QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents

Weiqi Wang, Yuxin Zhou, Mouxiang Chen, Siyuan Zhang, Yi Zhang, Yuyan Luo et al.

arXiv 2609.33848 24 pages Read the paper

The short version

The problem: Training an AI agent on tasks that take hours is difficult: some attempts run much longer than others, leaving expensive computers waiting, while retries and delegated subtasks create piles of repeated history. This affects teams trying to improve agents for large software projects, where one task can involve hundreds of interactions.
The idea: QWENGYRE shifts computing resources between running agent tasks and training as demand changes, without stopping tasks already in progress. It also turns each task’s branching history into a limited, organized set of training examples: for example, shared conversation history is kept once rather than counted repeatedly, and unfinished work can still be graded when its progress is measurable.
What it got: It made training up to 1.85 times faster than one comparison system and 1.78 times faster than another. On a large repository-building test, the agent’s success rate rose from 52.5% to 58.5% after 48 training steps.

Modules from this paper

Context-Faithful Trajectory Processing

Branching agent runs create overlapping, changing histories that are hard to train faithfully; QwenGyre preserves exact request contexts in shared-prefix trees, scores assessable work, and admits deduplicated targets per execution.

reinforcement-learningagentstrajectory-processingtraining-data

Elastic GPU Scheduling for Long-Horizon Agents

Long-running agent executions can outlast the GPU allocation that started them; QwenGyre shifts cells between rollout and training at safe boundaries while rerouting live requests and admitting satellites into an ongoing training stream.

reinforcement-learninggpu-schedulingagent-systemsdistributed-training