Editing with Canvas Instructions

Jinjing Zhao, Fangyun Wei, Yitong Wang, Xiuyu Wu, Yunuo Chen, Yang Yue et al. cross-listed In text-guided image editing, describing which object to change is awkward when several objects look alike. VibeEdit replaces the text prompt with a "canvas instruction": spatial marks and optional short notes drawn directly on the image, which together specify where to edit and what to change. It is built by adapting Qwen-Image-Edit with layer-decoupled conditioning, trained on 1.55 million source-target edit pairs using region-weighted supervised fine-tuning, and then refined with rubric-guided reinforcement learning. On a human-curated benchmark of 419 cases that focuses on choosing among similar objects, it scores 79.9 on a VLM rubric with 32.8 dB PSNR outside the edited region, against 67.4 and 24.0 dB for the strongest text-instructed baseline, FireRed.

arXiv:2610.12229 PDF

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

Dahyun Chung, Siyoon Jin, Hyunwook Choi, Honggyu An, Junyoung Seo, Hyunsung Kim et al. cross-listed Egocentric world models predict first-person video conditioned on an agent's actions, but most handle a single agent, and existing multi-agent versions use only coarse actions such as locomotion or camera control. ME-World generates synchronized first-person streams for multiple agents interacting through fine-grained actions. It jointly denoises all streams in one shared token sequence, conditions each stream on every agent's pose, and grounds generation in a shared environment memory. The authors also introduce metrics for environment, state-update, and identity consistency, and report improved shared-world consistency, action control, identity preservation, and video quality over existing methods.

arXiv:2610.12299 PDF

One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts

Adrian Bulat, Yassine Ouali, Georgios Tzimiropoulos cross-listed reViT applies a single Transformer block repeatedly in place of a full-depth vision encoder. To keep depth-specific behavior, the feed-forward network at each step is a convex combination of a small shared bank of experts, with the mixture set by a continuous normalized-depth coordinate. Trained from scratch, reViT-B/16 matches DeiT III accuracy on ImageNet-1k with about 70% fewer stored parameters. An 8-expert model distilled from DINOv2 keeps nearly all of the teacher's linear-probe accuracy and transfers to segmentation and depth prediction, and elastic-depth training lets a single checkpoint run at several depths.

arXiv:2610.12448 PDF

20 more specialized papers

Reinforcement Learning 23

Sample-Efficiency of Kolmogorov-Arnold Networks

Kevin Riehl, Shaimaa K. El-Baklish, Fan Wu, Anastasios Kouvelas Acquiring samples is a central cost in real-world deep reinforcement learning, and the study tests whether Kolmogorov-Arnold Networks (KANs) learn from fewer samples than multi-layer perceptrons. Systematic experiments on the Feynman dataset and the Gymnasium RL benchmark show that KANs reach similar performance with 40% fewer samples, with relative improvements of up to 50% during training. The gains hold under varying levels of reward noise.

arXiv:2610.10627 PDF

World-Model Policy Arbiter for Goal-Conditioned Reinforcement Learning

Junwei Quan, Evgenii Opryshko, Nicholas Rhinehart, Igor Gilitschenski In offline goal-conditioned reinforcement learning (GCRL), no single algorithm wins across environments, goals, or even phases of one task. World-Model Policy Arbiter (WMPA) treats a set of frozen policies as a portfolio and decides at test time which one should act. It imagines each policy's rollout in a learned world model, scores the imagined futures with a shared goal-conditioned value function, and runs the winner for a short commitment interval before choosing again. Without retraining any policy, it raises the macro-average success rate on 18 OGBench datasets from 44% (best single policy per dataset) to 58%, with gains of over 30 points on cube-double-play and scene-play.

