Slime: Asynchronous RL for Agentic Tasks

Introduced the training method that reduced GLM-5's hallucination rate by 35%.