A UK quantum computing company says it found a way to shrink one of the field’s biggest error problems without building a single new chip. Riverlane, a UK company whose Riverlane error correction software already runs in quantum labs, posted a research paper on November 4, 2026, describing a technique it calls DEM augmentation. Quantum Computing Report and Quantum Zeitgeist both covered the result within days, and the underlying paper is public on arXiv. Here’s what the paper actually shows, and what it doesn’t.
What a Detector Error Model Actually Does
Quantum computers make mistakes constantly, so engineers spread one reliable “logical” qubit of information across many physical qubits, a scheme called the surface code. The computer keeps measuring for signs that something slipped, called a syndrome. A separate piece of software, the decoder, reads that syndrome and decides what probably went wrong and how to undo it.
The decoder’s guess is only as good as its map of likely errors, called a Detector Error Model, or DEM. Most DEMs come from theory: physicists list the error types they expect, like a single qubit flipping at random. Real chips misbehave in messier ways. Two neighboring qubits can glitch together, an effect called crosstalk, and theoretical models tend to leave that kind of correlated noise out entirely. Whatever a DEM omits, the decoder simply can’t correct for.
Finding Errors the Theory Left Out
Riverlane’s four researchers, based at the company’s Cambridge, UK office, built a method that goes straight to the raw measurement data a device produces and hunts for error patterns the theoretical model missed. Checking every imaginable combination of errors would take far too long to compute, so the method instead starts from error types already known to exist and pairs them up to generate candidates. It then tests each candidate against how often different detectors actually fire together in the data, keeping only the pairings with a real, measurable correlation.
The important part, according to the paper, is that each newly found error pattern comes with a defined effect on the final answer, so it can be dropped straight into decoders already running in labs today, such as PyMatching, without rebuilding them. Decoding software has quietly become its own competitive front: IonQ, for one, has published its own real-time decoder built to keep pace with incoming error data on ordinary computer chips. Riverlane’s researchers say no earlier data-driven method had combined all three pieces, patterns pulled from real data, defined logical effects, and drop-in compatibility, until now.
Riverlane’s Error Correction Numbers: Simulation vs. a Real Chip
The team tested the method two ways, and the two results aren’t the same kind of evidence.
| Test | Where the Data Came From | Error Suppression Gain | Logical Error Rate |
|---|---|---|---|
| Simulation (code distances 7-15) | Modeled circuit noise plus an added two-qubit crosstalk effect | Λ up 20%, from 1.71 to 2.06 | 75% lower at distance 15 |
| Real hardware (code distances 5-7) | Google Quantum AI’s public surface-code dataset | Λ up 9%, to 2.02 | 14% lower at distance 7 |
Λ (lambda) measures how fast errors shrink as the code grows. A higher Λ means a future fault-tolerant computer would need fewer physical qubits for each reliable logical qubit. Source: arXiv:2611.06605 (Riverlane, Nov. 4, 2026).
The hardware test ran on only about 50,000 measurement samples per circuit, a small dataset by the standards of this kind of research. To stretch it further, the team used a trick called recurrence-edge aggregation, treating data from repeated measurement rounds as if it were additional samples. That matters because real quantum computers can only produce experimental data so quickly, so a method that still works with a thin dataset has a practical edge over one that needs months of machine time to gather enough evidence.
What This Doesn’t Show
This isn’t a demonstration of a more powerful quantum computer. It’s an upgrade to the software that interprets errors after they’ve already happened, and it was tested two different ways that shouldn’t be read as equivalent. The simulated numbers, the ones with the bigger percentage gains, come from a noise model the researchers built themselves, not necessarily how any specific chip actually misbehaves. The hardware numbers come from a dataset Google had already published, not from a fresh experiment on Riverlane’s own machine or a live, real-time decoding run; the team analyzed data that already existed, after the fact. The paper hasn’t been peer-reviewed yet. It was posted directly to the arXiv preprint server, and the researchers’ own claim to being first with this particular combination of techniques is their own assessment, not an independently verified title. The method has also only been tested against one family of decoders, correlated matching decoders such as PyMatching, so whether the gains carry over to other decoding approaches isn’t yet shown.
Why It Matters to You
Nobody will notice this in a quantum product this year, and probably not for several more. The more interesting part is where the improvement came from: a publicly available dataset other researchers had already collected, not a fresh booking on an expensive quantum computer. Other teams are chasing the same underlying problem from the hardware side instead, such as Quantinuum’s redesigned error-correction architecture, which aims to need fewer physical qubits per logical qubit in the first place. If software improvements like Riverlane’s can keep pace with hardware redesigns, the cost of pushing error correction forward doesn’t have to rise in lockstep with the cost of the chips themselves. That’s one of the quieter routes toward quantum computers reliable enough to eventually help with things like drug design or new battery chemistry, where the payoff for an ordinary person would show up as a product on a shelf long before it shows up as anything you’d touch directly.
What is a Detector Error Model?
It’s the map a quantum decoder uses to guess what went wrong inside a quantum computer. The decoder reads signals called syndromes and checks them against the DEM’s list of likely error types to work out a fix. A more accurate DEM means fewer mistakes slip through uncorrected.
Did Riverlane test this on a live quantum computer?
Only partly. The hardware figures come from analyzing a dataset Google had already published from its own superconducting experiments, not a new live run on Riverlane’s own hardware. The simulation figures, which show the larger gains, come from modeled noise rather than a real device.
Sources
- arXiv: Discovering correlated error mechanisms: data-driven augmentation of detector error models improving decoding (Schwartzman-Nowik, Stafford, Varbanov and Camps, Riverlane, submitted Nov. 4, 2026)
- Quantum Computing Report: Riverlane Demonstrates Data-Driven DEM Augmentation to Reduce Surface-Code Logical Errors (Nov. 7, 2026)
- Quantum Zeitgeist: Data-Driven Error Models Lift Quantum Error Correction By 20 Percent (Nov. 9, 2026)