Abstract: To address the limitations of conventional path-loss modeling approaches for VHF communications in complex mountainous environments, including the insufficient accuracy of empirical models, the high computational cost of deterministic methods, and the limited cross-scenario adaptability of both approaches, this paper proposes a physics–data collaborative path-loss modeling framework, termed the Kolmogorov–Arnold Network with physics-guided residual learning (KAN-PG). Firstly, a set of geometry-based propagation descriptors independent of absolute geographic coordinates is constructed from a digital elevation model, including the three-dimensional propagation distance, equivalent height difference between the transmitter and receiver, and the minimum relative clearance of the first Fresnel zone. Subsequently, a physics-consistent propagation baseline, consisting of free-space path loss and dominant knife-edge diffraction loss, is established to characterize large-scale attenuation. On this basis, a Kolmogorov–Arnold Network is employed to model and compensate for the residual path-loss components induced by complex terrain effects. The proposed method is evaluated using a WinProp-generated propagation dataset over a 25 km×25 km mountainous region. In the source-scene test, the proposed model accurately captures the spatial variation of path loss. In the cross-scenario adaptation experiment, after calibration using only 10% of the target-scene samples, the coefficient of determination (R²) for the proposed model reaches a maximum value of 0.9070, outperforming both the Random Forest and Backpropagation Neural Network models. Further analysis reveals that the physics-based baseline preserves the consistency of large-scale propagation characteristics across different environments, while the KAN-based residual model effectively enhances adaptability to previously unseen scenarios. The results demonstrate that the proposed framework does not rely on absolute geographic coordinates or high-precision geographic feature parameters, while maintaining physical interpretability and rapid cross-scenario adaptation capability. Consequently, it provides an efficient propagation modeling approach for the rapid deployment of emergency communication networks in mountainous regions.