2026.07.03 | 小模型本地化击败大模型；自主策略演化聚焦结构合成

【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] 🧩 Program-as-Weights: A Programming Paradigm for Fuzzy Functions（程序即权重：面向模糊函数的编程范式）[01:24] 🧠 EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments（EvoPolicyGym：在交互环境中评估自主策略演化）[02:24] 🧠 AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents（AgenticSTS：面向长时程LLM智能体的有界内存测试平台）[03:18] 🔍 Morphing into Hybrid Attention Models（变形为混合注意力模型）[04:07] 📊 AgenticDataBench: A Comprehensive Benchmark for Data Agents（AgenticDataBench：面向数据智能体的综合性基准测试）[05:12] ⚡ Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling（多分辨率流匹配：通过分阶段采样的无训练扩散加速）[05:52] 🎬 WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory（世界导演：构建具有持久动态记忆的可控世界模拟器）[06:49] 🏥 Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning（打破失败级联：面向医学多模态推理的步骤感知强化学习）[07:37] 🎨 Optimizing Visual Generative Models via Distribution-wise Rewards（通过分布级奖励优化视觉生成模型）[08:32] 🎯 SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use（SkillCoach：用于评估和增强智能体技能使用的自我演化评分标准）[09:21] 🖐 AGVBench: A Reliability-Oriented Benchmark of Data Augmentation for Vein Recognition（AGVBench：面向静脉识别的可靠性导向数据增强基准）[10:16] 🔬 From SRA to Self-Flow: Data Augmentation or Self-Supervision?（从SRA到Self-Flow：数据增强还是自监督？）[11:10] 🧠 Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads（对数几率贡献评分识别非字面检索头）[11:58] 🎯 AnyGroundBench: A Specialized-Domain Benchmark for Video Grounding in Vision-Language Models（AnyGroundBench：面向视觉语言模型中视频定位的专业领域基准）[12:46] 🤔 When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search（搜索代理何时应提问：面向澄清感知的深度搜索基准DiscoBench）【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