hee-seon.
AI RESEARCHER / KAIST CILAB
김희선 · Ph.D. Student in Electrical Engineering
Exploring the boundaries of
vision, language, and trust.
I study the robustness and safety of large vision-language models: where they fail, why they fail, and how we can better evaluate them. My work also spans video understanding and efficient model adaptation.
Hi, I’m Hee-Seon!
KAIST CILab
01 / RESEARCH
Recent work on multimodal robustness, alongside selected work in video and computer vision.
View figure ↗
An execution layer for safer mobile agents: checking risky actions before execution and reusing verified routes for recurring tasks.
View figure ↗
Locating vulnerable components inside vision encoders to evaluate LVLM robustness with a single universal perturbation.
View figure ↗
Studying whether adversarial images can break multimodal safety alignment even when the accompanying prompt is harmless.
View figure ↗
Efficient video recognition with selective state space models that capture spatial and temporal relationships.
02 / A LITTLE BACKGROUND
I’m a Ph.D. student in Electrical Engineering at KAIST, working in the Computational Intelligence Lab.
My current focus is large vision-language models, with a particular interest in adversarial robustness and multimodal safety.
NAVER Cloud
KAIST · In progress
KAIST
UNIST
LET’S CONNECT