Does Representation Matter? Exploring Intermediate Layers in Large Language Models