ViTs also have proven to be more effective for zero-shot generalization tasks due to their ability to capture global context and relationships in the input data

ViTs also have proven to be more effective for zero-shot generalization tasks due to their ability to capture global context and relationships in the input data, which CNNs struggle with. https://arxiv.org/abs/2304.02643 https://arxiv.org/abs/2304.02643