Research on LLM Usage Efficiency: How to Reduce Resource Consumption Through Prompt Design Optimization

This study was published by Thoericht on GitHub on May 20, 2026, focusing on the issue of LLM usage efficiency. Through the analysis of real conversation datasets and controlled experiments, it explores how prompt design and user interaction patterns affect resource consumption, and provides actionable optimization recommendations. The core goal is to improve the resource usage efficiency of LLMs, reduce costs and environmental burdens.

Background: LLMs have been integrated into daily work processes, but there are significant differences in the efficiency of user usage patterns, leading to unnecessary computational overhead, increased costs, and environmental burdens.

Core Questions:

1. How does prompt structure affect token consumption and response length?
2. Are there efficient topic or task types?
3. Can machine learning be used to model usage efficiency?

Data Collection: A dual strategy was adopted—real conversation datasets (similar to ShareGPT style) + synthetic prompt experiments (controlled comparison).

Analysis Framework: A four-stage process: Exploratory Data Analysis (statistics + novelty embedding) → Topic Modeling (Sentence Transformer + BERTopic) → Efficiency Modeling (regression model + SHAP analysis) → Controlled Experiments (quantify efficiency-quality trade-off).

Key Metrics: target_success (whether the first response requires no clarification), target_cost (minimum number of tokens for the first response).

Tool Stack: Python pandas/numpy, scikit-learn, matplotlib/seaborn, sentence-transformers, tiktoken.

Expected Outcomes:

1. Identify inefficient usage patterns;
2. Establish a prompt efficiency prediction framework;
3. Provide actionable prompt optimization guidelines.

Practical Significance: Help development teams and users reduce operational costs, minimize environmental impact, and turn resource efficiency into an engineering constraint.

Limitations: Using token count and interaction complexity as proxy indicators for resource consumption, without directly measuring energy usage, which may deviate from the actual carbon footprint.

Future Directions:

1. Integrate real energy consumption monitoring data;
2. Expand to more LLM providers and model architectures;
3. Develop real-time prompt optimization tools;
4. Explore cumulative efficiency optimization for multi-turn conversations.

In today's era of widespread LLM applications, resource efficiency has become an essential engineering constraint to consider. This study provides a systematic analysis framework, using data-driven methods to understand and optimize LLM usage efficiency, which has important reference value for reducing costs and environmental impact.