The Signal
Preprint reports benchmarked-noise ZNE beats standard method on hardware
A preprint reports that randomized benchmarked-noise zero-noise extrapolation produced more accurate results than standard zero-noise extrapolation on IBM and IQM quantum processors. The tests used the same total shot count for both methods, while the new approach added benchmark circuits to measure effective noise.

A randomized version of benchmarked-noise zero-noise extrapolation (bnZNE) produced more accurate results than standard zero-noise extrapolation in tests on IBM and IQM quantum processors, according to a preprint posted Nov. 12. The work by Joseph Harris, Kevin Lively and Peter Ken Schuhmacher of the German Aerospace Center reports tests across quantum simulation and other circuit applications, with both methods using the same total number of shots.
Source arxiv.org
Measuring noise for the extrapolation
Standard zero-noise extrapolation, or ZNE, runs versions of a circuit with deliberately increased noise and extrapolates the measured values toward a zero-noise estimate. The authors say this approach can be inaccurate when the nominal noise-scaling values do not track the noise actually experienced by the circuit.
BnZNE adds benchmark circuits that mirror the structure of the application circuits and can be simulated classically. Their measured outputs estimate effective noise at each scaled level; the extrapolation then uses those measured noise levels rather than the standard method’s scale factors. The randomized version also allows multiple circuit variants at each level, enabling techniques such as Pauli twirling and readout-error mitigation.
Tests across two processors
On IBM hardware, the experiments used the 156-qubit ibm_fez superconducting processor. The applications included a kicked Ising quantum simulation using 100 qubits, quantum Fourier transform circuits using 20 and 50 qubits, and a five-qubit Grover search. The researchers also tested a 30-qubit kicked Ising simulation on IQM’s 54-qubit Emerald processor.
For each application, the researchers used 32 randomized circuits at each of three noise levels. Standard ZNE used 2,048 shots per circuit, while bnZNE used 1,024 shots for each application and benchmark circuit; the total was 196,608 shots per application for either method. The experiments combined Pauli twirling, dynamical decoupling and twirled readout error extinction on IBM, while the IQM tests used randomized, hardware-tailored benchmark circuits.
Accuracy and limits
The authors report that bnZNE improved mean fidelity over standard ZNE in every IBM application they tested, with mean error often below 5%. They also report more accurate results for the IQM simulations, including circuits with between 5 and 40 Trotter layers. In the IBM tests, the paper says, greater circuit complexity was accompanied by higher variance in bnZNE estimates.
For one multi-qubit correlation analysis, the researchers excluded seven bnZNE datapoints because their variance was high. They say the method still improved on standard ZNE in noise-heavy, deep-circuit cases, but the low signal-to-noise ratio led to less stable fits. The paper describes the additional benchmark circuits as the method’s extra circuit cost, while its shot-count comparison held total shots equal between bnZNE and ZNE.