SWE-Prime: Fewer Trajectories, Better Performance for SFT Data Selection

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 trains on large-scale agent trajectory datasets of successful trajectories, but task success does not guarantee high-quality supervision: successful trajectories may contain ineffective, redundant, or risky steps, introducing noisy supervision and encouraging undesirable problem-solving behaviors. SWE-Prime progressively filters training data at the trajectory and segment levels. The first stage performs trajectory-level screening based on process quality, result quality, and data representativeness, selecting a high-quality subset of successful trajectories. The second stage groups consecutive steps into semantic segments and evaluates each segment's contribution to the final solution, learnability, and potential risk. 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.

论文概要

研究领域: NLP 作者: Dewu Zheng, Ruizhe Ye, Yanlin Wang, Yang Ye, Hongyu Zhang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jianxing Yu, Zibin Zheng 发布时间: 2026-08-27 arXiv: 2608.27449

Original 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, 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, and data representativeness.

Method Overview

Stage 1 — Trajectory-level screening: selects high-quality and representative successful trajectories based on process quality, result quality, and data representativeness.
Stage 2 — Segment-level selection: groups consecutive steps into semantic segments and evaluates each segment based on its contribution to the final solution, learnability, and potential risk.

Results

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

Key Takeaway

Success alone is not a sufficient filter for SFT supervision: filtering successful agent trajectories at multiple granularities (trajectory and segment) yields better software-engineering performance with far less training data.

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