ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Large Language Models #Training Data Optimization #Agentic #Supervised Fine-Tuning #AI Research

SWE-Prime: A Multi-Granularity Approach to Optimizing Training Data for AI Agents

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 21 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:SWE-Prime is a novel method for optimizing training data for AI agents, specifically designed to enhance the ability of large language models to resolve real-world software issues. It employs a two-stage selection process: first, it screens trajectories based on process quality, result quality, and data representativeness; second, it assesses semantic segments for their contribution to the final solution, learnability, and potential risks. Experiments on SWE-Bench Pro and SWE-Bench Verified demo


Background and Motivation

In the realm of enhancing large language models' ability to resolve real-world software issues, traditional supervised fine-tuning (SFT) methods rely on datasets of 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 training introduces noisy supervision and encourages models to imitate undesirable problem-solving behaviors.

The SWE-Prime Method

SWE-Prime introduces a multi-granularity, two-stage SFT data selection method:

  1. Trajectory-Level Screening: Screens trajectories based on process quality, result quality, and data representativeness to select a high-quality and representative subset of successful trajectories.
  2. Segment-Level Selection: Groups consecutive steps into semantic segments and assesses each segment based on its contribution to the final solution, learnability, and potential risks. During training, all segments remain in the sequence, but only selected segments contribute to the loss computation.

Experimental Results

Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on the 10% subset of trajectories selected by SWE-Prime outperforms training on the full dataset, achieving relative performance gains of up to 12.2% and 24.2%, respectively. This demonstrates the method's effectiveness in reducing noisy supervision and improving model performance in complex tasks.

Technical Highlights

  • Multi-Granularity Selection Strategy: Combines trajectory and segment-level screening for a more refined optimization approach.
  • Context Preservation: Retains all segments in the sequence during training, avoiding information loss while optimizing the loss computation.
  • Significant Performance Improvement: Demonstrates strong performance across multiple benchmarks, validating the method's effectiveness.

Industry Impact and Developer Recommendations

SWE-Prime offers a new approach to training AI agents for complex tasks, particularly in scenarios where model performance and reliability are critical. Developers can leverage its multi-granularity selection strategy to optimize their training data and enhance overall model performance. Additionally, this method can be combined with other training techniques to further improve the robustness and adaptability of models.


Source: ArXiv cs.AI (2026-08-27)

— END —

Tags: #Large Language Models #Training Data Optimization #Agentic #Supervised Fine-Tuning #AI Research

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…