Hi, welcome to my corner of the Internet!
I am a Ph.D. student in Computer Science at UT Austin, advised by Kevin Tian and Eric Price. Before joining UT Austin, I received my bachelor's degree in Software Engineering from Tsinghua University.
My research focuses on the theoretical foundations of generative models. I am also broadly interested in high-dimensional statistics, learning theory, and trustworthy machine learning.
Recently, I have been particularly interested in the question: How can we effectively and reliably evaluate and audit the behavior of modern generative models? As these models tackle increasingly complex and challenging tasks, much of the research focuses on expanding and improving their capabilities. This progress calls for principled evaluation and auditing methods that help us understand what these models can do, where they fail, and whether their behavior meets our expectations.
For example, my recent work explores the following questions:
When a metric relies on distributional assumptions, are standard estimators accurate under those assumptions, and can we do better?
For the widely used Fréchet Inception Distance (FID), the standard empirical estimator can exhibit substantial systematic bias even when the underlying Gaussian assumptions hold. Reproducible scores can therefore remain far from the population quantity the metric is intended to measure. In our work, we give a sharp theoretical characterization of this estimation problem and develop a computationally efficient estimator with optimal sample complexity under a Gaussian model. We also demonstrate improved estimation accuracy on ImageNet. Paper
Can we use the structure of modern generative models to measure distributional differences efficiently, without imposing parametric assumptions on their output distributions?
In our work on autoregressive models, we use their sequential prediction and sampling structure to develop efficient estimators of total variation distance with provable guarantees. We apply these methods to quantify differences between the output distributions of different LLM deployments, including vLLM and SGLang serving identical model weights. This provides a direct way to assess how deployment choices affect model behavior. Paper
Authors are listed in alphabetical order unless otherwise noted.
@misc{chen2026sampleoptimalestimationfrechetinception,
title={Sample-Optimal Estimation of the Fr{\'e}chet Inception Distance},
author={Ziyun Chen and Jerry Li and Kevin Tian and Yusong Zhu},
year={2026},
eprint={2610.07114},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2610.07114}
}
@misc{kumar2026highmagnetizationsamplinglowtemperatures,
title={High-Magnetization Sampling at Low Temperatures: Ising Models and Bayesian Sparse Linear Regression},
author={Syamantak Kumar and Purnamrita Sarkar and Kevin Tian and Yusong Zhu},
year={2026},
eprint={2609.08873},
archivePrefix={arXiv},
primaryClass={cs.DS},
url={https://arxiv.org/abs/2609.08873}
}
* First author; authors are not listed alphabetically.
@misc{modi2026revisitingprovableauditableprivacygap,
title={Revisiting the Provable-Auditable Privacy Gap of {DP-SGD}},
author={Saloni Modi and Srivi Balaji and Yusong Zhu and Gautam Kamath and Kevin Tian},
year={2026},
eprint={2608.28934},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2608.28934}
}
Workshop Preliminary version presented at the ICML 2026 Hypothesis Testing Workshop.
@misc{price2026totalvariationdistanceestimation,
title={Total Variation Distance Estimation in Autoregressive Models},
author={Eric Price and Kevin Tian and Zhiyang Xun and Yusong Zhu},
year={2026},
eprint={2607.19510},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2607.19510}
}
@misc{gu2026tractabilitylandscapesamplinginexact,
title={The Tractability Landscape of Sampling with Inexact Scores},
author={Anming Gu and Kevin Tian and Hubert Yang and Yusong Zhu},
year={2026},
eprint={2607.19004},
archivePrefix={arXiv},
primaryClass={stat.ML},
url={https://arxiv.org/abs/2607.19004}
}
@InProceedings{pmlr-v336-chen26b,
title={Separating Oblivious and Adaptive Models of Variable Selection (Extended Abstract)},
author={Chen, Ziyun and Li, Jerry and Tian, Kevin and Zhu, Yusong},
booktitle={Proceedings of Thirty Ninth Conference on Learning Theory},
pages={1113--1114},
year={2026},
editor={Hanneke, Steve and Lattimore, Tor},
volume={336},
series={Proceedings of Machine Learning Research},
month={29 Jun--03 Jul},
publisher={PMLR},
pdf={https://raw.githubusercontent.com/mlresearch/v336/main/assets/chen26b/chen26b.pdf},
url={https://proceedings.mlr.press/v336/chen26b.html}
}
@InProceedings{pmlr-v291-kumar25a,
title={Spike-and-Slab Posterior Sampling in High Dimensions},
author={Kumar, Syamantak and Sarkar, Purnamrita and Tian, Kevin and Zhu, Yusong},
booktitle={Proceedings of Thirty Eighth Conference on Learning Theory},
pages={3407--3462},
year={2025},
editor={Haghtalab, Nika and Moitra, Ankur},
volume={291},
series={Proceedings of Machine Learning Research},
month={30 Jun--04 Jul},
publisher={PMLR},
pdf={https://raw.githubusercontent.com/mlresearch/v291/main/assets/kumar25a/kumar25a.pdf},
url={https://proceedings.mlr.press/v291/kumar25a.html}
}
@inproceedings{NEURIPS2023_2b00b333,
author={Cheng, Xiang and Wang, Bohan and Zhang, Jingzhao and Zhu, Yusong},
booktitle={Advances in Neural Information Processing Systems},
doi={10.52202/075280-0587},
editor={A. Oh and T. Naumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},
pages={13374--13394},
publisher={Curran Associates, Inc.},
title={Fast Conditional Mixing of MCMC Algorithms for Non-log-concave Distributions},
url={https://proceedings.neurips.cc/paper_files/paper/2023/file/2b00b3331bd0f5fbfdd966ac06338f6d-Paper-Conference.pdf},
volume={36},
year={2023}
}
ICML 2026 (Gold Reviewer), FOCS 2026, and NeurIPS 2026.
Ph.D. student in Computer Science
Advisors: Kevin Tian and Eric Price
B.S. in Software Engineering
Minor in Statistics
