This paper proposes a multi-task learning approach for cross-problem generalization in vehicle routing problems (VRPs). The authors address the challenge of solving various VRP variants with different attributes using a single model, which can generalize to unseen combinations of attributes through attribute composition. The proposed model is trained using reinforcement learning and can solve multiple VRPs simultaneously. The model is evaluated on eleven VRP variants, benchmark datasets, and real-world logistics scenarios. The results show that the unified model outperforms existing methods, reducing the average gap to around 5% from over 20% and achieving significant performance improvements on benchmark datasets and real-world applications. The model's ability to generalize across different VRPs is attributed to its shared underlying attributes and attribute composition mechanism. The paper also discusses related work in neural combinatorial optimization (NCO), multi-task learning, and zero-shot learning, highlighting the novelty of the approach in addressing cross-problem generalization in VRPs. The model is implemented using PyTorch and trained on a single GPU, with experiments conducted on various VRP instances. The results demonstrate the effectiveness of the proposed method in solving a wide range of VRPs with different attributes in a zero-shot manner.This paper proposes a multi-task learning approach for cross-problem generalization in vehicle routing problems (VRPs). The authors address the challenge of solving various VRP variants with different attributes using a single model, which can generalize to unseen combinations of attributes through attribute composition. The proposed model is trained using reinforcement learning and can solve multiple VRPs simultaneously. The model is evaluated on eleven VRP variants, benchmark datasets, and real-world logistics scenarios. The results show that the unified model outperforms existing methods, reducing the average gap to around 5% from over 20% and achieving significant performance improvements on benchmark datasets and real-world applications. The model's ability to generalize across different VRPs is attributed to its shared underlying attributes and attribute composition mechanism. The paper also discusses related work in neural combinatorial optimization (NCO), multi-task learning, and zero-shot learning, highlighting the novelty of the approach in addressing cross-problem generalization in VRPs. The model is implemented using PyTorch and trained on a single GPU, with experiments conducted on various VRP instances. The results demonstrate the effectiveness of the proposed method in solving a wide range of VRPs with different attributes in a zero-shot manner.