IFML Seminar: 04/11/25 - Beyond Benchmarks: Building a Science of AI Measurement

Abstract: The widespread deployment of AI systems in critical domains demands more rigorous approaches to evaluating their capabilities and safety. While current evaluation practices rely on static benchmarks, these methods face fundamental efficiency, reliability, and real-world relevance challenges. This talk presents a path toward a measurement framework that bridges established psychometric principles with modern AI evaluation needs. We demonstrate how techniques from Item Response Theory, amortized computation, and predictability analysis can substantially improve the rigor and efficiency of AI evaluation. Through case studies, we show how this approach can enable more reliable, scalable, and meaningful evaluation of AI systems. This work points toward a broader vision: evolving AI evaluation from a collection of benchmarks into a rigorous measurement science that can effectively guide research, deployment, and policy decisions.

 

Bio: Sanmi Koyejo is an assistant professor in the Department of Computer Science at Stanford University and a co-founder of Virtue AI. At Stanford, Koyejo leads the Stanford Trustworthy Artificial Intelligence Research (STAIR) lab, which works to develop the principles and practice of safe and secure AI. Koyejo has received several awards, including a Skip Ellis Early Career Award, a Sloan Fellowship, and a PECASE. Koyejo serves on the Neural Information Processing Systems Foundation Board and as president of Black in AI.