FURI | Spring 2025

Reducing LLM Energy Waste via Automated Prompt Optimization

Large language models (LLMs) consume vast amounts of power, which has a significant impact on the environment through carbon emissions to generate the requisite power and water consumption to cool down the data centers used. This project uses a man-in-the-middle LLM to optimize a user’s request for quality short responses. This solution, depending on the request type, can significantly reduce output length, which is directly proportional to power consumption, reducing both environmental impact and cost while still maintaining user satisfaction. This solution creates a real incentive for companies to reduce environmental impact because it reduces cost.