By enabling decentralized model training, federated learning allows sensitive data to remain on local devices, significantly reducing privacy and security risks. In real-world applications, federated learning faces key challenges, such as high communication costs, device heterogeneity, and the varying computational capacities of devices, all of which negatively impact training efficiency and limit the generalization capabilities of the global model. In this paper, we propose a novel framework called Reinforcement Learning-Based Federated Pruning (RLBFP), including a Reinforcement-Based Generalization Pruning (RBGP) process and a Federated Sparse Aggregation (FedSA) process, which can dynamically adjust pruning rates and apply efficient sparse aggregation strategies. RLBFP can not only reduces communication costs but also improves model sparsity and performance. Comprehensive experiments conducted on benchmark datasets such as CIFAR-10, MNIST, and Fashion-MNIST demonstrate that RLBFP surpasses existing federated learning approaches in terms of sparsity, accuracy, and convergence speed. FedSA further enhances model sparsity by up to 18.9 \(\%\) without compromising performance.