Peihan Liu
CS PhD student, Columbia · peihanliu@cs.columbia.edu · Scholar
I'm a second-year CS PhD student in the Columbia theory group, advised by Rachel Cummings and Roxana Geambasu. I work on trustworthy machine learning — particularly the theoretical and empirical foundations of privacy, fairness, and modern machine learning.
Previously, I received my M.Eng. in CSE from Harvard, where I worked on algorithmic fairness with Cynthia Dwork and Juan Perdomo. Before that, I earned B.S. degrees in Mathematics and Statistics, with high honors and high distinction, from the University of Michigan, where I worked with Martin Strauss, Ranjan Pal, Shizhang Li, and Nuh Aydin on algorithmic fairness, simplicial algebra, and algebraic coding theory.
Beyond research, I walk my dog and (used to) play poker.
Experience
-
Fall 2025
Student Researcher, Google
hosted by Alex Bie and Lily Tsai -
Summer 2025
Student Researcher, Google
hosted by Alex Bie and Lily Tsai -
Summer 2023
Student Intern, OpenDP
supervised by Salil Vadhan and Wanrong Zhang
Publications
See Google Scholar for the up-to-date list.
-
ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?
Peihan Liu, Lucas Rosenblatt, Weiwei Kong, Natalia Ponomareva, Gautam Kamath, Rachel Cummings, Roxana Geambasu, Yu Gan, Lillian Tsai, Alex Bie
preprint, 2026. [arXiv] [code] [dataset] -
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
Ved Sriraman, Peihan Liu, Daniel Hsu, Adam Block
preprint, 2026. [arXiv] [code] -
Adaptive Target-Charging with Privacy Filters and Individual Accounting
Peihan Liu, Alison Caulfield, Mark Chen, Rachel Cummings, Roxana Geambasu, Mathias Lécuyer, Pierre Tholoniat
preprint, 2026. [arXiv] [code] -
Privately Fine-Tuned LLMs Preserve Temporal Dynamics in Tabular Data
Lucas Rosenblatt, Peihan Liu, Ryan McKenna, Natalia Ponomareva
ICML, 2026. [arXiv] -
Convex Relaxation for Solving Large-Margin Classifiers in Hyperbolic Space
Sheng Yang, Peihan Liu, Cengiz Pehlevan
TMLR, 2024. [arXiv] -
Safeguarding Data in Multimodal AI: A Differentially Private Approach to CLIP Training
Alyssa Huang, Peihan Liu, Ryumei Nakada, Linjun Zhang, Wanrong Zhang
preprint, 2024. [arXiv]
Blog
-
2026-05-14
Release of ContinuousBench
Does DP synthetic text actually transfer knowledge? A new benchmark says no — even at ε = 100. Read →