NaturalBench: Evaluating Vision-Language Models on Natural Adversarial Samples | Read Paper on Bytez