A structured review of large language models in metaheuristic optimisation

Metaheuristics are widely used to address complex optimisation problems where traditional exact methods are computationally infeasible or insufficiently flexible. With the rapid advancement of artificial intelligence, large language models, such as ChatGPT, Claude, Gemini, and LLaMA, have emerged as powerful tools capable of enhancing, automating, and adapting various stages of the optimisation process. This systematic literature review investigates the evolving role of large language models in metaheuristic optimisation, with a focus on algorithm generation, parameter tuning, hybridisation, constraint handling, and multi-objective optimisation. Following PRISMA guidelines, 25 studies from nine major scientific databases were selected and analysed. Through thematic analysis, a novel role-based taxonomy was developed that categorises large language model contributions into four functional roles: Advisor, Refiner, Enhancer, and Innovator. The findings demonstrate that large language models support the automation of metaheuristic workflows, enable dynamic adaptation, and contribute to the creation of novel heuristic strategies. Despite these advantages, the review also identifies persistent limitations, including prompt sensitivity, computational overhead, and scalability challenges. These issues highlight the need for more robust evaluation frameworks and benchmarking practices. This review offers a comprehensive synthesis of the current landscape, highlights research gaps, and provides actionable insights for researchers and practitioners aiming to integrate large language models into advanced optimisation systems across domains such as engineering, logistics, and computational design.