EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
The method trains an agent to rehearse both tool calls and the environment’s replies inside the model.
EnvACE alternates between action generation and “world rehearsal,” then optimizes both roles with task-success rewards. The paper says this lets the policy internalize action-response dynamics instead of relying on external environments or simulators during training. It reports stronger overall results than environment-scaling baselines across BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench. At test time, the model can privately rehearse before executing, which the authors say adds gains without extra external interaction. HF Daily Papers' note
EnvACE alternates between action generation and “world rehearsal,” then optimizes both roles with task-success rewards. The paper says this lets the policy internalize action-response dynamics instead of relying on external environments or simulators during training. It reports stronger overall results than environment-scaling baselines across BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench. At test time, the model can privately rehearse before executing, which the authors say adds gains without extra external interaction. HF Daily Papers' note
score 5