ROME: Evaluating Pre-trained Vision-Language Models on Reasoning beyond Visual Common Sense F i n d i n g s - E M N L P