In this paper, we propose a novel framework, Lyria, which combines genetic algorithms with large-scale language models (LLMs) to overcome the limitations of large-scale language models (LLMs). By combining the excellent semantic understanding ability of LLMs with the powerful global search and optimization abilities of genetic algorithms, we aim to solve complex problems such as multi-objective optimization, precise constraint satisfaction, and massive solution spaces. Lyria consists of seven essential components, and its effectiveness is demonstrated through extensive experiments using four LLMs for three types of problems. In addition, we systematically analyze the factors affecting the performance through seven ablation experiments.
