xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-10-01 · THURSDAY · 13:12 PDT

  1. 1 The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation 221 UPVOTES · HAO LI ET AL. · ARXIV 2609.36484 On-policy distillation trains a student to match a teacher's next-token distributions along the student's own trajectories, and generalized variants push past the teacher by extrapolating a reward in output space. The authors argue the language-model head attenuates most of what the teacher's hidden states encode, and work on representation residuals instead.
  2. 2 False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents 172 UPVOTES · MEIJIA CHEN ET AL. · RUTGERS UNIVERSITY · ARXIV 2609.39102 Self-evolving search agents build their own curricula by training a question proposer against a solver, and this paper names a failure mode in that loop: the pair converge on shared errors, so the internal reward rises without real gains. Audits against source evidence show the problem deepening over rounds as pseudo-label correctness stalls or declines.
  3. 3 Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence 148 UPVOTES · XI XIAO ET AL. · AMAZON · ARXIV 2609.34563 Latent visual reasoning lets multimodal models compute in continuous latent tokens rather than spelling out every step, which also makes those tokens hard to supervise. The authors measure a latent evidence-credit gap: the tokens barely react to image perturbations that change the correct answer, which they trace to the absence of explicit visual grounding.
  4. 4 AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks 109 UPVOTES · HONGJIN QIAN ET AL. · BEIJING ACADEMY OF ARTIFICIAL INTELLIGENCE · ARXIV 2609.38288 AREX-2 targets self-improving agents, defined here as the ability to refine a solution repeatedly at test time. The team splits that into reflection, which makes each round better than the last, and long-horizon execution, which keeps many rounds productive, and trains both on synthesized improvement trajectories from machine learning and algorithmic work.
  5. 5 Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents 90 UPVOTES · MINKI KANG ET AL. · NVIDIA · ARXIV 2609.39982 A terminal agent can generate a good command and still execute a bad one, and a single wrong action such as a botched package install can leave the environment worse for every later step. Mid-Harness spends test-time compute at the model-harness boundary, sampling and verifying candidate actions before they reach the shell.

2026-09-30 · WEDNESDAY · 13:10 PDT

  1. 1 In-Context Learning for Robots: Methods and Applications 251 UPVOTES · HAOJIAN HUANG ET AL. · KNOWIN AI · ARXIV 2609.36012 A 39-author survey organizes the literature on in-context learning for robots around the interfaces that connect contextual evidence to physical action, with neural parameters held fixed during deployment. It distinguishes four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. It is the day's most upvoted paper at 251.
  2. 2 Omni-IO Skills: Harnessing Your Agent Omni-Native 136 UPVOTES · YANLIN LI ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2609.31847 A National University of Singapore team presents Omni-IO Skills, a plug-and-play agent harness that makes an existing agent omni-native through hierarchical skills instead of model updates. They argue that adding modalities to a foundation model ties capability growth to costly retraining, while assembling specialist tools leaves procedures, dependencies, intermediate assets and cross-turn revisions uncoordinated.
  3. 3 VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models 114 UPVOTES · YANG XIAO ET AL. · THE UNIVERSITY OF MELBOURNE · ARXIV 2609.32607 VoxMem benchmarks memory in large audio language models, testing what a system recovers from earlier speech beyond the words: who spoke, how it was said and what was audible, none of which survives in a transcript. It also separates memory operations, from retrieving a single fact to integrating evidence across turns and tracking an evolving state, spread across distinct sessions rather than one conversation.
  4. 4 VisionHOPE: Visual Backbones as Self-Modifying Learning Systems 77 UPVOTES · SIRAN PENG ET AL. · MININGLAMP TECHNOLOGY · ARXIV 2609.33325 VisionHOPE is put forward as the first generic visual backbone formulated as a self-modifying learning system. The paper traces the progression from convolutional networks to vision transformers, state-space models and test-time training layers, and notes that visual computation has grown steadily more adaptive to each input while the rules governing that adaptation remain prescribed by the trained backbone.
  5. 5 LEGO-Anything: Coding Agents for 3D Scene Reconstruction 77 UPVOTES · XIRUI LI ET AL. · AMAZON WEB SERVICES · ARXIV 2609.36380 LEGO-Anything reconstructs a 3D scene from a single image as an explicit scene program rather than a rendering or a fixed 3D output, so the result can be inspected, edited and queried. A coding agent iteratively writes and runs Blender code, examines the scene and its renderings, and revises the program. They also introduce LEGO-Bench, a simulator-grounded benchmark of 208 images from 104 indoor and outdoor scenes.

