Bio
I am currently on the job market, for research scientist opportunities in multimodal post training and evaluation + interactive agents.
I am a postdoctoral researcher in Artificial Intelligence at Yonsei University, where I also received my Ph.D. in 2026, advised by Professor Seon Joo Kim. Previously, I was a research intern at Microsoft Research, LG AI Research, and Naver Cloud (HyperCLOVA X).
My research focuses on multimodal intelligence, especially reasoning, grounding, and agentic behavior. I develop models and evaluations for systems that reason over visual evidence and make decisions across multiple steps. I have published 17 first-author papers at venues including NeurIPS, ICML, ICLR, CVPR, ICCV, ACL, and EMNLP.
Research interests: multimodal reasoning, grounded AI, agentic systems, diagnostic evaluation, and representation learning.
Featured Research
MathLensWhat MLLMs Learn When They Learn Multimodal Reasoning
NeurIPS. 2026.
Work done as part of an internship in Microsoft
Developed MathLens, a benchmark that separates perception, reasoning, and multimodal-specific skills, revealing how RL and SFT post-training shape distinct capabilities in multimodal LLMs.
DIST2LossTeaching Metric Distance to Discrete Autoregressive Language Models
ICLR. 2026.
Introduced a training objective that teaches discrete autoregressive language models to represent metric distance, improving their ability to capture spatial and relational structure in applications as embodied AI.
Rethinking State Tracking in Recurrent Models Through Error Control Dynamics
NeurIPS. 2026.
Showed that affine recurrent models such as state-space models and linear attention cannot correct state-tracking errors, explaining why they learn finite-horizon solutions and predicting the horizon at which tracking collapses.
Research Experience
Yonsei University, CIP Lab Seoul, South Korea
Postdoctoral Researcher Sep. 2026 - Present
PI: Seon Joo Kim
Microsoft Research, AI Frontiers Redmond, U.S.
Research Intern May 2025 - Aug. 2025
Conducted research on diagnosing multimodal reasoning capabilities in MLLMs (MathLens; NeurIPS 2026), with the collaboration later continuing into process-level evaluation of web agents (WebStep; 2026).
LG AI Research, AML Seoul, South Korea
Research Intern Mar. 2024 - Jul. 2024
Conducted research on extending multimodal LLMs beyond language-only outputs, including embodied VLA, image generation, and mathematical reasoning (DIST2Loss; ICLR 2026).
Naver Cloud, Foundational Research Seongnam, South Korea
Research Intern Jul. 2023 - Jan. 2024
Contributed to HyperCLOVA X LLM development and conducted preliminary exploration on omnimodal language modeling.
Education
Yonsei University Seoul, South Korea
Ph.D., Artificial Intelligence Mar. 2023 - Aug. 2026
Advisor: Seon Joo Kim
Seoul National University Seoul, South Korea
M.S., Computer Engineering Mar. 2019 - Feb. 2023
Advisor: Gunhee Kim
Yonsei University Seoul, South Korea
B.Sc., Computer Science and Philosophy (Double Major) 2014 - 2018
Publications
Research outputs across venues. ‡ = equal contribution. Full list on Scholar
Mixture of Layers: Dynamic Layer Routing for Visual Reasoning
NeurIPS. 2026.