New Pennsylvania University paper finds, being rude to some LLMs leads to considerably shorter responses and higher accu…
Rohan Paul Twitter · Rohan Paul (@rohanpaul_ai) · 2026-08-02
A University of Pennsylvania study on 570 MMLU questions across seven prompt tones finds that rudeness or politeness can shift LLM output token usage by up to 44.3% and accuracy by up to 2.99 percentage points, with the optimal tone varying by model.
Extraction
Topics: prompt-engineeringllm-inference-costtone-effectsllm-behavior
Claims
- Prompt tone shifts LLM output token usage by as much as 44.3% while moving accuracy by at most 2.99 percentage points across the same 570 MMLU questions.
- For ChatGPT-4o, rude tone produced the highest accuracy (89.04%) and shortest responses (223 tokens average), making it cost-optimal for that model.
- For Gemini 2.5 Flash Lite, neutral tone achieved the best accuracy (88.25%) with fewest tokens, while rude tone degraded accuracy and increased token use by 35.2%.
- Rude prompts caused Gemini 2.5 Flash Lite to skip option checks and self-correction steps, suggesting tone affects reasoning depth rather than just verbosity.
- Tone is a model-specific cost and reliability variable that production systems should standardize and benchmark rather than treat as a purely user-experience concern.
Key quotes
the way you speak to a model change your inference bill?
tone is not merely a UX choice; it is a model-specific cost and reliability setting that production systems should standardize and benchmark.
Accuracy moved by at most 2.99% points within a model, but output-token use shifted by as much as 44.3%.