EnvACE：通过世界预演内化环境动态的智能体强化学习方法
原标题：EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
AI 摘要
EnvACE 是一种新的智能体强化学习方法，通过世界预演（world rehearsal）替代训练中的外部环境交互。策略在生成工具调用后扮演环境角色产生响应，并基于预演响应进行后续决策，两者通过任务成功奖励进行端到端联合优化。在 BFCL-v4、tau^2-Bench、VitaBench 和 FinMCP-Bench 基准上，EnvACE 取得了强且可迁移的性能，优于环境扩展基线。测试时，内部化的世界模型支持私有预演，在适度预演预算下无需额外外部交互即可获得进一步收益。代码已公开。
正文节选
EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning Abstract Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. We introduce EnvACE, an agentic reinforcement learning method that replaces external environment interaction during training with world rehears