Yuyuan Chen

I am a first-year math PhD student at Harvard University, where I am fortunate to be advised by Michael Albergo. I am interested in generative modeling and in designing principled, scalable algorithms for generative systems, including flow and diffusion models, robotics, and new alternatives to current generative paradigms.

Before Harvard, I completed a joint BA-MS in mathematics summa cum laude at the University of Chicago, with a minor in computer science. During my undergraduate years, I worked mainly on geometry and topology.

news

top
  1. I am starting as a research intern at FortyFive Labs this summer.
  2. Our paper Discrete Tilt Matching was accepted to ICML 2026.
  3. I am starting as a PhD student at Harvard University, advised by Michael Albergo.

publications

top

Beckmann Transport Models: From Autonomous Flows to One-Step Maps

Lee Cheuk-Kit, Florentin Coeurdoux, Yuyuan Chen, Sophia Tang, Peter Potaptchik, Yilun Du, Michael S. Albergo, Eric Vanden-Eijnden

Preprint

Beckmann Transport Models make flow matching a provably valid autonomous transport to sufficiently lower-dimensional data manifolds, then use an endpoint conservation law to learn one-step generative maps directly.

arXiv

From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models

Seunggeun Kim*, Jaeyeon Kim*, Taekyun Lee*, Yuyuan Chen*, Yilun Du, Sham Kakade, Sitan Chen

Preprint

Masked diffusion models advertise any-order generation, but then quietly decode left-to-right. Positional uncertainty is why, and insertion-based and latent-space masked diffusion are two ways out — a 7B FlexMDM for code and a 125M LatentMDM for GSM8K, freeing the model to decide where, not just what.

arXiv Code

Discrete Tilt Matching

Yuyuan Chen*, Shiyi Wang*, Peter Potaptchik, Jaeyeon Kim, Michael S. Albergo

ICML 2026

A likelihood-free fine-tuning method for masked diffusion language models that matches local unmasking posteriors under reward tilting, with control variates for improved stability.

arXiv Code Poster
Email Google Scholar GitHub