arXiv:2610.10932 PDF

Measuring and Mitigating Solution Mode Collapse in RLVR

Liv G. d'Aliberti, Marwa Abdulhai, Sofiia Druchyna, Peter Henderson, Manoel Horta Ribeiro Reinforcement learning with verifiable rewards (RLVR) gives the same reward for every correct answer, regardless of whether the model has produced that solution many times before. Using ModeBench, a benchmark of multi-solution tasks whose verifier reports both correctness and which solution mode was found, the authors show that RLVR concentrates probability onto fewer correct modes even as accuracy holds or improves, and that frontier models are already highly concentrated. Their fix, Re:Max, keeps one verified example per discovered mode in a replay buffer and trains on the stored modes uniformly. Across three model scales and two RL objectives, it improves both success rate and the number of distinct ways the model succeeds.

arXiv:2610.11064 PDF

DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks

Xinfei Wang, Shanchen Pang, Chenhao Zhang, Shudong Wang, Wenhao Ji, Haiyuan Gui et al. Offline multi-task reinforcement learning (Offline MTRL) depends heavily on the quality of pre-collected data. The authors identify three data-level bottlenecks: prompt lengths that do not match task complexity, irrelevant randomly sampled prompt segments, and misleading supervision from fragmented trajectories. DaCe-DT addresses each with length-gated prompt masking, retrieval-augmented prompt construction and value-adaptive return calibration. On Meta-World it improves on state-of-the-art methods by an average of 11.73% on optimal datasets and 13.34% on suboptimal ones.

arXiv:2610.11085 PDF

Balancing Reference Guidance and Free Generation in Trajectory Rollouts for Reasoning RL

Hanyu Wang, Nakul Agarwal, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari, Makoto Fukushima, Jinghui Chen et al. When training reasoning models with reinforcement learning, a verified reference solution can either be used directly or have a prefix of it used to guide the model's own continuation. The question is how much of the reference to supply. The authors derive the KL divergence between each approach and the ideal success-conditioned model distribution in closed form. The result shows that continuation success rate alone does not determine the best prefix length. Based on this, Adaptive Reference Guidance (ARG) learns a shared prefix selector and applies it to groups where every sample failed in Group Relative Policy Optimization (GRPO), achieving the highest aggregate pass@12 on five math benchmarks with Qwen3-4B and Qwen3-8B.

arXiv:2610.11128 PDF

Higher-Order Action Supervision Makes A Strong Policy Class

Peng Cheng, Yunxian Hou, Zhi Zhou, Qian Zhang, Chang Huang, Xianyuan Zhan cross-listed The authors attribute much of the control instability seen in imitation learning and reinforcement learning to supervising only the action labels themselves, ignoring how actions change over time. They propose a loss scheme, with formal guarantees, that supervises both zeroth-order actions and first-order action changes. It works with deterministic, stochastic, or flow policies and requires no architectural changes. As a plug-and-play addition to existing offline RL frameworks, it yields substantial performance and robustness gains across continuous-control tasks in OGBench and D4RL, and it improves out-of-distribution generalization when training data is scarce.

arXiv:2610.11175 PDF

Fed-GRPO: Reward-Signal-Driven Federated Group Relative Policy Optimization

Pengxin Guo, Shuang Zeng, Zonggen Li, Weiying Zheng, Mengting Liu, Liangqiong Qu Group Relative Policy Optimization (GRPO) assumes all training data is in one place, which privacy or regulatory rules can prevent. Fed-GRPO makes GRPO federated by reusing the reward statistics it already computes as free signals. Clients are weighted by reward standard deviation during aggregation, per-prompt objectives are reweighted by the gap between local and global reward, and bandwidth is allocated according to how informative each update is. On mathematical reasoning tasks it beats FedAvg and other federated methods and comes close to centralized training, while cutting communication 32x without loss and supporting up to 621x compression with gradual accuracy decline.

arXiv:2610.11502 PDF

Residual Advantage: Student-Relative Teacher Guidance for RL with Verifiable Rewards

