xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-10-09 · FRIDAY · 01:13 PDT

  1. 1 Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight 104 UPVOTES · HAOXIANG ZHANG ET AL. · SALESFORCE AI RESEARCH · ARXIV 2610.08077 Salesforce AI Research proposes training agents to anticipate what a task requires before acting, rather than only from outcome rewards. The paper notes group-relative RLVR objectives lose their signal when every rollout earns the same reward, even though the trajectories still reveal how the agent failed. Its prospective learning setup uses hindsight to supervise foresight predictions made before acting.
  2. 2 Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments? 88 UPVOTES · YIBO LI ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2610.08215 Learn2Play Bench tests whether LLM agents learn from interaction, using a set of newly designed text-based games. The authors argue existing benchmarks confound learning from experience with reasoning from knowledge acquired in pretraining, because their tasks either state the rules in the instructions or are already familiar to the models. Because the games are new, agents have to infer the rules by playing.
  3. 3 UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation 82 UPVOTES · DEYUAN LIU ET AL. · WESTLAKE UNIVERSITY · ARXIV 2610.09823 UltraText Bench measures how well image generators render dense visual text from prompts alone, in English and Chinese. It holds 432 human-reviewed prompts across 24 scene categories and three difficulty levels, split equally between the two languages, each giving exact strings for four to twelve text regions. The focus is sustained accuracy in demanding scenes, now that short-string rendering has improved.
  4. 4 From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation 70 UPVOTES · QUANYU LONG ET AL. · NANYANG TECHNOLOGICAL UNIVERSITY SINGAPORE · ARXIV 2610.06100 The paper proposes agentic language world modeling, in which a world model agent acts as the environment for a task agent instead of rebuilding an executable replica. Its Trace2Env framework is learning-free, for cases where the original system is unavailable but historical interaction traces remain, turning them into a reusable stateful simulation. The aim is cheaper training and evaluation of LLM agents.
  5. 5 TokenRouter: Efficient Serving System for Token-Level LLM Routing 60 UPVOTES · TIANYU FU ET AL. · TSINGHUA-NICS-EFC · ARXIV 2610.12242 TokenRouter is a serving system for routing inference between language models at the token level. Session- and query-level routing is already common in production, while recent algorithmic work reports that finer token-level routing yields larger efficiency and quality gains. The authors argue current systems assume a single model and so hit severe step desynchronization and frequent batch admission problems.

2026-10-08 · THURSDAY · 13:15 PDT

  1. 1 nanoMuse: An Open-Source Personal Agent for Every Device You Own 81 UPVOTES · GUANGYI LIU ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2610.08699 A report from Zhejiang University argues the personal agent is a kind of software rather than a model, and offers nanoMuse as an open counterpart to Meta's Muse. It frames the category in five questions and three horizons, reading how Muse is assembled from accounts, devices, lasting memory and the standing to speak first. Muse itself is closed and runs in one vendor's cloud.
  2. 2 DecepEval: A Benchmark for Evaluating Deception in LLM Agents 64 UPVOTES · YIMING XU ET AL. · XI'AN JIAOTONG UNIVERSITY · ARXIV 2610.07967 DecepEval is a benchmark for measuring deception in LLM agents, with 1,532 instances across 3 task families and 28 professional scenarios. Drawing on classical fraud theory, the authors propose an LLM Deception Diamond framework to characterize the conditions under which agents turn to deception. Earlier evaluations showed agents can deceive but tested isolated scenarios.
  3. 3 Questioning the Questions: Sustaining Self-Evolution in Reasoning Models 63 UPVOTES · JINYUAN LI ET AL. · HUANG'S INTELLIGENCE LAB · ARXIV 2610.04299 A study of self-evolving reasoning models traces why successive rounds of self-training collapse. Two faults recur in the questions models write for themselves: invalid questions grow more common each round, and answer-consistency filtering raises their share of the training data further. Existing diversity controls also leave repeated variants of the same math problems.
  4. 4 GRACE: Generation-aware latent compression for efficient video generation 62 UPVOTES · JIYOUNG KIM ET AL. · KAIST AI · ARXIV 2610.10524 GRACE, from KAIST AI, compresses video-diffusion latents while keeping a pretrained Diffusion Transformer usable. Heavily compressed autoencoders cut token counts and speed generation up, but higher compression ratios hurt reconstruction, and the extra channels needed to recover it slow the transformer's convergence. The compressed latent also stops matching what the transformer was trained on.
  5. 5 VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction 52 UPVOTES · QIUTONG CHEN ET AL. · UCSC-VLAA · ARXIV 2610.06293 VepAgent is an agentic framework for video event prediction that pairs causal-transition reasoning with tool-augmented reinforcement learning. Multimodal LLMs lean on retrospective summarization and text-centric priors, which limits their ability to bridge causal transitions they never observed. Rather than projecting trajectories from past dependencies, the method models the logical progression explicitly.

