Weiyang Liu

I am a machine learning researcher at Max Planck Institute for Intelligent Systems, Tübingen. In my PhD study, I was fortunate to be supervised by Adrian Weller and Bernhard Schölkopf. As a member of the advising team at MeshCapade, I also work closely with Michael J. Black. Previously, I spent wonderful time at Georgia Tech, Google Brain, Nvidia Research, and MERL. I work primarily on principled modeling of inductive bias in learning algorithms. My research seeks to understand how inductive bias affects generalization, and to develop "light-yet-sweet" learning algorithms: (i) light: conceptually simple in methodology and easy to implement in practice, (ii) sweet: having clear intuitions and non-trivial theoretical guarantees. Over the years, I always find myself fascinated by geometric invariance, symmetry, structures (causality, discrete knowledge) and how they can benefit generalization as a guiding principle. Recently, I have begun rethinking inductive bias in the era of foundation models, and developed a deep interest in large language models and generative modeling across visual, textual, and physical domains. More specifically, my current research focuses on (i) developing principled algorithms for training/adapting foundation models, and (ii) understanding how LLMs perform reasoning and eliciting it in formal/verifiable scenarios (math/symbolic reasoning). I always believe in two principles in my research: (i) insight must precede application, and (ii) everything should be made as simple as possible, but not simpler.