PReMM: LLM-Based Program Repair for Multi-method Bugs via Divide and Conquer
Linna Xie, Zhong Li, Yu Pei, Zhongzhen Wen, Kui Liu, Tian Zhang, Xuandong Li
Abstract
Large-language models (LLMs) have been leveraged to enhance the capability of automated program repair techniques in recent research. While existing LLM-based program repair techniques compared favorably to other techniques based on heuristics, constraint-solving, and learning in producing high-quality patches, they mainly target bugs that can be corrected by changing a single faulty method, which greatly limits the effectiveness of such techniques in repairing bugs that demand patches spanning across multiple methods. In this work, we propose the PReMM technique to effectively propose patches changing multiple methods. PReMM builds on three core component techniques: the faulty method clustering technique to partition the faulty methods into clusters based on the dependence relationship among them, enabling a divide-and-conquer strategy for the repairing task; the fault context extraction technique to gather extra information about the fault context which can be utilized to better guide the diagnosis of the fault and the generation of correct patches; the dual-agent-based patch generation technique that employs two LLM-based agents with different roles to analyze the fault more precisely and generate patches of higher-quality. We have implemented the PReMM technique into a tool with the same name and applied the tool to repair real-world bugs from datasets Defects4J V1.2 and V2.0. PReMM produced correct patches for 307 bugs in total. Compared with ThinkRepair, the state-of-the-art LLM-based program repair technique, PReMM correctly repaired 102 more bugs, achieving an improvement of 49.8%.
Ask Lune
Know where this paper stands in minutes, not an afternoon.
Lune can read the top-tier papers that cite it and the ones closest to it, and tells you what it claims, what came after and where it falls short, each point cited to the paper it comes from. No skimming, no AI slop.
2 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
its abstract · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
Cited by top-tier papers2
Related papers
- Automated Program Repair in the Era of Large Pre-trained Language ModelsChunqiu Steven Xia, Yuxiang Wei, Lingming ZhangICSE 2023 · 321 citations
- Repair Ingredients Are All You Need: Improving Large Language Model-Based Program Repair via Repair Ingredients SearchJiayi Zhang, Kai Huang, Jian Zhang, Yang Liu et al.ICSE 2026
- RepairAgent: An Autonomous, LLM-Based Agent for Program RepairIslem Bouzenia, Premkumar T. Devanbu, Michael PradelICSE 2025 · 54 citations
- The Plastic Surgery Hypothesis in the Era of Large Language ModelsChunqiu Steven Xia, Yifeng Ding, Lingming ZhangASE 2023 · 26 citations
- ThinkRepair: Self-Directed Automated Program RepairXin Yin, Chao Ni, Shaohua Wang, Zhenhao Li et al.ISSTA 2024 · 37 citations
