본문 바로가기
  • Home

Exploring creative variation in LLM literary translation: A multi-dimensional analysis of temperature and prompting effects

  • The Journal of Translation Studies
  • Abbr : JTS
  • 2026, 27(1), pp.163~191
  • DOI : 10.15749/jts.2026.27.1.005
  • Publisher : The Korean Association for Translation Studies
  • Research Area : Humanities > Interpretation and Translation Studies
  • Received : February 15, 2026
  • Accepted : March 16, 2026
  • Published : March 31, 2026

Yim, Jin ORD ID 1

1이화여자대학교

Accredited

ABSTRACT

This study explores the potential for inducing creative variation in Large Language Model (LLM) literary translation by examining the interaction between prompting strategies and the temperature parameter. Unlike traditional machine translation, LLMs offer a stochastic framework that allows users to modulate stylistic nuances. Focusing on Yi Sang’s modernist masterpiece The Wings, this research investigates how combinations of creativity-oriented instructions (P1–P4) and temperature settings (T0.2–T1.2) affect linguistic outputs. To ensure methodological rigor, the study utilized an API-based environment to isolate session histories and fix sampling variances, thereby ensuring reproducibility. The generated translations were analyzed using Principal Component Analysis (PCA) and Linear Mixed-Effects Models (LMM). Results indicate that higher temperatures, when combined with explicit creativity prompts, significantly increase lexical diversity and structural complexity, successfully leading to outputs that move beyond conventional baselines. Notably, while direct instructions (P2) produced the highest degree of novelty, few-shot examples (P4) acted as an anchor that limited creative divergence. Qualitative analysis further confirms that these statistical shifts translate into sophisticated literary nuances without compromising semantic acceptability. These results provide a methodological foundation for maintaining academic validity in the emerging discourse on AI-generated creative translation

Citation status

* References for papers published after 2024 are currently being built.