本周最显眼的现象是"决策模型"(System One / Jev 系列) 突然集中爆发——Jeva、VideoJev、MJev-Spatial、StartLux-Decision、DecisionRobust 等多篇论文几乎同时在做同一件事:把本该由大模型自回归生成的选择题,改造成一次前向打分,用来换取数量级的推理加速。与之呼应的是"在线策略蒸馏"(on-policy distillation) 这条线也在快速分化出新变体——Waypoint-OPD、CriticOPD、ActionOPD、VRSD 分别从早晚策略一致性、教师可恢复性、动作级对齐、熵保持四个角度修补同一个基础配方。Agent 方向的重心则从"能不能完成任务"转向"脚手架本身该不该自我演化",Self-Evolving Agent Harness、Harness²、WorkRover 都在给智能体的执行框架装上反馈闭环;FDE-Gym、AgentSWE 两篇基准论文则不约而同地给这股热情浇了冷水——即便功能通过率很高,真正可靠交付的比例仍然很低。机器人方向,VLA 与世界动作模型(WAM)的融合继续加深,GreenWAM、LeWAM、WAVE-0 都在探索"一个模型同时做预测与控制"。相比之下,智能眼镜/可穿戴方向本周明显安静,没有出现热度和技术突破都突出的工作。
预训练、后训练、推理与长上下文
Thinking Inertia: LLMs Keep Thinking When Told Not To Stanford University | 🔥 HF 250 赞 | 💻 代码 | arXiv:2610.11765
Vela: User Behavior as Supervision Sparks LLM Intelligence AllSpark Research | 🔥 HF 87 赞 | 💻 代码 | arXiv:2610.14709
Cortex: Consolidating Context into Parametric Memory ByteDance Seed | 🔥 HF 67 赞 | arXiv:2610.20399
UltraBERT: Beyond the Limit of BERT Pre-Training 🔥 HF 72 赞 | arXiv:2610.15028
Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders Nanyang Technological University | 🔥 HF 70 赞 | 💻 代码 | arXiv:2610.17771
LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation 🔥 HF 67 赞 | 💻 代码 | arXiv:2610.19975
Self-Distillation with Verbal Regularization King's College London | 🔥 HF 38 赞 | 💻 代码 | arXiv:2610.16354
DeltaTTG: Training Multi-Turn Agents with a Token-to-Go Progress Meter Nanyang Technological University | 🔥 HF 34 赞 | 💻 代码 | arXiv:2610.17666
Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards RUC-AIBOX | 🔥 HF 10 赞 | 💻 代码 | arXiv:2610.16533
LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators ETH Zurich | 🔥 HF 11 赞 | 💻 代码 | arXiv:2610.17246
VLM、图像与视频生成、文档与视频理解
OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding VCLab | 🔥 HF 578 赞 | 💻 代码 | arXiv:2610.18829
VideoJev: System One Model for Video Understanding and Interaction Multimedia Computing Group-Nanjing University | 🔥 HF 369 赞 | 💻 代码 | arXiv:2610.18615
MJev-Spatial: System One Spatial Intelligence via Native Attention Matching Mind Lab | 🔥 HF 355 赞 | 💻 代码 | arXiv:2610.22276
S-Space: Exploring Spatial Workspace in Multimodal Language Models MirroS | 🔥 HF 298 赞 | 💻 代码 | arXiv:2610.13527
AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards Microsoft | 🔥 HF 137 赞 | 💻 代码 | arXiv:2610.18860
Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation Shanghai AI Laboratory | 🔥 HF 43 赞 | 💻 代码 | arXiv:2610.20818
Multics: Generative Multi-Player World Model with Hierarchical Communication ARC Lab, Tencent | 🔥 HF 74 赞 | 💻 代码 | arXiv:2610.17201
LensErase: A Foundation Model for Erase Anything Microsoft | 🔥 HF 43 赞 | 💻 代码 | arXiv:2610.18072
WorldCast: Exploring Intelligent Visual Generation in World State Prediction LongCat | 🔥 HF 29 赞 | 💻 代码 | arXiv:2610.20731
InfoVis: Unified Slide Image Generation and Editing Beijing Academy of Artificial Intelligence | 🔥 HF 28 赞 | 💻 代码 | arXiv:2610.16784
工具调用、Computer Use、多智能体与编码 Agent
Player-1: Learning to Play from Synthetic Experience ARC Lab, Tencent | 🔥 HF 334 赞 | 💻 代码 | arXiv:2610.20584
OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use Tongyi-MAI | 🔥 HF 318 赞 | 💻 代码 | arXiv:2610.22200
FDE-Gym: The Last Mile of AI Delivery DeepLethe | 🔥 HF 284 赞 | 💻 代码 | arXiv:2610.22350
AgentSWE: Can Coding Agents Build the Agent You Actually Want? 🔥 HF 234 赞 | 💻 代码 | arXiv:2610.13361
Agent-Reflex: Teaching Language Agents to Act Reflectively NLPIR Lab @ RUC | 🔥 HF 213 赞 | 💻 代码 | arXiv:2610.17953
Self-Evolving Agent Harness ZGCAGI | 🔥 HF 203 赞 | 💻 代码 | arXiv:2610.18813
MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments Tongyi-MAI | 🔥 HF 156 赞 | 💻 代码 | arXiv:2610.20939
WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution Nex AGI | 🔥 HF 85 赞 | 💻 代码 | arXiv:2610.18067
Last Secure Code Benchmark: Is the Code You Vibe Secure? Qwen | 🔥 HF 74 赞 | 💻 代码 | arXiv:2610.14766
DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding ByteDance | 🔥 HF 36 赞 | 💻 代码 | arXiv:2610.19849
架构、优化、表征学习、语音与可解释性
Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation 🔥 HF 669 赞 | 💻 代码 | arXiv:2610.21601
Jeva: System One Models from First Principles National University of Singapore | 🔥 HF 419 赞 | 💻 代码 | arXiv:2610.18773
Learning Where to Split: Dynamic Band Splitting for Audio Source Separation Tsinghua Machine Learning Group | 🔥 HF 134 赞 | 💻 代码 | arXiv:2610.14993
TPPO: Tracking-Value PPO For Language Model Reasoning 🔥 HF 95 赞 | 💻 代码 | arXiv:2610.16794
DoSA: Dots Sparse Attention dots studio | 🔥 HF 81 赞 | 💻 代码 | arXiv:2610.13990
Behavioral Compositionality of Sparse Autoencoder Representations OpenSAE | 🔥 HF 107 赞 | 💻 代码 | arXiv:2610.21261
Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics ByteDance | 🔥 HF 70 赞 | 💻 代码 | arXiv:2609.38847
Your Transformer Is Secretly Elastic Video & Image Sense Lab | 🔥 HF 5 赞 | 💻 代码 | arXiv:2610.20588
RSI-Anything: Unlocking Recursive Self-Improvement with World Models PhAI Labs | 🔥 HF 50 赞 | 💻 代码 | arXiv:2610.13912
ProbabilityBench: Model Evaluation and Calibration with Exact Ground-Truth Uncertainty ishidalab | 🔥 HF 10 赞 | 💻 代码 | arXiv:2610.17016
VLA、操作与运动控制、人形机器人、仿真到真机
GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives XPENG AI | 🔥 HF 321 赞 | 💻 代码 | arXiv:2609.39601
Reflex: Visually Grounded Reactive Humanoid Control University of Washington | 🔥 HF 153 赞 | 💻 代码 | arXiv:2610.12772
GreenWAM: A World Foundation Model for Robotics Sber Robotics Center | 🔥 HF 107 赞 | 💻 代码 | arXiv:2610.18774
LeWAM: End-to-End World and Action Modeling with JEPAs LeWAM | 🔥 HF 78 赞 | 💻 代码 | arXiv:2610.12818
VLA-Δ: Learning Physical State Transition for Vision-Language-Action Models ZJU-OmniAI | 🔥 HF 55 赞 | 💻 代码 | arXiv:2610.21306
WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction 🔥 HF 47 赞 | 💻 代码 | arXiv:2610.18568
Towards Agentic Robots: Model, Data, Environment, and Harness 🔥 HF 34 赞 | 💻 代码 | arXiv:2610.20442
EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ Intern Robotics | 🔥 HF 32 赞 | 💻 代码 | arXiv:2610.18872
HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution SNU PI | 🔥 HF 27 赞 | 💻 代码 | arXiv:2610.02089
TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning 2077AI | 🔥 HF 19 赞 | 💻 代码 | arXiv:2610.10288
AR 眼镜、头显交互、egocentric 视觉、可穿戴感知
本周该方向动静不大,只精选了 8 篇真正相关、有实质内容的工作,没有为了凑数纳入边缘论文。
Hear the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation 🔥 HF 68 赞 | 💻 代码 | arXiv:2609.38748
Generative World Simulation with DAO and JING XGEN-Labs | 🔥 HF 8 赞 | 💻 代码 | arXiv:2610.19726
FlowAvatar: Real-Time Full-Body Avatars from Sparse Egocentric Inputs on Consumer XR Devices 热度未测得 | 💻 代码 | arXiv:2610.16277
VISTA: A Multimodal Egocentric Benchmark for Goal-Oriented Perception and Assistance to Blind and Low-Vision Users 热度未测得 | arXiv:2610.22822
EgoProbe: Benchmarking Long-Range and Exact Memory in Streaming Egocentric Videos 热度未测得 | arXiv:2610.17128
EgoWholeView: Camera-Only Head-Mounted Capture of Body and Bimanual Hand Motion for Embodied Demonstrations 热度未测得 | arXiv:2610.16256
PocketTSLM: Deploying an On-Device Soft-Prompt Time-Series Language Model for Privacy-Preserving Smartphone Reasoning 热度未测得 | 💻 代码 | arXiv:2610.21308
EgoOmniWorld: Unifying Egocentric Understanding via Ego-Action-World Interactions 热度未测得 | arXiv:2610.18499