A new approach to prompt engineering suggests that only the parts of a prompt resembling the desired output should be written natively in the target language, while instructions and machinery can remain in a language the model understands best. This method stems from the understanding that language models primarily imitate the style of the text they are continuing. Translated prompts often introduce "translationese," which includes dropped subject pronouns, awkward compound words, misplaced politeness, and unnatural sentence rhythms, leading to output that reads as translated. By separating output-like text from instructional text, prompts can be optimized for better model performance and more natural-sounding results. AI
IMPACT Improves prompt engineering by suggesting a nuanced approach to language selection for instructions versus output examples.
RANK_REASON The item discusses a novel technique for prompt engineering, offering advice on how to improve LLM output by considering the nature of prompts.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
