@misc{10498/37744,
year = {2025},
month = {10},
url = {http://hdl.handle.net/10498/37744},
abstract = {The carbon footprint of computing is increasingly shaped by software, yet existing programming tools and large language models (LLMs) remain largely energy-blind. We propose energy-guided code generation, a method that reranks LLM-generated programs based on direct energy measurements using CodeCarbon while ensuring functional correctness. Evaluating a benchmark of algorithmic and data-processing tasks, we show that energy-guided selection yields statistically significant energy reductions. It reduces consumption by an average of 44.69% compared to unguided Top-1 candidates and by an additional 1.86% compared to the fastest (Best-Time) implementations, all without runtime penalties or loss of accuracy. These results provide the first conclusive evidence that LLMs produce diverse implementations with substantial variation in energy use, and that energy-aware reranking can consistently surface verifiably greener solutions. By embedding energy as a first-class optimization signal in the act of code generation, this work establishes a foundation for green-by-design software generation systems, where sustainability is not an afterthought but a default property of programming tools.},
title = {Green by Design: Energy-Guided Reranking of LLM-Generated Programs},
author = {Xia, Yi and Aragón Jurado, José Miguel and Thawonmas, Ruck},
}