MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation

Traditional wireless foundation models typically pretrain on channel tensors using masked reconstruction. This approach often ignores the underlying physical characteristics of wireless propagation, such as reflections, scattering, and diffraction. The authors ask whether pretraining on the multipath propagation process itself—rather than the aggregate channel tensor—provides a more fundamental and reusable representation for wireless tasks.

MultiPathFormer treats wireless propagation as an ordered sequence of continuous-valued path tokens. Each transmitter-receiver link is represented as a sequence of paths, sorted by power. The model uses an autoregressive transformer architecture to perform next-path prediction, analogous to how large language models predict the next token in a sequence. To handle the complexities of physical environments, the authors introduce two key innovations:

1. Environmental RAG (Retrieval-Augmented Generation): A mechanism that retrieves relevant scene geometry to provide context for specific transmitter-receiver links.
2. First-Path Codebook: A KMeans-based prior that stabilizes the prediction of the dominant path, which is critical for accurate signal modeling.

The model was pretrained on 27 diverse environment scenarios and evaluated on downstream tasks including user localization, beam prediction, and line-of-sight (LoS) classification.

MultiPathFormer demonstrates that path-level pretraining captures reusable representations of wireless propagation. The Environmental RAG and codebook mechanisms significantly improve the estimation of path statistics, reducing delay mean absolute error by 38.7% and power mean absolute error by 59.2% compared to a baseline autoregressive model. In downstream evaluations, MultiPathFormer consistently outperforms state-of-the-art channel-based foundation models, achieving superior accuracy in beam prediction, localization, and channel estimation. The model also shows strong transferability, performing well on unseen users and adapting effectively to new environments through fine-tuning.