Woojung Song

Integrated M.S.-Ph.D. in Data Science, Seoul National University

Research Interests

I believe agents can already solve many of the problems we face in the real world. Still, people want better agents, and I don't think we get there just by making models code or complete tasks more accurately. We get there by understanding how agents actually behave, and what values they carry while doing so. My research covers these two sides: Agents and Value Alignment.

  1. Agents. I look beyond whether an agent succeeds or fails. I study the process itself: in which environments agents work well, where they break down, and how they behave along the way.
  2. Value Alignment. I build methods to uncover the values LLMs actually hold and how those values show up in real interactions. In the long run, I aim for human-like LLMs, with a growing interest in social agents and simulation of human behavior.

I'm currently looking for a summer internship where I can contribute to research on agents or value alignment. I'm also always open to collaboration. Reach me at opusdeisong@snu.ac.kr.

News

Recent Publications

See all on the Publications page.

Under Review

Agents Trust Tools Too Much: Measuring Reliance on Unreliable Tools

Hoyeol Yang*, Woojung Song*, Taewon Kim, Jonghyun Song, Seoyeon Park, Yohan Jo  Co-first

Tool-using agents are usually evaluated as if their tools always return the truth. We corrupt what web search, sub-agent delegation, and code execution return, and find that fourteen LLMs adopt the corrupted content at high rates, up to 68% for web search. Reasoning traces show agents often notice the conflict and recover the right answer internally, yet still report only the corrupted one, and none of our prompting, metadata, or post-training interventions fixes this across the board.

Paper Agents

Web-search evaluation design

EMNLP 2026Main

ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?

Woojung Song*, Nalim Kim*, Sangjun Song, Chaewon Heo, Jongwon Lim, Yohan Jo  Co-first

Role-playing agents should evolve with their character, not hold a fixed persona. ArcANE, an automatically built benchmark of 17 novels and 80 characters, shows that conditioning on a Character Arc, the narrative segmented into psychological phases, outperforms every other context strategy, most of all on scenarios the source text never explores.

Project Paper Code Agents Value Alignment

ArcANE construction pipeline

EMNLP 2026Main

Human Psychometric Questionnaires Mischaracterize LLM Behavior

Woojung Song*, Dongmin Choi*, Yoonah Park, Jongwook Han, Eun-Ju Lee, Yohan Jo  Co-first

The psychological profile an LLM reports on a questionnaire is not the one it shows when actually generating text. Familiar questionnaire items cue socially desirable answers, and persona effects that appear on questionnaires vanish on realistic user queries. We argue models should be profiled by what they generate, not what they self-report.

Project Paper Code Value Alignment

Questionnaire vs generation behavior