MULTIMODAL HALLUCINATIONS IN AI

Summary from the TLDR AI newsletter

This paper sheds light on why large vision-language models (LVLMs) sometimes incorrectly describe images, a phenomenon known as multimodal hallucination. Semantic shift bias, particularly around paragraph breaks, is a key factor.