A team led from the University of Sydney removed underperforming components from an error-correcting code on an IBM processor and cut the code’s logical error rate per round by up to 2.8 times, the researchers reported in a preprint posted July 13, 2026. The nine authors ran their experiments on ibm_miami, an IBM Nighthawk processor with 120 physical superconducting qubits arranged in a square lattice.
The team used an error-correcting code called the surface code to spread one logical qubit, the unit of information an algorithm works with, across a patch of physical qubits. A code’s distance is the smallest number of physical errors that can combine into a logical error. In its standard rotated form, a distance-3 surface code uses 17 physical qubits and a distance-5 code uses 49.
The team reduced logical errors per round by up to 2.8 times by excluding defects, but it got its larger code to outperform a smaller one only after discarding runs spoiled by qubit leakage, and only in one of two bases.
Superconducting qubits can leak out of the two energy levels that store information into higher levels. The two bases are two ways of preparing and reading out a logical qubit, each sensitive to a different type of error.
Data qubits in the patch store the encoded information. Measure qubits between them repeatedly check their neighbors for errors through couplers, the tunable connections between qubits.
Seven of the nine authors are at the University of Sydney, with one each from Sungkyunkwan University and IBM Quantum. The Intelligence Advanced Research Projects Activity (IARPA) supported three of the authors through its Entangled Logical Qubits program. Korea’s National Research Foundation and the Australian Research Council supported others.
Defects on ibm_miami
Citing IBM’s June 2026 calibration data, the authors wrote that the median two-qubit gate error on ibm_miami was about 0.283%. A two-qubit gate is a basic operation on a pair of linked qubits. The worst couplers had error rates up to about 16.9%, roughly 60 times the median.
The team flagged couplers with error rates above 5%, or lacking calibration data, as underperforming. For measure qubits, its threshold was an error rate of 7%. It counted a data qubit as defective when its relaxation and dephasing times, two measures of how long a qubit keeps its state, both fell below 80 microseconds, against device medians of 335 and 256 microseconds.
The authors wrote that “it is not possible to place a standard distance-5 rotated surface code on the device using this geometry while avoiding defective components.”
Memory and stability results
To exclude defects, the team removed flagged components from the circuit and merged the checks around them into larger “super-stabilizers” that avoid the damaged site. In a memory experiment, a test of how fast errors build up in a stored logical qubit, the team compared four strategies in one basis.
With the defects ignored, the logical error was 4.49% per round. With full calibration data given to the decoder, the software that infers errors from the check results, it was 4.22%. With the defective couplers excluded, it was 2.71%. With one defective data qubit also excluded, it was 1.62%. Removing that qubit took a row out of the patch and cut one of the code’s two distances from 5 to 4.
In the other basis, the team measured 2.03% and 1.89% with the two exclusion strategies. When the team excluded a healthy coupler as a control, its results were marginally worse.
The team also ran a stability experiment, which the authors called “a useful proxy for the reliability of operating a logic gate on the hardware.” With defects excluded, the team found that the failure rate fell by 6.3% with each added round of checks. Without exclusion, it observed no decay in the failure rate as more rounds of checks were added to the experiment.
Limits of the result
The preprint has not been peer reviewed. Without discarding data, the team measured error rates below 2% per round in both bases only for the fully defect-removed code, the version with one distance cut to 4. The authors described that result as “approaching, but not beating, the best distance-3 surface code.”
A mid-circuit reset returns a qubit to its starting state in the middle of a computation. The authors wrote that such resets of measure qubits are not available on ibm_miami. The team instead estimated each run’s leakage probability from raw measurement data and discarded runs above a 3% cutoff, a filter known as post-selection.
On the remaining data, the team measured 0.98% per round for the distance-5 code in one basis, against 1.26% for the distance-3 code. The authors did not observe a similar improvement in the other basis. They wrote that leakage post-selection “does not provide a scalable approach to improving QEC,” their abbreviation for quantum error correction.
In an appendix, the team reported that mid-circuit measurements on some qubits added excess noise to other qubits on the device. The authors wrote that IBM classifies ibm_miami as an exploratory processor.
RSA-2048 estimates
In a May 2025 preprint, Craig Gidney of Google Quantum AI estimated that a quantum computer with less than a million noisy, or physical, qubits could factor a 2,048-bit RSA integer in less than a week.
Gidney wrote that he kept the assumptions of his 2019 estimate with Martin Ekerå, including “a square grid of qubits with nearest neighbor connections, a uniform gate error rate of 0.1%.”
The Sydney authors did not discuss estimates of the quantum computing resources needed to break encryption. They wrote that earlier teams had run codes only on high-quality, uniform regions of a device, an approach that “can place unreasonable demands on fabrication yield,” the share of manufactured components that work as designed.
They also wrote that a scalable fix for leakage would need “additional device capabilities, such as the ability to reset into the qubit subspace,” the two energy levels that store information.
