RFShape: An IQ-Pretrained Wireless Foundation Model for Heterogeneous IQ & CIR Representations
In wireless systems, foundation models have been proposed as alternative artificial intelligence models to replace task-specific models with a shared pretrained model that can be reused across many downstream applications (source separation, interference detection, localization, etc.). These wireless foundation models promise to reuse learned representations across diverse tasks, but choosing a foundation model also means choosing its input representation. For example, a model pretrained on spectrograms expects spectrograms, while one pretrained on in-phase/quadrature (IQ) samples expects IQ, limiting which pretrained model can be used for a given downstream problem. To remedy this, we present RFShape, a 7.4 M-parameter foundation model that breaks this coupling by adapting how a signal is presented to a shared pretrained encoder. RFShape is pretrained solely on synthetic IQ waveforms, without real captures, labels, or channel impulse responses (CIRs), and uses deterministic representation operators to support heterogeneous downstream inputs. Across ten tasks and nine datasets unseen during pretraining, the same encoder supports classification, ranging-error regression, and two-dimensional positioning. Representation selection alone increases 23-class radar recognition from 23.0% to 69.5% and short-range technology recognition from 70.8% to 97.1%. RFShape also outperforms CIR-pretrained ultra-wideband baselines by up to 12.7 percentage points for (N)LOS classification despite never observing CIRs during pretraining. These results show that foundation-model reuse need not be limited by the representation or modality seen during pretraining, and adapting how a signal is presented can extend a single pretrained encoder across heterogeneous wireless tasks.