2026-09-29 · TUESDAY · 13:08 PDT

  1. 1 Post-Training Leaves Behavioral Shadows on Unrelated Decisions 249 UPVOTES · ZIYANG ZHANG ET AL. · PEKING UNIVERSITY · ARXIV 2609.29233 The paper reports that capabilities added by post-training leak into a model's answers on task-unrelated prompts. Active Taskless Distillation transfers capability using only a single word from the teacher per prompt, chosen at prompts where teacher and student diverge. Earlier subliminal-learning work needed extensive teacher outputs and targeted traits rather than skills.
  2. 2 Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence 90 UPVOTES · HONGCHENG GAO ET AL. · HEXAFUTURE · ARXIV 2609.35432 Vision-language-action models map observations straight to robot actions, so small changes in layout or viewpoint break the policy. The authors trace that fragility to representation: task requirements, progress and failure recovery stay implicit in action sequences and cannot be inspected or revised. They borrow the coding-agent pattern of explicit state and executable code instead.
  3. 3 Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue 81 UPVOTES · CHENGQIAN MA ET AL. · PEKING UNIVERSITY · ARXIV 2609.31948 Full-duplex speech models can listen while speaking, but existing benchmarks assume one designated user rather than an assistant sitting inside a conversation among several people. Duplex-MPE tests when such an assistant should answer, stay silent or stop speaking, across 2,000 scenarios with three or four human speakers. Each scenario is paired across explicit and implicit forms of addressing.
  4. 4 Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL 69 UPVOTES · JINHAO DONG ET AL. · XIAOMI MIMO · ARXIV 2609.32577 Reinforcement learning for code agents usually rewards a trajectory by whether executable tests pass, so GRPO hands every passing rollout in a group the same advantage regardless of implementation quality. GAGAR redistributes credit within the group according to a quality grade, giving the policy a reason to prefer clean, in-scope implementations over ones carrying unnecessary changes.
  5. 5 MassAlloc Attention: Let Attention Allocate Its Own Compute 64 UPVOTES · JINGZE SHI ET AL. · BEIJING ACADEMY OF ARTIFICIAL INTELLIGENCE · ARXIV 2609.32712 Dense attention kernels run the full post-score path for every QK tile, even where normalized mass is negligible. MALA is a fused primitive that keeps score access to every legal causal interaction but allocates post-score computation by normalized contribution, using the evolving online-softmax normalizer on the forward pass and the finalized one backward. A single tolerance governs training and inference.

2026-09-28 · MONDAY · 13:08 PDT

  1. 1 Disaggregated Quantization: Specializing LLM Prefill and Decode 29 UPVOTES · ANDREI PANFEROV ET AL. · IST AUSTRIA DISTRIBUTED ALGORITHMS AND SYSTEMS LAB · ARXIV 2609.26333 The paper proposes disaggregated quantization, which gives an LLM's prefill and decode phases their own numeric formats, weights and storage placement. On Qwen 3 and Gemma 3, dropping activation quantization only during decode improves accuracy on decode-heavy tasks without raising inference cost, while separate compute-native prefill weights speed up prompt processing.
  2. 2 Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors 10 UPVOTES · ELIZAVETA KOVTUN ET AL. · ARXIV 2609.24385 Tactile-JEPA learns representations from distributed electronic skin, whose sensing elements are sparse and irregularly arranged. Existing self-supervised work on touch targets camera-based sensors, leaving these skins to encoders trained from scratch on raw noisy signals, so the authors build the sensor topology into the learning objective instead.
  3. 3 Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem 8 UPVOTES · GUOMING LING ET AL. · THE CHINESE UNIVERSITY OF HONG KONG · ARXIV 2609.30216 Researchers at the Chinese University of Hong Kong analysed 2,170 public GitHub projects built on the Jev decision model, collected as of September 22, 2026. The study maps how the model is used across domains and how public attention relates to project distribution, and reports rapid early growth in both new projects and integrations into existing repositories.
  4. 4 Enhancing Photogrammetric Digital Surface Models with Pretrained Diffusion Models and Multimodal Conditioning 8 UPVOTES · ANTOINE LORENTZ ET AL. · THALES · ARXIV 2609.31199 Thales researchers use diffusion models to refine digital surface models derived from satellite stereo imagery, which are cheap to produce but carry noise, outliers and voids. Their modified Stable Diffusion 3 conditions on both the photogrammetric model and Pleiades imagery, with a pruned text stream and patch-wise normalization to keep training stable.
  5. 5 TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations 7 UPVOTES · AYUSH JAIN ET AL. · CARNEGIE MELLON UNIVERSITY · ARXIV 2609.30222 TrackEverything is a 3D point tracker from Carnegie Mellon that represents a video as persistent scene tracks in world coordinates. Treating video as a projection of an underlying 3D world ties model complexity to scene geometry rather than clip length, breaking the usual trade-off between tracking few points for long or all points briefly.

