Tutorial Speaker

Title: Towards Trustworthy Medical AI: Hallucinations in Multimodal LLMs

Abstract: Large language models (LLMs) and multimodal foundation models are increasingly used in healthcare applications that involve clinical text, medical images, and their combinations. While these models demonstrate strong capabilities in reasoning, summarization, and content generation, they are also prone to hallucinations, i.e., outputs that appear plausible but are factually incorrect, unsupported, or inconsistent with clinical reality. In medical settings, such errors can have serious implications for trust, reliability, and downstream decision-making. This tutorial provides a high-level, unified view of hallucinations in LLM-based healthcare systems, with a particular focus on multimodal scenarios involving both text and images. We discuss common hallucination patterns, why they arise in large-scale models, and how they manifest differently across interpretive and generative tasks. The tutorial also highlights challenges in evaluating hallucinations using standard performance metrics and motivates the need for more reliability-aware assessment and validation practices. By synthesizing insights across multiple modalities and application settings, this tutorial aims to equip researchers and practitioners with a clearer understanding of hallucination risks and emerging directions for building safer and more trustworthy medical AI systems.

The tutorial plans to cover (i) an overview of hallucinations in large language models and multimodal foundation models, (ii) how hallucinations manifest in healthcare-related text, image, and multimodal tasks, (iii) differences between hallucinations in interpretive tasks (e.g., understanding or summarizing inputs) and generative tasks (e.g., producing new content), and (iv) why standard evaluation metrics may fail to capture hallucination-related risks in medical AI.