Tag
This paper evaluates the zero-shot capability of multimodal large language models (MLLMs) for localized concept naming in images, proposing a reproducible evaluation protocol that achieves 62-88% object-level accuracy without training.