Grounded or Guessing? BICR Detects When Vision-Language Models Ignore the Image

arXiv·medium signal

BICR (Blind-Image Contrastive Ranking) is a model-agnostic confidence estimation method that detects when large vision-language models produce correct answers driven entirely by language priors, with the image contributing nothing. By comparing model behavior with and without the actual image, BICR identifies visually ungrounded predictions that existing methods miss. Practical tool for anyone deploying VLMs in production where visual grounding matters.