RF-GPT architecture: from RF waveform to spectrogram to language model

RF-GPT: Teaching Language Models to See the Invisible Spectrum

RF-GPT is the first radio-frequency language model — it converts RF signals into spectrograms and feeds them through a vision-language model, achieving 99.6% wireless technology recognition across 6 standards while general-purpose VLMs score near-random.

February 17, 2026 · 14 min · Hang Zou