本周 60 篇里最清晰的信号是蒸馏与自进化:同族 on-policy 蒸馏的缩放定律、把教师当"方向"而非"终点"的表示空间外插、多教师反馈的音量校准、统一多模态模型拿自评当特权信息自蒸馏、代码 Agent 的分组评审信用重分配,一路排到揭露自进化搜索里"出题器与解题器串通作弊"的诊断工作。研究重心明显不在堆更大的模型,而在把已有能力更省地搬到另一个模型或另一段推理里,同时防住自我改进过程中的指标幻觉。
第二个高频信号是"harness"这个词本身——Raven 做"harness 的 harness"、Omni-IO 给已有 Agent 外挂全模态手脚、Mid-Harness 在模型与执行层之间插一道动作验证、MaLiang-Harness 把图像生成组织成构建-检查-修订流程,还有机器人侧的递归 harness 蒸馏。共同点是不碰模型权重,改模型外面那层执行环境。世界模型这条线也出现了类似的收敛:实证表明收益来自"准备未来"而非"生成未来",一次前向就够。
两点需要如实说明:机器人方向本周 25 篇候选里有 24 篇拿到了社区投票,数据相当扎实;但智能眼镜方向本周没有任何以眼镜硬件或整机系统为主题的论文,25 篇候选里只有 5 篇有公开热度数据,那一章的排序主要靠主题相关性,不是靠热度。
预训练、后训练、推理与长上下文
这一批论文几乎都在同一个问题上发力:如何把强模型已有的能力更省、更准地搬到另一个模型或另一段推理里。蒸馏方向占了半壁江山,其余几篇则从注意力分配、量化、循环深度等角度压缩同等能力下的算力开销。
Scaling Properties of Same-Family On-Policy Distillation Zhejiang University | 🔥 HF 293 赞 | arXiv:2609.32722
Post-Training Leaves Behavioral Shadows on Unrelated Decisions Peking University | 🔥 HF 271 赞 | 💻 代码 | arXiv:2609.29233
The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation 🔥 HF 231 赞 | 💻 代码 | arXiv:2609.36484
Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation Nanyang Technological University | 🔥 HF 176 赞 | 💻 代码 | arXiv:2609.35347
Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs 🔥 HF 100 赞 | arXiv:2609.29845
EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery Minnesota NLP | 🔥 HF 97 赞 | 💻 代码 | arXiv:2609.40340
Disaggregated Quantization: Specializing LLM Prefill and Decode IST Austria Distributed Algorithms and Systems Lab | 🔥 HF 90 赞 | 💻 代码 | arXiv:2609.26333
MassAlloc Attention: Let Attention Allocate Its Own Compute Beijing Academy of Artificial Intelligence | 🔥 HF 74 赞 | 💻 代码 | arXiv:2609.32712
Improving Test-Time Scaling with Adaptive Looped Transformers Tsinghua-NICS-EFC | 🔥 HF 54 赞 | 💻 代码 | arXiv:2609.35748
Knowing When Thinking Is Not Enough: Teaching Small Reasoning Models to Reason Beyond Their Parametric Knowledge KAIST AI | 🔥 HF 40 赞 | 💻 代码 | arXiv:2609.34327
VLM、图像与视频生成、文档与视频理解
这一批的主线是"让多模态模型自己看见、自己验证、自己纠错"——无论是程序化生成后的渲染回检、潜空间推理的视觉证据监督,还是统一模型的自蒸馏,核心都在把视觉反馈接回训练与推理闭环;另一条副线则在重估架构地基:视觉骨干、视觉编码器、自编码器都被重新质询。
MaLiang-Harness: A Programmable Path to Image and Video Generation National University of Singapore | 🔥 HF 386 赞 | 💻 代码 | arXiv:2609.34309
VisionHOPE: Visual Backbones as Self-Modifying Learning Systems Mininglamp Technology | 🔥 HF 319 赞 | 💻 代码 | arXiv:2609.33325
UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement Stanford NLP | 🔥 HF 237 赞 | arXiv:2609.38721
Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence Amazon | 🔥 HF 207 赞 | 💻 代码 | arXiv:2609.34563
FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders USC Physical Superintelligence (PSI) Lab | 🔥 HF 157 赞 | 💻 代码 | arXiv:2609.31620
What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling Tsinghua-LeapLab | 🔥 HF 114 赞 | 💻 代码 | arXiv:2609.34981
WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents University of California, Santa Barbara | 🔥 HF 93 赞 | 💻 代码 | arXiv:2609.40325
Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies Amazon Science | 🔥 HF 88 赞 | 💻 代码 | arXiv:2609.38155
How Far Are We from Removing the Visual Encoder? Scaling Laws for Encoder-Free Multimodal Pretraining Tencent Hunyuan | 🔥 HF 64 赞 | arXiv:2609.35457
Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering KAIST AI | 🔥 HF 57 赞 | 💻 代码 | arXiv:2609.38177
工具调用、Computer Use、多智能体与编码 Agent
本周 Agent 方向的主线是"harness"——不改模型权重,而是改模型外面那层执行环境:自动构造、自动优化、在模型与 harness 之间加一层动作校验,再配上自我改进与反作弊机制。
Raven: The Harness of Harnesses for Composable Agentic Intelligence EverMind | 🔥 HF 501 赞 | 💻 代码 | arXiv:2609.33439
Omni-IO Skills: Harnessing Your Agent Omni-Native National University of Singapore | 🔥 HF 272 赞 | 💻 代码 | arXiv:2609.31847
False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents Rutgers University | 🔥 HF 190 赞 | arXiv:2609.39102
Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL Xiaomi MiMo | 🔥 HF 124 赞 | arXiv:2609.32577
AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks Beijing Academy of Artificial Intelligence | 🔥 HF 121 赞 | 💻 代码 | arXiv:2609.38288
LEGO-Anything: Coding Agents for 3D Scene Reconstruction Amazon Web Services | 🔥 HF 120 赞 | arXiv:2609.36380
Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents NVIDIA | 🔥 HF 103 赞 | arXiv:2609.39982
Follow the Entities: A Corpus Map for Agentic Search Microsoft | 🔥 HF 80 赞 | arXiv:2609.37226
RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement RSIGame | 🔥 HF 75 赞 | 💻 代码 | arXiv:2609.39045
LLMs are General Asynchronous Agents Yandex Research | 🔥 HF 67 赞 | 💻 代码 | arXiv:2609.35427
架构、优化、表征学习、语音与可解释性
这一批论文集中在"主干之外的基础层":音频与语音生成/评测、长上下文注意力的效率与隐患、数据管线系统、持续学习动力学与水印安全。共同点是都在拆解既有组件的内部机制——不是堆更大的模型,而是找出哪一层设计决定了效果与失效。
YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality Multimodal Art Projection | 🔥 HF 240 赞 | 💻 代码 | arXiv:2609.33757
VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models The University of Melbourne | 🔥 HF 136 赞 | 💻 代码 | arXiv:2609.32607
Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression ByteDance Seed | 🔥 HF 90 赞 | arXiv:2609.36322
Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue Peking University | 🔥 HF 89 赞 | arXiv:2609.31948
RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation Peking University | 🔥 HF 54 赞 | 💻 代码 | arXiv:2609.18703
Block Sparse Attention with Log-Linear Complexity ByteDance Seed | 🔥 HF 30 赞 | arXiv:2609.31093
EmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech Generation Meta | 🔥 HF 18 赞 | 💻 代码 | arXiv:2609.38157
Residual Transferability in Neural Image Watermarking 🔥 HF 12 赞 | arXiv:2609.32241
Learning Dynamics of Continual Learning: A Unified View of Data Attribution, Forgetting, and Plasticity Loss 🔥 HF 6 赞 | 💻 代码 | arXiv:2609.33620
Pretraining Transformers with Quantized Softmax in Attention University of Illinois at Urbana-Champaign | 🔥 HF 5 赞 | arXiv:2609.33591
VLA、操作与运动控制、人形机器人、仿真到真机
这一批工作几乎都在绕开"端到端策略一条路走到黑":把任务状态写成可检查的代码、把干预经验攒成可复用的手册、把感知拓宽到全景与空间音频、把潜在行为空间跨本体共享——核心都是让策略在新场景里还能站得住。
In-Context Learning for Robots: Methods and Applications Knowin AI | 🔥 HF 360 赞 | 💻 代码 | arXiv:2609.36012
PanoVLN: Towards Effective Panoramic Vision-and-Language Navigation Zhejiang University | 🔥 HF 152 赞 | 💻 代码 | arXiv:2609.34759
Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence HexaFuture | 🔥 HF 106 赞 | 💻 代码 | arXiv:2609.35432
Recursive Harness Distillation across Agents for Robot Manipulation Seoul National University | 🔥 HF 40 赞 | arXiv:2609.33378
OmniEcho: Spatial Audio Understanding for Embodied Agents PKU-VaLuE-Lab | 🔥 HF 35 赞 | 💻 代码 | arXiv:2609.23407
TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion Mathis Group @ EPFL | 🔥 HF 32 赞 | 💻 代码 | arXiv:2609.38653
Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies Galbot | 🔥 HF 28 赞 | 💻 代码 | arXiv:2609.38537
Tactile-JEPA: Topology-Aware Self-Supervised Representation Learning for Distributed Tactile Sensors 🔥 HF 22 赞 | 💻 代码 | arXiv:2609.24385
CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments 🔥 HF 21 赞 | arXiv:2609.38087
REALM: A Coarse-to-Fine Generative Framework for Embodied Reactive Listening Macquarie University | 🔥 HF 20 赞 | 💻 代码 | arXiv:2609.33095
AR 眼镜、头显交互、egocentric 视觉、可穿戴感知
本周没有真正以智能眼镜硬件或整机系统为主题的论文,这一批集中在第一人称视觉、手势与视线估计、XR 交互和可穿戴感知这几条更偏研究侧的线上。该方向公开热度数据极少,十篇里只有少数几篇拿到社区投票,其余均未测得,因此排序主要依据主题相关性而非热度。
APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants Shanghai Jiao Tong University | 🔥 HF 36 赞 | 💻 代码 | arXiv:2609.37559
InfiniHand: Streaming World-Space Hand Motion Estimation from Egocentric Video Shanghai AI Laboratory | 🔥 HF 9 赞 | 💻 代码 | arXiv:2609.35743
MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories Meta | 🔥 HF 2 赞 | arXiv:2609.40195
EyeTAG: Eye Trajectory-Aware Gaze Estimation 热度未测得 | 💻 代码 | arXiv:2610.00922
ESTHER: Egocentric Stereo Hand Estimation and Reconstruction in the Wild 热度未测得 | arXiv:2609.34817
Beyond Monoscopic Viewing: A Study on 3D Gaussian Splatting Quality in VR 热度未测得 | arXiv:2609.38525
VocalEyes: Speaker-Aware Augmented Reality Captioning through In-Conversation Registration 热度未测得 | arXiv:2609.32983
"Black Mirror?": Public Sensemaking of AI-Powered Lifelogging 热度未测得 | arXiv:2609.34950
Exo2EgoHOI: Hand-Object-Interaction Aware Exocentric-to-Egocentric Video Generation 热度未测得 | arXiv:2609.38615
One Sensor, Whole Body - 3D Body Pose from a Single Consumer Earbud IMU 热度未测得 | 💻 代码 | arXiv:2609.34978