Quantum computers are delicate. Their qubits drift as the surrounding environment shifts, and keeping them accurate normally means stopping the machine to retune it. That pause is manageable today, but it clashes with the ambitions of the field: useful quantum algorithms may one day need to run without interruption for days or even months.
A team at Google Quantum AI and Google DeepMind, reporting in Nature, has found a way to remove the pause. Instead of treating calibration and computation as separate jobs, they merged them.
The idea builds on quantum error correction, the routine that keeps a quantum computer honest by constantly checking for errors. Each check produces a stream of "error" or "no error" signals. Normally those signals are used only to repair the logical state. The researchers gave them a second job: feeding them to a reinforcement learning agent that reads the pattern of errors and quietly adjusts the machine's controls to compensate for drift, all while the computation keeps running.
Learning from its own mistakes
The agent oversees more than a thousand control parameters, the settings that translate an abstract error-correction circuit into the analog waveforms that actually steer the qubits. Tested on Google's Willow superconducting processor, the system improved the logical stability of the surface code 3.5-fold when the team deliberately injected drift, aided by complementary decoder steering.
Along the way, the group also set performance records for two leading error-correction schemes, reaching an average logical error of 7.72 x 10^-4 per cycle for the surface code and 8.19 x 10^-3 for the colour code.
Encouragingly, the agent reached high performance even when it started from randomised settings, hinting that the approach could eventually replace slower, traditional calibration altogether. Numerical simulations of much larger codes, with tens of thousands of parameters, suggested the method keeps working as machines grow, because its optimisation speed does not depend on system size.
The authors stress that the technique is not tied to one kind of hardware; they write that it is "directly applicable to any physical qubit modality and quantum error correction architecture." Their conclusion frames the shift plainly: the work "enables a new paradigm: a quantum computer that learns from its errors and never stops computing."