2026.08.28 | 视频模型概率校准不足；智能体数据需ACE平衡

【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：[00:30] 🎲 PAWBench: How Far Are We from Probabilistically Aligned World Modeling?（PAWBench：我们距离概率对齐的世界建模还有多远？）[01:09] 🤖 What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents（什么造就了好的智能体数据？面向LLM智能体的数据生成的ACE视角）[02:01] 🎯 TTPO: Test-Time Policy Optimization（TTPO：测试时策略优化）[02:49] 🤖 Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report（训练智能体与其运行框架协同进化：TaoLive数字人智能体技术报告）[03:41] 🏙 UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City（UrbanGround：从局部感知到真实尺度城市中的空间能动性）[04:34] 🎮 GameWAM: A World Action Model for Video Games（GameWAM：视频游戏的世界动作模型）[05:24] 🔄 PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents（PILOT在环：面向长时程智能体的实时自我改进）[06:03] 🤖 Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization（Zero-WAM：基于人类视频的上下文世界-动作建模实现开放式任务泛化）[06:55] ⚗ Self-OPD: On-Policy Distillation for Flow Matching Models without Teacher（Self-OPD：无需教师模型的流匹配模型在线策略蒸馏）[07:51] 🧬 Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO（理解面向LLM推理的进化策略：比GRPO更广泛的推理覆盖）[08:44] 🎮 Magpie: Real-Time World Renderer for Interactive Games（Magpie：用于交互式游戏的实时世界渲染器）[09:41] 🧩 Procedura: Agentic 3D Modeling with Procedural Control（Procedura：程序化控制下的智能体3D建模）[10:39] 🎬 Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning（思考镜头：基于智能体推理的一致性多镜头视频编辑）[11:40] 🧠 WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution（WikiSkill：将智能体经验编译为持久知识以促进技能进化）[12:46] 🔗 CaSKG: Counterfactual-Causal Skill Graphs for Scalable Agent Skill Retrieval（CaSKG：面向可扩展智能体技能检索的反事实因果技能图）【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