Xiaobing Chen, Zhiqi Pang Reinforcement learning with verifiable rewards (RLVR) gives one outcome label per response, while on-policy distillation gives token-level teacher signals that can be unbounded and can mislead a student whose solution paths differ from the teacher's. Residual Advantage treats the bounded gap between teacher and student probabilities as a one-step reward and subtracts the student's state value to form an advantage. It centers that advantage within each response, so the teacher only redistributes credit among steps and never changes the verifier's outcome label. A variant also updates a teacher LoRA adapter using the verifier signal. With Qwen3-1.7B-Base and Qwen3-4B-Base students and a Qwen3-8B teacher, it improves GRPO and REINFORCE++ in all 24 comparisons on three math benchmarks, raising Pass@8 by 3.9 to 6.3 points.

arXiv:2610.11519 PDF

Constrained Command-Conditioned Reinforcement Learning with Bandit Strategy Selection in Real-Time Strategy Games

Nick Leenders, Roy Lindelauf, Joost van Oijen, Boris Cule Deep reinforcement learning agents for real-time strategy games are strong but can break down against opponents unlike those they trained on. In MicroRTS, the authors separate strategy from execution: a constrained, command-conditioned Proximal Policy Optimization (PPO) executor carries out discrete commands covering economy, army composition, military posture, and worker policy, while a Thompson-sampling bandit chooses commands based on what it observes of the opponent's strategy during the game. Compared with a flat PPO baseline trained with the same architecture, budget, curriculum, and self-play league, this system wins significantly more often against three of the four strongest opponents, raising win rates from 0.55 to 0.97 and from 0.01 to 0.34 against the two strongest held-out opponents.

arXiv:2610.11663 PDF

Can Jev be Your Q or Policy in Reinforcement Learning?

Yi Ma, Tianpei Yang, Yaodong Yang, Weixun Wang, Hongyao Tang Foundation models can give reinforcement learning (RL) useful priors, but token-by-token generation makes querying them slow and sequential. Jev, a recently released decision model, generates no text and instead returns calibrated, typed answers in a single forward pass, and the authors study how it can serve as a frozen component of RL training. They show it can meet the requirements of most RL components except acting as a value function, and use it as a reference policy, an exploration judge, and a replay rater. Across nine MiniGrid tasks and three Atari games, training with Jev beats a standard RL learner, including on tasks where the learner alone makes no progress, while Jev itself stays untrained.

arXiv:2610.11692 PDF

GRPODropout: Less is More for Online Reinforcement Learning Rollouts

Hexuan Deng, Zihao Yan, Xuebo Liu, Shuo Nie, Yue Wang, Chen Wang et al. Addresses policy entropy collapse in GRPO-style reinforcement learning for LLM reasoning, where sampling diversity shrinks and exploration stalls. Instead of changing rewards or adding regularization, GRPODropout drops a small number of high-probability rollouts with positive advantage before each update and recenters the advantages of the rollouts it keeps; a rollout-level theoretical analysis guides how the threshold is chosen. The authors report higher accuracy and higher policy entropy than standard GRPO while updating on fewer rollouts, with negligible added compute.

arXiv:2610.11854 PDF

MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Core Team, Zongming Qiao, Ziyue Hua, Zirui Ou, Zihao Yue, Zihan Jiang et al. The MiMo-V2.6 technical report describes an omni-modal model family improved by scaling reinforcement learning (RL) compute after multimodal mid-training on a hybrid sliding-window-attention architecture. RL is scaled along three axes. Asynchronous training consumes 1,568 samples and 2.7 to 3.7B tokens per step at contexts up to 1M. Environments are more diverse, spanning code, general, visual, and cyber tasks across several agent harnesses. Groupwise agentic grading gives more accurate rewards on long-horizon tasks and pushes the model toward shorter solutions. Training is stabilized by freezing the mixture-of-experts router and layering defenses against reward hacking, and the authors open-source the training dynamics, RL environments, and the RL framework.

