Improving Adversarial Robustness in Vision-Language Models with Architecture and Prompt Design F i n d i n g s - E M N L P