Zachary Bezemek | Duke University

About Me

I am an AI research scientist working on generative models, with a focus on training objectives and inference algorithms grounded in stochastic analysis. I currently lead an independent generative AI research program at Duke University, where I am a Phillip Griffith Assistant Research Professor of Mathematics.

My work develops mathematically principled methods for diffusion and masked diffusion language models, improving sampling efficiency and downstream performance across text, code, and scientific generation tasks.

I received my PhD in Mathematics from Boston University, where I studied rare events for interacting particle systems using tools from stochastic control, PDEs, and optimal transport under the advisorship of Konstantinos Spiliopoulos. This background informs my approach to modern machine learning: designing theoretically grounded algorithms that translate into measurable empirical improvements.

Research Interests

Generative Modeling
Machine Learning
Stochastic Analysis

Publications

ICLR2026, Oral

Planner Aware Path Learning in Diffusion Language Models Training

Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R. Zhang, Michael Bronstein, Avishek Joey Bose, Alexander Tong

ICLR2026, Oral (top 1%)

PDF Code

DeLTa@ICLR2025

Path Planning for Masked Diffusion Model Sampling

Fred Zhangzhi Peng, Zachary Bezemek, Sawan Patel, Jarrid Rector-Brooks, Sherwood Yao, Avishek Joey Bose, Alexander Tong, Pranam Chatterjee

DeLTa@ICLR2025 Outstanding Paper Award, and rejected by ICML, NIPS2025, ICLR2026

PDF Code

Importance Sampling for the Empirical Measure of Weakly Interacting Diffusions

Zachary Bezemek, Max Heldman

Applied Mathematics & Optimization, Vol. 89, No. 7, 2024

PDF Code Project Page

Moderate Deviations for Fully Coupled Multiscale Weakly Interacting Particle Systems

Zachary Bezemek, Konstantinos Spiliopoulos

Stochastics and Partial Differential Equations: Analysis and Computations, Vol. 12, 2023

PDF Project Page

Large deviations for Interacting Multiscale Particle Systems

Zachary Bezemek, Konstantinos Spiliopoulos

Stochastic Processes and their Applications, Vol. 155, Issue C, 2023

PDF Project Page

Rate of Homogenization for Fully-Coupled McKean-Vlasov SDEs

Zachary Bezemek, Konstantinos Spiliopoulos

Stochastics and Dynamics, Vol. 23, No. 02, 2023

PDF Project Page

Other Awards

Lewis Blake Award for Excellence in Teaching at Duke University: “Your classroom environment puts students at ease and helps them to focus on their learning. They have written with appreciation and admiration of your clear lectures and helpful explanations. This department has appreciated having you as a reliably strong instructor, and your colleagues have benefited from your contributions to the atmosphere of the department.”
First Place Deep Learning Boot Camp Project (Erdős Institute, 2024, out of 26 teams) for RivusVox Editor – The world’s first near-live zero-shot adaptive speech editing system.
Top 5 Data Science Boot Camp Project (Erdős Institute, 2024, out of 58 teams) for Geo-Locator – A classifying model that can predict in which out of 17 cities a picture was taken.

Miscellanea

I grew up in Metro Detroit working for my parents’ janitorial repair company and got my B.S. from Michigan State University.
I spend most of my free time listening to music and attending live shows. I’m especially into jazz and hip-hop, though I go through phases of listening to anything from desert rock to ambient electronica. I’m into “crate digging” and making sample-based hip-hop and house music with my friends.
I also enjoy table-top role playing games. We mostly play Pathfinder and Starfinder these days.
I love nature and animals, and spend what time I can near running water. In 2026 my partner and I started fostering cats from a local sanctuary, so in addition to our 3 (Habanero, Poblano, and Jalapeño) it’s become common to have 2-6 feline guests living with us.