Yoojin Oh

I'm a M.S. student at KAIST AI, advised by Prof. Jong Chul Ye. I earned my B.S. in Artificial Intelligence from Ewha Womans University in Feb. 2026 (Summa Cum Laude).

My research focuses on the foundations of few-step generative models, particularly on stabilizing their training dynamics. I am also interested in their applications to image/video editing.

Email  /  Scholar  /  Github

profile photo

News

ยท [2026.02] I started M.S. at KAIST Graduate School of AI. (Advisor: Prof. Jong Chul Ye).

Publications
Beyond Softmax: Dual-Branch Sigmoid Architecture for Accurate Class Activation Maps
Yoojin Oh, Junhyug Noh
BMVC, 2025
project page / arXiv / code

To mitigate two fundamental softmax-induced distortions: Additive Logit Shift and Sign Collapse, we propose a simple, architecture-agnostic dual-branch sigmoid head that decouples localization from classification.

SteeringTTA: Guiding Diffusion Trajectories for Robust Test-Time-Adaptation
Jihyun Yu, Yoojin Oh, Wonho Bae, Mingyu Kim, Junhyug Noh
ICML PUT Workshop, 2025
arXiv

We propose SteeringTTA, an inference-only test-time adaptation framework that applies Feynman-Kac steering to diffusion-based input adaptation. Using pseudo-label driven rewards and multiple particle trajectories, it balances exploration and confidence through top-k probabilities and entropy scheduling.


Thank you for providing a source code.