Repairing Code Defects Through Multiple LLM Collaboration and Failure Information Capturing

In recent years, the advent of large language models (LLMs) has significantly advanced automatic program repair (APR). However, due to a lack of domain-specific software engineering knowledge, base LLMs are prone to hallucinations during the repair process, which compromises the correctness of the generated patches. Although training or fine-tuning domain-specific LLMs can mitigate this issue, it is prohibitively expensive. In this paper, we propose a novel APR model based on multiple LLM collaboration and failure information capturing. Without requiring fine-tuning or retrieval-augmented generation (RAG), our model enhances the APR capabilities of base LLMs. It is notable that we utilize sandboxed executions to extract failure contexts for bug root-cause analysis and design a hierarchical coarse-to-fine (i.e., method-to-block) bug localization to constrain edit scopes. By isolating patch generation from irrelevant code, our approach substantially mitigates incorrect patch generation. Evaluations on Defects4J (V1.2 & V2.0), QuixBugs-Java, and HumanEval-Java datasets demonstrate that the proposed DeepSeek-backed model outperforms baseline models, yielding 192, 182, 40, and 158 correct patches, respectively.