Researchers at QuTech, a quantum research institute run by Delft University of Technology and the research organization TNO, have built a leakage reduction unit into the measurement of superconducting qubits. Leakage is an error in which a qubit leaves the two states that represent 0 and 1. In a paper in PRX Quantum published on September 21, 2026, the team reported that the unit removed 98.4% of the leakage it had deliberately induced in a single qubit, with no added time for a standard two-level readout.
QuTech’s unit removes leakage during the measurement itself, without a separate step, but the team has tested it only in error-correction experiments on five and seven physical qubits.
The qubits are transmons, superconducting circuits that store information in their two lowest energy states. A transmon leaks when it is excited out of those states, most often into the second excited state, the next energy level up.
According to the authors, leakage can persist for several rounds of error correction and spread through two-qubit gates, operations that act on two qubits at once. The errors it causes are correlated in time and space, they wrote.
Error-correcting codes such as the surface code are designed to fix Pauli errors, flips of a qubit’s value, its phase, or both. The team wrote that “a leakage event cannot be mapped to the Pauli framework.”
Many earlier leakage reduction units need a dedicated step in each round of error correction, the authors wrote. They added that the idle time from that step can raise the physical error rate.
How the unit works
The team adapted double drive reset of population, a technique that a Yale-led group published in 2013. In QuTech’s version, one microwave signal, or drive, moves a leaked transmon from the second excited state down to the first.
A second drive goes to the transmon’s readout resonator, the microwave circuit used to measure the qubit, and fills it with photons only when the transmon is in the first excited state. The photons shift the transmon’s frequency enough that the first drive cannot excite it back into the leaked state.
In the preprint version of the paper, which the team posted on arXiv before peer review in November 2025, the authors reported that the combined operation took 500 nanoseconds in the single-qubit calibration. They described that duration as typical for a standard measurement on their processor. The measurement can also register the second excited state as a third outcome. The team uses that outcome in its decoder, the software that infers which errors occurred, to identify which qubit leaked and in which round.
For two-level readout, the team measured an average assignment fidelity of 99.2%, the share of trials in which the readout matched the prepared state.
The error-correction experiments
The researchers ran two small error-correction experiments and decoded the results with neural networks. They did not use post-selection, the practice of discarding runs in which an error is detected.
In a memory experiment, they stored an encoded, or logical, state in a distance-3 repetition code on five physical qubits, three holding data and two checking for errors. This repetition code spreads one bit of information across several qubits and can correct bit flips only. Its distance is the smallest number of physical errors that can change the stored information without being detected.
In a stability experiment, the team used seven physical qubits, four for measurement and three for data, to run a chain of four repeated parity checks, measurements that reveal errors without reading out the stored information. The authors described it as a test of moving encoded information through space, as in lattice surgery, a technique for operating on neighboring logical qubits. The team performed no operation between logical qubits in that experiment.
To stress the unit, the team raised leakage deliberately by adjusting the controlled-Z gates, two-qubit operations that entangle neighboring transmons.
Results
In the preprint, the authors reported that by reading out three levels instead of two, they lowered the logical error rate by 11.1% in the memory experiment at the processor’s baseline leakage. The authors reported that the unit reduced logical errors at every injected leakage level above that baseline.
For the stability experiment, the team fitted an error-suppression factor, γ, a measure of how quickly the chance of a logical error shrinks as rounds of checks are added. Compared with two-level readout without the unit, γ at baseline leakage rose 29% with the unit, 23% with three-level readout alone, and 54% with both, the authors reported. At an intermediate leakage rate of 3.7%, against a baseline of 1.2%, the increases were 61%, 33%, and 90%.
With the unit in place, γ stayed nearly constant as the team injected more leakage, the authors wrote. Without leakage removal, γ declined as the injected leakage grew.
In these experiments, the team used a 700-nanosecond measurement and reported in the preprint average leakage removal of 95% in the memory experiment and 94.2% in the stability experiment. The authors attributed the longer duration to differences in parameters between transmons.
Limits and next steps
In the memory experiment, the team obtained its best results across the full range of injected leakage by combining the unit with three-level readout. “Further research is required to see if this holds at larger code distances and for other codes,” the authors wrote.
They also wrote that the unit could be deployed across larger error-correcting codes in future work. The team had not fully characterized microwave crosstalk, unwanted interference between units operating in parallel, according to the paper.
The team has published scripts to reproduce its analysis, with links to the experimental data, on GitHub. Funders included the EU Quantum Flagship project OpenSuperQPlus100, Intel, the Dutch National Growth Fund, and the Netherlands’ Ministry of Economic Affairs.
