Physix FrontierPhysix Frontier · News Briefing Card (IT Home · Oct 12, 2026)
Alibaba Qwen Releases QwenGyre Online RL Framework for Long-Horizon Agents
KEY FACTS
- Alibaba Qwen released QwenGyre, an online reinforcement learning framework for long-horizon agent tasks, on October 12.
- The framework splits GPUs into compute units; once a task finishes, the freed unit switches to training.
- QwenGyre uses a trajectory tree to record context and only expands selected paths as training samples.
- Tasks that time out but are partially completed can provide feedback through Partial Scoring.
- In NL2RepoBench experiments, the model's pass rate rose from 52.5% to 58.5% after 48 training steps.
KEY DATA
52.5%→58.5%NL2RepoBench pass rate improvement
48 stepsTraining steps
about 700KTokens per rollout
1.78×End-to-end speedup vs. Async
PHYSIX OBSERVATION
Long-horizon agent training has long been trapped by long trajectories and sparse feedback. QwenGyre breaks task completion down into recoverable training signals, effectively letting idle compute continuously cash in. A 1.78× speedup is not disruptive, but combined with the design that also scores partial completion, the barrier for small and mid-sized teams to run ultra-long tasks will drop noticeably, and RL engineering has taken another step forward.
Source: IT Home report