Quantum Computer That Learns from Its Mistakes

Google Quantum AI (in collaboration with Google DeepMind) published a paper in Nature combining reinforcement learning (RL) with quantum error correction (QEC). Traditionally, calibrating quantum computers requires completely halting computations, which is unacceptable for long-duration algorithms. The authors proposed using QEC error detection data not only for decoding and correction but also as an active training signal for an RL agent that dynamically adjusts thousands of control parameters, compensating for drift. An experiment on a 105-qubit Willow processor with artificially introduced drift showed a 3.5-fold improvement in logical stability. Even after expert calibration, RL fine-tuning further reduced the logical error rate by 20%. Record-level performance was achieved: fewer than one logical error per thousand correction cycles for the surface code and one per hundred for the color code. Numerical simulations for hundreds of qubits confirmed that the number of required training iterations does not depend on system size, indicating the method is scalable.