Adaptive 3D-RoPE: Physics-Aligned Rotary Positional Encoding for Wireless Foundation Models

2026-05-01

AI Summary

Existing wireless channel modeling approaches rely on static or one-dimensional positional encodings, which fail to capture the intrinsic three-dimensional spatio-temporal-frequency structure of wireless channels, thereby limiting extrapolation and generalization capabilities. This work proposes a physics-aligned adaptive 3D rotational positional encoding (Adaptive 3D-RoPE) that explicitly models multidimensional phase dependencies through a learnable, axis-decoupled 3D frequency bank. A lightweight channel-conditioning controller is introduced to dynamically adjust positional priors according to heterogeneous channel environments. Notably, this method pioneers sample-adaptive positional encoding in wireless foundation models, shifting from static inductive bias to coherence-aware dynamic bias. Experiments demonstrate a 10.7 dB NMSE improvement under 8× antenna-scale extrapolation across 100 datasets, along with gains of 1.07 dB and 0.90 dB in unseen mobility scenarios and zero-shot tasks from sub-6 GHz to millimeter-wave bands, respectively.

Abstract

Positional encoding plays a pivotal role in determin?ing the extrapolation and generalization performance of wireless foundation models for channel state information (CSI) modeling, latent characterization, and task-specific prediction. However, existing CSI models inherit static or one-dimensional positional priors from natural language and vision architectures, which fundamentally misalign with the intrinsic physics of wireless channels by lacking explicit relative decay, collapsing the 3D spatio-temporal-frequency structure, and remaining scenario?rigid. This paper proposes Adaptive 3D-RoPE, a physics-aligned rotary positional encoding that establishes the structural corner?stone for wireless foundation models. The framework integrates a learnable, axis-decoupled 3D frequency bank to explicitly disentangle multi-dimensional phase dependencies, coupled with a lightweight channel-conditioned controller that dynamically modulates the prior via compact global CSI descriptors. This sample-adaptive mechanism transforms positional encoding from a static transformer component into a dynamic, coherence-aware inductive bias to resolve heterogeneous channel physics. Extensive experiments across 100 datasets demonstrate the superiority of the proposed scheme in both scale extrapolation and zero-shot generalization. Compared to the state-of-the-art, our method achieves up to a 10.7 dB reduction in normalized mean square error (NMSE) under 8 times antenna scale extrapolation. Given the same CSI input scales, our method can also improve zero-shot NMSE by 1.07 dB across unseen mobility scenarios and 0.90 dB in low-frequency-to-millimeter-wave tasks.

Problem

• positional encoding
• wireless foundation models
• channel state information
• 3D spatio-temporal-frequency structure
• physics alignment

Innovation

• Adaptive 3D-RoPE
• physics-aligned positional encoding
• wireless foundation models
• channel state information
• dynamic inductive bias