A group of researchers has published what it calls the first unified survey of wireless foundation models, the AI systems being proposed to run next-generation 6G networks.
The paper reviews how these models differ from today's wireless AI, which is typically trained separately for each job: one model for signal processing, another for localization, another for network optimization. Wireless foundation models instead learn generalized representations from large, varied wireless data, then adapt to new tasks with minimal extra training. The survey lays out a taxonomy covering model architectures, self-supervised pre-training methods, and parameter-efficient adaptation techniques, and catalogs the datasets and benchmarks used to evaluate them. It also surveys proposed applications spanning physical-layer signal processing, network intelligence, and cross-layer optimization.
This matters because 6G standards work is happening now, and "AI-native" is one of the buzzwords carriers keep attaching to it. A foundation-model approach would mean training once on broad wireless data and reusing that base across sensing, localization, and optimization tasks, rather than building bespoke models for each. That is the same bet that paid off for language and vision AI, now aimed at radio networks.
The authors are upfront that this is still mostly a roadmap: they flag data availability, generalization, interpretability, efficient edge deployment, and standardization as unresolved. A survey organizing a fragmented research field is useful. It is not the same as a wireless foundation model anyone is running in a live network.