SWE-Prime: Fewer Trajectories, Better Performance

Forum topic · 小凯 · 2026-08-30

Summary

SWE-Prime is a multi-granularity, two-stage supervised fine-tuning (SFT) data selection method for improving large language models' ability to resolve real-world software issues. Prior work builds large-scale agent trajectory datasets and fine-tunes on successful trajectories, but task success does not guarantee high-quality supervision: successful trajectories may still contain ineffective, redundant, or risky steps, introducing noisy supervision and encouraging undesirable problem-solving behaviors. SWE-Prime progressively filters training data at two levels. Stage one performs trajectory-level screening based on process quality, result quality, and data representativeness to select a high-quality, representative subset of successful trajectories. Stage two groups consecutive steps into semantic segments and evaluates each segment by its contribution to the final solution, learnability, and potential risk for segment-level selection. 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 gains of up to 12.2% and 24.2% respectively. Paper: arXiv 2608.27449.

Paper Overview

Field: NLP Authors: Dewu Zheng, Ruizhe Ye, Yanlin Wang, Yang Ye, Hongyu Zhang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jianxing Yu, Zibin Zheng Published: 2026-08-27 arXiv: 2608.27449

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.

Therefore, the authors propose SWE-Prime, 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.

Results

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

--- *Auto-collected on 2026-08-30*