This paper addresses the challenge of cross-problem generalization in vehicle routing problems (VRPs) using a multi-task learning approach. The authors propose a unified neural network model that can solve multiple VRPs simultaneously by treating them as different combinations of shared underlying attributes. This model leverages reinforcement learning to train without labeled solutions, enabling zero-shot generalization to unseen attribute combinations. Extensive experiments on eleven VRP variants, benchmark datasets, and real-world logistic scenarios demonstrate superior performance compared to existing methods, reducing the average gap from over 20% to around 5%. The proposed model shows strong generalization capabilities and outperforms single-task models, achieving significant performance boosts on both benchmark datasets and real-world applications. The source code is available at https://github.com/FeiLiu36/MTNCO.This paper addresses the challenge of cross-problem generalization in vehicle routing problems (VRPs) using a multi-task learning approach. The authors propose a unified neural network model that can solve multiple VRPs simultaneously by treating them as different combinations of shared underlying attributes. This model leverages reinforcement learning to train without labeled solutions, enabling zero-shot generalization to unseen attribute combinations. Extensive experiments on eleven VRP variants, benchmark datasets, and real-world logistic scenarios demonstrate superior performance compared to existing methods, reducing the average gap from over 20% to around 5%. The proposed model shows strong generalization capabilities and outperforms single-task models, achieving significant performance boosts on both benchmark datasets and real-world applications. The source code is available at https://github.com/FeiLiu36/MTNCO.