Date: October 23, 2026
Speaker: Ruqi Zhang, Assistant Professor, Department of Computer Science, Purdue University
Title: Continuous Methods for Discrete Spaces: From Sampling to Diffusion Language Models
Abstract: Discrete data is everywhere, from language and code to molecules and combinatorial structures. However, many powerful tools in modern machine learning, such as gradients, Langevin dynamics, and score-based diffusion, are built for continuous spaces. This talk explores how gradients of continuous extensions can bring these methods to discrete spaces, for both probabilistic sampling and generative modeling.
For sampling, I will present discrete Langevin samplers, which use gradients to make large, informed jumps for unnormalized discrete distributions. For generative modeling, I will focus on diffusion language models and explain why continuous diffusion underperforms on text. This analysis motivates CANDI, a hybrid discrete-continuous diffusion that enables guidance with off-the-shelf classifiers and few-step generation. Finally, I will introduce Sticky Jump Diffusions, which unify discrete, continuous, and hybrid diffusion as limits of a single process and reveal new design axes for diffusion language models.
Bio: Ruqi Zhang is an Assistant Professor in the Department of Computer Science at Purdue University. Her research focuses on probabilistic machine learning, generative modeling, and trustworthy AI. Prior to joining Purdue, she was a postdoctoral researcher at the Institute for Foundations of Machine Learning (IFML) at the University of Texas at Austin, and she received her Ph.D. from Cornell University. She has served as a key organizer of the Symposium on Probabilistic Machine Learning. Her work has been recognized with the AAAI New Faculty Highlights, Amazon Research Award, Spotlight Rising Star in Data Science, Seed for Success Acorn Award, and Ross-Lynn Research Scholar.