Abstract: To address the shortcomings of existing methods in traffic flow prediction regarding computational complexity, real-time performance, and the integration of local and global features, this paper proposes a traffic flow prediction model based on dual-scale adaptive token attention. The model incorporates a dual-scale adaptive token attention mechanism designed to extract complex spatio-temporal features while optimizing computational efficiency. Through dual-scale learnable pooling operations, the obtained tokens effectively capture long-term and short-term temporal features of the data. Furthermore, the adaptive token attention mechanism integrates global dependencies to enhance prediction accuracy and operational efficiency. Experimental validation on two public datasets demonstrates that the proposed method outperforms existing mainstream models in both prediction accuracy and computational efficiency. Particularly suitable for real-time traffic flow prediction scenarios, this approach provides an efficient and accurate solution for intelligent transportation systems, exhibiting significant theoretical and practical implications.