My research goal is to improve the generalization of generative models and deepen the fundamental understanding of deep learning.

• Diffusion Models

  • Theoretical foundations of diffusion model performance

  • Classifier-Free Guidance (CFG)

  • Integrating Reinforcement Learning with Diffusion Models

• Over-optimization in Diffusion Models

  • Mitigating Reward Hacking

My research goal is to develop a deeper understanding of optimization and learning dynamics in deep learning.

• Optimization

• Learning Dynamics

My research goal is to understand how Self-Supervised Learning works. 

• Self-Supervised Learning 

• Learning Dynamics 

• Role of individual loss terms 

• Desired Embedding Distribution

My research goal is to build trustworthy  AI by developing provable robustness certification

• Trustworthy AI

  • Provable Robustness

    • Randomized Smoothing

  • LLM Jailbreaking

  • Safety & Reasoning trade-off

My research goal is to build a trustworthy, interpretable ai by understanding how models learn.

• Trustworthy AI

  • Adversarial Robustness

    • Robustness Accuracy trade-off

    • Catastrophic Overfitting

  • LLM Jailbreaking

  • Certifiable defense

• Interpretable AI

  • In-Context Learning

  • Linear Transformer

  • Model Unlearning

  • Reasoning

My research goal is to understand why deep learning works.

• Diffusion model

  • Why diffusion models generalize well

  • (ICML 2026) GK and SL, Localizing Memorized Regions in Diffusion Models via Coordinate-Wise Curvature Differences

• Transformer

My research aims to develop improved machine learning methodologies by deepening the understanding of deep learning.

• Deep Learning

  • Optimization and Generalization

    • Implicit Bias

    • Training Dynamics

    • Inconsistency (Disagreement)

    • Sharpness

    • (ICML 2026) HK and SL, Gradient Descent with Large Step Size Restores Symmetry in Deep Linear Networks with Multi-Pathway

    • (ICML 2026) HK, Hyeonseong Kim, and SL, Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data

  • Large Language Models

    • (Linear) Transformers

    • In-Context Learning

My research focuses on developing an in-depth understanding of generative models.

• Deep Learning

• What are the factors influencing the generalization performance of diffusion models?

  • Diffusion Model Architecture

  • Training Dynamics and Generalization Trade-offs

  • Data Augmentation for Improved Generalization

  • Mathematical Framework for Generalization

• Over-optimization in diffusion model

  • Mode Collapse and Sample Diversity in Generalization

• Deep Learning

  • How can we build a better feature extractor?

    • Self-Supervised Learning

    • Optimization and Generalization

    • Advanced Learning Algorithms

    • (NeurIPS 2026) JK, {Yoonsoo Nam and SL}, Saddle-to-Saddle Dynamics in Self-Supervised Shortcut Learning

• Continual Learning

My research goal is to build interpretable and trustworthy LLMs.

• Interpretable LLMs

  • An interpretable analysis of training instabilities, such as loss spikes and gradients exploding, observed during LLMs pre-training.

  • (EMNLP 2025) JH and SL, Variance Sensitivity Induces Attention Entropy Collapse and Instability in Transformers

• Trustworthy LLMs

See Home.