Tag
This paper introduces a method to predict middle-layer attention in multimodal LLMs to prune visual tokens efficiently, using question-contrastive teacher selection and cross-modal attention distillation.