Alibaba Qwen Releases QwenGyre for Ultra-Long-Horizon Agent Reinforcement Learning
Alibaba's Qwen team has released QwenGyre, an online reinforcement learning framework designed specifically for ultra-long-horizon AI agent tasks. It features elastic GPU scheduling and advanced trajectory processing to transform complex, multi-hour agent workflows into efficient training data. Training models on ultra-long-horizon tasks traditionally suffers from severe GPU idling and trajectory redundancy bottlenecks. QwenGyre solves these efficiency issues, achieving significant end-to-end speedups and performance gains across complex coding and terminal tasks. QwenGyre utilizes an elastic scheduling mechanism that dynamically reallocates completed GPU compute units to training without interrupting live executions, while organizing outputs into trajectory trees using token-level tracking. In experiments like NL2RepoBench, it delivered a 1.78x end-to-end speedup compared to asynchronous baselines under the same GPU budget.
## BACKGROUND
Online reinforcement learning for AI agents involves models interacting with dynamic environments over hundreds of sequential steps to maximize rewards. Long-horizon tasks require processing massive token contexts over extended periods, which typically causes severe infrastructure bottlenecks and resource idling when training and environment rollouts are decoupled.