Google Quantum AI reports that Virtual Error Cancellation reduced residual logical errors by more than 1,600-fold in distance-7 surface-code simulations. The offline method reuses recorded stabilizer measurements, but has not yet been demonstrated on a quantum processor.
In numerical simulations, Google Quantum AI researchers reduced residual logical errors in a distance-7 surface code by more than 1,600-fold using software post-processing. Their Virtual Error Cancellation (VEC) protocol does not change the circuit run on a quantum processor: it uses recorded stabilizer measurements to reweight outcomes afterward. The method is described in the 8 October 2026 arXiv preprint Measure Now, Mitigate Later.
A software correction layer
Fault-tolerant circuits use stabilizer measurements to detect patterns of errors without directly reading out the encoded quantum information. A real-time decoder must interpret those measurement results quickly enough to guide Pauli-frame updates during execution. That deadline limits how much computation the decoder can perform.
VEC pairs that fast decision-maker with a more computationally intensive reference decoder that runs offline. The reported architecture uses minimum-weight perfect matching for real-time decoding and tensor-network maximum likelihood for the later reference analysis. The two decoders process the same syndrome record; their agreement or disagreement then informs each shot's weight in the final estimate. The division of labor lets the more demanding analysis run after the experiment, outside the tight timing constraints of real-time control.
This is not a method for preventing faults in the processor or improving the fidelity of individual operations. Instead, it aims to reduce bias in estimates from universal logical circuits after the run, using stabilizer records without changing the quantum circuit, prior noise characterization, or separate calibration runs.
Signed weights matter
VEC does not simply discard experimental shots when the decoders disagree. Agreement receives a positive weight greater than one, while disagreement receives a negative weight associated with the inferred X, Y, or Z fault direction. Averaging these contributions is intended to cancel residual logical Pauli error in the agreement results across observables.
That distinction matters: postselection throws away discordant data, which can waste processor time and introduce bias. Signed reweighting instead retains those records in the estimate. The contrast with capital-intensive hardware efforts is visible in a recent account of photonic factory financing: VEC's proposed gain is in how data are processed, not in a new processor or manufacturing line.
What the simulations show
In code-capacity simulations of a distance-7 rotated surface code at physical error rate p = 0.005, VEC reduced residual logical errors by more than 1,600 times relative to unmitigated decoding and by more than 200 times relative to agreement postselection. This is consistent with the preprint's characterization of the improvement as exceeding three orders of magnitude compared with the ordinary decoded logical error. The researchers also report an overhead scaling exponent of about 1.3 for VEC at distance 7, compared with about 4 for standard logical probabilistic error cancellation; they project that VEC's exponent approaches 1 as code distance increases.
These are numerical results, not a demonstration on a running quantum processor. The described scenario is presented as experimentally achievable, but that does not make it a hardware result. The simulations do not establish performance under the full range of hardware effects or show that universal logical circuits have been executed with VEC. The numerical gains are evidence for the decoding strategy, not a measured reduction in errors on a physical machine.
Evidence and limits
The method estimates noise parameters from recorded stabilizer syndromes using Walsh-Hadamard syndrome-fidelity linear inversion. The researchers say this removes the need for dedicated noise-characterization circuits or physical gate interventions. That is a software-level simplification, but the reported result remains a research protocol tested numerically: the available confirmation does not establish a hardware validation, independent replication, peer-reviewed publication, or deployment in an operating quantum processor.
The central engineering question is whether offline computation produces reliable estimates for realistic circuits and noise. Signed weights also make the reported overhead a central part of the proposal: a large error reduction is useful only if the number of samples and post-processing resources required remain manageable. The reported exponent is encouraging within the simulation, not proof of practical scaling.
VEC is a credible way to make syndrome records do more work, and the simulation numbers justify further testing. But the result is evidence for an error-mitigation protocol, not evidence that logical errors have been eliminated or that a universal fault-tolerant computer is ready. A hardware test would determine whether this separation between fast decoding and careful offline analysis survives real processor noise.
The work fits into, but is distinct from, Google Quantum AI's earlier surface-code research reported in Nature. Those studies concerned physical error correction, including suppression of errors as code distance increased and operation below the surface-code threshold; VEC instead reweights estimates after a computation. A logical qubit stores information across multiple physical qubits so that stabilizer measurements can reveal certain errors without directly measuring the encoded state. A decoder interprets those measurements and infers likely faults; it does not make the underlying physical operations error-free. VEC adds a later statistical correction to estimates derived from those records, making it error mitigation rather than real-time physical correction of every fault.
