From the source
Alibaba's Qwen team introduced QwenGyre, an elastic reinforcement learning system for extreme-long-horizon agent training that reallocates GPUs between rollout and training without interrupting live executions.
- The Elastic Scheduler achieves up to 1.85 times speedup over Colocate and 1.78 times over Async by shifting compute resources as demand changes.
- A Trajectory Processor converts branching executions into prefix-shared trajectory trees with bounded, deduplicated training samples, scoring partial progress even after timeout.
- Scaled to Qwen3.8-2.4T-A95B with roughly 700K tokens per rollout, QwenGyre lifted NL2RepoBench from 52.5 percent to 58.5 percent in 48 steps while cutting end-to-end training time by about 44 percent versus Async.
