The system was in charge of two logical qubits housed in a calibrated system. The two used different error correction schemes (a surface code and a color code). These were set to a specific state and then the error correction system was used with and without corrections driven by reinforcement learning. Having the system active led to a 20 percent increase in the ability to detect and correct errors in logical qubits.
Going in real time
The limitation of this approach is that it works only if the drift keeps the system reasonably close to the state in which it was trained. Fixes that could realign things from one state might not be effective when the system is in a significantly different state.
The solution to this is to constantly reevaluate the effectiveness of different changes. But this has an obvious problem: you can’t simply randomize all potential control settings in the middle of a calculation. Even with limited variation, the system will necessarily operate outside of its optimal error correction. So the question was whether frequent suboptimal bug fixing was worth it by preventing drift from causing even bigger problems. “Favourable resolution of the trade-off between exploration and exploitation would mean that the aggregate performance of all candidate policies in the sample, most of which are worse than (the optimal one), is still better than the performance without reinforcement learning guidance,” the researchers write.
Performing many simulations with a very small error-corrected qubit showed that the compensation worked, as long as the drift was slow enough. The team demonstrated that it could operate in real time with a large error-correcting qubit, in which the reinforcement learning system had control over approximately 40,000 parameters.
It is clear that this is not a solution at the moment; We can only keep systems running long enough to run simple, relatively short algorithms, so drift isn’t even a concern. Ultimately, our intention is to build hardware that can perform the kind of calculations where problems like this will be important. And there is some value in showing that something we know could be a problem can be solved.
Nature, 2026. DOI: 10.1038/s41586-026-10759-2 (About DOIs).
