Detecting Knowledge Boundary of Vision Large Language Models by Sampling-Based Inference