Green LLM Techniques in Action: Industry Energy Efficiency Evaluation

Authors report experimental evaluation of four energy-reduction techniques applied to an industrial LLM chatbot at Schuberg Philis.

• Main action: The paper evaluates Small and Large Model Collaboration, Prompt Optimization, Quantization (including 2-bit), and Batching across eight variations on a Schuberg Philis chatbot, reporting that some techniques reduced energy use by up to 90% but often caused unacceptable accuracy degradation; NPCC (Nvidia’s Prompt Task and Complexity Classifier) with prompt complexity thresholds was the only approach that delivered strong energy reductions without substantially harming accuracy or response time.
• Background and details: Submitted to arXiv on 5 Jan 2026 (arXiv:2601.02512) and accepted for publication at ICSE-SEIP’26; includes full experimental results on energy consumption, accuracy, and response time, and is released under CC BY 4.0 (arXiv DOI: https://doi.org/10.48550/arXiv.2601.02512).