avatar

Chenxiao Yang
杨 晨晓

1st Year PhD Student
Toyota Technological Institute at Chicago (TTIC)
chenxiao [at] ttic (dot) edu


🤗 Hi! I am Chenxiao, a PhD student at Toyota Technological Insitute at Chicago, an independent philanthropically endowed computer science research institute located on the University of Chicago campus. I am fortunate to be advised by Zhiyuan Li. Previously, I was a research intern at Amazon Web Services, mentored by David Wipf. I received both my M.S. and B.S. degrees from Shanghai Jiao Tong University, where I worked with Junchi Yan. I also occasionally visit Massachusetts Institute of Technology. For more details, see CV. My primary research interests include:

1️⃣ Next-Generation Generative Models: I am interested in analyzing the strengths and limitations of current generative models (e.g. LLMs, diffusion models) from a theoretical perspective. In doing so, I explore simple and novel approaches or fundamental principles that overcome these limitations, enabling models to reason and generate more efficiently and effectively, and work universally across domains such as natural langauge, math, coding, science, etc.

2️⃣ Theory of Modern Learning Paradigm: I am also curious about various modern learning paradigms that emerge in the era of LLMs and how they should be positioned in the landscape of machine learning theory to enrich our understanding. I am especially interested in understanding expressivity of various models and how they develop out-of-distribution generlization capabilities through large-scale training.

3️⃣ Learning on Combinatorial Structures (Graphs): I develop new algorithms or models that can better handle combinatorial structures and complex topologies such as graphs and sequences with some global constraints, which are common in domains like biology where current LLMs struggle, with accompanying theories.

I maintain side interests in theoretical computer science, causal learning, reinforcement learning, and machine-learning applications in various domains.

Recent News

Selected Publications

Full list of publications can be found in Google Scholar

  1. Preprint
    On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond
    Chenxiao Yang, Cai Zhou, David Wipf, Zhiyuan Li
    Preprint
    PDF

  2. ICML'25
    ✏️ PENCIL: Long Thoughts with Short Memory
    Chenxiao Yang, Nathan Srebro, David McAllester, Zhiyuan Li
    International Conference on Machine Learning (ICML), 2025.
    PDF Code Poster Slides Talk Blog 博客

  3. ICML'24
    How Graph Neural Networks Learn: Lessons from Training Dynamics
    Chenxiao Yang, Qitian Wu, David Wipf, Ruoyu Sun, Junchi Yan
    International Conference on Machine Learning (ICML), 2024.
    PDF Code Poster

  4. ICLR'23
    Graph Neural Networks are Inherently Good Generalizers: Insights by Bridging GNNs and MLPs
    Chenxiao Yang, Qitian Wu, Jiahua Wang, Junchi Yan
    International Conference on Learning Representations (ICLR), 2023.
    PDF Code Poster 视频 博客

Teaching

Services

Misc

In my free time I enjoy hiking, traveling, swimming, photography, and experimenting with AI-generated creative contents. I also like studying psychology and philosophy.

Feel free to leave anonymous feedback about me or my research here!