RAGEN

Why we included this project

RAGEN is for teams that train their own LLM reasoning agents with reinforcement learning and want to know why training drifts or collapses instead of just watching it happen. Rather than handing you a bare PPO or GRPO loop, it treats agent RL as a diagnosis problem: the StarPO framework optimizes multi-turn, trajectory-level behavior, and built-in diagnostics track entropy plus mutual information proxies to catch template collapse that plain entropy metrics miss. Ten ready-made environments, including Sokoban, WebShop, SearchQA, and Lean, let you benchmark against familiar tasks quickly, and the gym-compatible interface makes wiring in your own environment straightforward. The RAGEN-2 additions, such as SNR-adaptive filtering that drops noisy rollouts by reward variance, are the sort of lightweight interventions that stabilize long training runs without major rework. For researchers and engineers who want to understand agent training rather than just hope it works, this is a practical starting point.