SAVeS: Steering Safety Judgments in Vision-Language Models via Semantic Cues

Carlos Hinojosa, Clemens Grange, Bernard Ghanem
PDF

Abstract

We study how semantic cues can be used to steer safety judgments in vision-language models, proposing SAVeS as a method to improve the reliability of safety-relevant decisions made by multimodal models.

Type
Preprint
Publication
arXiv preprint arXiv:2603.19092
AI Safety Vision-Language Models Alignment
Avatar
Carlos Hinojosa
Research Scientist

I’m a computer scientist working in computer vision, machine learning, and AI safety. My research focuses on developing accurate, reliable, and efficient vision and AI systems.