EnvACE-Qwen3-8B
Model Description
EnvACE-Qwen3-8B is a tool-interactive agent model based on Qwen/Qwen3-8B (Thinking Mode), trained with the EnvACE framework for long-horizon tool use. Instead of interacting with real or synthesized executable environments during training, the model is trained by world rehearsal: it internalizes environment dynamics in its own parameters and acts as its own environment.
Training Process
EnvACE replaces external environment interaction with world rehearsal inside a single policy:
- Acting / rehearsal alternation: the policy first generates a tool call, then plays the role of the environment to produce the response induced by that action, and conditions its subsequent decisions on the rehearsed response.
- Joint end-to-end optimization: both roles share one set of parameters and are trained together with role-wise GRPO (a separate advantage baseline per role, one shared policy) from task-success rewards only — a pure RL stage, with no external simulator or executable environment in the loop.
- Agent world model: through rehearsal the policy internalizes the relationship between actions and environment responses, yielding a world model that directly supports decision making. At test time this enables private rehearsal before committed execution, giving further gains under a moderate rehearsal budget without additional external interaction.
The method is evaluated on BFCL-v4, τ²-Bench, VitaBench and FinMCP-Bench, where it shows strong and transferable performance against environment-scaling baselines; controlled studies show world rehearsal consistently improves policy learning across model scales. EnvACE-Qwen3-8B reaches Overall 32.91 across BFCL-v4 / tau^2-Bench / VitaBench and TF1 46.78 on FinMCP-Bench, outperforming Simulator-8B, TOUCAN-7B, EnvScaler-8B, AWM-8B/14B and ScaleEnv-8B at the same open-source scale. See the paper for full results.
How to Use
Basic Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Team-ACE/EnvACE-Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
# Qwen3 chat template + native function-calling format:
# tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True)
Recommended sampling parameters (inherited from Qwen3): temperature=0.6, top_p=0.95,
top_k=20.
With the EnvACE Framework
For training and for full integration with tool-interactive environments, see the EnvACE repository.
Related Resources
- Paper: EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
- Project page: https://within-yao.github.io/EnvACE/
- Code: Within-yao/EnvACE
- Collection: EnvACE models
- Other Models:
Citation
If you use this model, please cite our work:
@article{xu2026envace,
title={EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning},
author={Xu, Zishan and Yao, Zhiyuan and Chen, Yuxin and Guo, Yifu and Lu, Zhengxi and Lu, Yuquan and Huang, Jinyang and Xu, Yan and Wang, Yasheng and Zhang, Weinan and Zeng, Xingshan and Liu, Weiwen},
journal={arXiv preprint arXiv:2608.06197},
year={2026}
}
License
This model is licensed under the Apache 2.0 License, following the base Qwen3 model license.