arXiv:2610.11959 PDF

CausalDreamer: Learning Predictive World Models with Latent Disentanglement

Prince Jha, Nils Lukas, Kun Zhang, Salem Lahlou Generative world models such as Dreamer 4 use a video tokenizer trained only to reconstruct frames, so its latents do not separate what the agent controls from what matters for reward. CausalDreamer freezes the tokenizer and re-encodes its latent into four groups along two axes. Only the controllable groups receive the action, and the reward-relevant groups are trained to predict reward. The pretrained dynamics model is then fine-tuned on this factored representation. With model-predictive planning on 20 MMBench2 tasks, it scores 14% higher than the base world model on seen tasks and 25% higher on variants with a changed background, object, or maze layout. Neither model beats a random policy by a meaningful margin in entirely new environments.

arXiv:2610.12016 PDF

A Geometric Approach to Soft Actor-Critic with Zonotopes for Locomotion Learning

Panagiotis Roditis, Panagiotis P. Filntisis, Petros Maragos Off-policy actor-critic methods control value overestimation by taking the minimum of two critics everywhere, however the critics disagree. GeZo-SAC has each critic also predict a zonotope, a geometric set whose width serves as a state- and action-dependent pessimistic offset and as a measure of how much the critics disagree. The method blends the critics from a weighted average toward the minimum as disagreement rises, and the deployed policy is an unmodified SAC actor. Across four MuJoCo-v5 locomotion tasks it gets the best mean return on Ant-v5 and Hopper-v5, stays competitive elsewhere, and shows the lowest actuator work per metre with near-zero measured overestimation.

arXiv:2610.12113 PDF

Q-Shaped Options for Hierarchical Reinforcement Learning

Clarisse Wibault, Antoine Gorceix, Antonio L\'eon Villares, Alexey Zakharov, Evangelos Chatzaroulas, Michael Matthews et al. Hierarchical reinforcement learning (HRL) can shorten decision horizons and allow coarser state abstractions, but only if it learns the right action abstraction, and current methods either discard distinctions needed for optimal control or keep unnecessary ones. The authors set out three requirements for an action abstraction and introduce Q-Shaped Options (QSO), which gives each level of the hierarchy its own value functions and policies. The option space between levels is learned as a shared encoder shaped by two Q functions: the low-level Q function treats the option as a goal, which preserves the distinctions control needs, and the high-level Q function treats it as an action, which pushes out unnecessary ones. On offline goal-conditioned locomotion and manipulation tasks, QSO learns semantically meaningful option spaces, outperforms baselines, and achieves non-zero performance on tasks where every other evaluated algorithm fails.

arXiv:2610.12135 PDF

When KL Regularization Misfires in Group Policy Optimization

Fei Ding Removing the reference-policy KL-divergence penalty sometimes improves group policy optimization methods used to train language models, and the authors ask how reference-policy information should enter group-relative updates. They analyze seven failure modes in how the KL term interacts with rewards, including residual KL updates after reward clipping or in groups with identical rewards, KL growth with response length, KL concentrated on a few tokens, and sampling noise from the k1 estimator. Their method, Zero-Sum Calibrated Policy Optimization (ZCPO), uses relative drift measured by conditional KL to calibrate reward coefficients within each group. Experiments and ablations on mathematical reasoning support the design within the authors' settings.

arXiv:2610.12161 PDF

A Unified Bellman Operator for Safety-Critical Reinforcement Learning

Nishanth Arun Rao, Royina Karegoudra Jayanth, Benjamin Eysenbach, Jaime Fern\'andez Fisac Safe reinforcement learning methods tend to either need prior knowledge for strict guarantees, as safety filters do, or learn safety jointly with the task but satisfy constraints only on average. The authors propose a Bellman operator that combines task performance and safety into one joint value function, learned with two-timescale temporal difference learning: the safety value is estimated on the fast timescale and the joint value on the slow one. They prove convergence by treating the limiting dynamics as an occupation-averaged differential inclusion, and show that the converged policy maximizes task return while staying safe at all times. Experiments on continuous control tasks with neural function approximation show stable convergence and near-zero safety violations at test time.

