Skip to main content

Benchmarks

Leaderboard for wireless foundation model evaluations. Each entry links to the dataset protocol (splits/metrics) and the finetuning recipe used to generate the score. Scores are reported as mean ± std across 3 seeds.

Use filters to select a release and narrow the table; click column headers to sort. Multi-task datasets list subtasks as rows with their own metrics. Click a score to open the detailed evaluation and click recipe to open the finetuning command.

WavesFM v1.0

Multimodal release with a shared ViT backbone for IQ, spectrograms, CSI, and CIR, trained with masked wireless modeling.

  • Supports IQ and image-like modalities via modality-specific input embeddings.
  • Evaluated on RFP, INTD/INTC, HAS, RFS, RML2022, RADCOM, 5G NR positioning, UWB indoor/industrial localization, and DeepMIMO LoS/NLoS + beam prediction.
  • Finetuning regimes: linear probing, partial fine-tuning (FT2), and LoRA (R=32, alpha=64).
13 results
5G NR PositioningCSI5G NR PositioningMean Localization Error (m)
CommRad RFSpectrogramRF Signal ClassificationMean Per-Class Accuracy
DeepMIMO Beam PredictionCSIBeam PredictionMean Per-Class Accuracy
DeepMIMO LOS/NLOSCSILoS/NLoS ClassificationMean Per-Class Accuracy
EfficientFiCSIHuman Activity SensingMean Per-Class Accuracy
ICARUSIQInterference DetectionInterference Detection Accuracy
Interference ClassificationInterference Classification Accuracy
POWDERIQRF FingerprintingMean Per-Class Accuracy
RADCOMIQRadar Signal ClassificationRadar Signal Classification Accuracy
Modulation ClassificationModulation Classification Accuracy
RML2022IQModulation ClassificationMean Per-Class Accuracy
UWB Indoor Positioning and TrackingCIRUWB Indoor PositioningMean Localization Error (m)
UWB Industrial LocalizationCIRUWB Industrial LocalizationMean Localization Error (m)