Prompt Politeness Affects LLM Accuracy
A recent study investigates how the politeness of prompts affects the accuracy of large language models (LLMs). The research found that impolite prompts yielded higher accuracy compared to polite ones, challenging previous assumptions about tone and performance. This highlights the need for further exploration of the social dimensions in human-AI interactions.
- ▪The study created a dataset of 250 unique prompts with varying politeness levels.
- ▪Impolite prompts achieved an accuracy of 84.8%, while very polite prompts had an accuracy of 80.8%.
- ▪These findings suggest that newer LLMs may respond differently to tonal variations than previously thought.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2510.04950 (cs) [Submitted on 6 Oct 2025] Title:Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (short paper) Authors:Om Dobariya, Akhil Kumar View a PDF of the paper titled Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (short paper), by Om Dobariya and Akhil Kumar View PDF Abstract:The wording of natural language prompts has been shown to influence the performance of large language models (LLMs), yet the role of politeness and tone remains underexplored. In this study, we investigate how varying levels of prompt politeness affect model accuracy on multiple-choice questions.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
