Projects

- AOP-grounded mechanistic toxicity reasoning to validate LLM predictions and curb hallucinations.
- Currently under review

- Designing implicit harmonization methods that preserve dataset-specific clinical signals for downstream tasks.

- Analyzing how CXR representations (raw images vs. radiology reports) affects post-ICU mortality prediction combined with discharge notes.
- Currently under review

- Developing a compact pivot-language framework that encodes English reasoning into latent representations for multilingual medical QA.
- Currently under review

- Studying early detection of Alzheimer’s disease using real-world Korean patient speech and interview transcripts.

- Developed LLM-based aspect-oriented summarization pipelines for 30-day psychiatric readmission prediction using multi-hospital EHR data.
- Published at EMNLP 2025: “Aspect-Oriented Summarization for Psychiatric Short-Term Readmission Prediction.”

- Improving biomedical generative entity linking via preference optimization with negative samples to handle ambiguous mentions more robustly.
- Published at ACL 2025 Findings: “Learning from Negative Samples in Biomedical Generative Entity Linking.”
- Code: GitHub; checkpoints: Hugging Face.

- Exploring retrieval-augmented generation workflows where domain experts iteratively refine biomedical hypotheses with LLM assistance.

- Fine-tuning LLMs to provide guideline-grounded explanations for diabetes care as part of a broader precision-medicine platform.

- Developing an open-source Korean LLM that explains cultural heritage sites in faithful, accessible language for non-experts.