Yunjae Won

I'm an integrated M.S.+Ph.D. student at KAIST AI, advised by Minjoon Seo. I received my B.S. in Electrical Engineering with a minor in Computer Science from KAIST in 2024. Previously, I interned at Naver Place, where I implemented a distributed hyperparameter optimization framework and developed a lightweight, high-accuracy visual quality assessment model. I also interned at the NCSoft Speech AI Lab, where I built a SOTA-level vocal-track extraction model for K-Pop music.

Email  /  CV  /  Scholar  /  LinkedIn  /  Twitter

Research

I am primarily interested in preference optimization, continual learning, and large language models. My current research focuses on quantifying information gain to enhance continual learning and knowledge distillation.

We introduce Context-INformed Grounding Supervision (CINGS), a post-training supervision method where LLMs are trained to ground its responses on external context without overly relying on its own parametric knowledge.

By interpreting preference optimization as the process of learning differential information, we characterize the structure of preference data, justify DPO's log-ratio reward, and explain the resulting behaviors of trained policies.

Projects

By deriving a closed-form expression for the ideal distribution of rejected responses under the DPO framework, we show that it is possible to generate harmful responses from aligned policies.

Using projected gradient descent, we propose a method to learn from human preferences while minimizing forgetting of previous knowledge.

We formulate the metabolic reaction prediction task as a next token prediction problem, and propose using a transformer architecture that predicts the next reaction given the current reaction and its reactants.

Developed a Korean text recognition model on a dataset of handwritings from Korean children. Ranked 1st place out of 1,158 participants using only limited Google Colaboratory resources, outperforming an OCR corporate team backed by state-of-the-art GPUs.

This website's source code was borrowed from Jon Barron.