Prompt Politeness Affects LLM Accuracy
a day ago
- #politeness effect
- #model accuracy
- #LLM prompting
- A study investigated how prompt politeness affects LLM accuracy on multiple-choice questions.
- Using ChatGPT 4o, 50 base questions in math, science, and history were rewritten into five politeness variants.
- Contrary to expectations, impolite prompts outperformed polite ones, with Very Rude prompts achieving 84.8% accuracy.
- This contrasts with earlier studies linking rudeness to poorer outcomes, suggesting newer LLMs respond differently to tone.
- The findings highlight the importance of studying pragmatic aspects of prompting and social dimensions in human-AI interaction.