清华与通义千问团队论文:重构环境训练代码Agent优于轨迹模仿
New Tsinghua + Qwen Team's paper flips the usual agent-data recipe:
代码 Agent 训练的核心痛点是数据质量,这篇论文提出的“重构环境”思路直接解决了轨迹数据稀缺和不可复用的问题,实验提升显著,值得 Agent 开发者参考其数据构建逻辑。
New Tsinghua + Qwen Team's paper flips the usual agent-data recipe:
清华大学与通义千问团队的新论文颠覆了常规的智能体-数据配方:
Reconstructing and re-solving old agent workspaces trains better coding agents than imitating the original runs
重建并重新求解旧的智能体工作区,比模仿原始运行过程能训练出更优秀的编码智能体
A trajectory is like a recording of 1 coding agent fixing 1 bug: you can only copy what that agent did.
轨迹就像记录了一个编码智能体修复一个 bug 的过程:你只能复制该智能体的行为。
The paper instead rebuilds the actual code workspace from that recording, so you can give the same workspace new bugs, new features, or a stronger agent and generate many fresh training examples from it.
相反,该论文从该记录中重建实际的代码工作区,因此你可以为同一工作区添加新的 bug、新功能或更强的智能体,并从中生成大量新鲜的训练样本。
That is the core difference: trajectory = 1 frozen solution; environment = a reusable codebase that can produce many new verified tasks, and the paper finds the second is much better for training agents.
这就是核心差异:轨迹 = 一个冻结的解决方案;环境 = 一个可重用的代码库,能够产生许多新的已验证任务,而论文发现第二种方式对训练智能体效果更好。
Training Qwen3.5-27B on the resulting data moved Terminal-Bench 2.1 from 46.2% to 58.1% and EvoCode-Bench v2 MT@4 from 6.3 to 20.1 under the paper's Terminus2 setup.
在 Terminus2 设置下,使用该论文生成的数据训练 Qwen3.5-27B,使 Terminal-Bench 2.1 从 46.2% 提升至 58.1%,EvoCode-Bench v2 MT@4 从 6.3 提升至 20.1。
collect execution environments you can reuse, not just more traces you can imitate.
收集你可重用的执行环境,而非仅仅收集更多可供模仿的轨迹。
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