xClean Tools

AI Papers — Daily Top 5

UPDATED 2026-11-03 01:12 PST

2026-10-26 · MONDAY · 13:16 PDT

  1. 1 Reading, Writing, Sharing, and Interacting with the World through Video 211 UPVOTES · SHENGQIONG WU ET AL. · UNIVERSITY OF OXFORD · ARXIV 2610.31668 A survey from the University of Oxford adopts world modeling as a single lens on both video foundation models and video world models. Treating an observed video as a partial, lossy and temporally ordered projection of a latent world, it organizes the literature around four complementary operations: reading, writing, sharing and interacting. It drew 211 upvotes, far ahead of the rest of the day's pool.
  2. 2 Every Weight Is a Path of Shared Discrete Decisions 18 UPVOTES · AMMAR ALI ET AL. · ITMO UNIVERSITY · ARXIV 2610.31931 ITMO University proposes TWIG, or Tree-structured Weights with Inherited Gates, a representation in which every network weight is the endpoint of a path of small discrete decisions: a sign, a shift by a power of two, a zero gate and an index into a shared table. Weights that belong together share the upper part of their paths, so each decision is stored and applied once. Any trained network can be rewritten this way.
  3. 3 Dual Exam: Controlled Evaluation of Multimodal Agents and World Models 16 UPVOTES · ZHEFAN RAO ET AL. · ARXIV 2610.31066 Dual Exam pulls apart the evaluation of multimodal agents and world models, testing agents under fixed environment responses and world models under fixed action interventions across physical, digital, social and scientific domains. The agent track compares 10 configurations on 1,088 items in 15 reporting categories; the world model track holds 400 cases across eight scenarios.
  4. 4 DirCast: Live Esports Directing and Commentary with a Single Streaming Video-Language Model 15 UPVOTES · ZIXIAN GAO ET AL. · ARXIV 2610.31734 DirCast is a single streaming video-language model that both directs and commentates a live esports broadcast, taking every player's raw camera feed and working only from what has already happened. Earlier streaming models learned to commentate, but only over a broadcast a human director had already cut. Each second DirCast scans all cameras at low resolution before deciding what goes on air.
  5. 5 TenetBench: A Clause-Anchored Russian Corpus for Evaluating Language Models against Professional Codes of Ethics 12 UPVOTES · YEGOR ANICHKOV ET AL. · ARXIV 2610.29477 TenetBench tests whether an assistant breaches a specialist's professional code, or helps them breach it, without either party noticing. It pairs 3,433 requests that conflict with clauses of fifteen Russian-language professional codes against benign twins differing only in the legitimacy of what is asked, so each case points at a specific provision rather than a broad harm category.

2026-10-24 · SATURDAY · 15:23 PDT

  1. 1 WeVisOne-VL: Unified Visual Perception from Objects to Pixels 232 UPVOTES · TIANMING LIANG ET AL. · TENCENT · ARXIV 2610.31329 Tencent presented WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception through structured next-token prediction. One model handles detection, referring localization, counting, OCR, segmentation and matting, localizing at whatever granularity a user asks, from points and boxes to masks and alpha mattes. A lightweight pixel decoder attaches to the VLM through a special token.
  2. 2 Editing Images into Reusable Layers via Compositional Edit-Prior Distillation 175 UPVOTES · WEN XUE ET AL. · JINGDONG · ARXIV 2610.24738 LayerEditor couples instruction-guided image editing with the generation of reusable RGBA layers, so an edit returns separable elements rather than a flattened picture. Its In-Context RGB-Alpha Token Modeling jointly represents the edited canvas and an ordered sequence of layers. Existing editors flatten their output, and decomposition methods optimize layer recovery apart from the edit itself.
  3. 3 IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems 86 UPVOTES · A. JIAN YANG ET AL. · IQUEST · ARXIV 2610.30949 IQuest-Q1 is a sparse mixture-of-experts agentic foundation model with 320B total parameters and 15B activated, built for command-line agent systems. The 177-author report describes a three-stage recipe: pre-training with Stable SplitNorm, Muon and muP on a curriculum that raises the coding mix, fine-grained mid-training that extends context, then multi-scaffold post-training.
  4. 4 Toward Recursive Self-Improvement in Physical AI 50 UPVOTES · MENG LUO ET AL. · NATIONAL UNIVERSITY OF SINGAPORE · ARXIV 2610.29677 A position paper from the National University of Singapore asks what recursive self-improvement would require of robots, where feedback is partial, trials are costly and some mistakes cannot be undone. It sets out a ladder of claims from behavioral improvement through procedure improvement to recursion, each defined by the comparison that would establish it. No robot study yet isolates a gain from recursion.
  5. 5 Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models 41 UPVOTES · RUBIN WEI ET AL. · ARXIV 2610.25100 Deep Fusion Memory gives a language model external knowledge without retrieval at inference time. A causal memory predictor learns from retrieved chunk embeddings, cutting retrieval queries 32-fold and retrieved items by 2048 times while building training data, and the predicted memory enters intermediate hidden states through gated cross-attention with the backbone frozen.

