Alibaba's Qwen Introduces QwenGyre: Elastic RL for Extreme-Long-Horizon Agents

What happened

Alibaba's Token Hub Qwen team (with the University of Science and Technology of China and Tsinghua University) published the paper "QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents" (arXiv 2609.33848), an end-to-end online reinforcement learning framework for extreme-long-horizon agent tasks.

The framework addresses two challenges: an elastic scheduler that reallocates GPUs between rollout and training according to unfinished rollout work, without interrupting live executions; and a trajectory processor that organizes each execution's branching history into trajectory trees with shared prefixes, scores partial progress of timed-out runs, and caps each execution at 5 kept trajectories to bound training costs.

On the flagship Qwen 3.8 2.4T model, 48 training steps lifted the NL2RepoBench pass rate from 52.5% to 58.5%; the paper reports up to 1.85x end-to-end speedup over Colocate and 1.78x over Async across workloads. The experiments used Claude Code 2.1.220 as the harness; the paper does not link a public code release (per MarkTechPost).

Key facts

1. Paper: arXiv 2609.33848, by authors from Alibaba Token Hub, USTC, and Tsinghua University
2. Target tasks: Extreme-long-horizon agent tasks: single executions lasting hours, hundreds of model–environment interactions, ~700K tokens per rollout
3. Two mechanisms: An elastic scheduler that dynamically reallocates rollout and training GPUs; a trajectory processor that turns branching execution histories into bounded training samples (max 5 trajectories per execution)
4. Results: On Qwen 3.8 2.4T, NL2RepoBench rose from 52.5% to 58.5% in 48 steps; up to 1.85x speedup over Colocate and 1.78x over Async
5. Code: The paper does not link a public code release (per MarkTechPost)