How Do Prompt Variations Affect Energy Consumption in On-Device LLMs?

As Large Language Models (LLMs) move from cloud servers to mobile devices, energy efficiency has become a critical constraint due to limited battery life. While much research has focused on model-centric optimizations like quantization and pruning, this paper investigates whether prompt design itself can serve as a model-agnostic lever for improving energy efficiency. The authors conduct a comprehensive empirical study to understand how two specific prompt properties—cognitive load and phrasing pattern—influence the energy behavior of on-device LLM inference.

The researchers developed a systematic pipeline to isolate the effects of prompt variations. They defined two primary properties:

• Phrasing Pattern: Surface-level linguistic and structural variations (e.g., tone, persona, formatting) that maintain semantic equivalence.
• Cognitive Load: Variations in reasoning demand, categorized into intrinsic, extraneous, and germane load.

Using a set of base prompts from standard benchmarks (SVAMP, BoolQ, AI2-ARC), they generated controlled variants. These were validated for semantic consistency and property purity before being profiled on mobile devices (Pixel 7 and Pixel 8 Pro) using the MLC-LLM framework. The study utilized phase-level profiling to separate energy consumption into prefill (input processing) and decode (generation) phases.

The study reveals that prompt properties influence energy consumption through distinct mechanisms. Cognitive load primarily alters the per-token energy cost, particularly during the decoding phase, as it changes the computational intensity of the reasoning process. In contrast, phrasing patterns influence energy consumption largely by changing the total number of tokens processed or generated, rather than the cost per individual token.

Crucially, the authors demonstrate that prompt design can reshape the energy-quality frontier. Certain phrasing patterns allow for lower energy consumption while maintaining or even improving response quality. However, these effects are highly model-dependent, meaning there is no single optimal prompt style that works across all architectures. This highlights the necessity for model-aware prompt engineering in energy-constrained environments.