A quantum error-correction experiment or simulation produces a great many independent shots, and most of them fail. A preprint posted to arXiv at the end of September proposes an uncomfortable but useful idea: if you can cheaply predict which shots will fail, throw them away, and the rate at which your good answers go wrong collapses. On one family of codes, discarding 10% of the shots cut the logical error rate by a factor of roughly 3,700.
The mechanism is post-selection combined with a graph neural network that predicts decoding failure from error syndromes alone, without running the decoder at all. That last part is what makes it more than a statistical trick: the decoder is the expensive component, so a predictor that skips it is architecturally interesting. The measured figures come from classical simulation only, no hardware was involved, and every headline number is conditional on keeping 90% of shots — which is a real cost in wall-clock time and has to be stated alongside them.
Key takeaways:
- The GNN predicts decoder failure from syndromes alone, without running the decoder. For high-rate qLDPC codes it also draws on existing belief-propagation posteriors.
- Retaining 90% of shots: about a 3,700× logical error reduction on a
[[72,12,6]]bivariate bicycle code, and about 740× on[[144,12,12]]. - On a distance-5 surface code at physical noise
p = 0.003, the approach matches the leading complementary-gap method’s suppression without running the decoder. - All numbers are classical simulation. No quantum hardware was run.
- Post-selection helps far more on high-rate qLDPC codes than on surface codes. That asymmetry is the real finding.
- It is a preprint, not peer reviewed, and the physical error rate behind the two bivariate-bicycle figures is not stated in the abstract.
Why post-selection is not cheating
It is reasonable to be suspicious. Discarding failed attempts looks like reporting only the successes. Three things make it legitimate here.
First, the discard decision is made before the expensive verification, using only information available at decode time. You are not filtering on whether the answer turned out to be right — you are filtering on a cheap predictor. If the predictor were perfect, the error rate would be zero, which is exactly the point.
Second, the discard is not free, and the paper does not pretend otherwise. Keeping 90% of shots means discarding 10%, which means you need roughly 10% more shots to get the same number of accepted results. The wall-clock cost is real and proportional to the discard rate.
Third, this is a standard technique with a long history in experimental physics. Cut-based methods that select events on one detector and measure another are routine. The move here is doing the selection with a learned predictor instead of a threshold.
So the honest framing is: this trades compute for error rate, and the interesting question is the exchange rate.
Why syndromes-only is the architectural claim
A syndrome is the small piece of information a code emits about which qubits appear to have failed. It is cheap to obtain — it comes out of the stabiliser measurement you already did. A decoder is the expensive classical computation that turns that syndrome into the most likely error chain, and then into a correction.
The paper’s central claim is that a graph neural network reading syndromes alone can predict whether the decoder will succeed, well enough that skipping the decoder for the predicted failures costs little accuracy. At distance-5 surface code and p = 0.003, that is stated as comparable logical error suppression to the leading complementary-gap method, without running the decoder.
Why that matters in practice: decoding has to keep up with the syndrome stream in real time. If a cheap network can tell you which syndromes are hopeless, the expensive decoder only runs on the ones worth decoding. That is a throughput argument, and throughput is usually what actually limits error correction.
The numbers, with their conditions attached
| Code | Retained shots | Reported logical error reduction |
|---|---|---|
[[72,12,6]] bivariate bicycle | 90% | ≈ 3,700× |
[[144,12,12]] bivariate bicycle | 90% | ≈ 740× |
| Distance-5 surface code | — | Comparable to the complementary-gap method’s suppression, at p = 0.003 |
Three observations worth making about the shape of these numbers rather than their magnitude.
They are asymptotic, not constants. A factor of roughly 3,700 is a number at a particular code, a particular physical error rate, and a particular operating point. Move the operating point and it changes. It should be quoted as “reported at their chosen operating point”, never as a general multiplier.
They shrink with code size. 3,700× on the smaller code, 740× on the larger one. That trend is exactly what the mechanism predicts: the larger the code, the harder it is for a syndrome to look like a typical failure, so a cheap predictor does worse. The method degrades gracefully rather than falling off a cliff, but it degrades.
The 90% is doing real work. Every one of these figures is conditional on discarding 10% of shots. Read as “we reduced the error rate 3,700-fold” without that clause, it is misleading by a factor of ten on the compute budget.
Why the qLDPC result is the interesting one
Post-selection helps dramatically more on high-rate qLDPC codes, including bivariate bicycle codes, than it does on surface codes. The paper reports the distance-5 surface code case as comparable to a strong existing decoder method, which is a respectable result but not a step change.
That asymmetry matters because high-rate qLDPC codes are the family most often proposed as a route past the overhead problem. Every proposal for beating the threshold at useful scales leans on codes with much better rates than the surface code. If a technique that needs only cheap, pre-decoder information works disproportionately well on exactly those codes, it is aimed at the part of the field where it would count.
The mechanism is plausible: higher-rate codes have fewer logical qubits per physical qubit, so a given physical error produces more distinct, recognisable failure patterns, which is precisely what a syndrome-reading network can learn to spot.
What is not established
- No hardware. Every figure is a classical simulation result. Writing that this reduces error rates “on a quantum computer” would be false.
- Not peer reviewed. It is a preprint. The comparison against the complementary-gap method is the authors’ own and is asserted for one specific configuration.
- The discard cost is unquantified here. Discarding 10% of shots costs roughly 10% more shots, but the wall-clock and energy costs of running the GNN on every shot are not folded into the comparison.
- Noise model is uniform depolarising circuit-level noise. Coherent error and leakage are not covered.
- Codes are code memories. Logical qubits, not full fault-tolerant logical gate sequences. Every result shown here is a step short of what a fault-tolerant machine needs.
- Injectivity is assumed, not proved. Post-selection schemes need to guarantee that discarding based on syndromes cannot systematically discard the state you wanted. The paper does not claim otherwise, but a reader should not infer it.
The idea worth stealing
Strip the quantum specifics out and this is a general engineering pattern: when the verification step is expensive relative to the prediction step, learn to predict verification failure and skip the work.
That shows up everywhere. Run a cheap classifier before the expensive reprocessing. Predict which builds will fail before provisioning the full test environment. Screen incoming requests with a small model before invoking the large one — which is the same idea as the small decision models appearing inside agent architectures, applied at the verification boundary instead of the routing boundary.
The exchange rate is the interesting quantity, and the useful habit is asking for it explicitly. A method that improves a metric by discarding 10% of your work is a different engineering decision from one that improves it for free, and the paper is honest enough to attach the number.
A note on the date
Version 1 of this preprint is timestamped 30 September 2026, 18:04 UTC, and appears in arXiv’s 2 October listing because of the standard announce lag. If you have seen it dated 2 October, the submission date is the 30th.
References
- Carty, Noszko, Roffe and Bondesan, arXiv:2610.00504 [quant-ph], v1 submitted 30 September 2026
- arXiv:2610.00548 — a companion preprint from the same announcement window, on classical sampling of noisy circuits
- arXiv quant-ph recent submissions
Related Tools
Try these companion tools on DevStackIO Tools:
- statistics calculator — compute mean, variance, and standard deviation for quantum error analysis.
