M.S. in Data Science, Seoul National University (Sep 2025 ~)
B.S. in Information System, Hanyang University
I am interested in agents and value alignment. As AI agents become more integrated into everyday life, we need reliable ways to understand what values they hold and how they actually behave. On the value alignment side, I work on psychometric evaluation of LLMs: whether human questionnaires reliably measure model traits, how data contamination distorts those measurements, and how to build better evaluation instruments. On the agent side, I study how to make smaller models more capable tool users and how agents cope with realistic, uncooperative users.
I'm always happy to talk about shared interests. Feel free to reach out at opusdeisong@snu.ac.kr.
Full list on the Publications page.
Proposes PA-Tool, a training-free method that renames tool schema components to align with small language models' pretraining familiarity, achieving up to 17% improvement on tool-use benchmarks without model retraining.
Shows that LLM psychological profiles from standard questionnaires diverge from profiles derived from actual generation behavior on real user queries.
A systematic framework to measure data contamination across item memorization, evaluation memorization, and target score matching in psychometric LLM benchmarks.
Proposes the Value Portrait framework for evaluating LLM values using items from real user-LLM interactions, finding across 44 LLMs that they emphasize Benevolence, Security, and Self-Direction.