ViLT's fine-tuning follows the classic pretrain-then-fine-tune paradigm, performing end-to-end adaptation for different downstream tasks after pretraining. For classification tasks such as Visual Question Answering (VQA) and Natural Language Visual Reasoning (NLVR²), the model adds a lightweight multilayer perceptron classification head on top of the Transformer output, using specific input strategies (e.g., the "pair method" for NLVR²). For retrieval tasks such as image-text retrieval, it directly reuses the pretrained Image-Text Matching head, employing contrastive learning to pull positive pairs closer and push negative pairs apart. The entire fine-tuning process uses a small learning rate and updates all parameters (including the image projection layer), while innovatively introducing image augmentation to improve generalization. This achieves efficient transfer across a variety of tasks with a remarkably minimal adaptation approach.