Radio spectrum cognition, as the foundation for the efficient utilization of national strategic spectrum resources, is of paramount importance. It underpins the construction of a refined spectrum management system, directly supporting key applications such as dynamic spectrum sharing, intelligent spectrum control, and electromagnetic spectrum operations. With the increasingly complex and dynamic electromagnetic environment, traditional cognition methods struggle to meet the demands for high accuracy and strong generalization. Therefore, the introduction of generative artificial intelligence, with its powerful capabilities in data distribution fitting and generation, represents a revolutionary technological pathway. It holds significant potential for breaking through existing bottlenecks and is crucial for ensuring the secure and efficient operation of future wireless networks and the informatized society.

Significant progress has been made in generative AI-assisted radio spectrum cognition. From the initial adoption of generative adversarial networks (GANs) and variational autoencoders (VAEs) to the current dominance of diffusion models, generative paradigms have demonstrated superior performance in spectrum map reconstruction from sparse samples or transmitter parameters. Representative works like DeepREM, RME-GAN, WiFi-Diffusion, and RadioDiff have progressively enhanced estimation accuracy and robustness, even under extremely low sampling rates. Furthermore, the field is expanding into three-dimensional cognition (e.g., RadioDiff-3D), integrating multimodal channel features, and exploring novel architectures like Transformers (e.g., RadioFormer, CollaboRadio) and large-scale models. These advancements are supported by the development of increasingly sophisticated and large-scale open datasets (e.g., RadioMapSeer, SpectrumNet, UrbanRadio3D), which provide essential foundations for model training and benchmarking.

In conclusion, generative artificial intelligence has injected new vitality into radio spectrum cognition, significantly improving cognitive accuracy and offering novel solutions for complex electromagnetic environment modeling. However, challenges remain, including data scarcity, limited generalization in unseen scenarios, insufficient model interpretability, and security threats. Future research should focus on: 1) constructing large-scale, high-fidelity, and multi-scenario benchmark datasets; 2) developing physics-informed generative models to enhance generalization by embedding electromagnetic propagation principles; 3) improving model interpretability and traceability to build trustworthy cognitive systems; and 4) optimizing model efficiency and output stability for real-time applications. Through cross-modal knowledge fusion and the establishment of a credible evaluation framework, generative AI is poised to drive radio spectrum cognition toward higher precision, stronger generalization, and greater reliability, effectively supporting the next generation of wireless communications.