EnvACE trains an agent to simulate its own environment

EnvACE trains one policy to both act and simulate the tools it calls, per a paper posted to arXiv as 2608.06197. Training tool-calling agents with reinforcement learning normally means running two separate systems, the policy and a live environment for it to react against. EnvACE folds both into one model, and that alone accounts for a 5.5-point gain over standard GRPO on τ²-Bench.

The architecture trains with role-wise GRPO, separate advantage baselines for the acting and rehearsal roles, sharing every parameter between them.

On BFCL V4, the tool-calling benchmark, it scored 46.04%. τ²-Bench averaged 36.7%, ahead of every baseline that scaled up with more environment data instead of a better training method. Overall, EnvACE reached 32.91%.

The more interesting number shows up at inference. With the simulation built into the same model, the agent runs a private rehearsal before it commits to a real tool call.

That means checking its own predicted outcome against the move it's about to make. Two rehearsal passes pushed τ²-Bench from 36.7% to 40.9%, with no added training.