视频选题榜 2026-08-08
- https://arxiv.org/abs/2608.06197 · EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning · 同策略权重 prompt 切换扮演策略者+环境双角色,端到端优化绕开可执行环境,测试时私有预演作 CoT · round=R1
- https://arxiv.org/abs/2608.06257 · MASS: Multiplayer World Models with Authoritative Shared State · 题面 verbatim = MASS(4 字);abstract body = MAS(3 字 · 论文自身摘要笔误);Logic Engine + Rendering Engine 双通路解耦;1024 玩家 × 10000 步 Snake 稳定不崩 · round=R2
- https://arxiv.org/abs/2608.06305 · Beyond Top-K: Replacing Black-Box Retrieval with Interpretable Agentic Operations · RAG 接口层替换 + 结构化长文档单位丢失 + audit trail 一等公民 · round=R3