Prompt Wording Impacts On-Device LLM Energy Consumption

The deployment of Large Language Models (LLMs) on mobile and embedded devices is growing due to benefits like enhanced privacy and reduced network latency. However, a major hurdle for on-device inference is the high energy consumption on battery-powered, resource-limited hardware. While efforts have focused on model compression and runtime acceleration, the impact of prompt design on energy efficiency has largely been overlooked. This paper presents an empirical investigation into how prompt wording influences the energy consumption of on-device LLMs. Using actual power measurements collected from a smartphone, the study quantifies the relationship between linguistic features, particularly imperative keywords and instruction structure, and their effect on decoding length and overall energy usage. The findings demonstrate consistent energy differences across various verbs and tasks, indicating that prompt engineering is not just about performance but also a surprisingly effective and lightweight method for improving energy efficiency. This suggests that careful crafting of prompts can lead to substantial power savings for LLMs operating in constrained environments.