ABOUT THE TALK Software vulnerabilities are widespread in modern systems, threatening reliability, safety, and maintainability across sectors from healthcare to transportation, and imposing substantial economic burdens and developer effort. Automated Program Repair (APR) aims to mitigate these risks by generating fixes automatically, reducing both time and human effort. Recent Large Language Models (LLMs) have shown promising performance in APR and, in some cases, surpass traditional techniques. However, practical deployment exposes several gaps. First, LLM performance on APR tasks is suboptimal in data-scarce settings. Second, benchmark accuracy alone does not demonstrate genuine bug understanding, and the models’ black box nature hampers human interpretability, undermining trust when stated explanations diverge from the actual fix strategies. Third, domain-specific semantics, such as Rust and C/C++ interoperability and C-to-Rust translation, are underrepresented in general benchmarks, and their specific characteristics remain underexplored, which obscures real-world failure modes, hinders the development of LLMs and tools, and limits the adoption of APR. Together, these gaps restrict the reliability, interpretability, and practical use of LLM-based APR at scale.
To address these issues, this dissertation proposes to (i) introduce knowledge-guided prompt strategies that substantially stabilize LLM repair in low-resource settings, (ii) develop model-agnostic tasks and metrics to assess LLM explainability, (iii) design a context optimization strategy to improve both repair accuracy and explanation quality, (iv) conduct an empirical analysis of vulnerabilities in Rust and C/C++ interoperability tools to derive a fine-grained taxonomy and provide practical implications for practitioners, and (v) characterize error patterns in LLM-translated Rust code. |
SPEAKER BIOGRAPHY CAI Xuemeng is a third-year PhD candidate and a Research Engineer at the Centre for Research on Intelligent Software Engineering (RISE) at Singapore Management University (SMU). Her research interests focus on software engineering challenges such as automatic program repair, code translation, and the interpretability of Large Language Models. |