arXiv:2610.12420 PDF

FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems

Songyuan Zhang, Baljeet Singh, Sarthak Ranjeet Kaingade, Chuchu Fan, Bryan Trinh cross-listed Safe reinforcement learning often puts safety and task reward in the same objective, where they can conflict, while classical safety filters require analytic safety functions and dynamics models and become undefined when no safe action exists. FAITH is a model-free framework that learns an approximation of the optimal state-action safety value and amortizes minimal-intervention filtering in a feedforward network. The task policy optimizes return through the filtered dynamics, and when no action meets the learned safety condition, the filter picks the action with the lowest predicted peak harm. On a 29-degree-of-freedom humanoid it reaches a 99.95% safety rate while keeping 97% of unfiltered return in a walking-and-avoidance task, and the policies are also demonstrated on a real Unitree G1 humanoid.

arXiv:2610.12432 PDF

5 more specialized papers

Reasoning 18

Visible Reasoning Is Not a Universal Optimizer: Persona- and Thinking-Dependent Effects in Analytics Code Generation

Bhawani Shankar Leelar, Pawan Chorasiya, Davin Hill, Robert E. Tillman, Tamer Soliman A common assumption holds that visible Chain-of-Thought (CoT) helps code generation, especially when the reasoning matches the target language, as in "think in SQL" or "think in Python". The authors test this on a query-matched SQL/pandas analytics benchmark with execution-based scoring, varying persona phrasing, target language, CoT format, control prefixes, direct generation, and the model's internal-reasoning settings. They find neither a universal accuracy gain from visible CoT nor a consistent benefit from matching reasoning to the target language. The effects depend on persona, target, model configuration, and internal-reasoning setting, and the control ablations separate the effect of reasoning content from that of prompt format.

arXiv:2610.10639 PDF

Learning How to Search for Plans with Exponentially Less Space

Dominik Drexler, Simon St{\aa}hlberg, Markus Fritzsche, Blai Bonet Heuristic search for plans can store exponentially many states, even with a near-perfect heuristic. The authors instead learn one generalized policy per domain, with registers that hold objects and a choose rule that marks the only backtracking points. They prove that structural termination bounds every execution polynomially, so depth-first execution needs only polynomial space, with time exponential only in the number of real choices along an execution. A language model writes these policies in a counterexample-guided loop that certifies termination and verifies the training tasks. The learned policies solve 1,709 of 1,890 tasks from the IPC 2023 Learning Track and Autoscale Agile, more than LAMA, BFWS and Levitron, mostly within one second and 100 MiB.

arXiv:2610.10954 PDF

Invariant-Measure Reasoners: Stable Representations for Latent Reasoning

Yuto Inui, Takuya Konishi, Yoshinobu Kawahara Latent reasoning models apply the same recurrent block to a hidden state again and again, and that state may keep moving within a bounded region instead of settling on a fixed point, which makes predictions unstable across recurrence depths. Invariant-measure reasoners (ImR) instead predict from the expected output of the prediction head under the long-run distribution of latent states, which stays unchanged under further updates. Whether used to fine-tune only the prediction head of existing models or to train from scratch, ImR reduces instability and improves accuracy in many settings on maze and Sudoku tasks. Models trained with ImR can reach high accuracy even when their latent states never converge, a regime in which existing models fail.

arXiv:2610.10996 PDF

Lapras: Latent Reasoning for Time Series Language Models

Yuliang Chen, Yu Yvonne Wu, Patrick Langer, Arvind Pillai, Sudarshan Regmi, Martin Maritsch et al. Time Series Language Models (TSLMs) that reason with text Chain-of-Thought (CoT) often describe the input signal inaccurately, and those early errors carry through to wrong answers with plausible-sounding explanations. Lapras is a post-training framework in which the model reasons through continuous latent thoughts in a joint time series-language space and writes text only for the final answer. It learns this through self-distillation from a teacher trained on explicit CoT traces. Across four TSLM backbones and five time series question-answering benchmarks, Lapras improves average F1 by up to 10.79% over explicit CoT while generating 23.9x fewer tokens, and its latent thoughts can still be decoded into readable reasoning.

