Sample-Efficient Self-Interference Cancellation for In-Band Full Duplex Radios via In-Context Learning

Digital self-interference cancellation is a central bottleneck in realising in-band full-duplex radios for 6G. Modern cancellers can model power-amplifier nonlinearities accurately, but their coefficients must be re-estimated from short calibration-interval pilots whenever the amplifier, front-end, or self-interference channel drifts. We recast this online calibration problem as in-context learning (ICL): a transformer pre-trained over a distribution of amplifier operating conditions, channel, and front-end conditions predicts the interference from a few calibration-interval examples supplied as context. On a synthetic model derived from real data, the proposed method matches the analytical Bayes-optimal predictor across context lengths. With real measured PA samples passed through a simulated Rician self-interference channel and I/Q imbalance, ICL is roughly $4.4\times$ more sample-efficient than a widely-linear alternating-least-squares baseline. On USRP~N210 full-duplex captures, ICL reaches its NMSE floor in approximately $64\times$ fewer calibration samples than the strongest model-based baselines. When the context samples are appropriately normalized, we also show that ICL adapts zero-shot across nominally identical hardware units.