2026.09.03 | 代码库蒸馏成AI技能；长视频世界模型可扩展

【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43【目录】本期的 15 篇论文如下：[00:31] 🤖 Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills（从仓库到技能：将GitHub代码库蒸馏为AI4AI技能）[01:26] 🌍 SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models（SolarWM：面向长时程视频世界模型的开放数据与可扩展训练）[02:19] ⏱ EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction（EarlyEval：通过早期结果预测实现更廉价的智能体评估）[03:21] 🤝 It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning（两相匹配：基于强化学习的生成式检索器协同进化）[04:18] 🎯 Language Models Can Control Their Own Attention（语言模型能够控制自身的注意力）[05:10] 🖼 On the Design Fundamentals of Pixel Text Representation Learning（论像素文本表示学习的设计基础）[06:12] 🔍 Beyond Visual Similarity: Entity-Aligned Retrieval for Knowledge-Based Visual Question Answering（超越视觉相似性：面向知识型视觉问答的实体对齐检索）[07:14] 🤖 HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?（HarnessDev：大语言模型能否创建并演化自己的智能体执行框架？）[08:05] 🗜 ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes（ZipTok3D：利用紧凑令牌前缀的高保真三维令牌化）[09:04] 🎯 Cliff: Learning Process Rewards from the First Mistake（Cliff：从第一个错误中学习过程奖励）[10:05] 🎯 Aspire: Can Models Self-Evolve from Vague Goals?（Aspire：模型能否从模糊目标中自我进化？）[10:46] 🎭 Influence-Directed Distillation: Solving the Diversity Bottleneck in Sampled-Token On-Policy Distillation（影响导向蒸馏：解决采样词元在线策略蒸馏中的多样性瓶颈）[11:42] 🎮 S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?（S3Gym：大语言模型能否将自我测试与自我评判转化为自我改进？）[12:36] 👀 A Glance Is All You Need: Single-Pass Fine-Grained Image Captioning with SimLoss（一瞥足矣：基于SimLoss的单遍细粒度图像描述）[13:35] 🎙 VibeVoice-ASR-Streaming Technical Report（VibeVoice-ASR-Streaming技术报告）【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递在小宇宙查看该单集文稿