Research on Program Automatic Repair Method Combining Context Optimization Strategy and Large Language Models

Automated program repair techniques address software errors, vulnerabilities, and defects through automation. With the rapid development of deep learning, deep learning-based automated repair techniques have improved repair performance but still face challenges of high data demands and low accuracy. The emergence of large language models offers new solutions. This paper proposes a program automated repair method called CodeFixer, which combines contextual optimization strategies and large language models. This method utilizes a Tree-LSTM model to learn the relationship between erroneous statements and their contexts, providing relevant contextual information to the large language model. By employing prompt engineering, the program repair process is divided into error analysis and patch generation stages. During patch generation, the quality and reliability of patches are enhanced by integrating tree-of-thoughts and model fine-tuning, and candidate patches are optimized through an automated evaluation consistency strategy. To evaluate the performance of CodeFixer, experiments were conducted on 395 errors in the Defects4J V1.2 dataset, successfully repairing 92 errors. Validation results on the QuixBugs dataset in both Java and Python programming languages also outperformed existing automated repair techniques.