Qi Lei (雷琦)
 |
Assistant Professor of Mathematics and Data Science, and, by courtesy, Email: ql518 at nyu.edu |
Research Overview
My research aims to bridge the theoretical and empirical boundary of modern machine learning algorithms and in particular AI safety, with a focus on data privacy, distributionally robust algorithms, sample- and parameter-efficient learning.
Recent research highlights: (Weak-to-Strong Generalization), (Data Reconstruction Attack and Defense), (Data and Model Pruning), (Theoretical Foundations of Pre-trained Models)
(Curriculum Vitae, Github, Google Scholar)
Advertisement
I am actively looking for self-motivated and proactive students to work with. You are welcome to shoot me an email with your CV and short research plans/interests. (You may refer to this link to see whether our research interests match.)
For Ph.D. applicants, please apply to Courant Mathematics or Center for Data Science whichever you see fit and mention my name in your application. For students admitted through Courant CS, I can only co-advise rather than serve as sole advisor.
For prospective post-doc applicants, I encourage you to apply for the positions of CDS Faculty Fellows, Courant Instructors, and Flatiron Research Fellows.
News and Announcement
04/2026 Papers accepted at ICML 2026:
Chenruo Liu, Yijun Dong, Yiqiu Shen, Qi Lei. A Task-Centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
Yujie Shen, Zihan Wang, Jian Qian, Qi Lei. Data Reconstruction: Identifiability and Optimization with Sample Splitting
01/2026 Paper accepted at ICLR 2026:
Chenruo Liu, Yijun Dong, Qi Lei. Does Weak-to-strong Generalization Happen under Spurious Correlations?
01/2026 Invited talk at NJIT on weak-to-strong generalization
01/2026 Invited talk at LASR workshop@London on Data Reconstruction Attacks and Induced Optimal Defense
09/2025 Papers accepted at NeurIPS 2025:
Shikai Qiu, Zixi Chen, Hoang Phan, Qi Lei, Andrew Gordon Wilson. Hyperparameter Transfer Enables Consistent Gains of Matrix-Preconditioned Optimizers Across Scales
Victor Li, Baiting Chen, Yuzhen Mao, Qi Lei, Zhun Deng, Performative Risk Control: Calibrating Models for Reliable Deployment under Performativity
09/2025 Invited talk at Maryland Numerical Analysis Group on Virtues and Pitfalls of Weak-to-Strong Generalization: Intrinsic Dimension and Spurious Correlation
08/2025 Invited talk at the Inaugural Workshop on Frontiers in Statistical Machine Learning on Weak-to-Strong Generalization
07/2025 Invited talk at Inverse Methods for Complex Systems under Uncertainty Workshop on Data Reconstruction Attacks in AI models are Inverse Problems
Selected Papers
8. Yijun Dong, Yicheng Li, Yunai Li, Jason D Lee, Qi Lei, Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension, ICML 2025
7. Sheng Liu*, Zihan Wang*, Yuxiao Chen, Qi Lei, “Data Reconstruction Attacks and Defenses: A Systematic Evaluation”, AISTATS 2025
6. Zihan Wang, Jason Lee, Qi Lei. “Reconstructing Training Data from Model Gradient, Provably”, AISTATS 2023
5. Baihe Huang*, Kaixuan Huang*, Sham M. Kakade*, Jason D. Lee*, Qi Lei*, Runzhe Wang*, and Jiaqi Yang*. “Optimal Gradient-based Algorithms for Non-concave Bandit Optimization”, NeurIPS 2021
4. Jason D. Lee*, Qi Lei*, Nikunj Saunshi*, Jiacheng Zhuo*. “Predicting What You Already Know Helps: Provable Self-Supervised Learning”, NeurIPS 2021
3. Simon S. Du*, Wei Hu*, Sham M. Kakade*, Jason D. Lee*, Qi Lei*. “Few-Shot Learning via Learning the Representation, Provably”, The International Conference on Learning Representations (ICLR) 2021
2. Qi Lei*, Lingfei Wu*, Pin-Yu Chen, Alexandros G. Dimakis, Inderjit S. Dhillon, Michael Witbrock. “Discrete Adversarial Attacks and Submodular Optimization with Applications to Text Classification”, Systems and Machine Learning (sysML). 2019 (code, slides)
Press coverage: <Nature Story> <Vecturebeat> <Tech Talks> <机器之心>
1. Rashish Tandon, Qi Lei, Alexandros G. Dimakis, Nikos Karampatziakis, “Gradient Coding: Avoiding Stragglers in Distributed Learning”, Proc. of International Conference of Machine Learning (ICML), 2017 (code)