Alibaba introduces elastic RL for long-horizon agents
Alibaba’s QwenGyre framework reallocates training resources across difficult rollouts, reporting gains on long-horizon software-engineering tasks.
Alibaba’s Qwen team has introduced QwenGyre, a reinforcement-learning framework designed for training agents on extremely long-horizon tasks. The work targets a growing bottleneck in agent development: a single failed or unfinished rollout can consume substantial compute without producing a useful learning signal.
Elastic training for difficult tasks
QwenGyre uses an elastic scheduler to reallocate training resources according to the difficulty and progress of individual rollouts. Rather than assigning every episode the same budget, the framework can give more capacity to tasks that need additional reasoning or execution time, while avoiding excessive spending on trajectories that are clearly unproductive.
The accompanying research describes a second design for handling long-horizon reinforcement learning, where credit assignment becomes difficult because success may depend on a sequence of actions spread across hundreds of thousands of tokens. The framework was evaluated on software-engineering and other agentic tasks, including settings that require an agent to plan, use tools and recover from intermediate failures.
Scaled to Qwen’s 2.4-trillion-parameter flagship model, the authors report a gain on NL2RepoBench from 52.5% to 58.5% after 48 training steps, with rollouts reaching 700,000 tokens. Those figures come from the authors’ evaluation and should not be read as a general ranking of Qwen against other frontier systems.
Why the method matters
Long-horizon agents are expensive partly because their training process has to preserve useful feedback across long, uneven trajectories. If QwenGyre’s scheduling approach generalizes, it could make agent training less wasteful and improve the economics of teaching models to complete multi-step coding, research or operational tasks.
The release is also notable for coming from Alibaba’s open research ecosystem at a time when leading laboratories are competing over agent reliability rather than only short-answer benchmarks. The remaining question is reproducibility: the reported gains depend on large-scale infrastructure and task-specific training details, and it is not yet clear how well the method transfers across models, environments and failure modes. Still, QwenGyre addresses a practical constraint that will shape whether long-running agents can move beyond demonstrations into routine work.