Google Quantum AI Uses Reinforcement Learning to Self-Calibrate Willow Mid-Computation

Google Quantum AI and Google DeepMind published a Nature paper showing a reinforcement learning agent can continuously recalibrate a superconducting quantum processor using the error-detection events that quantum error correction already produces, rather than pausing for dedicated tune-up passes. Running on the 105-qubit Willow chip, the RL agent adjusted more than 1,000 hardware control parameters in real time, cutting the logical error rate by roughly 20% versus expert-tuned calibration and improving stability 3.5x when hardware drifted, setting a record surface-code logical error rate of 7.72×10⁻⁴ per cycle at distance 7. The 299-author study, led by Volodymyr Sivak and Alexis Morvan, is the first demonstration of RL-based error-correction control at the scale of a full error-corrected processor.