CV
Basics
| Name | Sumin Park |
| Status | M.S. Student in Computer Science |
| psmiz@kaist.ac.kr | |
| Phone | +82-10-3475-0804 |
| Url | https://psmiz.github.io |
Education
Work
-
2024.09 - Present Daejeon, South Korea
Master’s Student Researcher
Big Data Analysis & Learning Laboratory (BigDyL), KAIST
Research on LLM architectures, representation learning, explainable LLMs, and brain-inspired AI.
- Advisor: Noseong Park
-
2023.07 - 2024.08 Seoul, South Korea
Research Intern
Big Data Analysis & Learning Laboratory (BigDyL), Yonsei University
Research on graph representation learning and graph neural networks.
- Advisor: Noseong Park
-
2022.02 - 2024.04 Seoul, South Korea
Data Scientist
DNI Consulting
Data analysis and modeling for B2B marketing and customer segmentation.
-
2020.08 - 2021.01 Seoul, South Korea
Undergraduate Research Intern
Cell Physiology Laboratory, Seoul National University
Research on PFC-involved systems consolidation via electrophyisology.
- Advisor: Seokho Lee
-
2019.12 - 2020.02 Seoul, South Korea
Undergraduate Research Intern
Neurobiology Laboratory, Seoul National University
Research on dopaminergic reward circuitry.
- Advisor: Bongkiun Kaang
-
2019.09 - 2019.07 Seoul, South Korea
Undergraduate Research Intern
Molecular Neuroscience Laboratory, POSTECH
Research on extinction of cocaine-context association memory.
- Advisor: Junghun Kim
Publications
-
2026 Q-Delta: Beyond Key–Value Associative State Evolution
Sumin Park, Seojin Kim, Noseong Park
Under Review
-
2026 STAR: Rethinking MoE Routing as Structure-Aware Subspace Learning
Sumin Park, Noseong Park
UnderReview
-
2026 How Many Experts Are Enough? Towards Optimal Semantic Specialization for Mixture-of-Experts
Sumin Park, Noseong Park
AAAI 2026
-
2025 DARS: Robust Sparse Fine-Tuning with Regularized Subspace Disalignment
Sumin Park, Noseong Park
ICLR 2025 Workshop (SCOPE)
-
2024 PANDA: Expanded Width-Aware Message Passing Beyond Rewiring
Jeongwhan Choi, Sumin Park, Hyowon Wi, Sung-Bae Cho, Noseong Park
ICML 2024
Projects
- 2025.03 - 2025.12
Simulation-Based Defect Variable Identification from High-Resolution Images
Samsung–KAIST industry-academia project on simulation-based defect variable prediction using high-resolution images.
- Multi-task, multi-class classification framework
- ViT with hierarchical attention for high-resolution processing
- Achieved 99.41% classification accuracy