Source-linked AI summary
How Do Prompt Variations Affect Energy Consumption in On-Device LLMs?
Wei Hu, Xiaolong Tu, Dawei Chen, Yitao Chen, Kyungtae Han, Haoxin Wang
TL;DR
On-device LLM deployment makes energy efficiency important, yet the energy effects of prompt design are underexplored. The paper conducts a broad, phase-level empirical study of cognitive load and phrasing pattern across prompt variants, datasets, models, and devices, finding distinct and model-dependent energy-quality effects.
Problem
Prompt-design effects on on-device LLM energy consumption remain underexplored despite energy efficiency being a strict constraint under finite device batteries.
Method
The study evaluates cognitive load and phrasing pattern across prompt variants, datasets, models, and devices using phase-level profiling of prefill and decode energy.
Results
Cognitive load mainly affects per-token energy, while phrasing pattern primarily affects token usage; energy-quality trade-offs vary across models.
Takeaways & Limitations
Prompt design can shift energy burden between prefill and decode and requires model-aware consideration of energy-quality trade-offs.
Takeaways & Limitations
The study primarily covers cognitive load and phrasing pattern, while other prompt properties remain unexplored.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly deployed on mobile devices, making energy efficiency a key deployment constraint, yet the energy impact of prompt design remains underexplored. This paper aims to understand how two prompt properties, cognitive load and phrasing pattern, shape the energy behavior of on-device LLM inference. We conduct a broad empirical study covering prompt properties, datasets, models, and devices, with phase-level profiling that separates prefill and decode energy. We find that cognitive load primarily affects the energy cost per token, while phrasing pattern affects energy largely through token usage. Our energy-quality analysis further shows that prompt design reshapes the attainable frontier differently across models, highlighting the need for model-aware prompt design in energy-efficient on-device LLM inference. Code, datasets, and scripts are available at https://amai-gsu.github.io/PromptProperty/.
1 Introduction
On-device LLM inference makes energy efficiency a user-visible deployment constraint, but prompt-design effects remain difficult to isolate and generalize. This study addresses that gap through controlled, broad empirical analysis of prompt properties across datasets, models, and devices.
- On-device LLMs operate under finite battery constraints, making energy efficiency a key deployment bottleneck.
- Prompt-energy analysis is challenging because linguistic variations have no direct mapping to the computation and power they trigger.
- The study designs prompt variants to isolate specific properties while preserving comparable task intent, enabling attribution of observed energy differences.
- The evaluation spans prompt variants, datasets, models, and devices to identify patterns beyond isolated examples.
- Cognitive load mainly affects per-token energy variation, whereas phrasing patterns primarily affect token usage.
- Energy-quality trade-offs are model-dependent, supporting model-aware prompt design for energy-efficient on-device inference.
2 Prompt Properties and Datasets
The paper studies phrasing pattern and cognitive load as complementary prompt properties, covering expression form and reasoning demand. It combines an existing phrasing-variation dataset with a newly constructed cognitive-load dataset designed to preserve semantic intent.
- 2 Prompt Properties and Datasets: The section introduces prompt properties and datasets for analyzing how prompt variations affect on-device energy behavior.
- 2.1 Prompt Properties: The study focuses on Phrasing Pattern and Cognitive Load as representative prompt properties.
- 2.1 Prompt Properties: Phrasing pattern captures linguistic and structural variation under semantic equivalence, while cognitive load captures reasoning structure and cognitive demand.
- 2.2 Prompt Datasets: Phrasing-pattern variants come from an existing robustness dataset containing controlled stylistic variations of semantically equivalent prompts.
- 2.2 Prompt Datasets: Figure 1 organizes phrasing pattern into seven surface-level sub-properties and cognitive load into three sub-properties varying information density and reasoning demand.
- 2.2 Prompt Datasets: The cognitive-load dataset uses intrinsic, extraneous, and germane load variants generated from base prompts sampled from SVAMP, BoolQ, and AI2-ARC.
3 Empirical Study Pipeline
The empirical pipeline generates and validates controlled prompt variants before profiling their on-device inference behavior. It records phase-level energy, latency, token, and response-quality measurements across lightweight local models and mobile devices.
- 3 Empirical Study Pipeline: The workflow consists of prompt generation, prompt validation, and on-device energy profiling.
- 3.1 Prompt Generation: Cognitive-load variants target intrinsic, extraneous, and germane load while preserving task intent and separating sub-property effects.
- 3.2 Prompt Variation Validation: Validation checks property purity and semantic similarity to ensure variants reflect their target property without semantic drift.
