My research interests lie in understanding how large language models behave throughout their reasoning processes, particularly how and why failures emerge. I am especially interested in safety failures, including how unsafe behaviors develop during reasoning and how these insights can inform methods for detecting and preventing them. I am also interested in how model behavior varies across languages, users, and social contexts, as well as the underlying factors that shape these differences.
Education
M.S., Data Science — Seoul National University (Sep. 2023 - Feb. 2026)
Exchange Student — New York University (Sep. 2021 - Dec. 2021)
B.A., Political Science and Diplomacy — Yonsei University (Mar. 2018 - Feb. 2023)
Publications
TrASE: Trajectory-Aware Safety Evaluation for Large Reasoning Models
Under Review
Beyond Case Law: Evaluating Structure-Aware Retrieval and Safety in Statute-Centric Legal QA
ACL 2026 Oral • arXiv
From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment
arXiv preprint • arXiv
"Well, Keep Thinking": Enhancing LLM Reasoning with Adaptive Injection Decoding
ACL 2025 Findings • arXiv
* : Equal Contribution
TrASE: Trajectory-Aware Safety Evaluation for Large Reasoning Models
Under Review
Beyond Case Law: Evaluating Structure-Aware Retrieval and Safety in Statute-Centric Legal QA
ACL 2026 Oral • arXiv
What If TSF: A Benchmark for Reframing Forecasting as Scenario-Guided Multimodal Forecasting
Under Review
From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment
arXiv preprint • arXiv
ATAS: Any-to-Any Self-Distillation for Enhanced Open-Vocabulary Dense Prediction
ICCV 2025 • arXiv
"Well, Keep Thinking": Enhancing LLM Reasoning with Adaptive Injection Decoding
ACL 2025 Findings • arXiv
HyperCLOVA X 8B Omni
Technical Report • arXiv
* : Equal Contribution
Work Experience
Residency Program, NAVER Cloud — Oct. 2025 - Jul. 2026
NLP Research Engineer, Twigfarm — Oct. 2022 - Feb. 2023
Data Analyst Intern, UpennSolution — Feb. 2021 - Jul. 2021
Research Projects
Clinical Information Extraction — Oct. 2025 - Feb. 2026
- Constructed an LLM-based pipeline that extracts and structures key clinical information from epilepsy patient notes, ensuring robustness to heterogeneous physician documentation styles.
Legal RAG System — Mar. 2025 - Aug. 2025
- Constructed a legal retrieval corpus and an evaluation QA dataset for Korean fire safety law, and developed a domain-specific RAG system validated on this benchmark.
Face Blur Estimation — Jul. 2022 - Sep. 2022
- Developed and evaluated a regression model estimating blur levels in facial images to filter motion-blurred inputs for reliable face recognition.
Patents
Method and Apparatus for Adaptive Injection Decoding to Prevent Incomplete Reasoning in Large Language Models
Awards & Honors
Fellowships / Scholarships
Support for Next-Generation Researchers Program
Merit-based Scholarship
Jilli Scholarship (Merit-based Scholarship)
Awards
Honors