Federated Learning (FL) represents a distributed, privacy-preserving machine learning (ML) paradigm that enables decentralized model training across multiple clients. While traditional aggregation techniques, such as Federated Averaging (FedAVG), have demonstrated effectiveness, they often struggle in Not Independent and Identically Distributed (non-IID) scenarios, where data distributions vary significantly among clients. To address these limitations, this study introduces FedGP, a novel aggregation strategy based on Genetic Programming (GP). FedGP dynamically evolves aggregation functions, enabling adaptive and personalized model updates that better capture the heterogeneity inherent in distributed data. The proposed method is evaluated on the PathMNIST dataset, employing a comprehensive experimental design comprising 24 configurations, including 8 setups with FedAVG and 16 with FedGP. The comparative analysis highlights FedGP’s superior generalization capabilities and reduced biases, outperforming FedAVG in terms of accuracy. These results position FedGP as a robust and scalable solution for real-world FL applications, particularly in environments characterized by data heterogeneity.