arXiv:2610.11111 PDF

OpenProblemBench: Benchmarking AI on Open Problems in the Foundational Theoretical Sciences

Zhiyi Li, Sihan Hu, Tianning Xiao, Xiansheng Cai, Xiaojun Tan, Youjin Deng et al. OpenProblemBench tests whether AI can make progress on 82 unresolved problems from the mathematics and theoretical physics literature. Each problem comes with its research context and prior progress, and problems were chosen so that proposed solutions can be checked relatively clearly. Without reference solutions, four evaluator models judge each submission for correctness, completeness and progress. GPT-6-Astra reaches the highest mean judged solve rate at 14.0%, against 5.5-6.7% for full-size open models and 2.4-3.7% for Flash models. Case studies link the stronger results to reformulating problems and to general arguments that go beyond finite evidence.

arXiv:2610.11118 PDF

SynCo: Data Synthesis Co-Training for Self-Evolving LLMs via Multi-Agent Reinforcement Learning

Wei Yang, Shawn Li, Yuehan Qin, Yawei Wang, Mingxi Wang, Shixuan Li et al. Self-evolving LLMs usually train on static or separately generated datasets, so tasks drift out of step with the model's ability: some become trivial while others stay too hard to learn from. SynCo addresses this with multi-agent reinforcement learning, jointly training two separately parameterized agents. A Synthesizer builds tasks based on the Reasoner's current capability, and a Reasoner learns from those tasks. Each task produces several Reasoner rollouts whose outcomes reward both agents: correctness trains the Reasoner, while task quality, answer reliability, and teachability guide the Synthesizer. Across eight mathematical reasoning benchmarks, SynCo outperforms a broad range of synthetic-data methods, with most of its gains coming from previously unsolved problems.

arXiv:2610.11345 PDF

RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty

Gukhyeon Lee, SangKeun Lee Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in language models but can badly degrade calibration and make models overconfident. Existing calibration-aware training methods help but stay overconfident under distribution shift and cost reasoning performance. RL-ARC jointly uses reasoning confidence and answer confidence: reasoning confidence acts as a regularizer on answer confidence for correct answers and as an overconfidence penalty for incorrect ones. Across in-distribution and out-of-distribution settings, RL-ARC improves calibration without substantially sacrificing reasoning performance, and models learn to adjust their confidence to each question.

arXiv:2610.11352 PDF

From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers

Gerard Grau Garc\'ia, Arnau Padr\'es Masdemont, Niccol\`o Grillo, Jordi Ros-Giralt, Arash Behboodi, Victor Conchello Vendrell Chain-of-thought (CoT) gives language models extra computation but generates it token by token, and most latent-reasoning alternatives still produce their latent vectors left to right. LLoCoT instead refines a compact set of latent slots in parallel by reapplying a shared transformer for a few iterations. A probabilistic head then samples latent tokens that condition an autoregressive answer decoder. Training uses continuous representations derived from explicit CoT, along with answer and latent-likelihood losses. On HumanEval and MBPP it matches the accuracy of an explicit-CoT fine-tuned baseline while cutting time to the first answer token by about 36x and reasoning-phase latency by about 42x.

arXiv:2610.11472 PDF

ReTeach: Building a Self-Teacher through Multi-Round Reflection and Retry

Yafeng Tang, Hao Li, Hongsheng Yu, Qiang Fu Self-distillation improves reasoning without a stronger teacher model, but the self-teacher usually needs extra information such as reference answers to outperform the student. ReTeach builds the self-teacher by starting from a failed student attempt and alternating reflection with new attempts until one succeeds or the retry budget runs out. It uses only self-generated attempts and outcome-level verification. Examples are weighted differently depending on whether they were correct at first, fixed through reflection or left unresolved, and the student learns the teacher's token distributions through on-policy distillation while keeping single-pass inference. Across six benchmarks covering math, science question answering and tool use, it improves average accuracy over GRPO by 1.39 percentage points.

