Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs

Reinforcement learning (RL) is widely used to improve large language models on reasoning tasks, and asynchronous RL training is attractive because it increases end-to-end throughput.