Benchmarking Robustness of Adaptation Methods on Pre-trained Vision-Language Models (Supplementary) Shuo Chen 1,3 Jindong Gu2 Zhen Han 1 Y unpu Ma1,3

BART had higher robustness against text corruptions on the GQA dataset. This may be due to the different language encoders used in BART and T5.Among all adaptation The top row shows the robustness against image corruptions and the bottom row is results against text corruptions. We also choose different embedding dimensions for adapter-based methods. The relative robustness of adaptation methods based on CLIP-T5 is presented in Table 1. Figure 2