The Influence of Prompt Politeness on Response Quality in Large Language Models

Abstract

This research examines the hypothesis that prompt politeness influences the quality of responses generated by large language models (LLMs). Our theoretical framework analyzes how variations in politeness modify attention mechanisms and probability distributions in the transformer architecture. The proposed approach suggests that for semantically equivalent content, polite prompts activate attention patterns that process requests differently. To verify this hypothesis, we conducted controlled experiments with 150 pairs of prompts identical at the semantic level but differing in their politeness level. Results show that polite prompts primarily enhance linguistic sophistication of responses (14.1% improvement, p < 0.001), with more modest effects on reasoning coherence, overall quality, and factual accuracy. Attention analysis reveals that polite prompts significantly increase attention entropy (16.2%) and context utilization (10.2%). This study introduces insights into how the social dimensions of interactions with LLMs affect their processing mechanisms and response characteristics, with potential applications for optimizing human-machine interfaces and designing conversational AI systems.