This paper introduces channel configuration (CFG) information into the pre‑training of Radio Foundation Models (RFMs) and shows that the resulting CFG‑enriched embeddings act as strong inductive priors for downstream tasks when labeled data are scarce. We compare these embeddings against raw channel tokens and several baseline models across three data regimes, evaluating three downstream tasks: Line‑of‑Sight/Non‑Line‑of‑Sight (LoS/NLoS) classification, timing advance (TA) estimation, and beam selection. Experiments are conducted in both in‑distribution (seen) and out‑of‑distribution (unseen) configuration settings. The results consistently demonstrate that CFG‑enriched embeddings either outperform or remain competitive with baselines, even under unseen configurations.
Review: The study validates the benefit of incorporating environmental configuration into large‑scale radio models, offering a promising direction for label‑efficient wireless communication research.