RAGEN

RAGEN is a flexible reinforcement learning framework designed for training LLM agents with multi-turn reinforcement learning capabilities. It is based on the StarPO algorithm, supports trajectory-level optimization, and provides multiple built-in environments. The RAGEN-2 version introduces SNR adaptive filtering and inference crash diagnosis to address template collapse issues. The framework is suitable for training agents requiring multi-turn dialogue or task processing, helping developers better understand and improve the reinforcement learning training process.