Quantum computers need help spotting mistakes. But getting the right answer is only part of the job. Their error-checking helper must also answer quickly.
A study by Changwon Lee, Tak Hur and Daniel K. Park explores a different design for that helper. Published in npj Quantum Information on October 28, 2026, it combines tests on saved experiment records with computer simulations. It does not demonstrate the new helper running a quantum computer live.
Clues to a mistake
A quantum computer stores information in parts called qubits, or quantum bits. Unwanted changes, called noise, can cause errors.
To protect information, researchers spread it across a group of qubits. Repeated checks produce clues about possible mistakes. A program on an ordinary computer reads those clues and estimates what went wrong. This program is called a decoder.
Think of a detective reading clues. The clues do not come with the answer attached. Unlike a storybook detective, the decoder must make its estimate while the quantum machine keeps working.
Two kinds of speed
The decoder needs to keep up with new clues. Otherwise, unread clues can pile up.
It also needs to deliver each answer soon enough. Some steps in a quantum calculation must wait for that answer. While the machine waits, its qubits remain exposed to noise.
A cafeteria helps explain the difference. It might serve meals steadily, yet each child still waits a long time. Keeping the line moving and shortening each wait are different goals.
What the researchers tested
The team built a decoder using Mamba, a computer model that learns from examples. It keeps and updates a compact record of information instead of making many pair-by-pair comparisons between clues.
On saved records from Google's Sycamore quantum processor, Mamba matched the accuracy of the Transformer decoder tested by the team. A Transformer is another kind of learning model. This was a test of recorded data, not live control.
The researchers also ran simulations—computer-made versions of an error-correction task. They added extra errors to represent the effects of waiting for a decoder. With those assumed extra errors, Mamba performed better across the tested range. Without them, the two models performed similarly.
What still needs testing
Mamba reduces the work needed by one costly part of the decoder. Other parts still take time. Tests of individual parts do not tell us how fast a complete installed system will answer.
The results give researchers a reason to test Mamba further. They do not prove it is ready for a large, reliable quantum computer. The lesson is simple: an error checker needs both good judgment and good timing.
Read the research
Published study and its authors, October 28, 2026. Full accepted paper, with the methods and limits. An earlier version of the paper was posted on October 26, 2025; some results changed before journal publication.