Research

* denotes equal contribution and ** denotes alphabet order. An up-to-date list is available on Google Scholar.

2025

  1. Evaluating the unseen capabilities: How many theorems do LLMs know?
    [arXiv][Poster][Slides]
    Xiang Li, Jiayi Xin, Qi Long, and Weijie J. Su
    arXiv preprint arXiv:2506.02058, 2025
  2. Optimal estimation of watermark proportions in hybrid AI-human texts
    [arXiv][Code][Slides]
    Xiang Li, Garrett G. Wen, Weiqing He, Jiayuan Wu, Qi Long, and Weijie J. Su
    arXiv preprint arXiv:2506.22343, 2025
  3. On the empirical power of goodness-of-fit tests in watermark detection
    Weiqing He*, Xiang Li*, Tianqi Shang, Li Shen, Weijie J. Su, and Qi Long
    In Advances in Neural Information Processing Systems, 2025, 🌟 Spotlight
  4. Mitigating the privacy–utility trade-off in decentralized federated learning via f-differential privacy
    Xiang Li, Chendi Wang, Buxin Su, Qi Long, and Weijie J. Su
    In Advances in Neural Information Processing Systems, 2025, 🌟 Spotlight
  5. Corruption-robust variance-aware algorithms for generalized linear bandits under heavy-tailed rewards
    [HTML]
    Qingyuan Yu, Euijin Baek, Xiang Li, and Qiang Sun
    In Conference on Uncertainty in Artificial Intelligence, 2025

2024

  1. A statistical framework of watermarks for large language models: Pivot, detection efficiency and optimal rules
    [arXiv][HTML][Code][Slides]
    Xiang Li, Feng Ruan, Huiyuan Wang, Qi Long, and Weijie J. Su
    The Annals of Statistics, 2025, 🏛️ Invited talk at AoS invited paper session, JSM 2025
  2. Robust detection of watermarks in large language models under human edits
    [arXiv][Code][Poster][Slides]
    Xiang Li, Feng Ruan, Huiyuan Wang, Qi Long, and Weijie J. Su
    Journal of the Royal Statistical Society: Series B, 2025, 🏆 IMS New Researcher Travel Award
  3. Debiasing watermarks for large language models via maximal coupling
    [arXiv][HTML][Code]
    Yangxinyu Xie, Xiang Li, Tanwi Mallick, Weijie J. Su, and Ruixun Zhang
    Journal of the American Statistical Association, 2025
  4. Decoupled functional central limit theorems for two-time-scale stochastic approximation
    [arXiv]
    Yuze Han, Xiang Li, Jiadong Liang, and Zhihua Zhang
    arXiv preprint arXiv:2412.17070, 2024
  5. Finite-time decoupled convergence in nonlinear two-time-scale stochastic approximation
    [arXiv]
    Yuze Han, Xiang Li, and Zhihua Zhang
    arXiv preprint arXiv:2401.03893, 2024
  6. Uncertainty quantification of data shapley via statistical inference
    [arXiv]
    Mengmeng Wu, Zhihong Liu, Xiang Li, Ruoxi Jia, and Xiangyu Chang
    arXiv preprint arXiv:2407.19373, 2024

2023

  1. Variance-aware decision making with linear function approximation with heavy-tailed rewards
    [arXiv][HTML]
    Xiang Li, and Qiang Sun
    Transactions on Machine Learning Research, 2024, 🏛️ Invited to present in ICLR 2025
  2. Online statistical inference for nonlinear stochastic approximation with Markovian data
    [arXiv]
    Xiang Li, Jiadong Liang, and Zhihua Zhang
    arXiv preprint arXiv:2302.07690, 2023
  3. Convergence and inference of Stream SGD, with applications to queueing systems and inventory control
    [arXiv][Code]
    Xiang Li*, Jiadong Liang*, Xinyun Chen, and Zhihua Zhang
    arXiv preprint arXiv:2309.09545, 2023
  4. Asymptotic behaviors and phase transitions in projected stochastic approximation: A jump diffusion approach
    [arXiv]
    Jiadong Liang, Yuze Han, Xiang Li, and Zhihua Zhang
    Technical report, arXiv preprint arXiv:2304.12953, 2023

2022

  1. A random projection approach to personalized federated learning: Enhancing communication efficiency, robustness, and fairness
    [HTML]
    Yuze Han**, Shiyun Lin**, Xiang Li**, and Zhihua Zhang**
    Journal of Machine Learning Research, 2024, Extended version of the conference paper: Personalized federated learning towards communication efficiency, robustness and fairness
  2. Asymptotic behaviors of projected stochastic approximation: A jump diffusion perspective
    [HTML][Slides]
    Jiadong Liang, Yuze Han, Xiang Li, and Zhihua Zhang
    In Advances in Neural Information Processing Systems, 2022, 🌟 Spotlight
  3. Personalized federated learning towards communication efficiency, robustness and fairness
    [HTML][Code][Poster][Slides]
    Shiyun Lin*, Yuze Han*, Xiang Li, and Zhihua Zhang
    In Neural Information Processing Systems, 2022
  4. Statistical analysis of Karcher means for random restricted PSD matrices
    [arXiv]
    Hengchao Chen, Xiang Li, and Qiang Sun
    In International Conference on Artificial Intelligence and Statistics, 2023

2021

  1. Statistical estimation and online inference via Local SGD
    [arXiv][HTML][Slides]
    Xiang Li, Jiadong Liang, Xiangyu Chang, and Zhihua Zhang
    In Conference on Learning Theory, 2022
  2. Fedpower: Privacy‑preserving distributed eigenspace estimation
    [arXiv][HTML]
    Xiao Guo, Xiang Li, Xiangyu Chang, Shusen Wang, and Zhihua Zhang
    Machine Learning, 2024, Extended version of the conference paper: Communication-efficient distributed SVD via local power iterations
  3. A statistical analysis of Polyak-Ruppert averaged Q-learning
    [arXiv][Code][Poster][Slides][Video]
    Xiang Li, Wenhao Yang, Jiadong Liang, Zhihua Zhang, and Michael I Jordan
    In International Conference on Artificial Intelligence and Statistics, 2023

2020

  1. Communication-efficient distributed SVD via local power iterations
    [arXiv][HTML][Code][Poster][Slides]
    Xiang Li, Shusen Wang, Kun Chen, and Zhihua Zhang
    In International Conference on Machine Learning, 2021
  2. Finding near optimal policies via reducive regularization in Markov decision processes
    [arXiv][HTML]
    Wenhao Yang, Xiang Li, Guangzeng Xie, and Zhihua Zhang
    In Workshop on Reinforcement Learning Theory, ICML, 2021

2019

  1. On the convergence of FedAvg on non-iid data
    [arXiv][HTML][Code][Poster][Slides]
    Xiang Li*, Kaixuan Huang*, Wenhao Yang*, Shusen Wang, and Zhihua Zhang
    In International Conference on Learning Representations, 2020, 🎤 Oral presentation
  2. A regularized approach to sparse optimal policy in reinforcement learning
    [arXiv][HTML]
    Wenhao Yang*, Xiang Li*, and Zhihua Zhang
    In Advances in Neural Information Processing Systems, 2019
  3. Do subsampled Newton methods work for high-dimensional data?
    [arXiv][HTML]
    Xiang Li, Shusen Wang, and Zhihua Zhang
    In AAAI Conference on Artificial Intelligence, 2020
  4. Communication efficient decentralized training with multiple local updates
    [arXiv]
    Xiang Li, Wenhao Yang, Shusen Wang, and Zhihua Zhang
    Technical report, arXiv preprint arXiv:1910.09126, 2019