STAI Lab - KAIST AI
We study AI safety, agents, memory and retrieval at KAIST AI, led by Seong Joon Oh. Our approach separates the knowledge a model holds from its reasoning, so people can trace, update and control that knowledge.
We are recruiting interns, MSc and PhD students and postdocs. How to join
Make AI useful without losing privacy or human control.
Improve the whole system, from planning and tool use to coordination between agents.
Help agents learn from new information and user corrections without retraining for every update.
Connect models to relevant, current evidence across text and images.
ICML Oral 2026
Vision models can only recognise familiar parts in new combinations if their representations split into near-orthogonal pieces, one per concept. CLIP, SigLIP and DINO show this structure in part, and the models that show more of it generalise better.
Arnas Uselis, Darina Koishigarina, Seong Joon Oh - ICML Spotlight 2026
Hoyeon, Bálint, Seong Joon Oh - ICML Spotlight 2026
17 conference papers in 2026. All publications
Graduate students lead their own first-author projects, with regular one-to-one meetings with Seong Joon Oh. Undergraduates can start with an internship alongside a PhD student. Our alumni have gone on to Apple, Fraunhofer HHI and the University of Tübingen.
Open positions: Postdoc, PhD, MSc + PhD (integrated), MSc, Internship