- 3.2 Prompt Variation Validation: Figure 2 traces validated prompts from generation and filtering to deployment on lightweight LLMs for power, latency, and response-quality measurement.
- 3.3 On-Device LLM Inference Profiling: The profiling framework instruments prefill and decode timestamps and samples device current and voltage through Android Debug Bridge.
- 3.3 On-Device LLM Inference Profiling: Phase-level energy is estimated by integrating sampled power over each phase window, with run metadata including device, model, prompt, token counts, and timestamps.
- 3.4 Response Quality: Response quality uses Exact Match accuracy for ground-truth cognitive-load tasks and six reference-free dimensions for open-ended phrasing-pattern tasks.
4 Results and Analysis
The study separates per-token energy, token usage, and energy-quality effects across prompt properties, models, datasets, and devices. It finds that cognitive load changes per-token inference cost, while phrasing patterns mainly alter energy through token usage and produce model-dependent energy-quality trade-offs.
- RQ1: Per-token Energy Analysis: Prefill energy per token exceeds decode by 3–6× on LLaMA-3.2-1B and both Qwen models, while Gemma-2-2B shows the opposite pattern.On-device runtimes exhibit this inversion relative to server-scale inference because of limited compute, memory bandwidth, and absent parallelized prefill support.
- RQ1: Per-token Energy Analysis: Absolute per-token energy is primarily model-dependent, with larger differences between model curves than between prompt sub-properties within a model.This indicates that model architecture and scale dominate per-token execution cost across phrasing pattern and cognitive load conditions.
- RQ1: Per-token Energy Analysis: Cognitive load changes per-token inference cost more than surface phrasing, whereas phrasing-pattern variants remain close to the base prompt in normalized per-token energy.Energy analysis therefore distinguishes reasoning demand from linguistic form rather than relying only on aggregate energy or token counts.
- RQ2: Token Footprint Analysis: CoT and aggressive prompts increase token usage relative to Base, while format, polite, and technical prompts stay closer to Base and sometimes reduce tokens.These token effects are model-dependent and can alter inference computation even when per-token energy shifts are small.
- RQ2: Token Footprint Analysis: CoT and aggressive prompts often increase decode burden, while format prompts can increase prefill burden and reduce or weakly increase decode burden.Token-related energy burden is distributed unevenly across phases, reflecting different effects on input processing and response generation.
- RQ3: Energy-quality Trade-off Analysis: Phrasing patterns shift the energy-quality frontier differently across models, with some sub-properties achieving lower energy and higher quality than Base under fixed weights and decoding settings.For LLaMA-3.2-1B, Qwen-2.5-1.5B, and SmolLM2-360M, such points lie on the Pareto frontier; Gemma-2-2B favors format or conversational prompts along both axes.
5 Related Work
Prior efficient-LLM research emphasizes model-centric optimization, while prompt studies often do not isolate individual properties. This work instead links controlled linguistic variation to measured system-level energy behavior.
- Model-centric techniques such as quantization, pruning, compression, and distillation dominate prior on-device LLM efficiency research.These methods reduce model size or computational cost through parameter or architectural modification.
- Existing prompt resources often vary semantics, length, and reasoning simultaneously, limiting isolation of individual prompt-property effects.
- The paper measures on-device energy directly, connecting linguistic prompt analysis with system-level energy profiling.The analysis characterizes measured system-level behavior rather than isolating low-level execution mechanisms.
6 Conclusion
The paper shows that prompt variants substantially affect mobile LLM inference efficiency, with cognitive load, phrasing, and phase allocation shaping energy and response quality. These effects vary across models and metrics, making prompt design a practical energy-optimization lever.
- Prompt-level variants can substantially affect mobile inference efficiency across models, devices, and prompt variants.
- Cognitively demanding prompts tend to increase decode energy cost per token, while phrasing patterns can change token usage, phase-level energy, and response quality.
- Different prompts can shift energy between prefill and decode phases, while energy-quality trade-offs vary across models and metrics.
- Prompt design is a practical optimization lever for energy-aware deployment, especially in mobile and resource-constrained settings.
Limitations
The study is limited by its narrow prompt-property coverage, lack of independent prompt-length control, evaluation methodology, hardware coverage, and profiling granularity.
- The study focuses primarily on cognitive load and phrasing pattern, leaving other semantic and structural prompt properties unexplored.
- Prompt length is not independently controlled because intrinsic and extraneous load manipulations require added guidance or redundant information.
- Gemini-2.5-Pro is used for prompt-property scoring and response-quality assessment, while automatic quality proxies may miss nuanced human-perceived utility.
- Energy measurements cover a limited set of mobile SoC architectures and focus on end-to-end and phase-level behavior rather than lower-level mechanisms.
