VHELM: A Holistic Evaluation of Vision Language Models Tony Lee 1 Haoqin T u 2 Chi Heem Wong