Paper Overview
- Field: NLP
- Authors: Hitesh Mehta, Arjit Saxena, Garima Chhikara, Rohit Kumar
- Published: 2026-04-17
- arXiv: 2604.16275
- Model performance is highly influenced by tone, conversation history, and language.
- Polite prompts improve average response quality by up to ~11%, while impolite tones degrade it — but these effects are neither consistent nor universal across languages and models.
- English responds best to polite or direct tones; Hindi to reverent and indirect tones; Spanish to assertive tones.
- Across models, Llama is the most tone-sensitive (11.5% range), while GPT is more robust to adversarial tone.
- Politeness is a quantifiable computational variable affecting LLM behavior, but its impact is language- and model-dependent rather than universal.
Summary
This paper explores how Large Language Models (LLMs) respond to user prompts with varying degrees of politeness and impoliteness. The Politeness Theory by Brown and Levinson and the Impoliteness Framework by Culpeper form the basis of experiments conducted across three languages (English, Hindi, Spanish), five models (Gemini-Pro, GPT-4o Mini, Claude 3.7 Sonnet, DeepSeek-Chat, and Llama 3), and three interaction histories between users (raw, polite, and impolite).
The sample consists of 22,500 prompt-response pairs, evaluated across five politeness levels using an eight-factor assessment framework: coherence, clarity, depth, responsiveness, context retention, toxicity, conciseness, and readability.
Key Findings
Resources
To support reproducibility, the authors release PLUM (Politeness Levels in Multilingual Utterances), a publicly available corpus of 1,500 human-verified prompts spanning three languages and five politeness categories, along with supplementary analysis of six falsifiable hypotheses derived from politeness theory and empirically evaluated against the dataset.
*Auto-collected on 2026-04-21*