2026.05.08 | 全局速写助长文；技能库让智能体进化

【目录】本期的 15 篇论文如下：[00:23] 🧠 MiA-Signature: Approximating Global Activation for Long-Context Understanding（MiA-签名：面向长上下文理解的全局激活近似方法）[01:32] 🧬 Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning（Skill1：通过强化学习实现技能增强型智能体的统一进化）[02:14] 🎯 MARBLE: Multi-Aspect Reward Balance for Diffusion RL（MARBLE：面向扩散强化学习的多维度奖励平衡方法）[03:08] 🤖 When to Trust Imagination: Adaptive Action Execution for World Action Models（何时信任想象力：面向世界动作模型的自适应动作执行）[04:06] 🧠 Continuous Latent Diffusion Language Model（连续潜在扩散语言模型）[04:50] 🏆 RaguTeam at SemEval-2026 Task 8: Meno and Friends in a Judge-Orchestrated LLM Ensemble for Faithful Multi-Turn Response Generation（RaguTeam 在 SemEval-2026 任务8：基于裁判编排的大语言模型集成实现忠实的多轮响应生成）[05:36] 🧠 Nonsense Helps: Prompt Space Perturbation Broadens Reasoning Exploration（无意义文本助力：提示空间扰动拓宽推理探索）[06:13] ⚡ Continuous-Time Distribution Matching for Few-Step Diffusion Distillation（连续时间分布匹配用于少步扩散蒸馏）[06:48] 🎬 Audio-Visual Intelligence in Large Foundation Models（大型基础模型中的音视频智能）[07:24] 🤖 Auto Research with Specialist Agents Develops Effective and Non-Trivial Training Recipes（利用专家智能体进行自动研究，开发高效且非平凡的训练方案）[08:12] 🤖 A$^2$TGPO: Agentic Turn-Group Policy Optimization with Adaptive Turn-level Clipping（A²TGPO：基于自适应回合级裁剪的智能体回合组策略优化）[09:12] 🧩 UniPool: A Globally Shared Expert Pool for Mixture-of-Experts（UniPool：面向混合专家模型的全局共享专家池）[09:58] 🧠 SkillOS: Learning Skill Curation for Self-Evolving Agents（SkillOS：学习技能策展以实现自我进化智能体）[10:49] 🚗 ReflectDrive-2: Reinforcement-Learning-Aligned Self-Editing for Discrete Diffusion Driving（ReflectDrive-2：面向离散扩散驾驶的强化学习对齐自编辑方法）[11:46] 📊 TabEmbed: Benchmarking and Learning Generalist Embeddings for Tabular Understanding（TabEmbed：面向表格理解的通用嵌入模型的基准测试与学习）【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