This paper introduces a novel Self-Improved Learning (SIL) method to enhance the scalability of neural combinatorial optimization (NCO) for solving large-scale combinatorial optimization problems. Existing NCO methods struggle with large-scale problems due to the lack of labeled data and high computational costs. SIL addresses these challenges by developing an efficient self-improved mechanism that enables direct model training on large-scale problem instances without any labeled data. The method uses an innovative local reconstruction approach to iteratively generate better solutions as pseudo-labels, which guide the model training. Additionally, a linear complexity attention mechanism is designed to handle large-scale combinatorial problems efficiently with low computational overhead. Extensive experiments on the Travelling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes demonstrate the superior scalability and performance of the proposed SIL method. The results show that SIL can achieve state-of-the-art performance on large-scale VRPs, outperforming both learning-based and classical solvers.This paper introduces a novel Self-Improved Learning (SIL) method to enhance the scalability of neural combinatorial optimization (NCO) for solving large-scale combinatorial optimization problems. Existing NCO methods struggle with large-scale problems due to the lack of labeled data and high computational costs. SIL addresses these challenges by developing an efficient self-improved mechanism that enables direct model training on large-scale problem instances without any labeled data. The method uses an innovative local reconstruction approach to iteratively generate better solutions as pseudo-labels, which guide the model training. Additionally, a linear complexity attention mechanism is designed to handle large-scale combinatorial problems efficiently with low computational overhead. Extensive experiments on the Travelling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes demonstrate the superior scalability and performance of the proposed SIL method. The results show that SIL can achieve state-of-the-art performance on large-scale VRPs, outperforming both learning-based and classical solvers.