Peng Zhao

peng_zhao.jpg

pzhao [at] udel [dot] edu

214 Townsend Hall

Newark, DE 19716

I am an Assistant Professor in the Department of Applied Economics and Statistics at the University of Delaware and a resident faculty member of the Data Science Institute.

My research develops a theory of deattenuation and anti-shrinkage. I study when statistical procedures systematically attenuate recoverable signal, and when undoing—or even reversing—that attenuation improves prediction or estimation because the signal-recovery gain exceeds the accompanying variance cost.

Before joining Delaware, I was a postdoctoral researcher in Statistics at Texas A&M University, where I worked with Bani K. Mallick, Anirban Bhattacharya, and Debdeep Pati. I received my Ph.D. in Statistics from Florida State University, advised by Yiyuan She and Yun Yang.

My research is currently supported in part by Microsoft Research through the Graph Machine Learning project, where I serve as a co-PI.

For a complete record, please see my publications or download my CV.

News

Sep 24, 2026 Our paper When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation has been accepted to NeurIPS 2026. Congratulations to Nischal!
Sep 16, 2026 Welcome Ira to our research group!
Aug 31, 2026 Our lab has been approved to join Anthropic’s Claude Team plan for Scientists. Thank you to Anthropic for supporting our research!
Aug 14, 2026 New blog post: What Is Negative-Shifted Gradient Descent—When the Useful Estimator Is Not an Endpoint — how a finite-time signed path passes smoothly through the would-be endpoint pole and creates head anti-shrinkage with controlled lower-spectrum exposure.

Selected Publications

  1. When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation
    Nischal Subedi, Cencheng Shen, and Peng Zhao
    In Advances in Neural Information Processing Systems (NeurIPS), 2026
    Accepted
  2. Beyond Negative-Ridge Endpoints: Mixed-Sign Spectral Regularization via Negative-Shifted Gradient Descent
    Peng Zhao
    arXiv preprint arXiv:2607.22474, 2026
  3. De-floored Principal Component Regression: When Rank Selection Alone Is Insufficient for Prediction
    Peng Zhao
    arXiv preprint arXiv:2607.16638, 2026
  4. Structured Optimal Variational Inference for Dynamic Latent Space Models
    Peng Zhao, Anirban Bhattacharya, Debdeep Pati, and Bani K. Mallick
    Journal of Machine Learning Research, 2024
  5. High-dimensional Linear Regression via Implicit Regularization
    Peng Zhao, Yun Yang, and Qiao-Chu He
    Biometrika, 2022