The Qwen-Gyre paper proposes elastically allocating GPUs between sampling and training when learning from long agent trajectories. Shared paths are stored as trees to reduce duplication, and trajectories reaching time limits can still be used partially. In Qwen3.8 experiments, the NL2Repo score rose from 52.5 to 58.5 after 48 training steps, with up to 1.85 times the efficiency reported under the same GPU budget. These are results from the researchers’ specific environment; public repository availability and reproducibility on other tasks were not confirmed.