EnvACE Trains Agents Without an Environment by Making One Policy Play Both Actor and Simulator — Then Reuses That World Model for Private Rehearsals at Test Time

arXiv / HuggingFace Daily Papers (39 upvotes)·medium signalXBlueskyLinkedInCopy link

EnvACE (arXiv 2608.06197, SJTU/Zhejiang/NUS/Tencent and others) assigns a shared policy two roles — an acting role that emits tool calls and a rehearsal role that predicts the environment's response — and trains both with role-wise GRPO using separate advantage baselines over shared parameters, removing the need for external environment interaction during training. Results: 46.04% on BFCL V4, 36.7% average on τ²-Bench (+5.5% over GRPO), 16.0% on VitaBench, 46.78% TF1 on FinMCP-Bench, and 32.91% overall, beating every environment-scaling baseline. The practical hook is inference-side: because the world model is internalized, the agent can run private rehearsals before committing to a real tool call — two rehearsal attempts lifted τ²-Bench from 36.7% to 40.9%.