Tsinghua University researchers propose critic-based asynchronous reinforcement learning to fix GRPO limitations in long-running AI agents · Digg