Google Quantum AI: обучение с подкреплением втрое стабилизировало квантовый чип Willow

On July 22, 2026, Google Quantum AI presented a system in the journal Nature that uses reinforcement learning to self-calibrate the Willow quantum processor directly during computation — without stopping — tripling the stability of the error-correction code.

How the chip learns from its own mistakes

Google Quantum AI built a reinforcement learning algorithm directly into the quantum error correction (QEC) loop. An autonomous agent monitors error-detection events on the Willow processor and adjusts thousands of control parameters in real time, compensating for hardware drift — all without interrupting the computation. Previously, calibration had to be run separately, stopping the processor's operation; now the system "tunes itself on the fly."

• Publication — July 22, 2026, journal Nature
• Platform — superconducting quantum processor Willow
• 3.5-fold increase in the logical stability of the error-correction code
• Additional 20% reduction in logical error rate after training
• Record: fewer than 1 error per 1000 cycles (surface code), about 1 per 100 (color code)

How much did errors decrease

Additional training via reinforcement learning delivered a 3.5-fold increase in the logical stability of the error-correction code and further reduced the logical error rate by 20%, Google Quantum AI reports. According to the Nature publication, the system reached record-low figures: fewer than one error per thousand cycles for surface code and around one per hundred for color code.

For decoding, Google uses the neural-network decoder AlphaQubit and the algorithmic decoder Tesseract. Reinforcement learning operates on top of them: it is not responsible for decoding errors, but for keeping the "hardware" itself in an optimal state while the count is running.

"We found a way to tune the instruments while the music is playing," is how the authors describe continuous calibration during active computation in the

Google Research blog.

Does the approach scale?

The approach scales to larger quantum machines: according to Google Quantum AI, the number of training iterations does not depend on system size. In numerical simulations with hundreds of qubits and tens of thousands of control parameters, the number of RL training rounds remained roughly constant — meaning the method will not "choke" when scaling up to the thousands of qubits needed for practical fault-tolerant computation.

The lead authors of the work are Google Quantum AI researchers Volodymyr Sivak and Paul Klimov. According to the team, it is precisely this scale-independence of training cost that makes the method suitable for future large quantum computers, not just for today's Willow.

What does this mean

Reinforcement learning turns a quantum computer from a fragile instrument requiring constant manual tuning into a system that stabilizes itself. This removes one of the key barriers on the path to fault-tolerant quantum computing — the continuous hardware drift that for years had to be compensated for manually.

Frequently Asked Questions

What is reinforcement learning in quantum error correction?

It is an AI agent that learns from error signals and itself adjusts thousands of the processor's control parameters during computation. In Google Quantum AI's work, it delivered a 3.5-fold increase in logical stability without stopping the count.

What is the Willow processor?

Willow is Google's superconducting quantum processor on which the method was tested. It was on this chip that the system reached the record: fewer than one logical error per thousand cycles for surface code.

Does this replace the AlphaQubit and Tesseract decoders?

No. AlphaQubit (a neural-network decoder) and Tesseract (an algorithmic decoder) still decode errors, while reinforcement learning operates on top of them — it calibrates the hardware, it does not decode.