M.S. student in Artificial Intelligence, KAIST
I am an M.S. student at KAIST AI, advised by Prof. Se-Young Yun. In 2026, I was a visiting researcher at the Vector Institute and the University of Toronto, working with Prof. Rahul G. Krishnan. I received my B.S. in Electrical and Computer Engineering from Seoul National University.
I want to close the gap between what people do easily and what models still cannot.
Rotate a Sudoku board and a person sees the same puzzle; a language model that just solved it often fails. A person knows where a thrown ball will land without picturing every instant in between; most world models still generate every frame.
People reason instead of matching patterns, learn from very little, and act in a physical world they understand. I work on bringing models closer to each: reasoning, efficient learning, and world models for embodied agents.
* equal contribution, random order
EMNLP 2026, Main Conference (Oral)
A masked diffusion language model can be post-trained on under 4% of its tokens. Supervising only the few commitments that most reduce uncertainty beats full-sequence fine-tuning and online RL at a fraction of the compute.
NeurIPS 2025 (Spotlight)
A training-free sampler that uses the KL divergence between denoising steps to find tokens that are safe to unmask in parallel. Up to 2.78× faster generation while improving over greedy decoding, across text, image, and molecular generation.
Under review. Earlier version at the ICML 2026 Workshop on Foundations of Deep Generative Models
Accuracy on one presentation of a problem hides failures on logically equivalent ones. We measure whether correctness survives answer-preserving rewrites, and a post-training recipe raises the all-view correctness of LLaDA-8B from 47.7% to 86.1%.
RSS 2026 Workshop on Trustworthy Embodied Foundation Models
A repair module that adds route guidance to a CBF-QP safety filter, so a vision-language-action policy can rejoin its task trajectory after avoiding an obstacle.
A world model that predicts the next physical event, its elapsed time, and the resulting state for each candidate action, instead of rolling out fixed-interval frames.
Three questions still stand between the two banks.
Hover or tap a stone to read its question.
Where does pattern matching end and reasoning begin?
Pattern matching is right near what a model has seen. Reasoning is using the structure underneath to be right where it has never been.
Why do models need so much more than people do?
Scale works, but every gain asks for more data than the last. People learn a new skill from a handful of examples; models have to learn more from less.
Can a model keep up with a world that does not wait?
The physical world keeps moving while a model computes. Acting in it takes a model that predicts what its own actions will cause, skips the details that do not matter, and corrects itself when it is wrong.
If you are working on one of these gaps, or want to, I would be glad to talk: sunwoo@kaist.ac.kr
Outside research, I play the drums and like watching trees and clouds. My favorite films are WALL-E and Boyhood; for music, Pachelbel's Canon, the Beach Boys' Pet Sounds, and U2's The Joshua Tree. In high school, on a trip to Boston, I snuck into a Harvard lecture. I understood almost none of it. Judging by the faces around me, I was not alone.