2026.05.20 | 反自蒸馏优化推理；可验证环境测评智能体

【目录】本期的 15 篇论文如下：[00:24] 🧠 Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information（基于点互信息的反自蒸馏用于推理强化学习）[01:08] 🖥 OpenComputer: Verifiable Software Worlds for Computer-Use Agents（OpenComputer：为计算机使用智能体构建可验证的软件世界）[01:53] 🧠 GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment（GoLongRL：面向能力的长上下文强化学习与多任务对齐）[02:49] 🔬 Process Rewards with Learned Reliability（具有学习可靠性的过程奖励模型）[03:44] 🤖 AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration（AutoResearchClaw：基于人类-人工智能协作的自我强化自主研究）[04:48] 🎭 When Vision Speaks for Sound（当视觉为声音代言）[05:50] 🏭 EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL（EnvFactory：通过可执行环境合成与鲁棒强化学习扩展工具使用型智能体）[06:45] 🎬 CogOmniControl: Reasoning-Driven Controllable Video Generation via Creative Intent Cognition（CogOmniControl: 基于推理驱动的可控视频生成与创意意图认知）[07:40] 🎯 Active Learners as Efficient PRP Rerankers（主动学习器作为高效的成对排序提示重排序器）[08:24] 🎥 Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos（Artifact-Bench：评估多模态大语言模型在检测与评估AI生成视频伪影方面的能力）[09:14] 🎬 Aurora: Unified Video Editing with a Tool-Using Agent（Aurora：使用工具型代理的统一视频编辑框架）[10:12] 🎯 CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization（对比证据策略优化：基于强化学习与可验证奖励的自蒸馏方法）[11:01] 📱 OmniGUI: Benchmarking GUI Agents in Omni-Modal Smartphone Environments（OmniGUI：在全模态智能手机环境中评估GUI代理的基准测试）[11:51] 🎬 MSAVBench: Towards Comprehensive and Reliable Evaluation of Multi-Shot Audio-Video Generation（MSAVBench：迈向全面且可靠的多镜头音视频生成评估）[12:44] 🎥 Video Models Can Reason with Verifiable Rewards（视频模型可通过可验证奖励进行推理）【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