Reproduce “Algorithm Evolution Using Large Language Model” on OpenResearch

About this paper

Fei Liu, Xialiang Tong, Mingxuan Yuan, Qingfu Zhang

AEL integrates large language models into an evolutionary computing framework to automatically generate and refine optimization algorithms, demonstrating improved performance over hand-crafted heuristics and better generalization across problem sizes compared to trained neural network solvers on the Traveling Salesman Problem.

Key insights

• The evolutionary framework effectively guides LLMs to iteratively refine and improve algorithm performance, demonstrating clear convergence and the discovery of complex, non-trivial heuristics.
• LLMs, when integrated into an evolutionary loop, can synthesize sophisticated algorithmic logic (e.g., multi-factor scoring functions and implicit hyperparameters) rather than merely retrieving or splicing known solutions.
• By performing evolution at the algorithm level, AEL-generated algorithms overcome the scalability and generalization limitations often observed when LLMs are used as low-level search operators or when neural solvers are trained on specific problem dimensions.

Summary by alphaXiv.