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

Researchers propose MAGE, a meta-reinforcement learning framework that enables Large Language Model agents to strategically explore and exploit in multi-agent environments. The framework uses multi-episode training with interaction histories and reflections, showing superior performance compared to existing baselines and strong generalization to unseen opponents.

• →MAGE addresses limitations of current LLM agents in adapting to non-stationary environments through meta-reinforcement learning.
• →The framework integrates interaction histories and reflections into context windows for strategic learning.
• →MAGE combines population-based training with agent-specific advantage normalization for stable learning and diversity.
• →Experimental results demonstrate superior performance in both exploration and exploitation tasks compared to baselines.
• →The framework shows strong generalization capabilities when facing previously unseen opponents.