SPIN: Sparsifying and Integrating Internal Neurons in Large Language Models for Text Classification F i n d i n g s - A C L