Hi! I'm a Ph.D. student at KAIST AI under the supervision of Prof. Seong Joon Oh.
My research lies in the field of AI Safety, with a particular interest in building models that are fair, responsible, and transparent.
Machine Unlearning: Models trained on large-scale data often absorb personal information or copyrighted content that should later be removed, and retraining from scratch is too costly. Unlearning aims to remove the influence of specific data from a trained model to protect user privacy, copyright, and data ownership. I’m particularly interested in making sure the data is truly forgotten and not just hidden, while keeping the model’s overall performance intact.
Red Teaming: Before we can fix a model, we need to know how it fails. Red teaming finds vulnerabilities in AI systems, such as jailbreaks, harmful or biased outputs, and privacy leakage. I believe uncovering these problems is a crucial step toward building safer AI, since it shows us where safety work is most needed.
Openings for Interns: One opening in AI Safety
News
09/2026: Served as a program committee member for two NeurIPS 2026 workshops: PALM and ATTRIB.
09/2026: Started my PhD journey in STAI group at KAIST AI.
05/2026: 2 papers accepted @ ICML 2026 Workshop on the Impact of Memorization on Trustworthy Foundation Models.
01/2026: 1 paper accepted @ ICLR 2026.
11/2025: 🏆 Granted AI SeoulTech Graduate Scholarship by Seoul Scholarship Foundation ($3.5K).
09/2025: 1 paper accepted @ NeurIPS 2025.
07/2025: 1 paper accepted @ COLM 2025.
01/2025: Served as a visiting scholar at the Tübingen AI Center (Host: Prof. Dr. Seong Joon Oh).
03/2024: 🏆 Granted Digital Human & Entertainment Scholarship by Smilegate ($27K).
03/2024: 🏆 Granted Albatross Fellowship by Sogang University ($21K, top 10% rank honor).
Selected Publications
*Equal contribution, ^Co-corresponding Authors, ᐩMentored Interns
[8] 🐝 Cross-Pollination Retrieval: Finding Transferable Ideas Across Fields in Scientific Literature [pdf]
Seokwon Jung, Yerang Kimᐩ, Yejin Kim, Alexander Rubinstein, Arnas Uselis*, Sangwoo Park*, Seong Joon Oh, Sangdoo Yun.
Preprint.
[7] 🌷TULIP: Targeted LLM Unlearning at Layers Identified Per-input [pdf]
Yejin Kim, William F. Shen, Seokwon Jung, Daeun Parkᐩ, Seong Joon Oh.
Preprint.
[6] Knowledge Vector Weakening: Efficient Training-free Unlearning for Large Vision-Language Models [pdf]
Yejin Kim, Dongjun Hwang, Sungmin Cha^, Junsuk Choe^.
Preprint.
[5] Break the Output Geometry for Large Language Model Unlearning [pdf]
Yejin Kim, William F. Shen, Seokwon Jung, Seong Joon Oh.
ICML 2026 Workshop on the Impact of Memorization on Trustworthy Foundation Models.
[4] Alignment-aware Data Selection for Unlearning in Contrastive Vision-Language Models [pdf]
Dongjun Hwang, Yejin Kim, Beomyun Kwon, Junsuk Choe.
ICML 2026 Workshop on the Impact of Memorization on Trustworthy Foundation Models.
[3] Enhancing Multi-Image Understanding through Delimiter Token Scaling [pdf] [code]
Minyoung Lee, Yeji Park, Dongjun Hwang, Yejin Kim, Seong Joon Oh, Junsuk Choe.
ICLR 2026.
[2] OVS Meets Continual Learning: Towards Sustainable Open-Vocabulary Segmentation [pdf] [code]
Dongjun Hwang, Yejin Kim, Minyoung Lee, Seong Joon Oh, Junsuk Choe.
NeurIPS 2025.
[1] Improving Fisher Information Estimation and Efficiency for LoRA-based LLM Unlearning [pdf] [code]
Yejin Kim*, Eunwon Kim*, Buru Chang^, Junsuk Choe^.
COLM 2025.