SWE-Prime: Fewer Trajectories, Better Performance for SWE Agent SFT

Forum topic · 小凯 · 2026-08-29

Summary

SWE-Prime is a two-stage, multi-granularity data selection method for supervised fine-tuning (SFT) of large language models on software engineering tasks. The authors observe that task success does not guarantee high-quality supervision: successful agent trajectories may still contain ineffective, redundant, or risky steps, and training directly on them introduces noisy supervision. Stage one filters trajectories based on process quality, result quality, and data representativeness. Stage two groups consecutive steps into semantic segments and evaluates each segment's contribution to the final solution, learnability, and potential risk; all segments remain in the training sequence for context, but only selected segments contribute to the loss. Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on a 10% trajectory subset selected by SWE-Prime outperforms training on the full resolved dataset, yielding relative improvements of up to 12.2% and 24.2% respectively. Paper: arXiv 2508.11370.

Paper Overview

Field: NLP Authors: Dewu Zheng, Ruizhe Ye, Yanlin Wang Published: 2026-08-28 arXiv: 2508.11370

Abstract

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.

Method: SWE-Prime

SWE-Prime is a multi-granularity, two-stage SFT data selection method that progressively filters training data at the trajectory and segment levels:

1. Trajectory-level screening (Stage 1): Selects a high-quality, representative subset of successful trajectories based on process quality, result quality, and data representativeness. 2. Segment-level selection (Stage 2): Groups consecutive steps into semantic segments and evaluates each segment by its contribution to the final solution, learnability, and potential risk. During SFT, all segments are kept in the sequence to preserve context, but only selected segments participate in the loss computation.

Results

Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on a 10% trajectory subset selected by SWE-Prime outperforms training on the full resolved dataset, achieving relative performance improvements of up to 12.2% and 24.2%, respectively.