This paper presents a review of recent advancements in metaheuristic algorithms, emphasizing their broad applicability across research domains and the performance improvements achieved through their derived variants. By mapping these algorithms to a proposed unified taxonomy, the review identifies the most generative and rapidly evolving category within the field. This paper also explores the emerging and fast-moving intersection between metaheuristics and Large Language Models (LLMs). This conceptual extension highlights a transformative convergence in which LLMs enable automated algorithm generation and optimization, while metaheuristic methods offer avenues to enhance the adaptability and efficiency of LLMs. Despite substantial progress and promising results, challenges remain regarding interpretability, reliability, computational demand, and ethical implementation. These findings underscore the need for continued, rigorous research into both metaheuristic methodologies and their evolving relationship with modern AI systems.