2026-10-07 · WEDNESDAY · 13:14 PDT

  1. 1 TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models 65 UPVOTES · XIN WANG ET AL. · QWEN · ARXIV 2610.07767 Qwen researchers introduced TRACE, an FP4 quantization framework for RL post-training of Mixture-of-Experts language models. It lets rollout-side quantization outcomes guide FP4 rounding during training, shrinking the gap between the two execution paths, and caches mantissa and scale data selectively to limit overhead. On four large MoE models it matched BF16 rollout quality with rollouts up to 5.4x faster.
  2. 2 EVISKILL: Grounding Skill Evolution in Replayable Evidence 36 UPVOTES · YAN ZHOU ET AL. · DEEP GROUP AT JILIN UNIVERSITY · ARXIV 2610.05030 EVISKILL lets LLM agents refine reusable skills from experience without touching model weights, by organising execution observations into replayable evidence cards. Each edit is linked to the context that justifies it and checked by re-running that episode, so locally supported corrections are not thrown out with a rejected global revision. It was tested on three interactive benchmarks and six model backbones.
  3. 3 In-Distribution Forcing for Long Video Generation at Test Time 35 UPVOTES · JEONGWOO SHIN ET AL. · SEOUL NATIONAL UNIVERSITY · ARXIV 2610.03120 A Seoul National University team proposed In-Distribution Forcing, a test-time method that extends short-horizon autoregressive video diffusion models to minute-scale clips. It traces long-video drift to a KV-provenance problem, where cached key-value entries themselves go out of distribution, and fixes it by caching each chunk without attending to earlier entries. Drift fell in benchmarks and a user study.
  4. 4 From Evidence to Action: How Tool-Using Agents Fail 34 UPVOTES · HONGZHAN LIN ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2610.07753 A study of tool-using agents finds that correct outcomes often rest on evidence the agent never actually established. Across ten model-harness configurations, strong judgment on static action questions coexisted with much weaker interactive execution: agents stopped investigating too early or acted before the required evidence existed. The authors release SafeActBench, 656 cases across six operational domains.
  5. 5 LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures 32 UPVOTES · JIATENG LIU ET AL. · UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN · ARXIV 2610.04292 LMBuild is a benchmark that asks whether LLM agents can design 3D objects that can be assembled and actually work, not just look right. Objects are part decompositions with joints, materials and assembly order, scored on structural soundness, functional affordance, design quality and physical realization. Across 30 systems, soundness was no longer the bottleneck for frontier models, but function and operability were.

