Survey Maps How LLMs Supercharge Evolutionary Optimization

This survey by Yisong Zhang and colleagues (Ran Cheng, Guoxing Yi, Kay Chen Tan) systematically reviews the integration of large language models (LLMs) into evolutionary computation for optimization tasks—dubbed evolutionary optimization. Unlike earlier surveys that examine isolated LLM roles, this work provides a unified, workflow-oriented framework connecting two primary stages: optimization modeling (defining the problem) and optimization solving (finding solutions). The authors organize recent literature under this lens, offering a clear taxonomy that bridges traditional evolutionary methods with LLM-enhanced approaches. They also include benchmark systematization, baseline comparisons, and practical guidance for practitioners, making the paper both a comprehensive reference and a roadmap for future work.

Within the solving stage, the survey identifies three distinct paradigms based on the LLM's role: (1) stand-alone optimizers where the LLM directly generates solutions, (2) low-level components embedded within existing evolutionary algorithms (e.g., as mutation or crossover operators), and (3) high-level managers that select or generate entire algorithms. The paper reviews representative methods for each paradigm, discusses their technical limitations, and clarifies relationships with traditional optimization. It also surveys interdisciplinary applications—from natural sciences and engineering to machine learning—and outlines future research directions toward dynamic, self-evolving, and agentic optimization ecosystems. An up-to-date collection of related literature is maintained on GitHub.