MAGE

Meta-Reinforcement Learning for Language Agents toward Strategic Exploration and Exploitation

Brief Introduction

This repository contains the code of MAGE, a Meta-Reinforcement Learning method for language agents toward strategic exploration and exploitation.

Contents

• Training
• Citation

Training

To train the LLM Agent with MAGE:

bash examples/kuhnpoker/mage.sh

See the examples folder for more examples.

Acknowledgements

This work is built upon lamer, verl, verl-agent, reflexion, RAGEN. We thank the authors and contributors of these projects for sharing their valuable work.

Citing

If you find our code useful, please consider citing:

@misc{yang2026magemetareinforcementlearninglanguage,
      title={MAGE: Meta-Reinforcement Learning for Language Agents toward Strategic Exploration and Exploitation}, 
      author={Lu Yang and Zelai Xu and Minyang Xie and Jiaxuan Gao and Zhao Shok and Yu Wang and Yi Wu},
      year={2026},
      eprint={2603.03680},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2603.03680}, 
}