Towards a quantum computer that learns from its errors

Google Research · 1 month ago · 9

Google researchers demonstrated a reinforcement learning framework that allows quantum computers to continuously recalibrate their control parameters during computation by learning from quantum error detection events, rather than stopping for manual recalibration. The RL agent improved logical stability 3.5-fold on their Willow processor and reduced logical error rates to fewer than one per thousand cycles in surface codes, with an additional 20% improvement after expert calibration. This enables quantum computers to maintain reliable operation during long continuous computations lasting days or months without interrupting calculations for tuning.