Adaptive 3D-RoPE: Physics-Aligned Positional Encoding for Wireless AI

A recent study introduces Adaptive 3D-RoPE, a physics-based rotary positional encoding tailored for wireless foundation models. Current CSI models adopt static or one-dimensional positional assumptions from NLP and vision, which misalign with the physics of wireless channels due to their lack of relative decay, the collapse of 3D spatio-temporal-frequency structures, and inflexibility across scenarios. The proposed framework features a learnable, axis-decoupled 3D frequency bank to separate multi-dimensional phase dependencies, along with a lightweight channel-conditioned controller that adjusts the encoding dynamically. This innovation lays a foundational structure for wireless models, enhancing extrapolation and generalization in CSI modeling, latent characterization, and task-specific predictions.