xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-10-17 · SATURDAY · 13:11 PDT

  1. 1 Speculative Forcing for Efficient Autoregressive Video Generation 77 UPVOTES · HYUNWOOK CHOI ET AL. · NAVER AI LAB · ARXIV 2610.22347 NAVER AI Lab proposes Speculative Forcing, which stops denoising a video token once it already matches its full-step output. Intermediate clean estimates serve as drafts, and a verifier - training-free or learned from the generator's features - picks which tokens to freeze early, while their cached keys and values still provide context. On LongLive it matches four-step sampling at about 1.7 steps per token.
  2. 2 Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL 55 UPVOTES · SKYLAR ZHAI ET AL. · MOHAMED BIN ZAYED UNIVERSITY OF ARTIFICIAL INTELLIGENCE · ARXIV 2610.21439 An MBZUAI-led team proposes MRTPO for models trained against several verifiable rewards at once, such as a tool call's correctness and format. Rather than spreading each reward's advantage over the whole response, the policy replays its answer with and without each verifier's information and uses the per-token Jensen-Shannon distance to find the tokens that reward actually judges. No extra models or labels needed.
  3. 3 VLA-Δ: Learning Physical State Transition for Vision-Language-Action Models 54 UPVOTES · QI LU ET AL. · ZJU-OMNIAI · ARXIV 2610.21306 A Zhejiang University team proposes VLA-Delta, which supervises vision-language-action models on the physical change between current and future states rather than only on demonstrated actions. The authors build a large pretraining set annotated with those transitions, covering task semantics and motion. It matches the state of the art on LIBERO and RoboTwin, beats pi-0.5 on a real robot, and adds no inference cost.
  4. 4 CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation 52 UPVOTES · XIAOLI XU ET AL. · ALIBABA · ARXIV 2610.20603 Alibaba describes CommerceGen-Studio, the pipeline it runs in production to generate e-commerce main product images at millions per day with seconds of latency. An agent loop distills SKU metadata into selling points, a layout stage aligns the rewriter with rendered-image quality, and SKU-DiT conditions on the white-background shot, the trademark crop and a structured layout. A live-traffic benchmark is released too.
  5. 5 Mosaic: Multi-Session Agentic Reinforcement Learning 52 UPVOTES · YILONG XU ET AL. · LARK LAB@HKUST (GZ) · ARXIV 2610.22174 A team at HKUST (Guangzhou) formalizes multi-session agentic RL, where context compaction and subagent forking split one trajectory across several sessions and break the single-session assumption that training usually makes. Their method, Mosaic, folds session-usage signals into trajectory-level advantage estimation and shapes each session's advantage by its estimated marginal contribution.

2026-10-16 · FRIDAY · 13:14 PDT

  1. 1 OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use 264 UPVOTES · HANZHANG ZHOU ET AL. · TONGYI-MAI · ARXIV 2610.22200 Tongyi-MAI is opening the whole stack for what it calls a generalist computer-use agent, one that picks for itself how to observe and act across visual and programmatic interfaces. The authors argue that frontier systems such as the computer-use model behind GPT-6 Astra stay closed, which blocks systematic study of how they are designed.
  2. 2 FDE-Gym: The Last Mile of AI Delivery 227 UPVOTES · DONGXU YANG ET AL. · DEEPLETHE · ARXIV 2610.22350 FDE-Gym benchmarks agents on the job of a forward-deployed engineer: 95 end-to-end delivery tasks drawn from 20 customer scenarios across 18 sectors, each adapted by industry experts from real projects. Agents work against scripted stakeholders and modifiable customer systems, so they must draw out vague requirements instead of working a ticket.
  3. 3 RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment 197 UPVOTES · QINGCHUAN MA ET AL. · MAC-AUTOML · ARXIV 2610.17665 RoTA-Bench asks whether vision-language models can turn 2D observations into 3D understanding precise enough to act on. A model sees a rotated object and must choose successive rotations to recover a target silhouette within a limited action budget, with the test objects built automatically from 2D icons.
  4. 4 MJev-Spatial: System One Spatial Intelligence via Native Attention Matching 139 UPVOTES · LUOZHENG QIN ET AL. · MIND LAB · ARXIV 2610.22276 Spatial judgments have compact answers - one of a few options, a scalar, or a 2D point - yet vision-language models pay for autoregressive decoding and answer parsing to produce them. MJev-Spatial instead handles spatial relations, distance and placement in a single classification model, so the answers come at interaction rates.
  5. 5 Learning Where to Split: Dynamic Band Splitting for Audio Source Separation 130 UPVOTES · BINGZHE ZHANG ET AL. · TSINGHUA MACHINE LEARNING GROUP · ARXIV 2610.14993 Band-split separators hold down cost by cutting the spectrum into predefined frequency bands and encoding each as a token, but the same boundaries are used for every mixture. Dynamic Band Splitting learns boundary positions and band widths end to end, so tokens are spent where the spectral detail actually is.

