SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs | Read Paper on Bytez