2026-10-23 · FRIDAY · 01:08 PDT

  1. 1 Spectral-Spatial Interpretation 127 UPVOTES · HAOTIAN MA ET AL. · ARXIV 2610.21114 Spectral-Spatial Interpretation attributes a model's decision jointly across spatial positions and spectral channels, instead of separating where a model looks from which channels matter. For each position the method fixes the surrounding context, defines a conditional game over that position's channels, and applies the Shapley value. The target is multispectral data, where channel relevance is spatially localized.
  2. 2 Deep Search in Real-World Web Environments 110 UPVOTES · XIAOMIN YU ET AL. · ARXIV 2610.30181 A 25-author team introduces OpenDS, a task for evaluating deep search agents end to end on the live open Web rather than in static retrieval settings. Real web evidence is spread across heterogeneous sources, modalities and formats, often behind interaction mechanisms an agent must work through itself. The benchmark scores multi-step search, access and evidence integration.
  3. 3 Making MLLM Latents Matter: Factor-Aware Predictive Learning for Cinematic Video Generation 102 UPVOTES · SONGLIN YANG ET AL. · TENCENT · ARXIV 2610.30926 Tencent researchers propose Factor-Aware Predictive Learning, which pushes a video diffusion transformer to actually use multimodal LLM representations instead of bypassing them through its native conditioning paths. The method targets cinematic generation, where a model must infer motion from sparse guidance and control character, scene, action and camera as separate factors.
  4. 4 Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds 54 UPVOTES · KUAN ZHANG ET AL. · TSINGHUA SPATIAL INTELLIGENCE LAB · ARXIV 2610.28698 A Tsinghua group ran 120 training configurations to test whether the scaling gains seen in physical embodied learning carry over to virtual worlds with different visual styles, dynamics, controls and objectives. The controlled sweep spans four model sizes from 0.8B to 27B, six data budgets from 50 to 1,600 hours, and five data mixtures starting from Minecraft alone.
  5. 5 Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills 32 UPVOTES · LETIAN ZHU ET AL. · NJU-LINK LAB · ARXIV 2610.17860 Socratic Skill Learning interleaves expert teaching with hands-on practice so an agent's reusable skills hold up beyond the task they were learned on. An Educator guides the investigation, but the Student stays the only operator of the environment and the sole author of each skill, free to question, test and adapt the advice. The authors release SkillTransferBench, with 35 task families and 408 executable cases.