2026-09-26 · SATURDAY · 13:09 PDT

  1. 1 Rufus-Air: An Open LLM Post-Training Recipe 12 UPVOTES · CHIA-YUAN CHANG ET AL. · AMAZON · ARXIV 2609.29421 Amazon published a reproducible eight-stage post-training recipe for GLM-4.5-Air-Base (106B-A12B), running from SFT through reasoning, coding, instruction-following and agent RL to RLHF. It documents the data, reward design, infrastructure and stage order, built on open components and public data with no new human annotation. The result improves on the official GLM-4.5-Air post-trained release.
  2. 2 IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis 9 UPVOTES · XINGYU WU ET AL. · REAL LAB · ARXIV 2609.29444 IterSynth splits a deep-search agent into a Planner that identifies information needs and a Synthesizer that folds evidence into an evolving summary, replacing the single ReAct policy. Role-Decoupled Policy Optimization adds turn-level rubrics and role-specific advantages to the outcome reward. IterSynth-8B averages 50.7 on five long-horizon benchmarks, 4.2 percent over the best prior agent at 8B or under.
  3. 3 ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds 9 UPVOTES · MING ZHANG ET AL. · TENCENT HUNYUAN · ARXIV 2609.30199 Tencent Hunyuan's ExplorationBench tests whether AI systems discover rules or merely recall them, using Alien Worlds whose rules are executable, so answers check exactly, and which clash with familiar knowledge. AlienCode and AlienLogic hold 55 discovery targets and 140 tasks. Across 10 systems the strongest acquire unfamiliar rules, but scores swing by trajectory and further exploration can stall or reverse gains.
  4. 4 Coding Agents for Generalized Task and Motion Planning Problems 9 UPVOTES · MATTEO MERLER ET AL. · FONDAZIONE BRUNO KESSLER - NLP UNIT · ARXIV 2609.30233 Coding agents given only a task description and simulator access were asked to synthesize programs that generalize across task-and-motion-planning instances, work that normally needs hand-built TAMP engineering. Over 980 programs and 98,000 evaluation episodes, all three agent configurations beat hand-engineered planners, at 56 to 95 percent mean success against 47 percent, on an order of magnitude less compute.
  5. 5 Learning to Discover Interesting Mathematics 8 UPVOTES · NIKET PATEL ET AL. · NEW YORK UNIVERSITY · ARXIV 2609.28603 The paper defines a theorem's intrinsic interestingness as the ratio of proof length to statement length, and shows the measure correlates strongly with an extrinsic measure of downstream usefulness. A 27B model trained to predict proof difficulty beats frontier general-purpose models, and optimizing for the metric cut substantial or full overlap with Mathlib from 91.9 to 30.6 percent.

