Zeyi Huang
I am a fifth-year PhD student in the Department of Computer Sciences at the University of Wisconsin–Madison, advised by Prof. Yong Jae Lee. Prior to UW–Madison, I obtained my M.S. degree from the Robotics Institute at Carnegie Mellon University, where I worked with Dr. Dong Huang, Prof. Haohan Wang, and Prof. Eric Xing. I received my B.E. degree from Nanjing University of Science and Technology.
I have interned at Microsoft with Yelong Shen; at Meta with Daniel Bolya, Christoph Feichtenhofer, and Miao Liu; and at Adobe with Zongze Wu and Eli Shechtman.
Research Interests
My research focuses on building more capable and computationally efficient large language models, focusing recently on recurrent Transformer architectures, pre-training, and post-training. LRT introduces cross-token latent recurrence to reduce language-modeling loss at matched training compute, with one forward pass per generated token. LRT with Thought Tokens scales test-time computation through latent thinking steps, offering a parameter-efficient alternative to increasing model depth.
My broader research includes reinforcement learning for selecting from a large set of visual tools (VisTA) and enabling large language models to understand images and videos through structured textual representations (SVG-LLM, VideoMindPalace).