A Framework for Integrating Large Language Models into Memetic Algorithms

Memetic algorithms achieve strong optimization performance by combining population-based global search with local refinement operators, yet their effectiveness critically depends on the design and management of memes. Local search strategies are typically handcrafted, problem-specific, and fixed prior to execution. This paper proposes a fourth-generation memetic framework in which Large Language Models (LLMs) are embedded directly into the optimization loop as adaptive generators of local search operators. At each triggering point, the LLM receives a structured state vector encoding the current search dynamics and generates a candidate meme in the form of an executable Python function. Generated operators are subject to a two-stage validation procedure combining semantic similarity assessment and implementation-level comparison, ensuring that only sufficiently novel and syntactically correct operators are admitted to a dynamically growing meme library. A cooldown-regulated triggering mechanism balances periodic and stagnation-based generation, while a probability-weighted selection policy prioritizes newly generated memes without discarding previously validated ones. The proposed framework is evaluated on the CEC 2017 benchmark suite for continuous black-box optimization and compared against classical memetic algorithms. Experimental results demonstrate that the LLM-driven approach consistently outperforms non-LLM memetic baselines, confirming the viability of generative language models as adaptive heuristic components within population-based optimization.