Ziming Liu

Hi! I’m Ziming Liu (刘子鸣). I am a tenure-track Assistant Professor in College of AI at Tsinghua University.

My lab is called LIU lab, which stands for Learning, Intelligence, and Universe.

• Learning is our mindset –the first principle of research is curiosity-driven learning, not publication
• Intelligence is our goal – understanding and building intelligent systems
• Universe is our methodology – we study AI as if we’re studying our Universe, i.e., breaking & building things from first principles, balancing between experiments and theory.

I did my postdoc in AI + Neuroscience at Stanford with Andreas Tolias. I did my PhD in AI + Physics at MIT with Max Tegmark. Before that, I obtained my B.S. in physics from Peking University in 2020. Get my CV here.

I’m currently devoted to two related things:

• “Physics of AI” – using first principles to understand the structure (“Google map”) of AI research/idea/design space.
• “AI for AI” – building an AI agent that can intelligently navigate through the map of ideas, like human AI researchers.

I believe both the AI agent and the “physics of AI” knowledge base will become key ingredients of a safe and efficient AGI, which naturally supports curiosity-driven continual learning.

Over a longer time scale, my research interests generally lie in the intersection of AI and Science:

• Science of AI: Understanding AI using science. I’m interested in understanding intriguing network phenomena (e.g., grokking, neural scaling laws).
• Science for AI: Advancing AI using science. When the scaling of the current paradigm plateaus, it is time to focus back on fundamental science (for AI).
• AI for Science: Advancing science using AI. I’m excited about inventing accurate/interpretable AI4Science models and curiosity-driven AI scientists.

Consider reading these Quanta articles if you want to get some taste of my research:

• Science of AI (Grokking): How Do Machines ‘Grok’ Data?
• Science for AI (Poisson Flow): The Physical Process That Powers a New Type of Generative AI
• AI for Science (KAN): Novel Architecture Makes Neural Networks More Understandable

Jun 22, 2026	Discovering 108 tricks to accelerate grokking
May 06, 2026	AI4AI needs its own "world model"
May 05, 2026	How does memorization affect l2?

1. 

   KAN: Kolmogorov-arnold networks

   Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark

   arXiv:2404.19756, 2024

2. 

   Seeing is believing: Brain-inspired modular training for mechanistic interpretability

   Ziming Liu, Eric Gan, and Max Tegmark

   Entropy, 2023

3. 

   Scientific discovery in the age of artificial intelligence

   Nature, 2023

4. 

   The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks

   Ziqian Zhong, Ziming Liu, Max Tegmark, and Jacob Andreas

   In Thirty-seventh Conference on Neural Information Processing Systems , 2023

5. 

   Omnigrok: Grokking Beyond Algorithmic Data

   Ziming Liu, Eric J Michaud, and Max Tegmark

   In The Eleventh International Conference on Learning Representations , 2023

6. 

   Poisson Flow Generative Models

   Yilun Xu, Ziming Liu, Max Tegmark, and Tommi S. Jaakkola

   In Advances in Neural Information Processing Systems , 2022

7. 

   Machine Learning Hidden Symmetries

   Ziming Liu, and Max Tegmark

   Phys. Rev. Lett., 2022

8. 

   Machine Learning Conservation Laws from Trajectories

   Ziming Liu, and Max Tegmark

   Phys. Rev. Lett., 2021