Preprint 2026Under review
AgentHabit: Characterizing Distinct Behaviors of Agents on Everyday Tasks
Agents show distinct behavioral tendencies that remain recognizable across everyday tasks.
I am a Ph.D. student in Artificial Intelligence at Seoul National University, advised by Prof. Yohan Jo in the Human-Oriented Language Intelligence (HOLI) Lab.
Research Interests
-
Understanding agents
- How can we assess the values and behavioral tendencies agents express in realistic interactions? [ACL ’25] [EMNLP ’26] [arXiv ’26]
- When do agents fail in interactions with users and tools? [ICLR ’26] [arXiv ’26]
-
Shaping agents
- How can we build models that faithfully simulate human behavior across people and situations? [TACL ’26] [EMNLP ’26]
- How can we improve agents’ decisions and tool use through training and interaction design? [ACL ’26]
News
-
ArcANE and Human Psychometric Questionnaires Mischaracterize LLM Behavior accepted to EMNLP 2026 (Main).
-
Awarded the Presidential Science Scholarship for Graduate Students (2026–2027) by the Ministry of Science and ICT.
-
PA-Tool accepted to ACL 2026 (Main).
-
Psychometric Item Validation Using Virtual Respondents accepted to TACL.
Earlier news
-
Non-Collaborative User Simulators for Tool Agents accepted to ICLR 2026.
-
Quantifying Data Contamination in Psychometric Evaluations of LLMs accepted to EACL 2026 Findings.
-
Value Portrait accepted to ACL 2025 (Main).
Publications
All publications
Preprint 2026Under review
Agents' Overreliance on Unreliable Tools
Agents often repeat incorrect tool outputs despite noticing conflicts; verification prompts reduce this reliance.
EMNLP 2026Main
ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?
Tracking changes in a character's values, motivations, and relationships improves role-play fidelity.
EMNLP 2026Main
Human Psychometric Questionnaires Mischaracterize LLM Behavior
Questionnaire-based value and personality profiles diverge from behavior on realistic user queries.
