Changxiao Cai
 |
I am an assistant professor of Industrial and Operations Engineering and an affiliate faculty member of Electrical and Computer
Engineering at University of Michigan–Ann Arbor. Previously, I was a postdoctoral researcher at University of Pennsylvania, advised by T. Tony Cai and Hongzhe Li. I obtained my Ph.D. in Electrical Engineering at Princeton University in 2021, advised by Yuxin Chen and H. Vincent Poor. I have also worked with Sergio Verdú at Princeton. Prior to this, I received my B.E. in Electronic Engineering from Tsinghua University in 2016. I am interested in the theoretical and algorithmic foundations of machine learning, with a recent focus on generative AI and diffusion models, and their applications. My research is supported by an NSF CAREER Award (CCF) and an NSF grant (DMS). E-mail: cxcai [at] umich [dot] edu |
News
New paper on a fully continuous flow-based LM without a CE-trained decoder: ConvergeFlow: Language Flow with Provable Convergence to Token Embeddings
New paper on provable efficiency of confidence-based decoding for masked diffusion LMs: Confidence-Based Decoding is Provably Efficient for Diffusion Language Models accepted to NeurIPS 2026
New paper on low-dimensional adaptation in diffusion models: Diffusion Models Are Statistically Optimal for Learning Low-Dimensional Multi-Modal Distributions accepted to ICML 2026
New paper on accelerating diffusion sampling via forward-value discretization: Are First-Order Diffusion Samplers Really Slower? A Fast Forward-Value Approach accepted to ICML 2026
New paper on sampling convergence of masked diffusion LMs: Breaking AR's Sampling Bottleneck: Provable Acceleration via Diffusion Language Models accepted to NeurIPS 2025
New paper on (nearly) dimension-free sampling convergence of DDPMs: Dimension-Free Convergence of Diffusion Models for Approximate Gaussian Mixtures accepted to ICML 2026
New paper on statistical optimality of DDIMs: Minimax Optimality of the Probability Flow ODE for Diffusion Models
New paper on a provable training-free accelerated sampler for diffusion models: Provable Acceleration for Diffusion Models under Minimal Assumptions