D Experimental Setup Details
The experimental setup evaluates prompt-energy behavior across controlled combinations of prompt properties, datasets, models, devices, and repeated runs, alongside multiple response-quality metrics.
- The study covers 2 prompt properties, 10 sub-properties, 4 datasets, 404 prompts, 5 LLMs, 2 mobile devices, and 3 repeated runs per configuration.These configurations produce 7,620 total inference runs; cognitive-load prompts are profiled on Pixel 8 Pro only.
- Open-ended phrasing tasks use six response-quality metrics to evaluate linguistic and functional competence.
- Response-quality evaluation includes rubrics for relevance and correctness, coherence and completeness, and instruction adherence and internal consistency.
F Detailed Full Experimental Measurements
The appendix reports detailed measurements for phrasing pattern and cognitive load across five LLMs and multiple datasets, while clarifying the per-token energy weighting convention.
- Measurement conventions: Per-token energy in the text and Figure 3 uses equal weighting across questions, unlike ratios computed from aggregate means.Both quantities derive from the same measurements but differ in how questions are weighted.
- Phrasing pattern: Phrasing-pattern summaries cover five LLMs on Pixel 8 Pro and Pixel 7, with energy and quality columns encoded by shading.Darker green indicates lower energy, while darker blue indicates higher quality.
- Cognitive load: Cognitive-load summaries cover AI2-ARC, BoolQ, and SVAMP across five LLMs on Pixel 8 Pro.Energy columns use darker green for lower energy, and accuracy uses darker blue for higher accuracy.
G Additional Figures for RQ1
Additional RQ1 figures extend per-token energy analysis across datasets, devices, and prompt-property settings, separating prefill and decode costs.
- Figure conventions: Each figure presents absolute per-token energy in the top row and values normalized by the corresponding Base prompt in the bottom row.The normalization is applied separately to each sub-property.
- Cognitive load: The appendix adds per-token energy results for cognitive load on AI2-ARC and SVAMP using Pixel 8 Pro.The figures compare prefill and decode per-token energy across intrinsic, extraneous, and germane load variants.
- Phrasing pattern: Additional phrasing-pattern results on Pixel 7 report prefill and decode per-token energy across phrasing-pattern sub-properties.These results extend the corresponding main-text analysis.
H Additional Figures for RQ2
Additional RQ2 figures examine token usage and fixed-baseline prefill and decode burden across datasets and prompt-property settings.
- Scope: The appendix extends the main token-usage analysis to additional datasets and prompt-property settings.Its top row reports total token ratio relative to Base, while its bottom row reports fixed-baseline burden.
- Figure conventions: Bottom-row burden values are normalized by Base prompt and completion token counts after subtracting the corresponding Base burden.The appendix separately reports fixed-baseline prefill and decode burden relative to Base.
- Cognitive load: The appendix reports total token ratios relative to Base for cognitive-load variants on AI2-ARC, BoolQ, and SVAMP.The corresponding figures cover AI2-ARC on Pixel 8 Pro, BoolQ on Pixel 8 Pro, and SVAMP.
- Phrasing pattern: Additional phrasing-pattern results on Pixel 7 report total token ratios relative to Base across phrasing-pattern sub-properties.The figure also includes fixed-baseline prefill and decode burden.
I Additional Figures for RQ3
Additional RQ3 figures extend response-quality and energy-quality analyses to cognitive-load accuracy and further trade-off settings, while documenting dataset and software-use boundaries.
- Scope: The appendix complements the main-text RQ3 analysis with cognitive-load accuracy results and additional energy-quality trade-off settings.These additions are described as supporting the Section 4.3 analysis.
- Accuracy: Cognitive-load accuracy heatmaps report accuracy across sub-properties, datasets, and models under objective ground-truth evaluation.They cover intrinsic, extraneous, and germane load effects on task correctness.
- Cognitive-load trade-offs: Energy-accuracy figures for AI2-ARC, BoolQ, and SVAMP plot average total energy consumption against accuracy for cognitive-load sub-properties.Each point uses fixed model weights and decoding settings.
- Phrasing trade-offs: The phrasing energy-quality figure on Pixel 7 plots average total energy against averaged response quality across six reference-free evaluation dimensions.Each point represents a phrasing-pattern sub-property under fixed model weights and decoding settings.
- Scope boundaries: The study’s datasets, models, software, and hardware are used under their respective licenses, applicable terms, and data-use agreements.The stated boundaries include SVAMP, BoolQ, AI2-ARC, CLEF 2025, DeepEval, MLC-LLM, the listed open-source models, Gemini-2.5-Pro, and Pixel devices.