The system was responsible for two logical qubits placed in a calibrated system. The two used different error correction schemes (surface code and color code). These were set in a given condition, and the error correction system was then used with and without reinforcement-learning-based corrections. The activation of the system led to a 20 percent increase in the ability to detect and correct errors in logical qubits.
Real time is running
A limitation of this approach is that it only works if the drift keeps the system close enough to the state in which the system was trained. Adjustments that may adapt everything from one state may not be effective when the system is in a significantly different state.
The solution is to constantly reassess the effectiveness of various changes. But this has an obvious problem: You can’t simply randomize all potential control configurations in the middle of a calculation. Despite the limited variability, the system will necessarily operate beyond optimal error correction. So the question was whether correcting the often sub-optimal errors paid off by preventing the drift from causing bigger problems. “A favorable solution to the exploration-exploitation trade-off would mean that the overall performance of the majority of all policy candidates selected is worse than (optimal) still better than the performance without reinforcement learning management.”
Performing many simulations with a qubit tuned with a very small error showed that the exchange works when the drift is slow enough. The team showed that the reinforcement learning system can work in real-time with a large error-corrected qubit that controls about 40,000 parameters.
This is obviously not a solution for now; we can only run systems long enough to execute relatively short, simple algorithms, so drift is not a concern. Ultimately, our intention is to create hardware that can perform calculations where such matters matter. There is some value in demonstrating that something we know can be a problem can be solved.
Nature, 2026. DOI: 10.1038/s41586-026-10759-2 (About DOIs).
