One of the most persistent challenges in quantum computing is drift: the control parameters that keep a quantum processor running at peak performance shift over time due to temperature fluctuations, material aging, and other environmental factors. The standard solution is to stop the computation, run calibration routines, and restart, a process that wastes time and limits how long a logical qubit can be maintained.
A new Nature paper from Google Quantum AI and Google DeepMind, with first authors Volodymyr Sivak, Alexis Morvan, and Michael Broughton, demonstrates an alternative: reinforcement learning that uses the error-detection events already produced by quantum error correction as a continuous learning signal, steering the processor’s controls in real time without interrupting the computation.
How it works
In quantum error correction, error-detection events are generated when a syndrome measurement flags a potential error. These events normally serve a single purpose: telling the decoder which corrections to apply. The new approach gives them a dual role: the same events are also used as a reinforcement signal for a model-free policy gradient algorithm called PEPG (Parameter-Exploring Policy Gradients).
The algorithm maintains a probability distribution over control parameters and updates it based on the rate of error-detection events. Critically, the system exploits a sparse factor graph structure: each error detector is connected only to the control parameters of gates within its detecting region, with an average of 302 parameters per detector and 18 detectors per parameter. This sparsity allows the reinforcement learning to converge independently of system size, meaning the approach should scale to larger processors.
The experiments were run on Google’s Willow superconducting processor (105 qubits), with control over more than 1,000 parameters for the distance-5 color code and simulated scalability to roughly 40,000 parameters for a distance-15 surface code.
The numbers
On the Willow processor, reinforcement learning steering improved logical error rate (LER) stability by a factor of 2.4, a 24% reduction in LER. When combined with decoder steering, the improvement reached 3.5-fold, or 31% LER reduction.
The system achieved record LERs for the hardware: 7.72 x 10^(-4) for the surface code at distance 7 using the AlphaQubit2 decoder, and 8.19 x 10^(-3) for the color code at distance 5 using the Tesseract decoder. These represent roughly 20% additional LER suppression beyond what expert manual calibration could achieve.
The RL system could track step-like parameter drifts within roughly 130 epochs and suppress low-frequency LER fluctuations by approximately 4 dB. The critical drift frequency for effective real-time steering was around 1/150 epochs, meaning any drift slower than that timescale is automatically corrected.
Why it matters
Continuous recalibration is a practical necessity for building stable quantum processors. Current calibration routines require stopping the processor, which limits the coherence time of logical qubits and introduces overhead that scales with processor size. An RL system that eliminates the need to stop is a step toward fault-tolerant quantum computing at scale.
The approach also simplifies the hardware requirements: rather than needing exquisitely stable control systems that never drift, the processor can tolerate drift and correct for it autonomously. The authors note that the convergence rate of the RL system is independent of system size, which suggests the approach should continue to work as processors grow to thousands of qubits.
The paper was published in Nature on July 8, 2026, and was covered by Ars Technica’s John Timmer.
Sources
[1] Sivak, V., Morvan, A., Broughton, M., et al. “Reinforcement learning control of quantum error correction.” Nature (2026). DOI: 10.1038/s41586-026-10759-2
[2] Timmer, J. “Quantum error correction can constantly recalibrate a processor.” Ars Technica (2026). https://arstechnica.com/science/2026/07/quantum-error-correction-can-constantly-recalibrate-a-processor/