代理工具
EnvACE Lets Agents Role-Play Their Own Tool Environments; Qwen3-8B Averages 32.91% Across Three Benchmarks
During training, EnvACE does not query external environments. Instead, the same policy alternates between generating tool actions and imagined environment responses. This reduces reliance on executable training environments, but also directly absorbs simulation errors into the agent’s parameters.
Training long-horizon tool-using agents typically requires executable websites, databases, or service simulators, whose construction and validation costs rise quickly as the range of tasks expands. EnvACE introduces “world rehearsal”: the same language model first assumes the agent role to generate a tool call, then switches to the environment role to predict the tool output or user response that the action should produce. The next decision reads this self-generated observation directly instead of actually executing the tool.
The two roles share parameters but use separate GRPO advantage baselines, with both ultimately updated by the task-success reward. This design differs from an independent simulator: the environment dynamics do not merely provide training data but are encoded into the actual decision-making policy. At inference time, the model can also imagine multiple trajectories in parallel or sequentially, compress the results into rehearsal memory, and then submit a single action to the real environment.
The authors trained Qwen3-8B for 470 steps using 16 NVIDIA H20 GPUs. They sampled four rollouts per prompt, capped trajectories at 30 turns, and used Qwen3-30B-A3B as the judge. EnvACE-8B achieved a three-benchmark average of 32.91% across BFCL V4, τ²-Bench, and VitaBench—0.99 points higher than EnvScaler-8B, which has complete results, but below AWM-14B, whose results are incomplete for some fields. On FinMCP-Bench, its tool F1 score was 46.78%, 3.10 points above EnvScaler-8B. In a controlled comparison, performance on τ²-Bench also rose from 31.2% with standard GRPO to 36.7%.
The open-source implementation includes training code and configurations for SGLang, verl, and Docker, although the model weights are still marked for later release. The most important risk is that “self-consistency” does not guarantee alignment with the real environment: the policy may hallucinate incorrect responses while simultaneously learning to score highly in that incorrect world. The research has so far been limited to an 8B model, and test-time rehearsal results have only been evaluated with a single execution. Future work should examine drift from real tool states, security-sensitive write operations, and whether errors compound as the number of imagined trajectories increases.