PhD Dissertation Defense by SUN Zhensu | Balancing Efficiency and Developer Needs for Sustainable AI-Supported Software Development

ABOUT THE TALK

Large Language Models have moved software development into a human–AI collaborative ecosystem, delivering immediate gains but raising a long-term question: can the practice be sustained as it scales? This dissertation frames that question around three recurring costs — machine computation, developer effort, and human oversight — and presents eight contributions, each of which isolates one cost, makes it measurable, and either eliminates it or exposes the risk it creates. Part I redesigns code representation for AI readers: removing human-oriented formatting cuts 24.6% of input tokens with no loss in generation quality, while SimPy and Token Sugar together remove over 22% more, all reversibly. Part II changes when the model is invoked and how its output reaches the developer, through utility-based request filtering (FrugalCoder), instruction-aware fill-in-the-middle (IFIM), and overlapping generation with execution (Eager). Part III measures what delegation already costs, demonstrating end-to-end compromise of deployed systems through the user-feedback channel (FDI), and showing that LLM-based runtime repair reports roughly twice as many recoveries as it actually achieves (Healer).

ABOUT THE SPEAKER

Zhensu SUN is a Ph.D. candidate at Singapore Management University, advised by Professor David Lo. His research focuses on intelligent software engineering, aiming to bridge the gap between artificial intelligence and the current human-centered software development ecosystem. His work has been published in top-tier conferences and journals, including ICSE, ASE, FSE, ISSTA, CACM, WWW, and TOSEM. His research contributions have received three ACM SIGSOFT Distinguished Paper Awards (one at ISSTA 2024 and two at ICSE 2026) as well as one Distinguished Paper Nomination (ICSE 2022). In addition, his research has been honored with the ByteDance Scholarship (2024).