Dahee Kwon

I am a PhD student at the KAIST Graduate School of AI under the supervision of Prof. Jaesik Choi. My research focuses on the internal mechanisms that shape the behavior and capabilities of visual AI systems. My research spans mechanistic interpretability and generative modeling, with a particular focus on how learned representations and computational pathways govern what models can express, which outcomes they favor during inference, and how these choices affect diversity, creativity, and generalization. I seek to understand why powerful AI systems often rely on a limited set of dominant computational patterns, and how this reliance shapes their behavior. By uncovering the principles behind these patterns, I aim to develop a more systematic understanding of visual AI systems and ultimately translate this mechanistic understanding into principled methods for building more generalizable and reliable AI models.

Research Interest

• Generative AI
• Representation Learning
• Interpretability
• Vision-Language Models
• Computer Vision

Education

• KAIST, PhD Candidate in Artificial Intelligence, Aug 2022 - Present
• KAIST, M.S. in Artificial Intelligence, Aug 2020 - Aug 2022
• Yonsei University, Bachelor's degree in Applied Statistics, Mar 2016 - Aug 2020

Awards

• Insung Scholarship, 2025
• KAIST Breakthroughs Spring 2026, selected for KAIST Breakthroughs 50

Selected Publications

See all publications

Talks

• Understanding Deep Neural Networks Decision-Making Through Exploring Learned Features, AI EXPO KOREA 2024 Workshop
• Analyzing the Attribute-relevant Featuremaps in Stable Diffusion Models, KCC XAI Workshop
• Understanding Diffusion-based Generative Models, KAIST XAI Tutorial Series
• Enhancing Creativity in Text-to-Image Generation, Samsung AI Forum