Seungwon Lim
Hi, I'm Seungwon Lim. I'm a researcher at Yonsei University, MIRLAB (Multimodal Intelligance Research Lab) advised by Youngjae Yu. I received my bachelor's degree in Computer Science, and I am currently pursuing an integrated MS/PhD program in Computer Science.
My research question centers on developing reliable agent systems. To achieve this, I am currently focusing on agent’s reasoning, action-decision, and human-centric AI.
Publications
Verifying the Verifiers: Unveiling Pitfalls and Potentials in Fact Verifiers
COLM2025
When AI co-scientists fail: SPOT-a benchmark for automated verification of scientific research
Under Review
TLDR; We introduce SPOT, a benchmark for automated verification of scientific research, and show a substantial margin exists for AI-assisted academic verification.
VisEscape: A Benchmark for Evaluating Exploration-driven Decision-making in Virtual Escape Rooms
Under Review
TLDR; We introduce VisEscape inspired by Escape Room games, and evaluate the Reasoning and Decision-making of diverse MLLMs in exploration-driven and dynamic environments.
Persona Dynamics: Unveiling the Impact of Persona Traits on Agents in Text-Based Games
ACL2025 Main
TLDR; We introduce PANDA, which incorporates Human Personality Traits into AI agents for Text-based Games and examines how these traits impact their behavior and performance.
Do LLMs Have Distinct and Consistent Personality? TRAIT: Personality Testset designed for LLMs with Psychometrics
NAACL2025 Findings
TLDR; We introduce a psychometric-based benchmark TRAIT to measure the personality revealed in the Behavior Patterns of LLMs along with verification of Reliability and Validity.
MASS: Overcoming Language Bias in Image-Text Matching
AAAI2025 Main
TLDR; We introduce MASS, a Training-free framework that improves Visual Accuracy and reduces Bias in Image-Text Matching for pretrained visual-language models.
Can visual language models resolve textual ambiguity with visual cues? Let visual puns tell you!
EMNLP2024 Main
TLDR; We introduce UNPIE, a new benchmark crafted to evaluate how multimodal inputs influence the Resolution of Lexical Ambiguities.
CLARA: Classifying and Disambiguating User Commands for Reliable Interactive Robotic Agents
ICRA2024
TLDR; We introduce CLARA, a LLM-empowered method for robots to estimate Uncertainty of user commands and to Disambiguate them via question generation for clarification.