Phrasing pattern
Seven semantically equivalent variants spanning tone, structure, persona, formatting, and reasoning cues.
01 / Abstract
Prompts have a measurable energy signature.
Large language models are increasingly deployed on mobile devices, making energy efficiency a key deployment constraint, yet the energy impact of prompt design remains underexplored.
We study how two prompt properties—cognitive load and phrasing pattern—shape on-device inference energy. Across prompt variants, datasets, models, and phones, phase-level profiling separates prefill from decode energy and reveals two distinct mechanisms: cognitive load primarily changes energy cost per token, while phrasing pattern changes energy largely through token usage.
Prompt design reshapes the attainable energy–quality frontier differently across models, making energy-efficient prompting a model-aware problem.
02 / Prompt properties
Two ways a prompt can change.
We separate surface-level expression from reasoning demand so their computational effects can be measured rather than conflated.
Cognitive load
Controlled variants that change information density and reasoning demand while preserving task intent.
03 / Method
From prompt variation to power trace.
A controlled three-stage pipeline validates prompt properties before profiling lightweight LLM inference directly on smartphones.
Five models
Gemma-2-2B, LLaMA-3.2-1B, Qwen2.5-0.5B, Qwen2.5-1.5B, and SmolLM2-360M.
Two phones
Pixel 7 and Pixel 8 Pro, instrumented for synchronized latency and power collection.
Four sources
SVAMP, BoolQ, AI2-ARC, and CLEF 2025 ELOQUENT prompt variants.
Phase-level energy
Power is integrated over prefill and decode windows, then paired with token usage and quality.
04 / Findings
One prompt, two energy levers.
Reasoning demand changes the cost of tokens. Linguistic form changes how many tokens the model uses.
RQ1 · Cost per token
Cognitive load changes decode cost; surface phrasing largely does not.
Absolute per-token energy is dominated by model choice. Within a model, phrasing variants remain close to the base prompt, while cognitive-load variants create larger, phase-dependent shifts.
1.29×
decode energy per token under extraneous load for Qwen2.5-1.5B on BoolQ
RQ2 · Token footprint
Phrasing changes energy mainly by changing model generation behavior.
Chain-of-thought and aggressive prompts consistently expand token usage, while format, polite, and technical prompts stay closer to—or sometimes below—the base. The same phrasing produces different expansion across models.
CoT often raises decode burden
Format can shift cost toward prefill
RQ3 · Energy × quality
There is no universally optimal phrasing style.
Prompt variants can move a fixed model to lower energy and higher quality, but the winning style changes across architectures. Energy-efficient prompt selection therefore needs to be model aware.
Lower-energy, higher-quality variants appear for four of the five evaluated models.
05 / Open artifact
Trace the result back to the prompt.
The repository contains the prompt datasets, power collection utilities, processed measurements, and staged analysis notebooks used to reproduce the paper's figures and tables.
Browse repository06 / Citation
Build on this work.
If this study or artifact is useful in your research, please cite the paper.
@inproceedings{hu2026prompt,
title = {How Do Prompt Variations Affect Energy
Consumption in On-Device LLMs?},
author = {Hu, Wei and Tu, Xiaolong and Chen, Dawei and
Chen, Yitao and Han, Kyungtae and Wang, Haoxin},
booktitle = {Proceedings of the 2026 Conference on Empirical
Methods in Natural Language Processing},
year = {2026}
}