This paper proposes a novel Self-Improved Learning (SIL) method for scalable neural combinatorial optimization (NCO), enabling direct training of NCO models on large-scale problems without labeled data. The method introduces an efficient self-improved mechanism that iteratively generates better solutions as pseudo-labels to guide model training. It also designs a linear complexity attention mechanism to handle large-scale problems with low computational overhead. Comprehensive experiments on the Travelling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes in both uniform and real-world distributions demonstrate the superior scalability of the proposed method. The SIL method outperforms existing approaches in terms of solution quality and efficiency, achieving state-of-the-art performance on large-scale VRPs. The method's linear attention mechanism significantly reduces computational cost, making it more suitable for solving large-scale combinatorial optimization problems. The results show that the proposed SIL method can achieve outstanding performance even for very large-scale problem instances. The method is trained on small-scale instances and then generalized to larger scales, with the model continuously improving through iterative local reconstruction and training. The method is compared with various baselines, including classical solvers, insertion heuristics, constructive methods, two-stage methods, and decomposition-based methods. The results show that the proposed method outperforms most of these methods, particularly on large-scale instances. The method's performance is evaluated on both synthetic and real-world datasets, demonstrating its effectiveness and scalability. The paper also discusses the computational complexity of the proposed method and compares it with existing approaches. The results show that the proposed method has significantly lower memory usage compared to existing methods, making it more efficient for large-scale problems. The method is trained on different scales of instances, with the model continuously improving through iterative local reconstruction and training. The method is evaluated on both synthetic and real-world datasets, demonstrating its effectiveness and scalability. The paper concludes that the proposed SIL method is a promising approach for scalable NCO, with potential for further improvements in future work.This paper proposes a novel Self-Improved Learning (SIL) method for scalable neural combinatorial optimization (NCO), enabling direct training of NCO models on large-scale problems without labeled data. The method introduces an efficient self-improved mechanism that iteratively generates better solutions as pseudo-labels to guide model training. It also designs a linear complexity attention mechanism to handle large-scale problems with low computational overhead. Comprehensive experiments on the Travelling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes in both uniform and real-world distributions demonstrate the superior scalability of the proposed method. The SIL method outperforms existing approaches in terms of solution quality and efficiency, achieving state-of-the-art performance on large-scale VRPs. The method's linear attention mechanism significantly reduces computational cost, making it more suitable for solving large-scale combinatorial optimization problems. The results show that the proposed SIL method can achieve outstanding performance even for very large-scale problem instances. The method is trained on small-scale instances and then generalized to larger scales, with the model continuously improving through iterative local reconstruction and training. The method is compared with various baselines, including classical solvers, insertion heuristics, constructive methods, two-stage methods, and decomposition-based methods. The results show that the proposed method outperforms most of these methods, particularly on large-scale instances. The method's performance is evaluated on both synthetic and real-world datasets, demonstrating its effectiveness and scalability. The paper also discusses the computational complexity of the proposed method and compares it with existing approaches. The results show that the proposed method has significantly lower memory usage compared to existing methods, making it more efficient for large-scale problems. The method is trained on different scales of instances, with the model continuously improving through iterative local reconstruction and training. The method is evaluated on both synthetic and real-world datasets, demonstrating its effectiveness and scalability. The paper concludes that the proposed SIL method is a promising approach for scalable NCO, with potential for further improvements in future work.