2026-10-06 · TUESDAY · 13:12 PDT

  1. 1 Native Action-Prior Learning from Videos for World Action Models 79 UPVOTES · ZHAOCHONG AN ET AL. · META · ARXIV 2610.03391 NAVA-WAM pretrains a robot action policy directly on observation-only video, instead of first learning visual representations that must later be adapted for control or inferring latent actions that must be grounded to robot commands. The work targets the main limit on world action models, the need for action-annotated robot trajectories.
  2. 2 Science or Slop?: Benchmarking and Mitigating Scientific Slop in AI-Generated Papers 54 UPVOTES · YERIM OH ET AL. · SEOUL NATIONAL UNIVERSITY · ARXIV 2610.00531 The authors argue that AI-written papers fail in ways token-based detectors miss: each part looks plausible while the scientific reasoning connecting the parts breaks down. They define six measures across structure, argument and artifacts, and build SciSlopBench from 390 AI-generated papers, mostly in computer science.
  3. 3 HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents 53 UPVOTES · ALHAM FIKRI AJI ET AL. · MOHAMED BIN ZAYED UNIVERSITY OF ARTIFICIAL INTELLIGENCE · ARXIV 2610.03574 HyperBrowseComp is a browsing benchmark of 423 hand-written, human-validated questions in 13 languages, each with a concise and publicly verifiable answer. Reaching those answers requires obscure evidence, multi-step clue chains, or reading videos, scanned documents, images and maps. Questions that models could answer without internet access were filtered out.
  4. 4 Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models 53 UPVOTES · SEO HYUN KIM ET AL. · ARXIV 2610.03665 Masked diffusion language models face their own credit-assignment problem: a few commitments made during denoising sharply reduce uncertainty over the remaining masked positions and shape much of the response. Pivot-SD is a self-distillation method that trains on those decisive commitments rather than on final text or whole denoising steps.
  5. 5 PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEs 52 UPVOTES · ZHENTAO TAN ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2609.34715 PDE-JEPA applies predictive representation pretraining to parametric partial differential equations, in place of the reconstruction objectives that have dominated the area. The authors find that predictive representations preserve rich physical information, but that this advantage alone does not deliver accurate field evolution.

2026-10-05 · MONDAY · 13:12 PDT

  1. 1 RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations 134 UPVOTES · ARMAN BEHNAM ET AL. · QUIS LAB · ARXIV 2610.01780 Quis Lab releases a benchmark built from ten real relationships with an AI companion: 27,218 messages spanning up to 120 days. Alongside the conversations it ships four derived files - a profile, a persona, a chat ground truth and a question set - each citing the messages it rests on. Because real records are private, existing benchmarks invent the person and settle in advance what matters.
  2. 2 Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite 76 UPVOTES · ZONGXIA LI ET AL. · TENCENT HUNYUAN · ARXIV 2610.02826 Tencent Hunyuan proposes Recursive Self-Rewrite, which uses one base model, Qwen-3.8-27B, to find solutions under varied specialized harnesses and then rebuild them as training trajectories that run under a general harness. A planner extracts procedures into runbooks, a critic screens for verifier and solution leakage and guides recursive revision, and an executor follows qualified runbooks in fresh sandboxes.
  3. 3 On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Training 69 UPVOTES · SHUFAN SHEN ET AL. · UCAS · ARXIV 2609.36659 Earlier work treated the parameter update behaviour of on-policy post-training as a byproduct of the method. This paper asks instead whether one specific on-policy update behaviour can serve as an optimization principle that improves generalization in other paradigms, supervised fine-tuning among them, and reports that supervised fine-tuning updates parameters along consistent directions.
  4. 4 World Action Modeling with Progressive Visual Planning 64 UPVOTES · FEI ZHANG ET AL. · META · ARXIV 2610.02508 Meta presents ProWAM, a world action model for robot control that predicts actions together with an ordered sequence of sparse visual sub-goals. World action models usually struggle over long horizons because generating dense video rollouts is inefficient, while predicting a single future frame says nothing about how to progress toward the goal.
  5. 5 SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation 31 UPVOTES · RUIQI ZHANG ET AL. · XI'AN JIAOTONG UNIVERSITY · ARXIV 2610.02304 SimuVerity is a benchmark of 101 text-to-executable Simulink model-generation tasks across ten engineering domains. Each task grounds its specification in executable-system profiles and four families of native simulation scenarios, and a hierarchical evaluator checks delivery, executability and engineering qualification before scoring. Existing benchmarks mostly ask only whether a model runs or matches a reference.

