LLM-Driven Metaheuristic Innovation: An Adaptive Hybrid COA–LSHADE–MPA Algorithm

The use of large language models (LLMs) as generators of metaheuristic algorithms introduces a new paradigm for automated algorithm design. This study proposes an LLM-generated algorithm called hybrid CLM, inspired by the crayfish optimization algorithm (COA), Success-History Based Adaptive Differential Evolution with Linear Population Size Reduction (LSHADE), and marine predator algorithm (MPA), and designed using GPT-4o via the evolution of heuristics prompting technique. The hybrid CLM switches between COA, LSHADE, and MPA features to guide population convergence and choice of optimization values. An ablation study of hybrid CLM comprising configurations was performed, followed by evaluation against COA, LSHADE, MPA, whale optimization algorithm (WOA), and sailfish optimization algorithm (SFO) across 64 CEC2017, CEC2022, and mathematical benchmark functions. Results reveal that the hybrid CLM algorithm delivers speed and accuracy, outperforming or matching rivals in 24 out of 29 CEC2017 benchmarks, 7 out of 12 CEC2022, and 12 out of 23 mathematical functions. In CEC2017, hybrid CLM’s execution time marked an overall 38% speed-up, six times faster than MPA, with insignificant fitness differences. Similar performance is attained on CEC2022 and mathematical functions. Overall, results demonstrate that this LLM-generated hybrid CLM can compete with human-designed optimizers while cutting execution time by up to 40%.