Google Shows RL Cuts Quantum Memory Errors Below 0.1% as Drift Stabilization Scales
Updated
Updated · Google Research · Jul 22
Google Shows RL Cuts Quantum Memory Errors Below 0.1% as Drift Stabilization Scales
2 articles · Updated · Google Research · Jul 22
Summary
Google Quantum AI said a reinforcement-learning controller can keep a quantum computer calibrated during computation, avoiding the stop-and-recalibrate cycle that now interrupts long runs.
On Google’s Willow superconducting processor, the system used quantum error-detection signals to steer thousands of control parameters, improving logical stability 3.5-fold under injected drift and cutting logical error rates another 20% after expert calibration.
The combined setup pushed quantum-memory logical errors to record lows—fewer than 1 per 1,000 error-correction cycles in the surface code and about 1 per 100 in the color code.
Simulations with hundreds of qubits and tens of thousands of parameters suggested the RL training effort does not grow with system size, supporting use on larger fault-tolerant machines.
Google said faster feedback between the agent and hardware, plus stronger ML methods, could further improve a system designed to keep quantum computers running for days or months.
On July 22, 2026, Google Quantum AI and DeepMind published a landmark study in Nature introducing AlphaQubit, an advanced reinforcement learning-based quantum error decoder. This breakthrough enables uninterrupted quantum computing by continuously calibrating and stabilizing quantum systems in real time, directly addressing the persistent issue of hardware drift and performance degradation. By eliminating the need for frequent, disruptive recalibrations, AlphaQubit allows quantum processors to run longer and more complex computations. This innovation marks a major step forward in quantum error correction, paving the way for more reliable and scalable quantum computers.