2026-10-15 · THURSDAY · 13:15 PDT

  1. 1 Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation 233 UPVOTES · SHI-YU TIAN ET AL. · JINGDONG · ARXIV 2610.13521 Routing data to specialist teachers is not enough for multi-teacher on-policy distillation, this paper argues, because updates from every domain still land on the student's shared parameters. When teachers differ sharply in style and expertise, supervision that works in isolation can leave persistent gradient conflicts and degrade capabilities during joint training. The authors propose cross-teacher guarding.
  2. 2 Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents 92 UPVOTES · YUYAO GE ET AL. · INSTITUTE OF COMPUTING TECHNOLOGY · ARXIV 2610.20344 Doc2Harness is a two-stage benchmark for building the harness around a coding agent rather than working inside a fixed one. In the first stage a build model runs in a fixed builder harness and gets only an interface contract skeleton plus a development document. The authors note that existing benchmarks measure downstream task success under a harness nobody is asked to design.
  3. 3 WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution 75 UPVOTES · HANG HE ET AL. · NEX AGI · ARXIV 2610.18067 WorkRover trains long-horizon cowork agents by having construction agents refine their own reinforcement learning environments recursively. Its harness-driven synthesis keeps a task, its work context and its verifier aligned as defects are corrected, which the authors say existing construction pipelines support poorly.
  4. 4 ExeWorldBench: Benchmarking Code-Based Executable 3D World Modeling 67 UPVOTES · XINGE PENG ET AL. · UNIVERSITY OF SCIENCE AND TECHNOLOGY OF CHINA · ARXIV 2610.15166 ExeWorldBench tests whether coding agents can build worlds that run, not just worlds that look right. It scores two capabilities together: reasoning about a scene's structural, mechanical and causal principles, and turning that understanding into code whose behaviour emerges from physical interaction. Existing benchmarks, the authors write, cover only fragments such as physical reasoning or visual fidelity.
  5. 5 LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation 64 UPVOTES · ZEFENG HE ET AL. · ARXIV 2610.19975 LatentLoop probes looped mixture-of-experts models, where one shared block is applied twice, and finds the two visits are not equal. Learned residual coefficients come out smaller on the first visit, and fixed-coefficient runs show that shrinking the first visit preserves or improves performance while the same cut to the second hurts it. The paper builds an asymmetric compute budget on that result.

2026-10-14 · WEDNESDAY · 13:11 PDT

  1. 1 Agent-Reflex: Teaching Language Agents to Act Reflectively 192 UPVOTES · GUANTING DONG ET AL. · NLPIR LAB @ RUC · ARXIV 2610.17953 The authors train a language agent to catch its own mistakes instead of pressing on. Agent-Reflex perturbs tool feedback to mine matched success-failure pairs, distills them into reusable reflection skills, then reinforces the behavior on-policy. Built on Qwen3-8B, it reports 58.6 across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench.
  2. 2 VideoJev: System One Model for Video Understanding and Interaction 189 UPVOTES · YUHAN ZHU ET AL. · MULTIMEDIA COMPUTING GROUP-NANJING UNIVERSITY · ARXIV 2610.18615 VideoJev swaps autoregressive answer decoding in multimodal LLMs for a head that scores time intervals and answer options against the video. One interface then covers both temporal grounding and multiple-choice video QA, with visual computation cached across queries. The paper reports mean mIoU gains of 6.7 and 10.8 points over its backbones at 2.8-3.1x faster inference.
  3. 3 GreenWAM: A World Foundation Model for Robotics 80 UPVOTES · GENNADY SIDOROV ET AL. · SBER ROBOTICS CENTER · ARXIV 2610.18774 One model is trained both to predict how an action changes a scene and to choose the action, across 20 robot families mapped into a shared joint-coordinate action space. GreenWAM reports leading trajectory, depth and 3D accuracy on WorldArena 1.0 and places among the top entries on the RoboTwin 2.0 leaderboard, and the team adapts it to a real humanoid.
  4. 4 AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards 77 UPVOTES · XIN GAO ET AL. · MICROSOFT · ARXIV 2610.18860 Code models that emit slides, posters and dashboards cannot see how layout and color land on the canvas. AesCode uses an image generated from the same prompt as aesthetic guidance and scores output through a design graph, pairing deterministic verifiers with a VLM judge under reinforcement learning. Its 8B and 32B models gain 31.1 and 22.4 points of visual quality over Qwen3-VL.
  5. 5 Harness^2: Recursive Agent Harnessing for an Open World 74 UPVOTES · RUIYAO XU ET AL. · GOOGLE · ARXIV 2610.14224 Harness^2 rewrites an agent's execution harness at test time without ground-truth verifiers. Task-level recursion edits the harness for the task at hand by proposing, probing and composing changes, while a second level banks that experience for future tasks. A single step lifts GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench.

