Our lab develops probabilistic methods for modeling both safe and unsafe distributions in AI, with the goal of controlling generation toward safe and reliable outcomes. Our research centers on diffusion and flow-matching frameworks for image, video, and language models, and we are actively extending these ideas to action models. We also study the fine-tuning of foundation models, including vision-language models and large language models, with particular attention to mitigating overfitting. Across these areas, trustworthiness and reliability serve as core principles shaping our research.
Mingyu will serve as an Area Chair for the NeurIPS 2026 Position Track, contributing his expertise in support of a rigorous and constructive peer-review process for the research community.
A new paper titled “LLM-as-UI: A Preliminary Exploration of Fine-Tuned Language Models as Intelligent Interfaces in Modular Ship Design” has been accepted for publication in Ocean Engineering. Ocean Engineering is a prestigious journal published by Elsevier, ranked in the top 5% across all scientific journals and recognized as one of the top-tier venues in the ocean and marine engineering field. Mingyu contributed as the corresponding author, collaborating with Minjoo Choi, Jaekyeong Lee, and Stein Ove Erikstad. Congratulations to all co-authors!
Mingyu’s research was highlighted through a CIFAR interview feature and selected for CIFAR Headliner News on March 5, 2026. CIFAR (the Canadian Institute for Advanced Research) is an internationally recognized organization that advances frontier research and supports transformative work in AI and other scientific fields. Read the article.
