Towards Robust And Scalable Large Language Models