Assessing the Impact of LLM Output Variability on Energy Consumption
Explore how output-length and response variability in LLMs influence inference energy use and sustainability trade-offs.
The same prompt can produce different LLM answers in length, structure, and reasoning depth. This thesis explores how that output variability affects energy use.
The Problem
Energy studies on LLMs often report averages across tasks and models. In practice, however, energy use can fluctuate strongly between runs because model outputs are variable: longer responses, different reasoning trajectories, and extra decoding steps can change runtime and power demand.
Without understanding this variability, comparisons between prompting strategies, models, or system settings can be misleading from a sustainability perspective.
What This Thesis Is About
This thesis investigates how LLM output variability impacts inference energy consumption. The goal is to link output-side characteristics (for example token count, structure, and response style) to energy patterns, and to identify which variability factors matter most in practice.
This is a strong fit for students who want to combine AI systems, empirical analysis, and sustainability into a concrete and timely research project.
Why This Topic Is Exciting
- You study a practical issue that affects benchmarking, deployment cost, and environmental impact of LLM systems.
- You gain hands-on experience with LLM experimentation, instrumentation, and robust statistical analysis.
- You produce actionable insights for designing greener prompting and inference strategies.
Objectives
- Characterize output variability across selected models, prompts, and decoding settings.
- Measure energy consumption per run and relate it to output properties such as response length and structure.
- Quantify sensitivity and uncertainty in reported energy metrics caused by output variability.
- Propose practical recommendations for more stable and energy-aware LLM evaluation and usage.
What You Will Do
| Phase | Work |
|---|---|
| Literature & design | Review LLM energy measurement and output variability studies; define workloads and metrics |
| Experiment setup | Build an experimental pipeline for repeated LLM runs with controlled decoding settings |
| Data collection | Execute runs, log outputs and runtime behavior, and collect energy/power measurements |
| Analysis | Model relationships between output variability and energy; identify dominant drivers and uncertainty ranges |
| Outputs | Deliver thesis, reproducible scripts, and practical guidelines for energy-aware LLM experimentation |
What You Gain
- A portfolio project at the intersection of Green AI and LLM systems.
- Practical skills in experimental design, instrumentation, and quantitative data analysis.
- Experience producing reproducible research artifacts useful for PhD or industry paths.
Who Should Apply
Required courses:
- Programming in Python (INF-22306)
- Machine Learning (FTE-35306)
- Big Data (INF-33806)
Required skills & mindset:
- Solid programming and data analysis skills
- Interest in AI systems and sustainability
- Curiosity for empirical experimentation and benchmarking
- Motivation to work independently while discussing progress regularly
Key References
- Verdecchia, R., Sallou, J., & Cruz, L. (2023). A systematic review of Green AI. WIREs Data Mining and Knowledge Discovery, 13(4), e1507. doi:10.1002/widm.1507
- Luccioni, A. S., Jernite, Y., & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? FAccT ‘24. doi:10.1145/3630106.3658542
- Kim, B., Choi, Y., Mei, H., et al. (2025). The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization. arXiv:2505.06371. arXiv:2505.06371
- Henderson, P., Hu, J., Romoff, J., Brunskill, E., Jurafsky, D., & Pineau, J. (2020). Towards the systematic reporting of the energy and carbon footprints of machine learning. Journal of Machine Learning Research, 21(248), 1-43.