Quantum computers have long faced an awkward trade-off: to keep errors low enough for useful computation, engineers must periodically stop the machine to recalibrate its control settings against environmental drift. A new study describes a way to eliminate that pause entirely, using artificial intelligence to steer the hardware while it continues computing.
The work, carried out on Google's Willow superconducting processor, unifies calibration and computation by treating quantum error correction's own error-detection signals as a training source for a reinforcement learning agent. Rather than only using these detection events to correct the logical quantum state, researchers repurposed them to teach the agent how to continuously adjust the machine's control parameters.
Quantum error correction (QEC) works by digitizing the fragile, analogue behavior of qubits into discrete "error" or "no error" events, which can then be decoded to protect a logical quantum state. But QEC only functions if the physical error rate stays below a threshold of roughly 10⁻³ to 10⁻², meaning the underlying hardware must be precisely tuned. Because that tuning is inherently analogue, it drifts over time — and previous experiments dealt with this by halting the entire QEC process for recalibration, an approach incompatible with the long runtimes — potentially days or months — that future quantum algorithms will require.
A self-correcting machine
In the experiment, the researchers deployed a reinforcement learning agent managing more than 1,000 control parameters on distance-5 and distance-7 surface codes and a distance-5 colour code. Against deliberately injected drift, the agent improved logical stability 2.4-fold on its own, and 3.5-fold when combined with complementary decoder steering. When applied to a processor that was already well calibrated, the RL fine-tuning delivered a further 20% reduction in logical error rate, surpassing what traditional physics-based calibration and human expert tuning could achieve. Notably, the agent could reach strong performance even when starting from randomized initial control parameters.
By combining these advances, the team reported record performance for both code types: an average logical error per cycle of 7.72(9) × 10⁻⁴ for the surface code and 8.19(14) × 10⁻³ for the colour code.
Numerical simulations extended the findings further, showing the framework scales to a distance-15 surface code with tens of thousands of control parameters, with optimization speed that does not slow as system size grows. The approach relies only on error-detection signals and tunable controls, meaning it should apply to other physical qubit platforms and QEC architectures beyond the superconducting circuits used here.
The authors describe the result as establishing reinforcement learning as a promising path toward automating the control of large-scale, error-corrected quantum systems — machines that, in their words, learn from their own errors and never stop computing.