2026.07.27 | 技能自我对弈推动模型能力前沿；智能体上下文管理优化成本与推理。

【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 14 篇论文如下：[00:32] 🔄 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills（技能自我对弈：通过协同进化技能推动大语言模型能力前沿）[01:22] 🧠 Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems（智能体上下文管理：将代理记忆与成本视为生命周期与架构问题）[02:18] 🤖 Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning（Molt：一个可扩展的、原生PyTorch的智能体强化学习训练框架）[03:07] 🧪 DataPrep-Bench: Benchmarking LLMs as Training Data Preparators（数据准备基准：将大语言模型作为训练数据准备工具的基准测试）[04:04] 📊 Scaling Native Multimodal Pre-Training From Scratch（从头开始扩展原生多模态预训练）[04:54] 🎯 Three-Body Scattering for Generative Modeling（用于生成建模的三体散射）[05:48] 🌐 LAMAR: An Open Language-Aware Multilingual Alignment Reranker（LAMAR：一种开放的语言感知多语言对齐重排序器）[06:46] 🧠 Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making（多头潜在控制：大语言模型智能体决策的统一接口）[07:43] 🎛 Spectral Prior for Reducing Exposure Bias in Diffusion Models（用于减少扩散模型曝光偏差的频谱先验）[08:34] 🧠 IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation（IDEAgent：面向研究创意生成的主体性质量-多样性搜索）[09:27] 🎯 SceneActBench: Can Agents Act on the 3D Scenes They See?（场景动作基准：智能体能否对所见的三维场景采取行动？）[10:37] 🔄 Closing the Loop: Training-Free Revisit Consistency for Autoregressive Generative Rendering（闭环：无需训练的循环一致性自回归生成渲染）[11:31] 🔊 Multimodal Speaker Verification as a Threat to Speaker Anonymization（多模态说话人验证对说话人匿名化的威胁）[12:26] 🧠 VisCo: Leveraging Large Language Models as Intrinsic Encoders for Visual Token Compression（VisCo：利用大语言模型作为视觉标记压缩的内在编码器）【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