-ˋˏ ｡ﾟ 　 Mingyu Kim 　 ﾟ｡ ˎˊ-

I'm an integrated course graduate student at Yonsei University in Seoul, Korea. I'm majoring in Statistics and Data Science, working as a researcher at MLAI advised by Prof. Kyungwoo Song.

My research interests are mainly on robustness of Large Language Models (LLMs) and Causal Machine Learning. Moreover, I also have interests in statistical and theoretical analysis of machine learning methods.

Email  /  CV  /  Github  /  LinkedIn

Always open to any research opportunities for related topics.

News

[Feb. 2025] I graduated as the top student (summa cum laude) in my department!

[Oct. 2024] I was qualified for the BS/MS integrated program in the department of statistics and data science!

Education

M.S./Ph.D. in the Department of Statistics and Data Science, Yonsei University, Seoul, Korea
    Mar. 2025 - Present

B.S. in the Department of Applied Statistics, Yonsei University, Seoul, Korea
    → GPA: 4.25/4.3 (overall), 4.27/4.3 (major) - Top Student in the Department
    Mar. 2020 - Feb. 2025

Research Interests

1. Statistical Inference on LLMs
To make the robust algorithm that learns invariant representations under harsh distribution shift, we need to develop the methodology based on statistical inference. I especially interested in application of the statistical methods such as conformal inference, and so on.

Publications & Preprints

This paper proposes FANS (Function And Noise Separation), a unified framework for Causal Mechanism Shifts (CMS) by disentangling changes in causal functions from changes in noise. Checking the independence between estimated noise and parent variables in new environments, it determines whether shifts are function-driven or noise-driven, while handling complex and diverse noise changes (e.g., higher-order moment shifts).

Projects

This project aims to detect and reduce translationese in English–Korean translations. Building on findings that translationese inflates performance metrics, we propose a fine-tuned encoder–decoder model using semi-synthetic dataset. The goal is to develop a style transfer system that rewrites translationese Korean into more natural Korean while preserving meaning.

This project aimed to improve wine recommendations using graph-based recommender systems on a highly sparse dataset from VIVINO. We benchmarked several models for handling sparsity including LightGCN, KGAT, MKR, MCCF, and GFormer. Despite limited interaction data and knowledge graph complexity, we successfully identified robust solutions for sparse data cases.

For MBTI personality classification, it proposes multi-stage transfer learning approach with KoELECTRA and GPT-4o. Pipeline includes 3 phases: task-specific fine-tuning with official MBTI texts, masked language modeling with GPT-guided masking, and classification. By integrating domain knowledge and contextual embeddings, it showed classification improvement.

Based on FADING, the diffusion based model for aging and GOAE, the 3D rendering model with only one picture, we made two-step face aging architecture especially for Korean missing children. The addressed pre-process techniques such as upscaling and labeling and the results from this project are expected to contribute to related research for face aging.

Using the Mask R-CNN architecture for face detection, we made the model that applies mosaic effect for all detected faces Fine tuned by the WIDER FACE dataset, it performed well both for image and video. Although it was conducted as simple project for reviewing, this project helped me to broaden the technique related to AI research.

Solving the problem of not being able to distribute the amount of transportation compared to the high demand for late-night buses, project presented a new route by analyzing the actual amount of movement and bus boarding during the night time. It is quite meaningful in terms of suggesting a new plan for a circular route rather than a simple end-to-end route.

To make the agent morality in RL, previous research aims to implement the conflict of ethics in terms of voting system. In this project, we aimed to implement the moral determinations based on Bayesian RL concept with multi-agent. Also, we implement the situation of trolley problem where the agent has to choose trade-off option using the safe gymnasium enviorment.

Based on the fact that the child protection zone in Osan City is insufficient to the number of traffic accidents occured, we analyzed the actual area status. For the analysis, spatial analysis APIs and multilinear regression model with diagnostic methods were used to increase their accuracy and intuition.

Talks

Lightly presented the recently developed methodologies for Parameter Efficient Fine-Tuning in LLM, including LoRA, Prefix-tuning, and their variants.

Lightly presented the concept and flow of various methodologies derived from DDPM, and introduced the core of DDPO which combines generative model and reinforcement learning methodology.

Teaching

Deep Learning (STA3140) - Professor: Jaewoo Park
Student Tutor on Spring 2024

SW Programming (YCS1002) - Professor: Jaekyung Kim
Course Tutor on Fall 2021

Awards and Scholarships

Yonsei Academic Track Scholorship - Spring 2025
Awarded 5 out among total graduate student freshmen in department.

Guwon Scholorship - Spring 2024 ~ Spring 2025, Spring 2022
Awarded 2 out of total 300+ department students. Detailed information can be found here.

Academic Awards at Yonsei Univ. - Multiple Semesters
Top 3+% over 120 students

Extracurricular Activities

Yonsei Artificial Intelligence
14th member (Jul. 2024 - Jun. 2025)

Yonsei Data Science Lab
12th member & head of academic team (Jul. 2024 - May. 2025)

Website templete adapted from Jon Barron.