Please note: This seminar will take place in DC 1304.
Keyon Vafa, Postdoctoral Fellow
Harvard University
Large language models (LLMs) and other generative models in AI have achieved remarkable performance on benchmarks and other sandboxed tasks, yet this success often fails to predict how a model will behave in real-world settings. This gap raises a broader question: what would it mean for a generative model to acquire a structural understanding of its domain, i.e., a good implicit world model?
In this talk, I will present three computational notions of an implicit world model, ranging from formal definitions grounded in language theory to behavioral criteria based on human understanding. These notions motivate new methods for testing and improving world models, which I will illustrate in domains ranging from medical diagnosis to Newtonian mechanics. I will conclude by discussing how these ideas fit into a broader agenda of building generative models that do what we think they do.
Bio: Keyon Vafa is a postdoctoral fellow at Harvard University and an affiliate with the Laboratory for Information & Decision Systems at MIT. His research focuses on building tools to make generative models in AI more robust in real-world settings.
Keyon completed his PhD in computer science from Columbia University, where he was an NSF GRFP Fellow and the recipient of the Morton B. Friedman Memorial Prize for excellence in engineering. He has organized the NeurIPS 2024 Workshop on Behavioral Machine Learning and the ICML 2025 Workshop on Assessing World Models, and serves on the Early Career Board of the Harvard Data Science Review.