QwenGyre targets the training bottlenecks of long-horizon agents
A Qwen research paper introduces QwenGyre, an end-to-end framework for online reinforcement learning on extreme-long-horizon agent tasks. The paper describes executions lasting hours, spanning hundreds of model-environment interactions and approaching one million tokens per rollout. Its central systems mechanism is elastic GPU allocation between rollout generation and training without interrupting live executions. A trajectory processor reconstructs branching histories, scores partial progress, and deduplicates redundant paths.
At scale with Qwen3.8 2.4T and 700,000-token rollouts, the authors report a 6.0 percentage-point improvement on NL2RepoBench, from 52.5% to 58.5%, in 48 steps. Across evaluations on diverse training-data domains, they report speedups of up to 1.85× over Colocate and 1.78× over Async.
These are results reported in the paper, not an independent replication. The supplied arXiv record does not establish implementation availability. The work’s immediate contribution is a training architecture aimed at reducing GPU idle time and redundant trajectory cost under long, variable executions.