[论文] SWE-Prime: Fewer Trajectories, Better Performance

论文概要

研究领域: NLP 作者: Dewu Zheng, Ruizhe Ye, Yanlin Wang 发布时间: 2026-08-28 arXiv: 2508.11370

中文摘要

为提升大语言模型解决真实软件问题的能力，先前工作专注于构建大规模智能体轨迹数据集，并对成功轨迹进行监督微调（SFT）。然而，任务成功并不能保证高质量监督：成功轨迹仍可能包含无效、冗余或高风险的步骤。直接使用此类轨迹进行SFT会引入噪声监督，并鼓励模型模仿不良的问题解决行为。因此，我们提出SWE-Prime——一种多粒度、两阶段的SFT数据选择方法，在轨迹和段级别逐步过滤训练数据。具体而言，第一阶段基于过程质量、结果质量和数据代表性进行轨迹级筛选，选择高质量且有代表性的成功轨迹子集。第二阶段通过将连续步骤分组为语义段来进行段级选择，并根据每段对最终解决方案的贡献、可学习性和潜在风险进行评估。在SFT期间，所有段保留在序列中以保持上下文，但只有选定的段参与损失计算。在SWE-Bench Pro和SWE-Bench Verified上的实验表明，使用SWE-Prime选择的10%轨迹子集进行训练优于使用完整已解决数据集进行训练，分别获得高达12.2%和24.2%的相对性能提升。

原文摘要

To improve large language models' ability to resolve real-world software issues, prior work has focused on constructing large-scale agent trajectory datasets and performing supervised fine-tuning (SFT) on successful trajectories. However, task success does not guarantee high-quality supervision: successful trajectories may still contain ineffective, redundant, or risky steps. Directly using such trajectories for SFT can introduce noisy supervision and encourage models to imitate undesirable problem-solving behaviors. Therefore, we propose SWE-Prime, a multi-granularity, two-stage SFT data selection method that progressively filters training data at the trajectory and segment levels. Specifically, the first stage performs trajectory-level screening based on process quality, result quality, ...

#论文 #arXiv #NLP #小凯