2026-10-22 · THURSDAY · 13:14 PDT

  1. 1 StreamEval: Evaluating Interactive Streaming Video Generation Models 305 UPVOTES · STREAMEVAL TEAM ET AL. · ARXIV 2610.28402 The StreamEval Team released a benchmark for interactive streaming video generation, spanning avatar dialogue and performance, world generation, instruction editing and reference editing across 300 scripted cases of 15 to 300 seconds. Across 37 models it finds that visual quality and identity retention do not track reliable request execution: the best following score among avatar-performance models is 36.90 of 100.
  2. 2 SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes? 128 UPVOTES · ZHAOXI ZHANG ET AL. · PEKING UNIVERSITY · ARXIV 2610.26659 A Peking University team introduced SpatialScout, a benchmark asking whether vision-language model agents can build metric 3D maps by exploring a scene rather than answering questions about it. Agents steer a camera through 80 indoor scenes to place 3,140 objects in a shared coordinate frame. The strongest model, Claude Opus 5.5, reached 41.71 percent mean overlap, with only 3.25 percent of targets above 0.9 IoU.
  3. 3 DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations 124 UPVOTES · MINGZE SUN ET AL. · ROPEDIA · ARXIV 2610.29997 Ropedia presented DONIC, which it calls the first generalist two-handed tracker for dexterous hands, trained jointly across operation categories and human-motion sources. It ships with BiHOI, a multi-view dataset of 1,575 sequences over 229 objects, and pairs policy optimization with behavioral cloning that feeds verified successes back into training. One checkpoint hits 86.79 percent success on unseen trajectories.
  4. 4 Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents 70 UPVOTES · JIAMING ZHANG ET AL. · ARXIV 2610.11942 This paper argues GUI agent safety is a property of the transition, not the screen: a login page after a Sign in tap is expected, the same page after View order is an attack, so a monitor that inspects only screens can be beaten by reusing a legitimate one. Its LGWM predicts the next screen's representation directly, trained on 1.85 million real transitions, reaching 0.987 AUC at 17 ms per decision.
  5. 5 PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis 48 UPVOTES · QIUSHI SUN ET AL. · NLP GROUP OF THE UNIVERSITY OF HONG KONG · ARXIV 2610.20538 A University of Hong Kong NLP group targets multimodal code intelligence, where a sample must be correct as a program, well built, and visually right once rendered - three things no single existing signal covers. Their PrismaForge pipeline synthesizes a decomposed rubric used as both data filter and RL reward, yielding the PrismaCoder-2M dataset and released 9B and 35B models.

2026-10-21 · WEDNESDAY · 13:11 PDT

  1. 1 Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds 176 UPVOTES · JIHAI ZHANG ET AL. · ARXIV 2610.27999 Matrix-Game 4.0 separates game execution from visual generation: a lightweight engine holds world state and interaction rules, a multimodal coding agent turns player ideas into game code, and an autoregressive video renderer converts low-poly proxies into detailed frames. The authors report 480p gameplay at 60 FPS on one RTX 5090 with 112 ms key-press-to-display latency, and release code, weights and assets.
  2. 2 Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience 149 UPVOTES · TIANQI LIU ET AL. · ROPEDIA · ARXIV 2610.26855 Ropedia W0 is presented as the first egocentric loco-manipulation world foundation model spanning humans and robots of differing embodiments, trained on 10,000 hours of egocentric human data and 3,000 hours of robot data with unified annotations. It reports state-of-the-art results on five cross-embodiment benchmarks and, after distillation, supports real-time minute-long rollouts for live VR interaction.
  3. 3 Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition 61 UPVOTES · WENJIE LOU ET AL. · INTERN SCIENCE · ARXIV 2610.28578 Sage-A1.5 is a 35B Mixture-of-Experts agentic model that builds on A1 with three additions: orchestration that delegates work to subagents with isolated contexts, verification through progress checks and evidence validation, and open scientific discovery. The report frames it as a path from long-horizon task execution to exploratory research with inspectable artifacts.
  4. 4 VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse 59 UPVOTES · SICHENG HE ET AL. · ZJU-OMNIAI · ARXIV 2610.25043 Flow-matching vision-language-action models execute only a short prefix of each action chunk and throw away the rest, regenerating from fresh noise. VLA-Rolling reuses that discarded tail, estimates per-action confidence and allocates a denoising budget, leaving the policy frozen. Across four models and four benchmarks it cut inference cost in all 16 pairs, by up to 1.48 times, and raised success rates in 14.
  5. 5 CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow? 59 UPVOTES · JIAQI SU ET AL. · ARXIV 2610.25078 The paper targets the brain-body mismatch in recursive self-improvement, where an evolving harness demands behavior the model never learned. CoEvo-RSI has a meta-agent inherit in both directions from a shared archive, pairing harness designs with model checkpoint selection or merging. Using only supervised fine-tuning on its own rejection-sampled traces, it puts Qwen3.5-9B ahead of DeepSeek-V4-Pro on TraceElephant.

