Scalable and Efficient AI Lab
Korea Advanced Institute of Science and Technology (KAIST) · Advisor: Prof. Sehoon Kim
- Improving reasoning capabilities and accelerating inference in recursive architectures
Sep 2026 – Present
Heejun Kim
Korea Advanced Institute of Science and Technology (KAIST) M.S. in Artificial Intelligence
I am currently pursuing an M.S. in Artificial Intelligence at KAIST, conducting research on Reasoning and Efficient AI systems at the Scalable and Efficient AI Lab under Prof. Sehoon Kim's supervision. AGI will usher in a new phase of human labor, and what counts as good AI judgment ultimately rests on human values. My work is dedicated to building AGI agents whose behavior aligns with those values.
Previously, I worked with Prof. Taesup Kim at Seoul National University and Prof. Sundong Kim at Gwangju Institute of Science and Technology, researching reasoning and LLM behavior.
I also worked as a visiting researcher at Forschungszentrum Jülich's INM-7 institute, directed by Prof. Simon B. Eickhoff, through the Korea-Germany Junior Research Fellowship, exploring the intersection of neuroscience and machine learning.
Education
-
Korea Advanced Institute of Science and Technology (KAIST)
M.S. in Artificial Intelligence
Sep 2026 – Present -
Gwangju Institute of Science and Technology (GIST)
B.S. in Electrical Engineering and Computer Science · GPA 4.2/4.5 (magna cum laude)
Feb 2020 – Aug 2026Military service leave (2 years) -
University of California, Berkeley (UC Berkeley)
Summer Session · GPA 4.0/4.0
Jun 2024 – Aug 2024
Research Experience
Learning and Adaptation Algorithm Lab
Seoul National University (SNU) · Advisor: Prof. Taesup Kim
- Conducting research on Agentic Memory Using LLM Uncertainty Estimation
- Studied internal behavioral mechanisms of LLMs underlying reasoning, hallucination, and memory recall
Jan 2026 – Aug 2026
Data Science Lab
Gwangju Institute of Science and Technology (GIST) · Advisor: Prof. Sundong Kim
- Researched test-time adaptation with reinforcement learning for self-improving reasoning capabilities of large language models on the Abstraction and Reasoning Corpus (ARC)
- Investigated metacognition-driven strategies for helping large language models escape local minima in reasoning
Feb 2025 – Aug 2026
Jan 2024 – Jun 2024
Jan 2024 – Jun 2024
Forschungszentrum Jülich
Host Institute: INM-7, directed by Prof. Simon B. Eickhoff, and contributed to the Mathematical Neuroscience Group led by Dr. Oleksandr Popovych
- Developed preprocessing pipeline for multi-modal MRI human connectome data using High-Performance Computing (HPC) and HyperThreading technology
- Korea-Germany Junior Research Fellowship 2024 organized by Max Planck Korea
Sep 2024 – Feb 2025
Publications
*Equal contribution · †Corresponding author
ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals
Under review Paper Project page
From Noise to Diversity: Random Embedding Injection in LLM Reasoning
NeurIPS 2026 ICML 2026 Mechanistic Interpretability Workshop Spotlight · 2.8% Paper Project page
Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer
NeurIPS 2026 ICML 2026 Mechanistic Interpretability Workshop Paper
On the Role of Proposal Support in Diffusion-Based Offline RL for Sequential Decision-Making
ICML 2026 DEMO Workshop Paper
Honors & Awards
Government Scholarship, GIST
Mar 2020 – Aug 2026
Government STEM Scholarship
Mar 2020 – Aug 2026
Dean's List, College of Information and Computing, GIST
Spring 2025, Fall 2025
Korea-Germany Junior Research Fellowship Program 2024
Fund: 12,000,000 KRW (8,197 EUR) by National Research Foundation of Korea (NRF), No. RS-2023-00254343
Sep 2024 – Feb 2025
GIST International Internship Support Program
Sep 2024 – Feb 2025
Hyundai AI Challenge
Top 25 of 330 teams
Sep 2023 – Oct 2023