2026.08.21 | 动态环境适配弱点；任务合成保留源意图

【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] 🌍 EnvHarness: Awakening Static Worlds for Agent Learning（EnvHarness：为智能体学习唤醒静态世界）[01:28] 🖥 FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis（FACET：在终端任务合成中保留源意图与可执行状态）[02:29] 🧪 SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?（SWE-bench Science：编码智能体能否解决科学领域的工程任务？）[03:15] 🕺 4DAnyone: Create Anyone in 4D from a Casual Monocular Video（4DAnyone：从随意单目视频创建任意人物的4D形象）[04:11] 👥 WithEveryone: Unified Planning and Identity Grounding for Group Image Generation（WithEveryone：面向群像生成的统一规划与身份锚定）[05:01] 🧠 MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use（MemTrapBench：大语言模型记忆使用中的认知陷阱基准测试）[05:52] 🔄 SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback（SkillEvo：来自多轮交互反馈的自我更新进化梯度）[06:50] 🎮 ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models（ForgeWM：面向少步动作条件视频世界模型的渐进式因果训练）[07:48] 🧩 Repo0: Design-Driven Zero-to-All Code Generation（Repo0：设计驱动的从零到全代码生成）[08:46] ⚡ FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving（FlashPrefill V2：面向长上下文大语言模型服务的块稀疏预填充注意力）[09:41] 🧠 Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization（注入、对齐、恢复：面向无检索文档知识内化的分阶段后训练）[10:42] 🤖 EXIMO: VLM Guided Exploration of VLA Policies（EXIMO：视觉语言模型引导的VLA策略探索）[11:36] 🧠 Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See（用低资源语言思考：SFT构建了什么，RL修复了什么，准确率无法看到什么）[12:18] 🎯 Towards Quantifying Benchmark Optimization in ASR Models（面向ASR模型中基准优化的量化研究）[13:20] 🛡 PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents（PolicyGuide：从守护单一动作到引导策略合规型LLM智能体的整个工作流）【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