Researchers at Google Quantum AI and Google DeepMind reported in Nature on July 8, 2026, that a reinforcement-learning (RL) agent can adjust a quantum processor’s control settings between error-correction runs. Reinforcement learning is a machine-learning method in which a program, called an agent, improves its actions by trial and error, guided by a reward. The team used the error-detection signals that error correction already produces as the agent’s learning signal, in place of halting the processor to recalibrate it.
On Google’s Willow superconducting processors, the team used the agent to steer the control settings across repeated runs, or shots, of short error-correction experiments, with the quantum state re-prepared for every shot. It demonstrated steering within a single long computation, which future machines will need, only in numerical simulation.
This article draws on the fourth version of the authors’ preprint on arXiv, a public server for research papers. The authors posted that version on June 19, 2026, after Nature accepted the paper on June 4.
Why recalibration interrupts error correction
Quantum error correction (QEC) spreads information across many physical qubits to catch and fix errors. Some physical qubits, called measure qubits, repeatedly check groups of data qubits, which store the encoded information. The surface code and the color code are two error-correcting codes of this kind.
Each round of checks is one error-correction cycle. A change in a check’s outcome from one cycle to the next, called a detection event, indicates a likely error.
The authors wrote that keeping errors rare enough for QEC requires constant adjustment of the processor’s control settings as conditions drift. Drift is slow, unwanted change in the hardware’s behavior over time. Experimenters now stop the entire computation to recalibrate, an approach the authors called “incompatible with the long runtimes of future quantum algorithms.”
Useful future algorithms “require continuous execution times on the order of days or even months,” the authors wrote. They cited a 2017 estimate for chemistry simulations and Craig Gidney’s 2025 estimate for factoring a 2048-bit RSA key.
How the agent learns
The agent manages more than 1,000 control parameters, settings that “specify how an abstract QEC circuit is translated into analog waveforms,” the authors wrote. They counted 1,582 parameters for a distance-5 surface code and 2,657 for a distance-7 surface code. Code distance is the smallest number of physical-qubit errors that can corrupt a code’s stored information without being detected.
The agent samples batches of candidate settings for all the parameters at once. It gives a higher reward to candidates that produce fewer detection events and updates its sampling distribution toward them.
The team used the average rate of detection events as “an efficiently-computable local proxy” for the logical error rate, the rate at which the stored information is corrupted. The authors wrote that the logical error rate itself “is unsuitable for real-time calibration and steering during a quantum computation.”
Results on Willow hardware
The team first calibrated its processor with what the authors described as an exhaustive conventional process. In five independent runs on each of two code types, the distance-5 surface and color codes, further RL fine-tuning “consistently yields about 20% additional suppression of the LER,” or logical error rate, the authors wrote.
To test steering, the team injected artificial drift into several control parameters of a distance-5 surface code at once. With RL steering alone, the team measured a 24% average reduction in the logical error rate and a 2.4-fold gain in its stability, defined by the spread of its values.
The team also adjusted the decoder, the software that infers errors from detection events. With both adjusted, it measured a 31% average reduction in the logical error rate and a 3.5-fold gain in its stability.
Without injected drift, the team analyzed its fine-tuning runs and estimated about 4 decibels of suppression of slow fluctuations in the logical error rate. For that estimate, the authors treated the agent’s adjustments as a filter on those fluctuations. They cautioned that this filter model is not fully justified.
The team combined the agent with its other recent advances and reported an average logical error per cycle of 7.72 × 10⁻⁴ for a distance-7 surface code decoded with the AlphaQubit 2 neural-network decoder. For a distance-5 color code decoded with Tesseract, it reported 8.19 × 10⁻³.
The team ran the distance-7 experiment on a different Willow processor from the distance-5 work. The authors attributed both figures to the combination of all their technical advances and called the results “record QEC performance demonstrated across any physical qubit modality.”
What ran in simulation
In numerical simulations, the team scaled the method to a distance-15 surface code with about 40,000 control parameters. The authors reported an optimization speed “independent of system size.”
The team also simulated real-time steering of a distance-3 surface code under sinusoidal drift, a regular wave-like rise and fall. The authors wrote that “in the future this steering must be done in real time during the single-shot execution of a long logical algorithm.”
Other uses and availability
Google researchers also applied the fine-tuning in a magic state cultivation experiment that they first reported in a December 2025 preprint. Magic state cultivation is a method for preparing the special resource states that fault-tolerant computers consume to run most algorithms.
In cultivation, any attempt in which the checks detect an error is discarded, a step called postselection. The Nature paper’s authors wrote that in the cultivation experiment, the fine-tuning “provided an order of magnitude improvement in the cultivation error and the postselection rate.”
The team posted its data on Zenodo, a public repository for research data. The authors wrote that the code is Google’s proprietary property and that author-affiliated entities have filed patent applications in the United States and internationally.
Update, November 13, 2026: Science published the Google researchers’ magic state cultivation paper on October 15, 2026.
