Leveraging Large Language Models for Automated Bug Fixing
Shatha Abed Alsaedi, Amin Yousef Mohammad Noaman, Ahmed A. A. Gad-Elrab, Fathy Elbouraey Eassa, Seif Haridi · International Journal of Advanced Computer Science and Applications · 2024
Bug fixing, which is known as Automatic Program Repair (APR), is a significant area of research in the software engineering field. It aims to develop techniques and algorithms to automatically fix bugs and generate fixing patches in the source code. Researchers focus on developing many APR algorithms to enhance software reliability and increase the productivity of developers. In this paper, a novel model for automated bug fixing has been developed leveraging large language models. The proposed model accepts the bug type and the buggy method as inputs and outputs the repaired version of the method. The model can localize the buggy lines, debug the source code, generate the correct patches, and insert them in the correct locations. To evaluate the proposed model, a new dataset which contains 53 Java source code files from four bug classes which are Program Anomaly, GUI, Test-Code and Performance has been presented. The proposed model successfully fixed 49 out of 53 codes using gpt-3.5-turbo and all 53 using gpt-4-0125-preview. The results are notable, with the model achieving accuracies of 92.45% and 100% with gpt-3.5-turbo and gpt-4-0125-preview, respectively. Additionally, the proposed model outperforms several state-of-the-art APR models as it fixes all 40 buggy programs in QuixBugs benchmark dataset.