“Scientific Inference with Diffusion Generative Models”

Diffusion models have revolutionized generative modeling in vision and language. But can they also serve as tools for scientific inference? I present a Bayesian view on diffusion models as tools for solving inverse problems—inferring hidden states or parameters from noisy observational data—with applications ranging from climate modeling to imaging. Unlike creative applications, scientific use cases demand more than photorealism: they require calibrated uncertainty, faithful distributions, efficient conditional sampling, and the capacity to represent extreme and heavy-tailed data. I will highlight four recent advances addressing these needs: Variational Control for expressive conditional generation (ICML ’25), Heavy-Tailed Diffusion Models for sparse and extreme-valued data (ICLR ’25), Conjugate Integrators for fast conditional sampling without retraining (NeurIPS ’24), and Generative Uncertainty for assessing epistemic uncertainty in generation (UAI ’25). Together, these results point to diffusion models as principled engines of scientific inference.

Stephan Mandt is an Associate Professor of Computer Science and Statistics at the University of California, Irvine. His research focuses on the foundations and applications of generative AI, including probabilistic generative modeling of 2D, 3D, and sequential data; statistical inference algorithms; resource-efficient learning; and AI for scientific discovery. He is a Chan Zuckerberg Initiative Investigator and AI Resident, and has received the NSF CAREER Award, the UCI ICS Mid-Career Excellence in Research Award, and a Kavli Fellowship. Prior to joining UCI, he led the machine learning group at Disney Research and held postdoctoral positions at Princeton University and Columbia University. He frequently serves as a Senior Area Chair for NeurIPS, ICML, and ICLR, and has recently been Program Chair (AISTATS 2024) and General Chair (AISTATS 2025).