MathVerse: A New Benchmark for Evaluating the Visual Mathematical Abilities of Multimodal Large Language Models

Apple Proposes MAD-Bench Benchmark to Address Hallucination Issues in Multimodal Large Language Models

Apple's research introduces the MAD-Bench benchmark to address the vulnerability of MLLMs in handling misleading information. MAD-Bench consists of 850 image-prompt pairs, assessing the capability of MLLMs in maintaining consistency between text and images. The study indicates that GPT-4V performs better in scene understanding and visual confusions, providing strategic insights for AI model design. The MAD-Bench benchmark can enhance the robustness of AI models, paving the way for more reliable MLLMs in future research.