Eunkyu (Eunice) Park

Ph.D. Candidate @ SNUVL

Pittsburgh, Pennsylvania

Hello! I am a Ph.D. student in Artificial Intelligence at Seoul National University, advised by Professor Gunhee Kim. Before SNU, I received my B.S. in Computer Science from Columbia University and a B.A. in Mathematics from Bard College. I am currently visiting Carnegie Mellon University’s Human-Computer Interaction Institute (HCII) as a Collaborative Researcher vising Professor Motahhare Eslami’s lab.

My research centers on multimodal AI, with a focus on the trustworthiness and interpretability of Vision-Language Models (VLMs). Previous works range from (1) bias-driven hallucinations in VLMs—examining how biases in model components can lead to unfaithful outputs to (2) how VLMs make moral judgments in multimodal contexts, and how their outputs align (or misalign) with human values. Most recently, my work investigates how to scaffold intermediate reasoning in order to better evaluate and calibrate user trust in model outputs. Across these directions, I design benchmarks, annotation pipelines, and preference learning techniques to improve the transparency, safety, and alignment of multimodal systems.

In my free time, I enjoy playing golf , watching football, and savoring good food and wine with good company .

news

Jul 27, 2025	Our latest work CoCoT! is out and live! Paper written during my first month at CMU!
Jun 30, 2025	Our dataset HalLoc V1 is finally out on HuggingFace!
May 28, 2025	I’ll be spending June through the end of the Fall 2025 semester as a Visiting Scholar at Carnegie Mellon University’s Human-Computer Interaction Institute (HCII).
Feb 28, 2025	Our paper, HalLoc: Token-level Localization of Hallucinations for Vision Language Models is accepted to CVPR 2025

publications

HalLoc: Token-level Localization of Hallucinations for Vision Language Models

Eunkyu Park^*, Minyeong Kim^*, and Gunhee Kim

CVPR, 2025

Abs

Hallucinations pose a significant challenge to the reliability of large vision-language models, making their detection essential for ensuring accuracy in critical applications. Current detection methods often rely on computationally intensive models, leading to high latency and resource demands. Their definitive outcomes also fail to account for real-world scenarios where the line between hallucinated and truthful information is unclear. To address these issues, we propose HalLoc, a dataset designed for efficient, probabilistic hallucination detection. It features 150K token-level annotated samples, including hallucination types, across Visual Question Answering (VQA), instruction-following, and image captioning tasks. This dataset facilitates the development of models that detect hallucinations with graded confidence, enabling more informed user interactions. Additionally, we introduce a baseline model trained on HalLoc, offering low-overhead, concurrent hallucination detection during generation. The model can be seamlessly integrated into existing VLMs, improving reliability while preserving efficiency. The prospect of a robust plug-and-play hallucination detection module opens new avenues for enhancing the trustworthiness of vision-language models in real-world applications.
Cognitive Chain-of-Thought: Structured Multimodal Reasoning about Social Situations

Pre-print, Underreview, 2025

Abs

Chain-of-Thought (CoT) prompting helps models think step by step. But what happens when they must see, understand, and judge-all at once? In visual tasks grounded in social context, where bridging perception with norm-grounded judgments is essential, flat CoT often breaks down. We introduce Cognitive Chain-of-Thought (CoCoT), a prompting strategy that scaffolds VLM reasoning through three cognitively inspired stages: perception, situation, and norm. Our experiments show that, across multiple multimodal benchmarks (including intent disambiguation, commonsense reasoning, and safety), CoCoT consistently outperforms CoT and direct prompting (+8% on average). Our findings demonstrate that cognitively grounded reasoning stages enhance interpretability and social awareness in VLMs, paving the way for safer and more reliable multimodal systems.

education

MS/PhD in Artificial Intelligence, Seoul National University, 2021~
B.S. in Computer Science, Columbia University Fu Foundation School of Engineering and Applied Science, 2017-2020
B.A. in Mathematics, Bard College, 2014-2017