Systematic literature review of multi-objective hyper-heuristics: a human-in-the-loop large language model methodology

Multi-objective hyper-heuristics (MOHHs) have emerged as a powerful paradigm in computational intelligence, which enables the dynamic selection or generation of low-level heuristics to solve complex optimisation problems involving multiple objectives. Despite growing academic interest and a wide range of applications, there has been limited comprehensive analysis of the field’s evolution, methodologies, and challenges. This study presents a systematic literature review of 236 peer-reviewed publications on MOHHs published between 2005 and 2025, supported by a human-in-the-loop process that utilises large language models (LLMs) to assist screening and analysis. The review categorises application domains, characterises heuristic management strategies, maps learning mechanisms, and identifies emerging research themes. The findings reveal a marked shift from heuristic selection to generation-based and hybrid approaches, an increasing integration of reinforcement learning, and growing attention to adaptive, user-centric, and explainable optimisation. Methodological trends are also discussed in relation to benchmark use, performance evaluation, and theoretical grounding. The paper concludes with a thematic roadmap that outlines multiple future research directions, including LLM-guided MOHHs, many-objective optimisation, and preference-aware systems. This comprehensive review provides a foundation for advancing MOHH research and supports its application in challenging real-world problems.