Conical Visual Concentration for Efficient Large Vision-Language Models | Read Paper on Bytez