Qwen发布QwenGyre：面向超长任务的弹性强化学习框架
Qwen released QwenGyre - RL framework for training AI agents on extremely long complex tasks
推荐理由
Agent训练是当前的热点难点，QwenGyre提供了具体的弹性调度和轨迹处理工程方案，数据详实且开源，对做长程Agent训练的同学有直接参考价值。
Introducing QwenGyre, bringing elastic RL to extreme-long-horizon (XLong) agent training.
介绍 QwenGyre，将弹性强化学习（RL）引入超长视界（XLong）智能体训练。
It launches with two core designs:
它包含两项核心设计：
- Elastic Scheduler: reallocates GPUs between rollout & training as demand shifts—live executions never interrupted. Up to 1.85× speedup over Colocate, 1.78× over Async.
- Trajectory Processor: turns branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples. Partial progress after timeout still gets scored.
- 弹性调度器：根据需求在 rollout 与训练之间重新分配 GPU——实时执行永不中断。相比 Colocate 方案最高提速 1.85 倍，相比 Async 方案提速 1.78 倍。
- 轨迹处理器：将分支执行过程（压缩、子智能体、重试）转化为前缀共享的轨迹树，并生成有界且去重的训练样本。即使超时后仍有部分进展，也会获得评分。
Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps, cutting end-to-end training time by ~44% vs Async.
在 Qwen3.8-2.4T-A95B 模型上扩展至每次 rollout 约 70 万 token，QwenGyre 在 48 步内将 NL2RepoBench 指标从 52.5% 提升至 58.5%，端到端训练时间较 Async 方案缩短约 44%。
🔗 Read the paper:
🔗 阅读论文：
arxiv.org/abs/2609.33848
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力