To measure this, Vector developed the Multimodal Massive Multitask Understanding (MMMU) benchmark, which evaluates a model’s ability to reason about images and text across both multiple-choice ...