Honest Health OOC Dataset

The rapid proliferation of health-related misinformation, particularly in multimodal formats combining text and images, poses significant risks to public health and trust in medical systems. Existing datasets and evaluation frameworks inadequately address the unique challenges of assessing honesty in multimodal large language models (MLLMs) within the health domain. In this paper, we introduce the Honest OOC Dataset, a specialized dataset designed to evaluate the honesty of MLLMs in out-of-context (OOC) scenarios. The dataset includes 8,016 real health-related image-caption pairs, with both similarity-based and LLM-generated falsified captions that closely mimic real-world misinformation patterns. To complement the dataset, we propose a comprehensive benchmark that assesses model honesty across three dimensions: (1) Truthful Representation, evaluating faithfulness to input information; (2) Honest Uncertainty, examining the model's ability to express knowledge limitations; and (3) Evidence Honesty, assessing the accuracy of judgments based on external evidence. We evaluate five different MLLMs using this benchmark, revealing the varying levels of honesty exhibited by the models from multiple perspectives. Our findings highlight the importance of accurately interpreting external evidence and adhering to reliable information, which are crucial for effective OOC detection in the health domain.