(Perhaps) Beyond Human Translation: Harnessing Multi-Agent Collaboration for Translating Ultra-Long Literary Texts

• https://arxiv.org/abs/2405.11804
• Minghao Wu, Yulin Yuan, Gholamreza Haffari, Longyue Wang

The authors created a virtual translation agency with a bunch of specialized LLMs, such as CEO, senior/junior editor, translator, proofreader, etc. They showed that this team could translate a long document successfully.

They introduce “addition-by-subtraction collaboration”. In this collaboration “algorithm”, one agent acts as “addition agent” and the other as “subtraction agent”. The addition agent generates a detailed response first, and then the subtraction agent review it and remove redundant information. This two answers are then appended to the conversation history. When they agree, the conversation is done.

There is also “trilateral collaboration”, with an action agent, a critique agent, and judgment agent. It is similar to the addition-by-subtraction, but the judge decides whether the quality is good and terminates the conversation.