I am a Ph.D. candidate at the Intelligent Computational Imaging Lab (advised by Prof. Se Young Chun) in the Department of Electrical and Computer Engineering at Seoul National University. I received my B.S. in Electrical Engineering from UNIST.
My research focuses on selective and efficient post-training of foundation models, enabling models to acquire new capabilities, remove unwanted knowledge, and adapt their behaviors while preserving existing performance. My work spans continual learning, machine unlearning, and reinforcement learning, with experience across visual, generative, and multimodal foundation models. Recently, I am interested in multimodal LLMs, agentic AI, and vision-language-action models for real-world applications.
* denotes equal contribution.
Transfers textual alignment from a strong, well-aligned donor modality to a weak target modality in multimodal LLMs via a closed-form weight-space update, without further fine-tuning.
Scales selective concept erasure to 2,000+ concepts while preserving the remaining ones, via optimal transport and mixture-of-experts updates.
Memory-free continual learning for gigapixel pathology slides, preserving both diagnosis and localization across a stream of organs.
Training-free, spatially localized concept erasure that removes only the target while keeping the rest of the image intact.
A residual attention gate that erases target concepts from diffusion models while preserving the generation of remaining concepts.
Future-aware online continual learning that improves stability and plasticity through doubly perturbed training.
Hierarchy-aware memory management for online continual learning under streaming, label-expanding data.