Understanding Tone-Dependent Inference Cost in Large Language Models
Summary
This paper investigates how the tone of prompts affects both the accuracy and the inference cost (output token consumption) of large language models, finding that output token length can vary by up to 44.3% across tones while accuracy changes are smaller, and identifies optimal tones for different models.
View Cached Full Text
Cached at: 07/28/26, 06:30 AM
# Understanding Tone-Dependent Inference Cost in Large Language Models Source: [https://arxiv.org/abs/2607.23915](https://arxiv.org/abs/2607.23915) [View PDF](https://arxiv.org/pdf/2607.23915) > Abstract:We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output\-token consumption\. Experiments were performed to understand the trade\-offs between accuracy and inference cost on a 570 Question MMLU dataset for LLM models prompted in seven different tones from sycophantic to threatening\. Our results show that the output\-token\-length variation substantially exceeded accuracy variation across all models\. Output\-token consumption varied by up to 44\.3% across tone conditions\. We also analyzed the tradeoff between the accuracy of the answers and the average output token length in the reasoning process\. For the ChatGPT models 4o and 5\-nano, the rude tone is quite dominant\. For the Gemini models 2\.5 Flash and 2\.5 Flash Lite, the rude and neutral tones are dominant on the Pareto\-optimal frontier\. We find that prompt tone influences not only answer quality but also the amount of billable inference resources consumed by modern LLMs\. ## Submission history From: Akhil Kumar \[[view email](https://arxiv.org/show-email/9e7be428/2607.23915)\] **\[v1\]**Mon, 27 Jul 2026 01:08:23 UTC \(1,357 KB\)
Similar Articles
Mind Your Tone: Does Tone Alter LLM Performance?
This paper investigates how tonal variations in prompts affect LLM accuracy on multiple-choice questions, finding systematic but model-dependent effects. The study uses multiple models and datasets to demonstrate that tone can significantly alter performance, cautioning against assuming tone-robust reliability.
Understanding the Energy Scaling of Large Language Model Inference Across Context Lengths and Attention Architectures
This paper presents a systematic empirical study of energy consumption in large language model inference, analyzing how attention architectures like Multi-Head Attention, Grouped Query Attention, and Sliding Window Attention affect energy scaling across context lengths and workloads.
Optimal Model Activation Policies for Inference Networks of Large Language Models
The paper introduces inference networks, a graph-based framework for optimizing the use of multiple LLMs in inference, with optimal activation policies that minimize cost while meeting performance targets.
Sentence-Level Contextual Entrainment in Large Language Models
This paper extends contextual entrainment from token-level to sentence-level, showing that even counterfactual sentences in prompts increase their probability during inference. The effect decreases with model size and is driven by 2-4% of attention heads, which can be ablated without performance loss.
Scaling of Capability and Efficiency at Inference Time in Large Reasoning Models
This paper evaluates the capability and efficiency of large language models using Chain-of-Thought reasoning, finding that capability gains diminish with increasing model size, while efficiency shows little improvement.