QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents
QwenGyre is built to keep RL training moving when agent rollouts stretch toward 1M tokens.
The paper says xlong-horizon agent runs create GPU idle time and redundant branching trajectories. QwenGyre reallocates GPUs between rollout and training without stopping live executions, then reconstructs and deduplicates trajectory histories. On Qwen 3.8 2.4T, the authors report a 6-point NL2RepoBench gain, from 52.5% to 58.5%, after 48 steps. They also report speedups of up to 1.85x over Colocate and 1.78x over Async. HF Daily Papers' note
The paper says xlong-horizon agent runs create GPU idle time and redundant branching trajectories. QwenGyre reallocates GPUs between rollout and training without stopping live executions, then reconstructs and deduplicates trajectory histories. On Qwen 3.8 2.4T, the authors report a 6-point NL2RepoBench gain, from 52.5% to 58.5%, after 48 steps. They also report speedups of up to 1.85x over Colocate and 1.78x over Async. HF Daily Papers' note
score 5