CLIP

Vision-Language Models Struggle to Bind Attributes to Objects When Inputs Are Shuffled.

SABER is a controlled benchmark for multimodal retrieval that evaluates attribute-object binding across image …