On the Advance of Making Language Models Better Reasoners