MUARF: Leveraging Multi-Agent Workflows for Automated Code Refactoring

Refactoring is crucial for maintaining a project, but it requires developers to understand code structure and system design principles well. Recent research on Large Language Models(LLMs) has shown their great capability for handling complex tasks, making them a possible solution for overcoming thes...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Proceedings (IEEE/ACM International Conference on Software Engineering Companion. Online) S. 226 - 227
1. Verfasser: Xu, Yisen
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 27.04.2025
Schlagworte:
Code Refactoring
Codes
Contextual Retrieval-Augmented Generation
Few shot learning
Large Language Model
Large language models
Measurement
Multi-Agent Communication
Prompt engineering
Retrieval augmented generation
Software engineering
System analysis and design
ISSN:2574-1934
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Refactoring is crucial for maintaining a project, but it requires developers to understand code structure and system design principles well. Recent research on Large Language Models(LLMs) has shown their great capability for handling complex tasks, making them a possible solution for overcoming these challenges. In this paper, we propose MUARF, an LLM-based solution designed to automate method-level code refactoring, aiming to generate correct, high-quality, and human-like refactored code. MUARF leverages Contextual Retrieval-Augmented Generation to search for similar refactoring samples for few-shot learning, uses Multi-Agent Workflow to simulate the human refactoring process, and integrates advanced software engineering tools (e.g., RefactoringMiner, PurityChecker, StyleChecker) to assist refactoring. Evaluation results show that MUARF achieves a compilation pass rate of 86.5% and a test success rate of 83.8% for the refactored code it generates. Additionally, metrics such as CodeBLEU score and AST Diff accuracy-which compare human-refactored code with the output of MUARF -highlight the generated code is human-like. The ablation results show that RefactoringMiner and Agentware made the greatest contribution to MUARF.
ISSN:2574-1934
DOI:10.1109/ICSE-Companion66252.2025.00071