Alibaba QwenGyre validation shows reinforcement learning training time reduced by approximately 44%

Article

Alibaba QwenGyre will 2.4T shorten the time required for model reinforcement learning 44%Beating newsflash,AI Alibaba's Qwen team has introduced QwenGyre, a AI Agent reinforcement learning framework designed specifically for ultra-long tasks.

As AI Agents begin taking on end-to-end software engineering tasks, a single run may last for hours and involve hundreds of model calls and tool operations. Models need to learn from the actual outcomes of these tasks, but conventional training systems can easily become bogged down in long waits and redundant computation.

The first issue is GPU utilization. Conventional synchronous training must wait for a batch of AI Agents to all complete their tasks before updating the model collectively. Task durations vary significantly, and tasks that finish earlier leave some GPU idle for extended periods. Although asynchronous training separates task execution from model training, it fixes the GPU allocation of both, making flexible scheduling difficult.

QwenGyre allows GPU to dynamically switch roles based on task progress. As tasks are completed in succession, the system shifts some GPU to model training, while the remaining GPU continue supporting AI Agents whose tasks are not yet complete. The entire process does not require interrupting tasks, and workspace and tool states can be retained. Newly added GPU can also directly participate in training for the current batch.

Another issue is redundant training data. When AI Agents repeatedly retry, compress historical records, or call sub-agents, they leave behind many overlapping execution paths. QwenGyre records the actual inputs and outputs of every model call and organizes different paths into a trajectory tree with shared history. During training, it selects only a limited number of paths, so the same model output is not computed repeatedly. Even if a task times out, completed and verifiable partial results can still be used for training.

The team validated the system on Qwen3.8-2.4T, with each task cumulatively processing approximately 700000 Token. Under the same GPU budget, completing 48 reinforcement learning steps took QwenGyre 75.42 hours, while the conventional asynchronous approach required 134.55 hours, reducing training time by approximately 44%. After training, the model's pass rate on the code-repository test-generation NL2RepoBench increased from 52.48%to 58.54%. The paper also shows that QwenGyre achieved training scores close to those of conventional training approaches.

Original link https://m.theblockbeats.info/flash/371383

Key points

QwenGyre is a AI Agent reinforcement learning framework for ultra-long tasks.
QwenGyre allows GPU to dynamically switch between task execution and model training based on task progress.
QwenGyre records the inputs and outputs of model calls and organizes them into a trajectory tree with shared history.
In Qwen3.8-2.4T validation, 48 reinforcement learning training steps took 75.42 hours, while the conventional asynchronous approach took 134.55 hours.
Under the same GPU budget, QwenGyre reduced training time by approximately 44%。
After training, the NL2RepoBench pass rate increased from 52.48%to 58.54%。
The paper shows that QwenGyre achieved training scores close to those of conventional training approaches.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

QwenGyre reduces waiting and redundant computation in reinforcement learning for ultra-long AI Agent tasks by dynamically allocating GPU and reusing execution trajectories. The experiments in the evidence show that it requires less time to complete training under the same GPU budget.

Why it matters to readers

If these results can be reproduced, QwenGyre may help reduce the training time and resource consumption of reinforcement learning for ultra-long AI Agents.

It demonstrates how AI systems can improve training efficiency by using GPU more flexibly and preserving task processes.

Risks and unknowns

The evidence comes from a media newsflash and does not provide the original paper or code link.
The evidence does not state whether the experimental results can be reproduced across other tasks, models, or scales.
The evidence provides limited experimental conditions for the reduction in training time and improvement in pass rate, and does not fully explain them.

Alibaba QwenGyre validation shows reinforcement learning training time reduced by approximately 44%

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center