Researchers Identify Visual Dominance Flaw in AI Models and Propose PRISM Framework

Vision-language models (VLMs) struggle with personalized safety in sensitive settings where missing user context can lead to hazardous advice. A new research preprint introduces MPS-Bench, a benchmark featuring 5,181 scenarios from 584 real-world images across 12 high-risk domains. Evaluations of eight leading VLMs revealed that models answered directly between 86% and 99% of the time instead of requesting missing information, with none scoring above 2.6 out of 5 in personalized safety.

The researchers found that visual inputs overpower textual risk signals during early fusion layers, a phenomenon termed visual dominance. Because early visual data alters internal text representations, late-stage remediation fails to prevent unsafe outputs.

To address this issue, the team developed PRISM, a lightweight input monitor using bidirectional cross-modal modulation. PRISM predicts when an AI query requires deferral, achieving a 0.978 AUC score and improving safety-utility trade-offs across all tested models.