阿里千问发布 QwenGyre 面向超长程智能体任务的在线强化学习框架Alibaba's Qwen Releases QwenGyre, an Online Reinforcement Learning Framework for Ultra-Long-Horizon Agent Tasks


AI free token/AI免费积分
感谢IT之家网友 Domado、Autumn_Dream 的线索投递!

IT之家 10 月 12 日消息,阿里千问今日正式发布 QwenGyre,这是一款面向长程智能体任务的在线强化学习框架。

阿里千问发布 QwenGyre 面向超长程智能体任务的在线强化学习框架

据官方介绍,当智能体承担复杂工作,一次执行往往持续数小时,模型与环境之间有数百轮相互依赖的交互,累计处理百万级 token。QwenGyre 借助弹性调度与轨迹处理,让这类超长程任务成为有效的训练来源,同时带来能力与效率提高。

QwenGyre 可以把 GPU 分成若干计算单元,任务每结束一批,空出的单元就转入训练,就像项目组里谁先交差、谁先加入复盘。还在跑的任务完全不受打扰,进度和状态原地保留,连之前算过的缓存都总体搬走,不必重头再来。新空出的单元直接加入本轮训练,谁算得快谁多干,最后由核心单元汇总结果,完成一次参数更新。

阿里千问发布 QwenGyre 面向超长程智能体任务的在线强化学习框架

IT之家注意到,QwenGyre 可借助 Token-in Token-out 记录每段输出生成时的真实上下文,按共享前缀组织成轨迹树,只把被选中的路径展开为训练样本。同时,超时但部分完成的任务也能借助 Partial Scoring 给出有用反馈,不浪费任何一步努力。

阿里千问发布 QwenGyre 面向超长程智能体任务的在线强化学习框架

跑分维度,官方在 NL2RepoBench 在线强化学习实验中,以每次 rollout 约 70 万 token 的任务开展训练,48 步训练后模型借助率从 52.5% 提高至 58.5%。相同 GPU 预算下,QwenGyre 相对 Async 达成 1.78× 的端到端加速,训练分数与 Async、Colocate 接近。在 Qwen3.6-122B 上的跨任务验证中,NL2RepoBench、DeepSWE、TerminalBench 三类任务均取得端到端提速。

阿里千问发布 QwenGyre 面向超长程智能体任务的在线强化学习框架

本文来源:https://www.ithome.com/1/012/066.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT Home, October 12 — Alibaba's Qwen team officially released QwenGyre today, an online reinforcement learning framework designed for long-horizon agentic tasks.

According to the official introduction, when agents take on complex tasks, each execution often lasts for several hours, involving hundreds of interdependent interaction rounds between the model and the environment, accumulating millions of tokens in processing. QwenGyre leverages elastic scheduling and trajectory processing to make such ultra-long-horizon tasks effective training sources, while simultaneously delivering improvements in both capability and efficiency.

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

QwenGyre can divide GPUs into multiple compute units. Whenever a batch of tasks finishes, the freed units shift into training, much like team members joining a retrospective as soon as they deliver their work. Tasks still running are left completely undisturbed, with their progress and state preserved in place, and previously computed caches migrated wholesale, so there is no need to start over. Newly freed units immediately join the current training round; whoever computes faster takes on more work, and finally a core unit aggregates the results to complete one parameter update.

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

IT Home noted that QwenGyre can use Token-in Token-out to record the actual context when each output segment is generated, organize trajectories into a tree by shared prefixes, and expand only the selected paths into training samples. At the same time, tasks that time out but are partially completed can still receive useful feedback through Partial Scoring, ensuring no effort is wasted.

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

In terms of benchmarking, in the official NL2RepoBench online reinforcement learning experiment, training was conducted on tasks with approximately 700,000 tokens per rollout, and after 48 training steps the model assistance rate increased from 52.5% to 58.5%. Under the same GPU budget, QwenGyre achieved a 1.78× end-to-end speedup over Async, with training scores close to those of Async and Colocate. In cross-task validation on Qwen3.6-122B, all three task categories—NL2RepoBench, DeepSWE, and TerminalBench—achieved end-to-end speedups.

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

Source: https://www.ithome.com/1/012/066.htm If there are any copyright issues, please notify us by email at 13951431913#139.com or submit a comment on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry