Details
- Qwen introduced QwenGyre, an online reinforcement-learning framework for training agents on extreme-long, or xLong, tasks.
- Its Elastic Scheduler dynamically reallocates GPUs between rollout generation and model training as workloads change, while keeping live agent executions running.
- A trajectory processor reconstructs branching execution histories, scores partial progress, and removes redundant paths to reduce training overhead.
- The framework targets tasks lasting hours, involving hundreds of model-environment interactions and approaching 1 million tokens per rollout.
- In reported tests, QwenGyre delivered up to 1.85× the speed of colocated training and 1.78× the speed of an asynchronous baseline under matched GPU budgets.
- On Qwen 3.8 2.4T with 700,000-token rollouts, NL2RepoBench performance increased from 52.5% to 58.5% after 48 training steps.
- The work addresses two operational bottlenecks in long-horizon agentic RL: idle or poorly balanced GPUs and rapidly expanding, repetitive trajectory data.
Impact
QwenGyre focuses on the infrastructure bottleneck behind increasingly capable coding and tool-using agents rather than introducing another model. If its reported efficiency gains generalize, elastic scheduling and branch-aware data processing could lower the compute cost of long-horizon reinforcement learning and make larger training runs more practical. The results also strengthen Qwen’s position in open research on agent training systems, although the comparisons are against internal scheduling baselines rather than directly against competing commercial platforms.