Workflow matters

Abstract Large language models (LLMs) have shown significant potential in translation tasks but often struggle with literary texts. This study compares professional human translations with translations produced by two AI-driven systems that coordinate multiple LLM-based agents. The first system mimics professional human translation practice, with distinct drafting and revision phases. The second redesigns the process specifically for LLMs’ capabilities, breaking translation into granular steps with specialized AI agents handling strategic planning, stylistic refinement, and coherence checking. Expert evaluations revealed that both AI systems achieved accuracy comparable to professional human translators. The LLM-capability-driven system produced translations with superior stylistic qualities and poetic language, though it occasionally added extraneous content. Meanwhile, the practice-derived system delivered concise translations but sometimes lacked cohesive flow. Blind evaluations showed that the translations from both AI systems were frequently preferred over human translations, particularly in terms of fluency. This study demonstrates that rethinking translation workflows around LLM capabilities can yield exceptional results, sometimes surpassing human performance in certain aspects.