Woojung Song
Integrated M.S.-Ph.D. in Data Science, Seoul National University
Research Interests
I'm interested in AI agents and value alignment. As LLMs become more capable, they are increasingly embedded in everyday life through tool use and direct interaction with people. With this comes the need for agents that can handle diverse real-world tasks reliably, and for evaluation methods that go beyond surface-level benchmarks to capture what values these systems actually express in practice.
- Tool-Use Agents. Enabling language models to use tools effectively so they can explore broader environments and help users solve real-world tasks across diverse scenarios.
- Value Alignment. As LLMs take on more societal roles, questions around safety and pluralistic values are becoming central. Yet we still lack adequate instruments to measure what values models hold and how they manifest in context. I work on building better evaluation frameworks for this.
If any of this resonates, I'd love to chat. Reach me at opusdeisong@snu.ac.kr.
Selected Publications
See all on the Publications page.
EMNLP 2026Main
ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?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.
Paper Agents Value Alignment
EMNLP 2026Main
Human Psychometric Questionnaires Mischaracterize LLM BehaviorThe 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—so we argue models should be profiled by what they generate, not what they self-report.
ACL 2026Main
Don't Adapt Small Language Models for Tools; Adapt Tool Schemas to the ModelsSmall models often hallucinate tool names that follow their pretraining conventions rather than the schema they were given. PA-Tool flips the fix: instead of retraining the model, it renames schema components into naming the model already knows, cutting such errors by 80% and raising tool-use accuracy by up to 17%.
ICLR 2026
Non-Collaborative User Simulators for Tool AgentsA user simulator that behaves like real users at their worst: asking for the unavailable, digressing, growing impatient, and leaving information out. Facing these users on MultiWOZ and τ-bench, state-of-the-art tool agents degrade sharply, hallucinating more and breaking down mid-dialogue.
ACL 2025Main
Value Portrait: Assessing Language Models' Values through Psychometrically and Ecologically Valid ItemsA value benchmark built from real user-LLM interactions, with every item psychometrically validated against human raters' own value scores rather than annotator intuition. Across 44 LLMs, models consistently favor Benevolence, Security, and Self-Direction over Tradition, Power, and Achievement.
