Exploring Activation Patterns of Parameters in Language Models