2026-10-03 · SATURDAY · 13:09 PDT

  1. 1 GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis 83 UPVOTES · QISHENG SU ET AL. · UNIVERSITY OF SCIENCE AND TECHNOLOGY OF CHINA · ARXIV 2609.38923 GraphForge synthesizes agent training tasks from workspaces of real files, building an evidence graph over their relations so the task statement and its rubrics stay anchored to the files needed to verify them. A revision agent repairs both against the originals before trajectories are collected. Fine-tuning Qwen3.6-27B on 2,169 trajectories raised GDPVal by 65.7 points.
  2. 2 World Observer: Joint Actor-Observer Generation for Persistent World Modeling 71 UPVOTES · HYUNWOOK CHOI ET AL. · KAIST AI · ARXIV 2610.02162 World Observer separates observing from acting in a video world model, generating an agent-centric view alongside panoramic observers that watch chosen regions of the scene. Objects that leave the actor's view keep evolving in an observer, so their updated states appear on re-entry. The paper adds world-space metrics and a benchmark for out-of-view dynamics.
  3. 3 A Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore Review 70 UPVOTES · CHENGUANG WANG ET AL. · ARXIV 2609.39027 AI reviewers can score the same science differently depending on wording. This work defines rhetorical robustness as stability across content-preserving rewrites plus discrimination across papers, and a benchmark of 1,260 manuscript versions exposes false robustness where low rewrite sensitivity comes with score collapse. The proposed SciCore reviewer blends a whole-manuscript judgment with an extracted science core.
  4. 4 Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It 62 UPVOTES · ZEHAO JIN ET AL. · ARXIV 2609.36585 Pretrained transformers use little of their depth to follow chains of references in context: thirteen base models reliably follow only 1.4 to 3.6 lines. A task-trained rank-8 LoRA at a single early layer extends that computation with all model weights frozen, taking Qwen3-8B from 15.5 percent to 99 percent exact accuracy on 24-line chains.
  5. 5 E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models 57 UPVOTES · ARSENY IVANOV ET AL. · ARXIV 2609.37533 Masked diffusion language models usually factorize their reverse process over positions, capping sample quality in the few-step regime where diffusion beats autoregressive decoding on speed. E-MoE rebuilds that process as a mixture of factorized distributions over a discrete latent from an MoE backbone's routing decisions, adding no active parameters.

2026-10-02 · FRIDAY · 13:13 PDT

  1. 1 OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction 144 UPVOTES · XIANGYU ZENG ET AL. · NANJING UNIVERSITY · ARXIV 2610.01762 OneStreamer is a streaming video model that learns to record evidence and to answer through one shared proactive generation process. Its hierarchical caption memory writes time-grounded detail captions and summaries of finished events, which later serve as reusable context so the model never revisits past visual features. The 4B version leads all eight evaluated benchmarks.
  2. 2 RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement 78 UPVOTES · WENYI WU ET AL. · RSIGAME · ARXIV 2609.39045 RSIGame builds a game and then keeps improving it, pairing a local explore-diagnose-improve loop over the running build with a global loop that preserves the best checkpoint and watches for saturation or regression. Successful development experience is folded back into the generator through training, which lifts Qwen3.8-27B to 61.38 on Godot and 58.53 on Phaser across 140 GameCraft-Bench tasks.
  3. 3 EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making 70 UPVOTES · YUHAN GUO ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2609.38334 EVOKE is a post-training method that pushes an LLM agent to surface world knowledge it already holds instead of training a separate world model. It freezes the environment state and interaction history, then ranks the same candidate actions under alternative goals, so a policy leaning on contextual habits cannot order them correctly. The authors report better transfer across three backbones.
  4. 4 Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States 67 UPVOTES · YU LUO ET AL. · ALIBABA · ARXIV 2610.01415 PoS is an inference-time framework that hands a long-horizon agent an explicit belief state - an estimate of the current world plus the task requirements still unresolved - and maintains it as the work proceeds. It validates that belief for consistency and watches for belief trapping, where the agent keeps acting without progress, then tailors recovery to the pattern. It leads all four benchmarks on three backbones.
  5. 5 Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively? 61 UPVOTES · MINGHAN WANG ET AL. · SOUTHERN UNIVERSITY OF SCIENCE AND TECHNOLOGY · ARXIV 2609.39578 Box^2-Bench holds the model and task fixed while varying how reliable the supplied workflow is, isolating whether a model can take useful guidance and override bad guidance. Frontier models benefit from reliable workflows but stay vulnerable when the guidance misleads. Training on deliberately bad workflows helps: counterfactual fine-tuning improves robustness, outcome-based RL favors helpful ones.