PAR: Prompt-Aware Token Reduction Method for Efficient Large Multimodal Models | Read Paper on Bytez