Multi-Stage Code Translation using LLMs for Enhanced Program Transformation

This paper studies a new method for code translation with multis-sage way, using large language model, like Qwen2.5-Coder-32B-Instruct, to make better and strong program translation between different coding languages.Old methods have problems with keeping meaning and making good translation.Our new method combines direct translation, understanding and rewrite, making unit test and translation, and reverse translation steps to make sure high accuracy in computing, good readability, and same meaning.We show design and how we do this pipeline, focusing on PHP to Go translation, and listing its benefits compared to usual ways.The results show that using large language models in different steps can make more trustworthy and easier to maintain code moving between languages.