PhD Dissertation Proposal by LYU Yunbo | Effective AI for Software Engineering: From Human Insights to Agentic Solutions

ABOUT THE TALK

AI is changing not only how software is written, but also who—or what—participates in writing it. AI coding assistants are now routine fixtures of the IDE, while a newer class of systems—software engineering (SE) agents—can plan, edit, execute, and revise code across entire repositories with limited human oversight. Yet adoption has outpaced understanding. Benchmarks measure whether a system completes a predefined task, but reveal much less about what users value, how practitioners build these systems, or where established automation fails on real development histories.

This dissertation proposal argues that effective AI for software engineering must begin with an understanding of the humans who use and build these systems. Human insights identify what matters in practice, empirical analysis makes the limitations of existing automation concrete, and agentic techniques provide a means of responding to both. Four studies develop this argument: a large-scale analysis of user reviews showing that a central complaint about AI coding assistants concerns obtaining the right context rather than interpreting the code already provided; a mixed-methods study of SE-agent builders, finding that cheaper implementation relocates engineering bottlenecks rather than removing them; an evaluation of SZZ algorithms against more than 76,000 developer-recorded Linux kernel links, exposing the limits of their standard blame-based heuristic; and AgentSZZ, an LLM-based agent that investigates repository history iteratively and improves F1 by up to 27.2% over the previous best-performing method.

ABOUT THE SPEAKER

LYU Yunbo is a PhD candidate in Computer Science at the School of Computing and Information Systems, Singapore Management University, supervised by Prof. David Lo. His research focuses on the intersection of Software Engineering and AI, taking a user- and developer-centered view of AI4SE systems such as AI coding assistants and SE agents: he empirically studies what users need and where builders struggle, then designs agentic techniques to close the gaps. His research has led to multiple publications at top-tier SE and AI venues, including ICSE, ASE, ISSTA, TSE, and ACL. His work has been recognized with several academic honors, including ACM SIGSOFT Distinguished Paper Award (ASE’25), IEEE Computer Society TCSE Distinguished Paper Award (SANER’25), and an OpenAI Researcher Access Program Grant. At SMU, he has also received the SMU PDF (2024), the PhD REA (Tier 1) (2025, 2026), and the SCIS Dean's List Award (2026).