A preprint submitted to arXiv (arXiv:2610.00504, quant-ph) by Conor Carty, Tamas Noszko, Joschka Roffe, and Roberto Bondesan (Imperial College London / University of Edinburgh) applies a cheap, single-pass neural "pre-screen" to quantum error correction. The headline number: with 90% of measurement results retained, the logical error rate is suppressed by a factor of 3,716.6 on one BB code configuration.
Where it sits in the stack
Quantum error correction encodes one logical qubit into many physical qubits. Decoding infers which physical qubits flipped from measured syndromes and issues corrections — and decoders themselves sometimes fail. The authors train a model that looks only at the syndrome (plus an existing decoder posterior for qLDPC codes) and predicts whether the decoder will succeed for that shot, outputting a confidence score. Shots below threshold are discarded before decoding.
The model produces no corrections and does not replace the decoder; it only decides which shots are worth decoding. It uses:
- Surface codes: a Message Passing Neural Network (MPNN) on the detector graph, requiring only the syndrome; filtering happens before MWPM decoding.
- BB qLDPC codes (e.g. [[72,12,6]], [[144,12,12]]): a factor-graph MPNN that additionally consumes belief-propagation posteriors; filtering happens before LSD decoding.
- Surface codes: detector graph, 6 features per detector, hidden dim 64.
- BB codes: bipartite graph over detectors and fault mechanisms (a fault can trigger 3+ detectors), 6 features each — 3 static (prior, prior LLR, degree) and 3 per-shot (BP posterior LLR, posterior-prior difference, BP hard decision); hidden dim 32.
- Global mean-pooling and re-broadcast after each message-passing round improves accuracy.
- The BB model avoids enumerating 2^k joint logical outcomes: it pools per logical observable, outputs k marginal success probabilities (k=12), and sums their logs as the acceptance score.
- Rare-event sampling: fewer failures at low error rates make training sets extremely imbalanced; reweighting and importance sampling mitigate but do not solve it.
- Decoder dependence: labels come from a specific decoder; changing decoders requires retraining. For qLDPC, the decoder must be belief-propagation-based.
- Narrow scope: three rotated surface codes, two BB codes, memory experiments only, uniform depolarizing noise, physical error rates of 0.001 and 0.003. Distillation/cultivation workloads were not tested.
- Future work targets microsecond-scale (or lower) inference, citing low-latency FPGA GNN work from high-energy physics. Code: github.com/conorcarty/MPNN4QEC.
- Carty, Noszko, Roffe, Bondesan, *Graph Neural Post-selection for Quantum Error Correction*, arXiv:2610.00504, quant-ph, 19 pages, 5 figures. Preliminary, not peer-reviewed.
- Points flagged in the paper as unstable (<5 accepted failures) include the highest suppression figures; all limitations above follow the authors' own statements.
Both models run a single GPU forward pass per shot — no repeated decoder calls, which is the dividing line from prior work.
The headline numbers
For the BB code [[72,12,6]] at physical error rate 0.001:
| Rejection rate | Acceptance | Logical error rate | Suppression | |---|---|---|---| | 1% | 99% | 1.24e-4 → 6.41e-7 | not listed | | 10% | 90% | ~3.3e-8 | 3,716.6x |
At 10% rejection, 18 of ~74,290 total failures were accepted — 99.976% of failures were filtered out by the pre-screen.
For [[144,12,12]] at physical error rate 0.003: logical error rate dropped from 6.02e-4 to 8.15e-7 (~738.2x).
Important caveats: the 3,716.6x figure carries a 95% confidence interval of −1,365.4 to +2,158.6, and the point is flagged as unstable (fewer than 5 accepted failures), meaning it should not be read as a robust estimate. The 8,495.0x figure at 20% rejection has an even wider interval (−4,379.7 to +9,041.7).
Comparison with existing methods
Against detector density, second-order Q_LLR expansion, and the complementary-gap method, the neural approach does not always win. Surface-code suppression at 10% rejection:
| Surface code | Neural post-selection | Complementary gap | |---|---|---| | d=5 | 96.5x | 104.2x | | d=7 | 70.5x | 463.3x | | d=9 | 20.5x | 2,096.1x |
The gap method runs the decoder twice (the second time under a complementary-logical-class constraint) and compares likelihoods — more informative at large code distance, but the neural method is ~2.8–3.3x faster end-to-end (e.g. d=9: 48.71 µs vs 138.11 µs), though hardware settings differ (A100 + CPU thread vs CPU thread only). The paper's precise claim: best among methods not requiring multiple decoder calls.
For BB codes, FMPNN takes ~1.09 ms per shot vs ~0.97 ms for Q_LLR, with a clear advantage only over multi-decode methods like 3R-LEC (~55.64 ms).
Design highlights
Acknowledged limitations
Context
Post-selection is not new (3R-LEC already uses it at the cost of multiple decoder calls). The contribution is making it affordable with a single forward pass. Structurally, it mirrors a broader trend of placing a cheap front-end decision model before an expensive computation — analogous to small routing/decision models unbundled from larger inference models. The 3,700x figure is best read as a real but probabilistic filtering gain: trading ~10% of results for an order-of-magnitude error reduction. The number to watch is whether the neural method's deficit vs complementary gap at large code distances (20.5x vs 2,096.1x) widens further.