Work

Drift: a calibration-free neural decoder that adapts to synthetic drifting noise from its own syndromes

Year: 2026
Type: preprint
Abstract: Quantum computers don't hold still. Their error rates drift: defects switch on and off, and cosmic rays briefly scramble groups of qubits. Most decoders are tuned to one fixed noise model, so when the noise moves they keep correcting for a chip that no longer exists. Drift is a neural decoder that watches its own data instead. Alongside every shot it gets a "fingerprint": how often each detector has fired over the last 2,000 shots it decoded. It never sees calibration data or the true noise, only what its own syndromes tell it. In circuit-level simulation at code distance 5, on large error bursts it was never trained on, a three-model ensemble reaches a logical error rate of 0.0457. Standard matching gets 0.0980, the best sliding-window reweighted matching 0.0896, and matching that knows the true noise 0.0426. A control experiment separates the two sources of this gain. With the fingerprint frozen, the same network scores 0.0576 (paired McNemar p = 3×10⁻²⁴⁷), close to a network trained without one. So the decoder genuinely adapts to synthetic drift, but most of its lead over matching comes from being a better decoder in the first place. The same network can also point to which qubits went bad. It does this better than a simple baseline on bursts like those it trained on, but not clearly on larger, unfamiliar ones. Four architectural extensions were judged by decision rules set before each run; all four failed, and the paper reports them. Every result uses three training seeds and five independent test streams. Every version-1 number can be recomputed from the released per-shot data, and a claim-by-claim audit was run before release. All results are from simulation at distance 5 or less. v2: adds a pre-registered physics-based drift evaluation (SI1000 device noise with TLS, cosmic-ray, slow and coupler drift). Without retraining, all four ensembles beat MWPM calibrated to the device (logical error rate 0.0016–0.0017 against 0.0022); the live fingerprint gives no gain on this drift. Round-10 streams regenerate from the manifest in the repository.
Language: English
Open Access status: green
OpenAlex ID: W7236175690
Cites:
Cited by:
Related to: 10