Cross-Model Comparative Loss for Enhancing Neuronal Utility in Language Understanding