Seojin Kim

M.S Student at KAIST AI.

profile_pic.JPG

I am a M.S student at KAIST Kim Jaechul Graduate School of AI.

I am now a member of BigDyl(at Daejeon, South Korea), advised by Noseong Park.

My research focuses on understanding the mechanisms of modern language models to improve their efficiency and scalability. I am particularly interested in efficient language modeling, including state space models, KV cache optimization, and long-context inference. I also work on sequence modeling problems in time series analysis and recommendation systems.

I obtained my Bachelor’s degree in Applied Statistics and Computer Science at Yonsei University, Seoul, Republic of Korea.

News

Sep, 2026 Two papers have been accepted to NeurIPS 2026! :sparkles: :smile:
May, 2026 One paper has been accepted to ICML 2026! :sparkles: :smile:
Feb, 2026 I will be joining Microsoft Research Asia as a research intern this March 🇨🇳
Feb, 2026 One paper has been accepted to ICLR 2026!
Sep, 2025 One paper has been accepted to NeurIPS 2025! :sparkles: :smile:

Selected Papers

* Equal contribution

  1. Yehjin Shin*, Seojin Kim*, and Noseong Park
    In The Fourteenth International Conference on Learning Representations, 2026
  2. Global Importance Estimation for KV Cache Eviction
    Seojin Kim and Noseong Park
    2026
  3. Attention Sinks as Spectral Spikes: A Mechanism Analysis of Gated Attention
    Seojin Kim*, Yehjin Shin*, and Noseong Park
    2026
  4. Seojin Kim*, Yehjin Shin*, and Noseong Park
    In ES-FoMo III: 3rd Workshop on Efficient Systems for Foundation Models, 2025