Hanbum Ko

I am a Ph.D. student in Artificial Intelligence at Korea University, advised by Prof. Sungwoong Kim and Prof. Sungbin Lim, as a member of the AGI Laboratory, and I received my M.S. in Artificial Intelligence from UNIST.

My research is on AI for Science — building large language models that reason about chemistry, ground themselves in molecular structure, and can be trusted on real scientific tasks.

Email  /  CV  /  Scholar  /  Github  /  LinkedIn

Updates

Research

My research is on AI for Science, primarily large language models for chemistry. Three lines of work: eliciting explicit reasoning for retrosynthesis and reaction prediction; grounding multimodal LLMs in molecular structure so that molecular graphs are used rather than ignored; and building the data pipelines and benchmarks needed to train and evaluate these models on realistic tasks.

I also work on neural combinatorial optimization for industrial problems such as circuit routing and truck-and-drone delivery. In both areas I work on settings with real-world constraints — feasible solutions that are hard to find, semi-structured data, and evaluation that reflects how a model would actually be used.

First-author works are highlighted below.

Trains an LLM to reason explicitly about strategic bond disconnections — which bond to break, and why — before proposing precursors, so a retrosynthesis prediction can be inspected rather than merely ranked.

A benchmark that places LLM agents in scenario-grounded design tasks — an objective and its constraints — and measures whether the molecules they propose actually satisfy them.

Rather than mapping reactants to products in a single pass, the model works through the transformation step by step, which helps most where pattern matching over SMILES fails.

Multimodal molecular LLMs are given molecular graphs but rarely use them. Molecular structure Preference Optimization (MolPO) optimizes preferences between correct and perturbed structures, so the model exploits structure and outperforms both generalist and specialist models on property and reaction prediction.

Rectilinear Steiner tree packing, the problem behind circuit routing, is hard even to find a feasible solution for. Our hierarchical optimizer decomposes it into high- and low-level sub-problems with far smaller search spaces, then composes their solutions.

A framework for gathering, augmenting, and restructuring unstructured scientific documents with LLMs, applied to chemical reactions for document augmentation and knowledge-graph completion.

A hybrid attention-encoder / LSTM-decoder model for routing a truck and a drone in tandem, improving both solution quality and computation time over purely attention-based models.

A generator and benchmark producing rectilinear Steiner tree packing instances at scale, with controllable size and difficulty, for training and comparing learned routing solvers.

A drone-swarm sensing and search system for disaster response, resilient to the loss of individual drones.

Structural augmentations to a UAV airframe that reduce the noise it radiates in flight.

Education

Experience

Selected Projects & Awards

Template borrowed from Jon Barron.