2026-10-20 · TUESDAY · 13:11 PDT

  1. 1 Scaling Parameter-Efficient Fine-Tuning for Brain Encoding 231 UPVOTES · ZHUOFAN JOSH YING ET AL. · MIND LAB · ARXIV 2610.26802 Brain encoding models usually fit a small readout on a frozen pretrained backbone, because neural recordings are too scarce to train or fully fine-tune a large model. The authors show that parameter-efficient fine-tuning adapts very large backbones with few trainable parameters, and exploit the backbone's causal structure to predict several timepoints at once. It tops today's pool with 231 upvotes.
  2. 2 OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners 122 UPVOTES · PEISONG WANG ET AL. · DATA SCIENCE & ARTIFICIAL INTELLIGENCE LAB AT HKUST · ARXIV 2610.24087 Operations research benchmarks usually start from a self-contained problem description, skipping the step where a practitioner turns raw client materials into a model. OR-License instead draws 31 instances from 10 industrial scenarios in paid engagements worth 3.68 million dollars, keeping the original requirements, input data and deliverable formats intact.
  3. 3 MemT-1.0: Towards Memory-Native Evolution 108 UPVOTES · ZHIYU LI ET AL. · MEMTENSOR · ARXIV 2610.26825 MemTensor introduces MemT-1.0, an architecture that makes memory part of the model's computation rather than an external store. Built on Metis, it uses Gated DeltaNet-2 for token-level memory updates within a turn and a Mixture-of-Storage scheme to scale native memory capacity, trained through a two-stage pipeline.
  4. 4 Holi-World: Building Holistic World Intelligence from Human-Ego Experience 104 UPVOTES · HAO LI ET AL. · ROPEDIA · ARXIV 2610.26549 Egocentric recordings capture how the world is arranged and how it changes under interaction, but turning them into training data normally means heavy manual annotation. Holi-World is an ecosystem that links experience capture, world representation, automated data construction, model learning and independent evaluation so the pipeline scales with the footage.
  5. 5 ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation 84 UPVOTES · YIHUI YANG ET AL. · ARXIV 2610.25690 ScaleNav addresses language-conditioned local navigation, where a robot must reach a target named in text and visible in its first view. Real-robot data is expensive and simulation leaves a sim-to-real gap, so the authors learn goal-directed visual dynamics from Internet egocentric video and use those dynamics to generate robot actions.

2026-10-19 · MONDAY · 13:11 PDT

  1. 1 HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents 228 UPVOTES · HONGCHENG GAO ET AL. · HEXAFUTURE · ARXIV 2610.23479 HexaFuture released HexaShell, a thin harness that lets a VLM coding agent drive a two-armed robot through one command line and three calibrated cameras. After each failure the agent reads the log and writes or repairs small triangulation tools, keeping what it learns as code. It reports 39.7 percent success across all 42 RoboDojo tasks, a state of the art, while tight insertion stays at 8.3 percent.
  2. 2 LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets 110 UPVOTES · JUN ZHAO ET AL. · FUDAN UNIVERSITY · ARXIV 2610.09872 Fudan University researchers introduced LiveMACEBench, which evaluates persistent LLM agents in live financial markets by inspecting their decision traces rather than only their returns. Across 30 days of live evaluation, five frontier models showed a pronounced outcome-capability gap: similar returns arose from markedly different use of tool access, memory, rule following and collaboration.
  3. 3 Speculative Sampling in Interactive Video Generation 70 UPVOTES · YUBO HUANG ET AL. · UNIVERSITY OF SCIENCE AND TECHNOLOGY OF CHINA · ARXIV 2610.20556 Researchers proposed Speculative Sampling, which pairs the quality of large bidirectional windows with the responsiveness of small chunks in interactive video generation. It pads the modeling window with noise latents and caches those predictions for reuse while the condition holds, so the uncommitted future stays replaceable. That removes the waiting-time floor set by chunk boundaries.
  4. 4 Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation 49 UPVOTES · XUANYU LIU ET AL. · ZJU-OMNIAI · ARXIV 2610.23121 Scene-Designer is an agentic framework that learns to both build and critique 3D indoor scenes from natural language. The team built two supervision datasets, Scene-OTA-35K for design and error correction and Scene-OTC-15K for critique, then trained with Scene-GRPO, a critic-guided variant that cuts needless trial and error. They report gains in physical validity, functional correctness and layout coherence.
  5. 5 OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools 38 UPVOTES · CHENYANG YAN ET AL. · NANJING UNIVERSITY · ARXIV 2610.14157 OpenMobile-2 is a mobile agent released with its full training environments and recipes. It is trained in MobileGym++, which adds 35 simulated commercial-style apps for a playground spanning over 110, on nearly 12K supervised trajectories and more than 2K verifiable RL tasks, and can interleave GUI actions with over 300 app-native tool calls. It scores 79.9 on AndroidWorld and lifts SPA-Bench from 31.9 to 59.6.