Geometric Over-Alignment Causes Hallucinations in Vision-Language Models

A new study from arXiv (2605.08245) identifies geometric over-alignment as the root cause of hallucinations in decoder-based Vision-Language Models (VLMs). The researchers trace failures to an over-alignment of visual embeddings with the text manifold, injecting linguistic bias that overshadows visual evidence. This is the first quantitative characterization of the phenomenon, showing that bias concentrates in top principal components. Prior work either aggressively closes the modality gap or uses expensive black-box decoding, but none addresses the underlying geometric cause. The findings have implications for high-stakes applications like medical imaging and autonomous systems.