2026-09-25 · FRIDAY · 13:11 PDT

  1. 1 Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs 53 UPVOTES · PAVEL TIKHONOV ET AL. · ARXIV 2609.29845 When two separate text streams are linearly combined at the input, a language model's output approximates a superposition of the two next-token distributions, a behaviour the authors name the Superposition Linearity Hypothesis. They argue it is a property of the Transformer architecture rather than something learned, and report that it weakens as pretraining goes on.
  2. 2 HappyWorld-Bench 40 UPVOTES · ZHIQI BAI ET AL. · ALIBABA · ARXIV 2609.24308 A benchmark for world models that scores not only the quality of the world a model generates but whether it stays consistent and responsive while an agent explores, interacts with and edits it. Six world capabilities are arranged in a hierarchy and run across three separate tracks covering video, spatial and embodied world models.
  3. 3 WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation 24 UPVOTES · YUBO ZHU ET AL. · ARXIV 2609.30221 WanPE is a 397-billion-parameter prompt enhancement model trained on 1.05 million real videos to turn a short prompt into a shot-level plan for action, camera movement, lighting and sound. The plans are built by reverse-constructing them from video, and the target is generators that now run to 30 seconds across multiple shots.
  4. 4 Schrödinger's Code Repository: Have LLMs Learned SWE-bench or Memorized It? 16 UPVOTES · SILIN CHEN ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2609.27891 Repository-level coding benchmarks are built on popular open-source projects that models have already trained on, so a high score may reflect memorized cues rather than reasoning. The authors propose SchrodingerRepo, which evaluates coding agents against dynamically instantiated versions of a test repository instead of its one canonical static form.
  5. 5 PACT: From Credit Assignment to Critic Alignment 15 UPVOTES · JIAYAN FU ET AL. · ALLSPARK RESEARCH · ARXIV 2609.26355 The paper gives token-level credit in reinforcement learning for language models a formal definition, proving that three conditions - completeness, prefix consistency and neutrality - determine it uniquely. That characterization is then used to explain behaviour seen in existing post-training algorithms and to derive an improved actor-critic procedure.

2026-09-24 · THURSDAY · 13:10 PDT

  1. 1 SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue 76 UPVOTES · HAOBO ZHENG ET AL. · ZHEJIANG UNIVERSITY · ARXIV 2609.26780 SpeakerMem-R1 proposes a speaker-centered dual-track memory for multi-party dialogue, where a system must track who said what, whom each statement concerns and how those relations change over time. The authors argue existing general-purpose LLM memory systems lose person and group relations and fail to combine clues spread across members and time.
  2. 2 Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World 41 UPVOTES · KAIXIANG YAO ET AL. · ZJU-OMNIAI · ARXIV 2609.23038 Spatial-Interactor trains vision-language models on interaction trajectories rather than static question answering, so they learn the local state transitions caused by object motion and viewpoint change. The authors argue current spatial training supervises object attributes and relations but gives little signal for tracking state over long trajectories.
  3. 3 Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models 26 UPVOTES · XIAOQIANG WANG ET AL. · UNIVERSITE DE MONTREAL · ARXIV 2609.24657 HyperQ attaches token-conditioned quantum residual branches to a frozen masked-diffusion language model. Each branch sits inside a transformer block, reads a token's hidden state, emits the coordinates of that token's quantum circuit, executes it, and adds the measured values back through a residual connection, leaving the backbone unchanged.
  4. 4 JEV-as-a-Judge: Accept When Confident, Escalate When Unsure 25 UPVOTES · YUBO LI ET AL. · CARNEGIE MELLON UNIVERSITY · ARXIV 2609.26550 The paper tests whether a decision-only judge can serve as a cheap first pass for LLM evaluation and flag the cases that need a stronger model. Against sixteen generative and reward-model judges with blinded human adjudication, jev-as-a-judge lands within three percentage points of the strongest comparator on preference and evidence-grounded factuality at 0.36 percent of its fee.
  5. 5 StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training 25 UPVOTES · BAO TANG ET AL. · KOLORS TEAM, KUAISHOU TECHNOLOGY · ARXIV 2609.26774 StableVQ examines why vector-quantized visual tokenizers are unstable to train and traces the problem to entanglement between encoder-decoder and codebook training. Because neither module can fulfill its own role in isolation, the authors argue, the system only works when the two happen to cooperate. The paper sets out practical guidelines for stable training.