2026-10-13 · TUESDAY · 13:16 PDT

  1. 1 Thinking Inertia: LLMs Keep Thinking When Told Not To 214 UPVOTES · DIANQIAO LEI ET AL. · STANFORD UNIVERSITY · ARXIV 2610.11765 Stanford researchers argue that telling a model not to think does not stop it reasoning. They split each response into a pre-answer trace and a final answer, then test six prompting interventions on six models: explicit inference persists under strict controls and grows as the answer space opens. Accuracy holds on Boolean and multiple-choice tasks, while open-ended ones trade compliance against accuracy.
  2. 2 AgentSWE: Can Coding Agents Build the Agent You Actually Want? 201 UPVOTES · JIAHAO WANG ET AL. · ARXIV 2610.13361 AgentSWE benchmarks coding agents on building other agents, across 25 tasks in three stages: creating an agent from a natural-language requirement, editing production agent codebases and optimizing toward a measurable target. Each task is a commission: the builder iterates on a few illustrative cases, while grading runs the frozen submission on held-out ones. No agent tested delivers what the requirement asks.
  3. 3 VIS-Ground: Video Interactive Storytelling with Contextual Grounding 75 UPVOTES · BINGXUAN LI ET AL. · GOOGLE · ARXIV 2610.09326 Google's VIS-Ground addresses interactive video storytelling, where a viewer's mid-generation request depends jointly on the source material and the video already rendered. It abstracts that heterogeneous context into an executable constraint model, then enforces them through planning, verification, revision and rendering. Across three backbones it beats the strongest per-backbone baselines by 10.3 points.
  4. 4 GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed 71 UPVOTES · QIZE YU ET AL. · XPENG AI · ARXIV 2609.39600 XPENG AI's GroundAnything is a 4B-parameter visual grounding model that drops left-to-right decoding for blockwise diffusion, letting spatial hypotheses emerge in parallel and be refined jointly. Its autoregressive variant averages 72.42 percent across 30 grounding benchmarks against 71.35 for GPT-6 Astra, and an optional self-speculative mode runs 4.51 times faster for a 0.74-point drop in COCO F1mIoU.
  5. 5 Multi-Stream Attention Residuals 34 UPVOTES · XIAOHAN QIN ET AL. · SHANGHAI JIAO TONG UNIVERSITY · ARXIV 2610.13630 MSAR extends Attention Residuals to multi-head attention along a network's depth axis, drawing stream diversity from multi-scale causal convolutions and mixing the retrieved streams back into a single-stream layer input. At 18B and 28B MoE scales it beats standard residuals, Hyper-Connections and AttnRes on all 12 downstream benchmarks, with an iso-loss compute speedup of 1.58x against 1.35x for AttnRes.

