Qwen Released QwenGyre: RL Framework for Extreme Long-Horizon AI Agents
The Qwen team has released QwenGyre, an elastic reinforcement learning framework designed specifically for training extreme-long-horizon AI agents. It features an Elastic Scheduler that dynamically reallocates GPUs without interrupting live executions, and a Trajectory Processor that turns complex branching executions into prefix-shared trajectory trees. This framework significantly accelerates and improves the training of AI agents on complex, long-duration tasks, addressing major bottlenecks like GPU idling and severe execution variance. By scaling to massive models and cutting end-to-end training time by roughly 44%, it paves the way for more capable and autonomous AI assistants. QwenGyre achieves up to a 1.85× speedup over Colocate methods and a 1.78× speedup over Async methods. When scaled to Qwen3.8-2.4T-A95B with approximately 700K tokens per rollout, it improved NL2RepoBench performance from 52.5% to 58.5% in 48 steps.
## BACKGROUND
Online reinforcement learning (RL) for large language models involves training agents by letting them interact with environments over multiple steps, using rewards to guide their behavior. When tasks become extremely long—spanning hours and generating nearly a million tokens per rollout—traditional RL setups suffer from severe computational inefficiencies and GPU starvation due to unpredictable execution times.