Scaling Vision-Language Models with Sparse Mixture of Experts F i n d i n g s - E M N L P