💡 This paper presents MME, the first comprehensive evaluation benchmark for Multimodal Large Language Models (MLLMs), covering 14 perception and cognition subtasks. It avoids data leakage via manually d