World rehearsal lifts an agent 5.5 points

EnvACE lifted an 8-billion-parameter agent’s average on τ²-Bench from 31.2% to 36.7%, a 5.5-point gain over standard GRPO, by training the same policy to generate both tool calls and rehearsed environment responses, according to the researchers’ paper. Two private rehearsals at test time then raised its combined τ²-Bench and BFCL Multi-Turn score from 36.7% to 40.9% without another external interaction. Agent teams should test this as a cheaper preflight layer, not a license to trust imagined tool state: the AISI evaluation-governance analysis shows why committed actions still need independent controls.