Two-stage pruning reduces redundancy in vision-language model inference

Original: When Attention Collapses: Stage-Aware Visual Token Pruning from Structure to Semantics

Source: arxiv.org ↗