Qwen Introduces QwenGyre: Elastic RL for Extreme-Long-Horizon Agent Training

Key Info

Qwen introduced QwenGyre, an elastic reinforcement learning framework designed for extreme-long-horizon (XLong) agent training, built on two core designs: an Elastic Scheduler and a Trajectory Processor.

Highlights

• Elastic Scheduler reallocates GPUs between rollout and training as demand shifts without interrupting live executions, delivering up to 1.85x speedup over Colocate and 1.78x over Async.
• Trajectory Processor converts branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples; partial progress after timeout still gets scored.
• Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps.
• End-to-end training time is cut by ~44% compared to Async.