Google: RL Algorithm Teaches the Willow Quantum Processor to Correct Its Own Errors, Stability Rises 3.5x

What is quantum error correction, and why does it need RL?

Quantum error correction is a technique that distributes a single logical unit of information across multiple physical qubits, to protect it from the noise that inevitably accompanies quantum computation. Google has now integrated reinforcement learning (RL) — an algorithm that learns optimal actions through trial and reward — into this process for continuous parameter tuning during the computation itself, not just beforehand.

Record-low error rates on Willow

On Willow, Google’s newest quantum chip, the RL approach improves logical stability by 3.5x compared to previous calibration methods. The system achieves fewer than one error per 1,000 correction cycles in the surface code, today’s standard approach to quantum error correction, and 1 error per 100 cycles in the color code, an alternative scheme that uses a different qubit layout. The comparison shows a clear advantage for RL: classical physical calibration relies on pre-measured noise models that quickly become outdated as the system changes, while RL adapts parameters on the fly.

Scaling to hundreds of qubits without retraining

Simulations show that the approach can scale to systems with hundreds of qubits and tens of thousands of control parameters, and training is independent of system size, meaning the model doesn’t need to be rebuilt for every new chip. This represents a shift from physical calibration models toward learning-based approaches, where AI takes over a task previously performed by engineers through manual tuning. Google states that this approach could accelerate the path toward practically useful quantum computers, since system stability directly determines how long complex algorithms can run before accumulating errors that invalidate the result.