RNNs trained on compositional tasks can exhibit functional and anatomical modularity when applying brain-inspired modular training, leading to improved performance and potential applications in neuromorphic computing and enhancing interpretability of neural network architectures.