Keynotes

The rapid ascent of Large Language Models (LLMs) has introduced what many perceive as a pivotal crossroads for software engineering: a choice between the unprecedented velocity of generative AI and the rigorous reliability of formal modeling. In this talk, I argue that the future of the field lies in a symbiotic "neuro-symbolic" relationship where these two paradigms are no longer viewed as alternatives, but as essential partners. We first examine an emerging shift in the modeling lifecycle: the transition from manual, high-overhead specification to AI-assisted synthesis. By using LLMs as catalysts to automatically generate high-fidelity artifacts, we can mitigate the modeling overhead that has historically been a challenge for static analysis and formal verification. Conversely, we explore the indispensable role of models as the "safety brain" for generative AI. Formal reasoning and program invariants can be leveraged to create runtime guardrails, transforming stochastic AI outputs into more verifiable, reliable systems. Through several case studies, this talk aims to provide a call-to-arms to realize a future where the scale and intuition of LLMs work in tandem with the precision and authority of models to define the next generation of trustworthy software engineering.

With GenAI able to do so much in building software, our field is facing an upheaval and questions on many levels. By looking at earlier revolutions in software development, we can identify potential answers: things that have weathered these previous storms and remained unchanged. Key invariants we will examine include human cognition, the power of language, the astonishing growth of computing power, the tyranny of scale, and the similarities and differences of human and machine. By lifting our eyes from extraneous details to the big picture, and delving deeper into the unchanging foundations of our endeavours, we can plot a well-founded course towards 2056.

Controlled experiments are a cornerstone of software engineering research, yet the reliability of their conclusions depends critically on how they are designed, executed, and analyzed. In this keynote, I will examine common pitfalls that undermine the credibility of software engineering experiments, focusing on key dimensions of validity, including internal, construct, and statistical conclusion validity. Drawing on examples, I will show how seemingly reasonable experimental choices—ranging from variable operationalization to subject selection, experimental design and data analysis—can lead to misleading or overconfident conclusions. These challenges arise across different kinds of experiments, from those involving human participants to those based on complex artifacts and toolchains, where small design decisions can have large, often unnoticed effects on results. Building on these observations, I will discuss practical strategies for designing and evaluating experiments that produce more trustworthy and informative results. The goal is not only to avoid common pitfalls, but to strengthen the scientific contribution of software engineering experiments and support more robust and cumulative progress in the field.