Not Every Mention Is a Claim: Rethinking Object Hallucination Detection in Vision-Language Models

Published in arXiv preprint, 2026

Hsin-Ling Hsu*, Neeraja Kirtane*, Kuan-Hao Huang

arXiv preprint, 2026

Introduces NAME-Bench, 25.5K object mentions from captions by five VLMs, separating real hallucinations from non-assertive mentions (e.g., "a bench for people to sit on"). All 12 detectors evaluated wrongly flag 14–80% of these mentions.