Google has developed a reinforcement learning system that continuously recalibrates its Sycamore quantum processor during computation, using error correction data to adjust roughly 1,000 control parameters in real time. The approach addresses calibration drift in transmon qubits — superconducting circuits controlled by microwave pulses — where hardware variations and thermal drift degrade performance during long algorithms. Unlike traditional calibration that halts computation, this method runs parallel to error correction, distinguishing calibration errors from random noise through deliberate parameter perturbations.

Testing on two logical qubits using surface code and color code schemes showed a 20 percent improvement in error detection and correction. The system scales to roughly 40,000 parameters for larger error-corrected qubits, though it requires drift to remain slow enough for the exploration-exploitation tradeoff to pay off. Simulations confirmed the aggregate performance of sampled policies outperforms uncorrected drift.

This solves a known bottleneck for future fault-tolerant quantum computers running complex algorithms like encryption-breaking routines. While current hardware operates too briefly for drift to matter, demonstrating continuous calibration viability removes a theoretical roadblock for scaled systems. The work appears in Nature 2026.