MMCOMPOSITION: Revisiting the Compositionality of Pre-trained Vision-Language Models: https://arxiv.org/pdf/2410.09733
