Last-Vit论文阅读怎么发现的问题?For instance, DINO [43] demonstrates that label-supervised ViTs suffer from an attention deficit [26], while CLIPSelf [37] observes that text-supervised ViTs fail to produce dense image features that are accurately aligned with textual cues in ope