arXiv:2610.11529 PDF

Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residuals

Pengxiang Li, Dilxat Muhtar, Di He, Guinan Su, Lu Yin, Shiwei Liu In looped Transformers, adding iterations can lower reasoning accuracy because noisy updates overwrite correct intermediate results. InfiLoop adds a residual connection designed for loops that combines content-based weighting with learned temporal decay, maintaining a running summary of past states in constant memory. A 7M-parameter model outperforms existing recursive architectures, reaching 97.9% exact accuracy on Sudoku-Extreme and 13.6% pass@2 on ARC-AGI-2. On Sudoku-Extreme, accuracy keeps improving beyond 20,000 effective test-time steps.

arXiv:2610.11570 PDF

Probing for Long-Horizon Deductive Reasoning Capabilities in Language Models with Prolog

Hadeel Al-Negheimish, Jasna Ilieva, Yoon Kim Frontier LLMs accept context windows of a million tokens or more, but it is unclear whether they can do more than retrieve facts from that much text. ProloNg is a synthetic testbed of deductive logic problems written in Prolog that systematically varies reasoning depth, up to a depth of 22 and a context length of 62k tokens. Eight reasoning models from five families of frontier LLMs were tested, and performance falls sharply as depth increases, with most models approaching chance beyond a reasoning depth of 10.

arXiv:2610.11592 PDF

NanoProof: Open and Efficient Automated Theorem Proving in Lean 4

Mat\v{e}j Kripner, Milan Straka NanoProof is a theorem prover for Lean 4 that is fully reproducible from open resources, with its training data, extraction tooling, training pipeline and weights all released. It follows the factorized, execution-guided design and comes with a released dataset of structured proof trees and a tool for programmatic interaction with the Lean verifier, and it was built with compute efficiency as a priority. It reaches 50.8% pass@16 on MiniF2F-Test, beating HyperTree Proof Search and ABEL with roughly 90 times and 7 times less compute, and using over four orders of magnitude less compute than AlphaProof. Stronger open-weight provers exist, but they are fine-tuned from large pretrained LLMs and do not release their data or training pipelines.

arXiv:2610.11605 PDF

LEVER: Adaptive Cost-Aware Proof Search Over AND/OR Graphs

Nihal Jain, Shuangjie Yao, Begum Cicekdag, Zhuo Zhang, Suman Jana LLM theorem provers usually search for any correct proof and only improve it afterward. LEVER instead makes the quality objective programmable and optimizes it during search over an AND/OR proof graph. It scores partial proofs by combining the realized objective so far with predictions for open subgoals, and the same mechanism can target compute cost, proof length, topical impurity, or weighted mixes of these, while the Lean kernel guarantees correctness. On PutnamBench in Lean 4 with matched budgets, it raises the solve rate from 80% to 96% at 34% lower cost than a strong single-conversation agent, and it cuts topical impurity more than post-hoc refactoring does (42% vs. 33%).

arXiv:2610.11862 PDF

Natural Language to First-Order Logic LLM-based Autoformalization

Andrea Brunello, Cristian Curaba, Luca Geatti, Michele Mignani, Angelo Montanari, Nicola Saccomanno LLMs have revived interest in autoformalization, but translating natural language into first-order logic (FOL) still lacks a shared task definition and a systematic survey. The authors define FOL autoformalization as two separate steps, Ontology Extraction and Logical Translation, and show that conflating them makes results hard to compare across studies. They review the existing datasets, evaluation metrics, and LLM-based methods, including fine-tuning, prompting, and verification-based refinement. They close by identifying open challenges in benchmarking, semantic evaluation, ontology-aware methods, and end-to-end applications.

arXiv:2610.12030 PDF

4 more specialized papers