Lianghe Shi (石亮禾)

About Me

Hi! I am a second-year Ph.D. student at EECS, University of Michigan, advised by Prof. Qing Qu. Before joining U-M, I received my undergraduate and master’s degrees from the Wuhan University, majoring in software engineering and computer science, respectively, advised by Prof. Weiwei Liu.
My long-term goal is to build general-purpose, robust, and efficient generative models, including diffusion models and vision-language models. I am particularly interested in addressing some limitations of current VLMs—such as visual hallucination, language bias, and model collapse—and in developing principled methods that improve generalization and robustness. I am also interested in the “competition” between diffusion LLM and autoregressive LLM.

A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
Lianghe Shi*, Meng Wu*, Huijie Zhang, Zekai Zhang, Molei Tao, Qing Qu.
ICML 2025 Workshop (Oral) and NeurIPS 2025 (Spotlight).
Website

Adversarially robust unsupervised domain adaptation
Lianghe Shi*, Weiwei Liu. The journal of Artificial Intelligence. AIJ.

A Closer Look at Curriculum Adversarial Training: From an Online Perspective
Lianghe Shi*, Weiwei Liu. Proceedings of the 38th AAAI Conference on Artificial Intelligence. AAAI 2024.

Adversarial self-training improves robustness and generalization for gradual domain adaptation
Lianghe Shi*, Weiwei Liu. 37th Conference on Neural Information Processing Systems. NeurIPS 2023.

Graduate Student Instructor of EECS 501: Probability and Random Processes (Fall 2025)