2026-10-12 · MONDAY · 13:14 PDT

  1. 1 S-Space: Exploring Spatial Workspace in Multimodal Language Models 200 UPVOTES · ZHIYUAN SHI ET AL. · MIRROS · ARXIV 2610.13527 The paper uncovers S-Space, a linear subspace in multimodal language models whose object-token activations encode continuous horizontal, vertical and distance coordinates. It is intrinsic rather than prompt-induced and readable even under non-spatial questions; causal interventions on it alter spatial judgments while sparing other answers. The bottleneck is transforming space across viewpoints, not representing it.
  2. 2 EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents 62 UPVOTES · JIE HUANG ET AL. · PKU-VALUE-LAB · ARXIV 2610.13021 PKU-VaLuE-Lab introduces EmbodiedCLBench, which it calls the first embodied-simulation benchmark for the continual learning of self-evolving agents. Agents learn basic tasks, then must recombine those skills on unseen advanced ones. Across several harnesses and learning methods, experience from basic tasks helps consistently, but the gain does not scale with more samples.
  3. 3 ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents 47 UPVOTES · CHENGYI YANG ET AL. · PROGRAMGYM · ARXIV 2610.13395 ProgramGym builds verifiable training environments for coding agents that write programs from scratch, repairing sandbox builds and black-box test quality in a feedback loop to yield 1,648 environments across C, C++, Go and Rust. Fine-tuning on trajectories from them lifts Qwen3.8-27B from 34.87 to 42.84 percent on ProgramBench, and its role-isolated agent workflow reaches 55.71 and 56.25 percent.
  4. 4 Continually Post-training Small Models via Synthetic Self-distillation 40 UPVOTES · WOOSUNG KOH ET AL. · LG AI RESEARCH · ARXIV 2610.13113 LG AI Research traces the failure of self-distillation fine-tuning on small models to distribution shift, which correlates linearly with weak acquisition and with forgetting. Its SynthSelf outsources the in-distribution rewrite to a slightly larger same-family aligner, adding off-policy supervision and rejection sampling. It beat six methods over four acquisition sets, 14 retention benchmarks and three scales.
  5. 5 Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction 39 UPVOTES · YUQI HUANG ET AL. · NEX AGI · ARXIV 2610.12984 Work2Skill turns recorded human desktop work into reusable agent skills. WorkRecorder aligns screen capture, input events and accessibility-tree data into WorkDemo-350, a synchronized set of 350 screened demonstrations, and an agentic step induces skills carrying workflow rules, step dependencies and completion checks. Four execution models gained 4.67 to 10.54 points on MyPCBench Bounded Action.

2026-10-10 · SATURDAY · 13:13 PDT

  1. 1 OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs 35 UPVOTES · YOU-ZHE XIE ET AL. · ALAYA LAB · ARXIV 2610.12461 OuroWorld is a mask-free framework that turns a static 3D Gaussian Splatting scene into a seamlessly looping 3D cinemagraph. A vision-language model infers the dynamics and steers a video model, whose reference video becomes multi-view supervision, while a Fourier-series deformation field guarantees the loop. Across 39 scenes it won 70.8 to 99.0 percent of user-study comparisons.
  2. 2 Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching 34 UPVOTES · LUPING LIU ET AL. · THE UNIVERSITY OF HONG KONG · ARXIV 2610.12421 FreeMatching drops the smooth-motion and rigid-geometry assumptions of classical dense correspondence, which break down in image editing and reference-guided generation. It pairs generative and semantic foundation representations with supervision from classical datasets, tracked video and synthetic scenes, refined by a teacher without dense labels. It also scores identity preservation in line with human judgment.
  3. 3 DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training 34 UPVOTES · JUNYAN LI ET AL. · ARXIV 2610.12468 DreamTrue is a cross-embodiment robot world model for action-faithful video prediction. It renders action trajectories into image-space conditions with offline calibration, and adds counterfactual post-training that edits recorded trajectories to cover unseen contacts, guided by a reward model trained on human-annotated defects. On AgiBot the human-assessed interaction defect rate fell from 48.12 to 6.25 percent.
  4. 4 MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers 33 UPVOTES · JIARUI CHEN ET AL. · TENCENT HUNYUAN · ARXIV 2610.06801 MC-Sparse traces the quality loss of sparse attention in diffusion transformers to token grouping, inaccurate interaction selection, and discarded attention mass. The training-free method picks individual key-value tokens while grouping similar queries into tile-aligned blocks, caching indices and residuals for reuse across denoising steps. It reports 1.80x denoising speedup on Minimax-H3-Base and 2.32x on 3D assets.
  5. 5 U-Space: Uncovering When and Why Uncertainty Arises in Language Models 33 UPVOTES · TOBIAS BRAUN ET AL. · MULTIMODAL AI LAB - TECHNICAL UNIVERSITY OF DARMSTADT · ARXIV 2610.09087 U-Space is a low-dimensional subspace that makes a language model's evolving uncertainty measurable, built from semantic anchors for doubt and certainty mapped into the residual stream. A U-Lens projection yields a token-level uncertainty map, readable directly or reduced to one score, with no correctness labels, repeated sampling or training. Its confidence score beat established baselines on reasoning benchmarks.