CV
Contact Information
| Name | Seojin Kim |
| Professional Title | M.S Student at KAIST AI |
| seojinkim@kaist.ac.kr | |
| Location | 291 Daehak-ro, Yuseong-gu, Daejeon, South Korea 34141 |
Professional Summary
M.S. student at KAIST AI working on efficient language modeling, studying the mechanisms of modern sequence models to build more scalable and practical AI systems.
Experience
-
2026 - 2026 Beijing, China
Research Intern
Microsoft Research Asia
Research Intern at MSRA Network Infrastructure Group, advised by Wenxue Cheng.
- Agentic AI for network diagnosis
-
2024 - Daejeon, South Korea
Graduate Researcher
BigDyl
Graduate Researcher at BigDyl, a research group at KAIST, advised by Noseong Park.
- Efficient Language Modeling (KV cache, State Space Models, Transformer Mechanisms)
- Time Series Analysis
- Recommendation System
-
2023 - 2024 Seoul, South Korea
Data Scientist Intern
Minds n Company (now GenOn)
Data Scientist Intern at Minds n Company, a startup company in Seoul, South Korea.
- Project: AI Banker (LLM Agent) project at Woori Bank
- Skills: Langchain, Natural Language Generation
-
2021 - 2022 Seoul, South Korea
Data Analyst Intern
Ecube Labs
Data Analyst Intern at Ecube Labs, a startup company in Seoul, South Korea.
- Project: Haulla call data analysis & referral marketing project
- Skills: Data Analytics with Python
Education
-
2024 - Daejeon, South Korea
M.S
KAIST Kim Jaechul Graduate School of AI
Artificial Intelligence
- Efficient Language Modeling, Time Series Analysis, Recommendation System
- KV cache
- State Space Models
-
2020 - 2024 Seoul, South Korea
B.A and B.S
Yonsei University
Applied Statistics and Computer Science
- Honors in 2020 spring semester (awarded to top 10 % of the students)
- Honors in 2024 spring semester (awarded to top 10 % of the students)
- Exchange student (Fall 2022, University of Glasgow, Department of Science and Engineering)
Interests
Publications
* Equal contribution
-
2026 Global Importance Estimation for KV Cache Eviction
The 40th Conference on Neural Information Processing Systems (NeurIPS 2026)
Seojin Kim, Noseong Park
A training-free KV cache eviction method using Personalized PageRank to estimate global token importance for memory-efficient long-context LLM inference.
-
2026 Attention Sinks as Spectral Spikes: A Mechanism Analysis of Gated Attention
The 40th Conference on Neural Information Processing Systems (NeurIPS 2026)
Seojin Kim*, Yejin Shin*, Noseong Park
Attention sinks are recast as sink-aligned spectral spikes, revealing that post-SDPA gating suppresses them through a unified forward spectral and backward value-path contraction mechanism.
-
2026 Q-Delta: Beyond Key-Value Associative State Evolution
The 43rd International Conference on Machine Learning (ICML 2026)
Sumin Park, Seojin Kim, Noseong Park
A query-aware delta rule that integrates mixed key-query prediction errors into state evolution, enabling jointly corrective dynamics.
-
2026 Graph Signal Processing Meets Mamba2: Adaptive Filter Bank via Delta Modulation
The Fourteenth International Conference on Learning Representations (ICLR 2026)
Yehjin Shin*, Seojin Kim*, Noseong Park
Reinterprets Mamba2 as an adaptive filter bank and proposes a hierarchical framework inspired by graph signal processing.
-
2025 TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation
The 39th Conference on Neural Information Processing Systems (NeurIPS 2025)
Yehjin Shin, Jeongwhan Choi, Seojin Kim, Noseong Park
Proposes time-variant convolutional filters for sequential recommendation to better capture temporal interactions while improving efficiency.
-
2025 Mitigating Over-Smoothing in Mamba2 via Spectral Domain Analysis
ES-FoMo III: 3rd Workshop on Efficient Systems for Foundation Models
Seojin Kim*, Yehjin Shin*, Noseong Park
Analyzes Mamba2 in the spectral domain and proposes a high-frequency enhancement method to reduce over-smoothing.
-
2025 Graph Signal Processing Meets Mamba2: Adaptive Filter Bank via Delta Modulation
ES-FoMo III: 3rd Workshop on Efficient Systems for Foundation Models
Yehjin Shin*, Seojin Kim*, Noseong Park
Reinterprets Mamba2 as an adaptive filter bank and proposes a hierarchical framework inspired by graph signal processing.
-
2024 Addressing Prediction Delays in Time Series Forecasting: A Continuous GRU Approach with Derivative Regularization
ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2024)
Sheo Yon Jhin, Seojin Kim, Noseong Park
Introduces a continuous-time GRU (neural ODE) with time-derivative regularization to mitigate prediction delay in forecasting.
-
2023 Long-term Time Series Forecasting based on Decomposition and Neural Ordinary Differential Equations
2023 IEEE International Conference on Big Data (IEEE BigData 2023)
Seonkyu Lim*, Jaehyeon Park*, Seojin Kim*, Hyowon Wi, Haksoo Lim, Jinsung Jeon, Jeongwhan Choi, Noseong Park
Proposes LTSF-DNODE, combining time-series decomposition with neural ODEs for long-term forecasting, improving over baselines on real-world datasets.
Awards
-
2023 2023 Yonsei Big Data Analysis Contest
Yonsei University
2nd place: Verification of AlleyWay Consulting and presentation of Growth Plans with LightGBM
-
2021 2021 Traditional Culture Youth Startup Idea Contest
Korea Craft and Design Foundation
Project: Moon-we: Interactive Matchmaking Platform for Collaboration between Intangible Cultural Properties-Company-Consumer and Crowdfunding Related to Intangible Cultural Properties