Prompt Wording Affects Energy Use in On-Device LLMs, Study Finds

A new study reveals that the choice of words in prompts significantly impacts the energy consumption of large language models (LLMs) running on mobile devices. Researchers from an undisclosed institution conducted empirical tests on a smartphone, measuring real power usage across different verbs and instruction structures. They found that imperative keywords and sentence complexity directly affect decoding length and total energy draw. The paper, titled "Keyword Matters: Unveiling the Energy Sensitivity of On-Device LLM Prompting," suggests prompt engineering as a lightweight method to improve energy efficiency without model compression or runtime acceleration. The findings highlight an underexplored factor in on-device AI deployment, where battery life is critical. The study was posted on arXiv under computer science and artificial intelligence categories on an unspecified date.