Wireless Foundation Models

Wireless foundation models aim to learn reusable patterns from radio channels, electromagnetic signals, network traces, and access-network telemetry, reducing the need to build a separate model for every task or environment. Common approaches include self-supervised masked or predictive pretraining, Transformers with physics-aware representations, and adaptation methods such as prompts, adapters, and low-rank fine-tuning. These models are being applied to channel prediction, localization, beam selection, sensing, anomaly detection, and resource allocation, including settings with limited labels or constrained edge hardware. A key open tension is whether reported transfer will hold across devices, environments, and tasks: physical structure can improve generalization and efficiency, but fragmented datasets and evaluation protocols make comparisons difficult, and larger models do not always perform better.