ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning

New policy optimization method solves training collapse in multi-step agentic RL tasks — more reliable agent training.