CS Special Seminar: Haoye Tian "Trustworthy Automatic Program Repair"

Trustworthy Automatic Program Repair

Haoye Tian
University of Melbourne
Google Scholar

Abstract: AI-driven automated program repair (APR) promises to boost software maintenance efficiency and reduce costs; however, current approaches often generate overfitting patches that, while passing test cases, may deviate from the intended logic and introduce unforeseen risks.

 In this talk, we aim to learn code change semantics to predict patch correctness. We first validate a key hypothesis: programs that fail similar tests often require similar code modifications. Based on this insight, we introduce BATS—a method that predicts patch correctness by statically comparing newly generated patches with previously validated ones. Furthermore, we present Quatrain, a supervised approach leveraging deep natural language processing to assess the correlation between bug reports and patch descriptions, ensuring that the patches effectively address the underlying issues.

Despite these advances, LLM-based program repair still faces systematic application challenges. Finally, this talk envisions achieving trustworthy automatic program repair through enhanced fine-tuning, hierarchical inference, and robust post-verification techniques.

Bio: Haoye Tian was a postdoctoral fellow at University of Melbourne. Prior to that, He was a postdoctoral researcher and pursued his Ph.D. degree in the SnT, University of Luxembourg. During his post in Luxembourg, he visited Carnegie Mellon University as a visiting scholar. He worked as a machine learning researcher at Tencent and Didi. His research interests encompass Software Engineering and Artificial Intelligence and AI for Software Engineering in general, and specifically in leveraging advanced AI techniques to systematically maintain, debug and verify software systems while ensuring security and privacy assurance. He has been collaborating with researchers globally and has published papers at top-tier venues, including ICSE, ASE, TOSEM, EMNLP, AAAI, ISSTA, etc.