Minchan Kwon

Stable-GFlowNet: Toward Diverse and Robust LLM Red-Teaming via Contrastive Trajectory Balance

Minchan Kwon, Sunghyun Baek, Minseo Kim, Dongyoon Han*, Junmo Kim*

ICML 2026 Spotlight (Top 2.2%) • Sep. 2025

Paper

We address the critical instability issues that arise when training LLMs with GFlowNet via a new loss function called Contrastive Trajectory Balance. The approach ensures stable LLM training while preserving GFN's diversity, discovering 7x more adversarial prompts than the original GFN.

Preference Distillation via Value based Reinforcement Learning

Minchan Kwon, Junwon Ko, Kangil Kim*, Junmo Kim*

NeurIPS 2025 • Sep. 2025

Paper | Code

Examining the distillation task on a DPO-style dataset from an RL perspective, we observed a multi-reward phenomenon. We propose a method that resolves this multi-reward issue and provide a mathematical proof for it.

StablePrompt: Automatic Prompt Tuning using Reinforcement Learning for Large Language Model

Minchan Kwon, Gaeun Kim, Jongsuk Kim, Haeil Lee, Junmo Kim

EMNLP 2024 • Sep. 2024

Paper | Code

Creating suitable prompts manually for each task is painful. We provide a method for generating prompts fully automatically using Online-RL, requiring only the task description, training dataset, the LLM to use the prompts, and the LLM to create them.

ConceptPrism: Concept Disentanglement in Personalized Diffusion Models via Residual Token Optimization

Minseo Kim, Minchan Kwon, Dongyeun Lee, Yunho Jeon*, Junmo Kim*

CVPR 2026 • Feb. 2025

Reducing the Content Bias for AI-generated Image Detection

Seoyeon Gye*, Junwon Ko*, Hyounguk Shon*, Minchan Kwon, Junmo Kim

WACV 2025 Oral (Top 18%) • Feb. 2025

FairASR: Fair Audio Contrastive Learning for Automatic Speech Recognition

Jongsuk Kim, Jaemyung Yu, Minchan Kwon, Junmo Kim

Interspeech 2025 • Aug. 2025

Revisiting Softmax Masking: Stop Gradient for Enhancing Stability in Replay-based Continual Learning

Hoyong Kim, Minchan Kwon, Kangil Kim

CoLLAs 2024 Workshop • Jun. 2024

Korea Advanced Institute of Science and Technology (KAIST)

Ph.D in Electrical Engineering • Mar. 2024 — Present

Advisor: Prof. Junmo Kim. Daejeon, Korea.

Korea Advanced Institute of Science and Technology (KAIST)

M.S, Graduate School of AI • Mar. 2022 — Feb. 2024

Advisor: Prof. Junmo Kim. Daejeon, Korea.

Gwangju Institute of Science and Technology (GIST)

B.S in Electrical Engineering and Computer Science • Mar. 2018 — Mar. 2022

Minor in Math and Economics. TGPA: 3.88/4.5, Major GPA: 4.21/4.5. Gwangju, Korea.

Naver

Research Intern • Jul. 2023 — Oct. 2023

Developed a Vision-Language Model based on LLM (Korean and Japanese) for the Naver Maps AI team. Studied model architectures with researchers, and refined Japanese/Korean vision-language data using CLIP score, aesthetic score, and other tools.