Experimental Stabilization of Quantum Error Correction using Reinforcement Learning

Long-term stability is a primary obstacle to fault-tolerant quantum computation, as environmental drift continuously degrades processor performance. The conventional solution, halting computation for recalibration, is unsustainable for long-running quantum algorithms. 

We present a new paradigm that unifies calibration and computation using reinforcement learning (RL). Our method repurposes error detection events from a running quantum error correction (QEC) protocol as a live feedback signal. An RL agent then actively steers thousands of control parameters in real-time to counteract system drift.

On a superconducting processor running a distance-5 surface code and color code, we experimentally demonstrate that this framework improves the stability of the logical error rate by a factor of 2.4 against injected drift. Crucially, the RL agent also pushes performance beyond the limits of state-of-the-art calibration routines, even after those calibrations have been exhaustively polished by human experts on both surface code and color code. This work establishes a path toward autonomous quantum computers that not only self-stabilize but also continuously self-optimize without interrupting computation.