Wireless Foundation Models: State-of-the-Art and Open Challenges

Summary

This survey systematizes the fast-growing literature on wireless foundation models (WFMs) for physical-layer applications. Using a PRISMA-informed search across seven databases covering 2024 to June 2026, the authors first formalize what qualifies as a WFM — pretraining on broad wireless data, reusable representations across tasks, and meaningfully distinct downstream tasks — and grade evaluation evidence as in-distribution, partial-shift, or out-of-distribution. The literature is organized into five task families (signal recognition and demodulation, channel representation learning, RF sensing and localization, beam management, spectrum sensing and monitoring) plus multi-task PHY models, each analyzed through uniform tables covering pretraining modality, objective, backbone architecture, adaptation strategy, and evaluation shift. Reported patterns include the dominance of masked-reconstruction pretraining (15 of 30 channel models, 7 of 16 beam models, 4 of 6 spectrum models), prevalence of Transformer backbones, and predominantly partial-shift evaluation with limited OOD evidence. The authors conclude that evidence for reusable wireless representations is growing but uneven, and that heterogeneous datasets and protocols prevent attributing transfer gains to specific design choices; they also observe that no surveyed work isolates gains from shared representations versus task-specific conditioning. Main limitations: the synthesis is qualitative with no normalized cross-model comparison; shift labels rest on implicit author judgment; the evidence base leans heavily on unreviewed preprints; and, as the paper's own Table IV shows, much of it relies on private datasets.