ABSTRACT Large language models (LLMs) are reshaping translation workflows, yet empirical evidence on student‐LLM interaction in authentic translation tasks remains limited. This study analyzes interaction logs from 33 Chinese EFL graduate students who completed a Classical Chinese‐to‐English translation on a custom‐built platform, yielding 133 prompts and corresponding model outputs. Through expert coding, we developed a task‐specific taxonomy of learners’ prompt strategies and traced their interaction patterns with the model. Students frequently used compound instructions and showed a predominant “generate‐then‐refine” pattern, while prompts explicitly targeting pre‐drafting analysis, terminology clarification, or language and cultural adaptation were relatively rare. Prompt quantity did not predict translation quality, whereas submissions involving human post‐editing were associated with higher scores, supporting a hybrid workflow in which model drafting and human revision complement each other. Post‐task questionnaires further showed perceived benefits, cautious trust, and learner agency, alongside frustrations with ineffective iterations and unstable outputs. Overall, the study offers an empirical account of how student translators work with LLMs and provides evidence for integrating prompt literacy into translator training.
Li et al. (Tue,) studied this question.