Google Uses Error Signals to Retune Willow's Quantum Controls
A reinforcement-learning experiment improves stability under injected drift and fine-tunes an already calibrated processor.
Willow retuning was tested under injected drift
- Reported
- Sivak and colleagues report reinforcement-learning control that reduces variation in logical error rates under deliberately injected drift.
- Supported observation
- Control steering improves stability 2.4-fold against a fixed policy; adding decoder steering raises this to 3.5-fold. Stability here refers to the standard deviation of logical error rates.
- Conditions
- The hardware experiment uses short quantum-memory runs on Willow with states prepared again for each shot. The decoder-steering method also estimates logical errors. The drift comparison uses control parameters fixed at their initial calibration.
This does not demonstrate uninterrupted execution of a long algorithm. The reported decoder steering has real-time scaling limitations; large-code scaling and continuous steering are studied in simulation.
- Reinforcement learning control of quantum error correctionNature · Peer-reviewed primary research paper · Source published Source checked
Google researchers are using Willow's quantum error-detection events as reinforcement-learning feedback to adjust the processor's control parameters. The peer-reviewed hardware experiment explores whether keeping calibration running during computation could reduce interruptions caused by hardware drift.
The paper reports a 2.4-fold improvement in logical-error-rate stability under injected drift, rising to 3.5-fold with additional decoder steering. In a separate test, fine-tuning an already calibrated processor reduced logical errors by about 20%.
Those results have different conditions: the 3.5-fold figure includes decoder steering whose reported implementation faces real-time scaling limitations, while the large-code results come from simulations. Further tests of sustained operation under natural drift and scalable decoder feedback would clarify how far those gains extend.