Automatically Generating Visual Hallucination Test Cases for Multimodal Large Language Models