TAIA: Large Language Models are Out-of-Distribution Data Learners