Exploiting Primacy Effect To Improve Large Language Models