University of Science and Technology of China

Alibaba Details QwenGyre to Accelerate Extreme-Horizon Agent Training

Alibaba's Qwen team, with USTC and Tsinghua, released QwenGyre, an elastic RL framework for extreme-long-horizon agent training. An Elastic Scheduler shifts GPUs between rollout generation and training to cut idling and trajectory redundancy; single rollouts often consume 700,000 to about 1 million tokens.