RadarChat: Advancing Radar Modulation Recognition Toward Knowledge-Driven Semantic Understanding with a Large Vision-Language Model

Radar modulation recognition (RMR) plays a critical role in enabling reliable operation in complex electromagnetic environments. However, current methods suffer from limitations in expert knowledge integration, data dependency, and low-SNR robustness. To address these challenges, we propose RadarChat, a framework integrating a large vision-language model (LVLM) with prompt engineering. This approach shifts RMR to a knowledge-driven paradigm, effectively eliminating the need for domain-specific pre-training and fine-tuning. Specifically, we design RadarPrompt via a ”three-step method” to systematically inject expert knowledge and implicit chain-of-thought (CoT) reasoning. Guided by RadarPrompt, RadarChat leverages the injected knowledge to achieve physically grounded semantic understanding and demonstrate strong generalization. Experimental results demonstrate that RadarChat achieves an average recognition accuracy of 92.12% (from –8 dB to 6 dB), maintains robust adaptability under low-SNR conditions, and exhibits strong generalization capabilities. Furthermore, the framework requires zero local trainable parameters and a mere 0.36 MB memory footprint, establishing it as a highly scalable and efficient solution for resource-constrained radar systems.