Exploring and Evaluating Multimodal Knowledge Reasoning Consistency of Multimodal Large Language Models