Keyword Matters: Unveiling the Energy Sensitivity of On-Device LLM Prompting

A new empirical study from arXiv reveals that prompt wording significantly affects energy consumption for on-device large language models (LLMs), with imperative keywords and instruction structure influencing decoding length and total energy. Using real power measurements on a smartphone, the researchers found consistent energy differences across verbs and tasks, suggesting prompt engineering as a lightweight method to improve energy efficiency on battery-powered devices.

arXiv:2607.22568v1 Announce Type: new Abstract: Large Language Models LLMs are increasingly deployed on mobile and embedded devices to improve privacy and reduce network latency. Yet on-device inference faces a fundamental constraint: high energy consumption on battery-powered, resource-limited hardware. While model compression and runtime acceleration have been widely studied, the effect of \emph{prompt design} on energy efficiency remains underexplored. This paper presents an empirical study of the relationship between prompt wording and energy consumption for on-device LLMs. Using real power measurements collected on a smartphone, we quantify how linguistic features, particularly imperative keywords and instruction structure, affect decoding length and total energy. Our results show consistent energy differences across verbs and tasks, indicating that prompt engineering is a lightweight lever for improving energy efficiency.