RFPrompt:
adapt a wireless model with prompts.
A signal-intelligence demonstration that adapts a frozen large wireless model with small expert prompts, then tests modulation classification with limited labels and shifted signal conditions.
Change the label budget. See adaptation.
Upload IQ samples, choose modulation classes, and compare a frozen model, partial fine-tuning, and RFPrompt as labels and signal conditions change.
Technical significance
On the Real-World IQ test set, RFPrompt peaks at 0.82 accuracy at N=400, exceeding the reported frozen and partially fine-tuned baselines.
The prompt-based adaptation uses about 0.34% trainable parameters in the reported comparison, while keeping the backbone frozen.
The reported sweep identifies P=16 as the best prompt length for the evaluated adaptation setting.
Figures are cropped from the submitted paper to show the system and the most relevant reported results.
Back to demos ↗