Research
Our research spans core machine learning algorithms, theoretical insights, and applied systems for perception, language, and decision making.
Deep Learning Application for Scientific Discovery
Dongwoo Kim · Seungbeom Lee
This research focuses on applying deep learning to the field of scientific discovery. Specifically, we aim to build model architectures tailored to each scientific domain while efficiently incorporating inductive biases, such as geometric priors, to address various scientific challenges. Our work has broad applications in areas such as molecule generation, molecular property prediction and drug-target binding pose estimation.
Effective Interaction between Users and LLMs
Dongwoo Kim · Youngbin Choi · Minjong Lee
People increasingly rely on LLMs in everyday scenarios, asking them a wide range of questions and tasks. However, LLMs often produce incorrect or unintended responses. To obtain the desired outcome, users interact with the model over multiple turns to iteratively refine its output. In this study, we explore ways to enhance the effectiveness of such interactions. For example, we propose in-place feedback, which moves beyond the conventional multi-turn setting by allowing users to directly edit the model’s previous response, leading to more efficient and natural interaction.
Graph Fouondation Models
Dongwoo Kim · Kyeongheung Yun
Recently, foundation models capable of processing diverse data and handling various tasks have emerged. Applying this concept to graphs allows us to envision models that can solve diverse graph types and tasks. However, graphs exhibit vastly different characteristics depending on domains like molecules or citations, making knowledge transfer extremely challenging. This research explores methods for understanding graphs, facilitating knowledge transfer across domains, and leveraging LLMs for graph processing. Ultimately, we aim to develop a foundation model that comprehends diverse graphs and excels at performing all types of graph tasks.
Graph Learning Enhancement by Structural Modifications
Dongwoo Kim · Jaeseung Heo
Using high-quality data is crucial for training accurate machine learning models. Graph data, which includes structural information in the form of edges, offers unique opportunities for improving model performance. Graph learning enhancement aims to refine this structural information by removing noise and emphasizing more informative patterns. Techniques such as edge rewiring, which modifies connections between nodes, node dropping, which removes less important nodes, and label smoothing, which adjusts labels to reduce overconfidence, all leverage the inherent relationships in the graph to improve model accuracy and generalization.
Theoretical Insights into How Neural Network Architectures Affect Performance
Dongwoo Kim · MoonJeong Park
Neural network architectures implicitly encode priors that influence how models understand data and generalize to new examples. Convolutional networks, for example, embed translational invariance suitable for visual perception, while graph neural networks assume relational smoothness between connected nodes. These structural biases improve performance in a parameter-efficient manner but can also introduce side effects, such as the oversmoothing phenomenon in GNNs. Therefore, a precise understanding of how architectural design influences learning is essential for developing robust and effective neural networks. This research aims to uncover the implicit biases underlying various neural network architectures and to analyze their effects from multiple theoretical perspectives, including learning dynamics, optimization stability, and generalization error. Such understanding reveals when and why models perform well or fail, guiding the design of more robust architectures.
Time Series Forecasting
Dongwoo Kim · Sunghyun Choi
Recent advances in deep learning have significantly improved time series forecasting, yet many models still struggle with non-stationarity, irregular sampling, and long-term dependency degradation. Our research addresses these challenges by developing robust forecasting frameworks that integrate diffusion-based generative modeling and temporal representation learning. Specifically, we explore how moving average mechanisms and denoising diffusion processes can enhance stability under distributional shifts and missing intervals. Through this approach, we aim to achieve more adaptive and generalizable forecasting models that can accurately predict complex temporal patterns in domains such as energy, transportation, and semiconductor manufacturing.
Advanced Planning for Robots with Multimodal Large Language Models
Jungseul Ok · Hyejin Park
Recent advances in AI have enabled robots to perform simple tasks based on language instructions. However, they struggle to handle long-term tasks in complex environments, due to difficulties in integrating visual observations with language instructions, and planning subsequent actions accordingly. In this study, we develop methods leveraging multimodal large language models to help robots interpret their visual surroundings and plan more effectively, ultimately enabling them to perform real-world tasks successfully.
Foundational Datasets for VLMs
Jungseul Ok · Hoyoung Kim
Vision-Language Models (VLMs) leverage large-scale multimodal datasets, such as image-text and video-text pairs, to learn cross-modal representations. While the abundance of data has driven significant advancements, it also presents challenges. First, excessive data volume increases computational costs and reduces training efficiency, making it difficult to extract high-quality representations. Additionally, such datasets contain noisy annotations, redundant samples, and domain biases, all of which can negatively impact model performance. Our research aims to evaluate the effect of dataset quality on VLM training, develop filtering techniques to eliminate noisy and redundant data, and explore strategies for generating high-quality synthetic data to enhance learning efficiency.
Hierarchical Planning for Efficient Exploration
Jungseul Ok · Youngsik Yoon
Hierarchical planning is essential for efficient exploration in Reinforcement Learning (RL), as it helps agents structure decisions across different levels. By focusing on high-level goals before refining low-level actions, it enables smarter exploration in complex environments. For example, in path planning tasks, hierarchical strategies can optimize route selection, while in LLM-based planning, they can help structure tasks into manageable subgoals. This approach is particularly effective in long-horizon tasks and sparse reward settings, where traditional methods often lead to inefficient exploration and slow learning. In this study, we explore how hierarchical planning can accelerate exploration and improve learning efficiency in RL.
Human-in-the-Loop Personalization
Jungseul Ok · Minhyeon Oh
Recently, we are moving beyond simply making AI systems more powerful and towards making them more human-centered—tailored to individual needs, preferences, and contexts. This requires AI to contextually learn and adapt based on user interaction, actively incorporating implicit user feedback and guidance into its response process. Therefore, we are developing AI that: (1) deeply understands dynamic user needs and contexts, leveraging implicit feedback; (2) exhibits context-aware learning, recognizing preference shifts across situations and adapting accordingly; and (3) provides interpretable personalization parameters, giving users more direct control. Participants in this study will contribute to our ongoing efforts to this development.
Post-hoc Explainable AI with LLMs
Jungseul Ok · Jaechang Kim
Post-hoc explainability is essential for improving the transparency and trustworthiness of machine learning models. Traditional methods, such as saliency maps, provide valuable insights into model decisions but often require domain expertise to interpret effectively. With the rise of LLMs, natural language explanations have emerged as a more accessible and flexible alternative, allowing users to engage with model reasoning in an intuitive way. This project explores how LLMs can generate contextually rich and faithful explanations while maintaining alignment with the underlying model’s decision-making process. We investigate techniques that enhance interpretability through structured reasoning and interactive dialogue, ultimately contributing to the development of more transparent and user-friendly AI systems.
Theoretical Analysis for Online Learning
Jungseul Ok · SeockBean Song
Online learning involves algorithms that make decisions and learn from data sequentially. We aim to develop a solid theoretical foundation to understand and evaluate the performance of these methods. In this project, students will study the core principles of online learning, including performance analysis, regret minimization, and the development of provably efficient algorithms.