Tianyang Hu

Assistant Professor, School of Data Science, The Chinese University of Hong Kong, Shenzhen.

Tianyang Hu presenting at SUFE in 2023

At SUFE, 2023. With thanks to Yixuan and Taiyun.

My research lies at the intersection of Statistics and AI. I develop statistical foundations and methodologies to make modern AI more capable and efficient. My interests include:

  • Statistical Machine Learning
  • Representation Learning
  • Generative AI
  • Foundation Models

I lead SAGA Lab at CUHK-Shenzhen, where we combine theoretical analysis, mechanistic investigation, and empirical research to develop better generative models for vision and language. A current focus is continuous diffusion language modeling. I also co-initiated Polyadic Research Group, a collaborative research collective advancing this direction through model development and open, reproducible research.

Join us. I am recruiting PhD and MPhil students and research assistants. See SAGA Lab openings for research opportunities and how to apply.

Before joining CUHK-Shenzhen, I was a Postdoctoral Fellow at the National University of Singapore, working with Prof. Kenji Kawaguchi, and a Research Scientist in the AI Theory Group of Noah’s Ark Lab. I received my Ph.D. in Statistics from Purdue University under the supervision of Prof. Guang Cheng, after earning an M.S. in Statistics from the University of Chicago and a B.S. in Mathematics from Tsinghua University.

Selected papers

A selection across my research interests. View all papers →

  1. arXiv
    Transformers Are Born Biased: Structural Inductive Biases at Random Initialization and Their Practical Consequences
    Siquan Li*, Yao Tong*, Haonan Wang*, and Tianyang Hu
    arXiv preprint, 2026
    The inspired methodologies are published in ICLR 2026 and ICML 2026.
    arXiv
  2. ICMLOral
    Any-Order GPT as Masked Diffusion Model: Decoupling Formulation and Architecture
    Shuchen Xue, Tianyu Xie, Tianyang Hu, Zijin Feng, Jiacheng Sun, Kenji Kawaguchi, Zhenguo Li, and Zhi-Ming Ma
    International Conference on Machine Learning (selected for Oral presentation), 2026
    Abridged in the ICML 2025 Workshop on Efficient Systems for Foundation Models (ES-FoMo III) (Selected for Oral presentation).
    arXiv
  3. ICLR
    PrefixMemory-Tuning: Modernizing Prefix-Tuning by Decoupling the Prefix from Attention
    Haonan Wang*, Brian Chen*, Siquan Li*, Xinhe Liang, Hwee Kuan Lee, Kenji Kawaguchi, and Tianyang Hu
    International Conference on Learning Representations, 2026
    Abridged in the ICML 2025 Workshop on Test-Time Adaptation: Putting Updates to the Test (PUT).
    arXiv
  4. NeurIPS
    Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation
    Yihong Luo*, Tianyang Hu*, Weijian Luo, Kenji Kawaguchi, and Jing Tang
    Advances in Neural Information Processing Systems, 2025
    arXiv
  5. JMLR
    Minimax Optimal Deep Neural Network Classifiers Under Smooth Decision Boundary
    Tianyang Hu, Ruiqi Liu, Zuofeng Shang, and Guang Cheng
    Journal of Machine Learning Research, 26 (136), 2025
    arXiv
  6. JMLR
    Random Smoothing Regularization in Kernel Gradient Descent Learning
    (α-β) Liang Ding, Tianyang Hu, Jiahang Jiang, Donghao Li, Wenjia Wang, and Yuan Yao
    Journal of Machine Learning Research, 25 (284), 2024
    arXiv
  7. NeurIPSSpotlight
    Complexity Matters: Rethinking the Latent Space for Generative Modeling
    Tianyang Hu, Fei Chen, Haonan Wang, Jiawei Li, Wenjia Wang, Jiacheng Sun, and Zhenguo Li
    Advances in Neural Information Processing Systems (selected for Spotlight presentation), 2023
    arXiv
  8. NeurIPS
    Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models
    Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang
    Advances in Neural Information Processing Systems, 2023
    arXiv
  9. ICLR
    Your Contrastive Learning Is Secretly Doing Stochastic Neighbor Embedding
    Tianyang Hu, Zhili Liu, Fengwei Zhou, Wenjia Wang, and Weiran Huang
    International Conference on Learning Representations, 2023
    arXiv
  10. NeurIPSSpotlight
    Understanding Square Loss in Training Overparametrized Neural Network Classifiers
    Tianyang Hu*, Jun Wang*, Wenjia Wang*, and Zhenguo Li
    Advances in Neural Information Processing Systems (selected for Spotlight presentation), 2022
    arXiv
  11. AISTATS
    Regularization Matters: A Nonparametric Perspective on Overparametrized Neural Network
    Tianyang Hu*, Wenjia Wang*, Cong Lin, and Guang Cheng
    In International Conference on Artificial Intelligence and Statistics, 2021
    arXiv