Alibaba's Qwen Team Releases QwenGyre, an Online Reinforcement Learning Framework for Ultra-Long-Horizon Agent Tasks

QwenGyre is an online reinforcement learning framework released by Alibaba's Qwen team, specifically designed to address the challenges agents face when executing complex long-horizon tasks. When an agent runs for hours at a time, interacts with its environment over hundreds of rounds, and processes millions of tokens cumulatively, traditional training methods struggle to effectively leverage such ultra-long trajectory data.

The framework's core mechanism is elastic scheduling and trajectory processing: GPUs are divided into multiple compute units, and as each batch of tasks completes, the freed units immediately transition to the training phase. Tasks still in progress remain undisturbed, with their states and caches migrated as a whole without needing to recompute from scratch. This design turns ultra-long-horizon tasks themselves into an effective source of training data while balancing capability and efficiency.

Training long-horizon agents has long been a challenge in reinforcement learning, as overly long trajectories make credit assignment difficult and lead to low compute resource utilization. QwenGyre's elastic scheduling approach offers an engineering-level solution to this bottleneck, potentially advancing agents' capabilities in complex multi-step tasks.

The release of this framework also reflects the trend of large model training evolving from short-horizon instruction following toward long-horizon autonomous decision-making. The combination of online reinforcement learning and agent tasks is likely to become a key direction in future technological competition.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.