This paper proposes TRANSAGENT, a multi-agent system based on a Large-Scale Language Model (LLM). TRANSAGENT improves the accuracy and efficiency of LLM-based code translation by correcting syntactic and semantic errors that arise during the translation of source code to other programming languages. Unlike existing rule-based methods, TRANSAGENT utilizes a learning-based approach utilizing parallel data and LLMs. It corrects errors through the synergy of four LLM-based agents: an initial code translator, a syntactic error corrector, a code aligner, and a semantic error corrector. Specifically, it minimizes the difficulty of error correction by preemptively identifying erroneous code blocks based on the execution alignment of the target and source programs, narrowing the scope of correction. We evaluated TRANSAGENT using a new benchmark and confirmed that it outperforms UniTrans, a state-of-the-art LLM-based code translation technology, in both translation efficiency and effectiveness. We also validated TRANSAGENT's generalization performance and the effectiveness of each agent through analysis of the contribution of various LLMs and agents.
