Pith. sign in

REVIEW 4 major objections 5 minor 54 references

Machine learning on raw syndromes can serve as a decoder-agnostic post-selection score for quantum error correction, outperforming syndrome-weight filtering on experimental magic-state distillation data and improving fidelity when combined

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Syndrome-only supervised learning can post-select quantum error correction runs, matching syndrome-weight filtering on simulations and outperforming it on experimental magic-state distillation data.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A useful, modest paper that applies an off-the-shelf binary classifier to syndrome ensembles for post-selection; the experimental magic-state distillation results are the real value, but several headline claims are under-supported. the 4 major comments →

arxiv 2607.19563 v1 pith:Q4DDY7GC submitted 2026-07-21 quant-ph cs.LG

Machine-learned syndrome post-selection for reliable quantum error correction

Tobias Haug , Askery Canabarro , Leandro Aolita This is my paper
classification quant-ph cs.LG
keywords quantum error correctionpost-selectionsyndrome datamachine learninglogical gapmagic-state distillationsurface codeqLDPC code
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Quantum error correction typically uses every run, but some runs are more likely to fail than others. This paper shows that a supervised classifier trained only to tell apart syndromes sampled at low vs high physical error rates produces a reliable 'abort score' for post-selection, without needing decoder-level information such as logical-error labels or logical-gap values. The method is validated in three settings: circuit-level simulations of a bivariate-bicycle qLDPC code, code-capacity surface-code simulations, and experimental magic-state distillation data from a neutral-atom processor. In the experimental data the learned score outperforms syndrome-weight post-selection, and when combined with logical-gap filtering it yields higher output fidelity than logical-gap filtering alone. The authors also identify a post-selection transition in the surface code distinct from the conventional decoding threshold, suggesting that syndrome-distribution learning captures structure beyond simple bit counts.

Core claim

The central claim is that syndrome-only learning is sufficient for effective post-selection: a classifier that outputs the probability a syndrome came from a high-noise ensemble can be used as an abort score, and this score carries more information than the number of flipped syndrome bits. On experimental magic-state distillation data, the learned score alone outperforms syndrome-weight post-selection, and a two-stage filter (machine-learning score followed by logical-gap post-selection) achieves higher fidelity than the logical-gap-only baseline at the same acceptance rate. In surface-code simulations, the classifier output exhibits a universal crossing near p_ML ≈ 0.0866, distinct from the

What carries the argument

The central object is a binary classifier C(s) that maps a syndrome record s — a vector of parity-check outcomes, possibly including multiple rounds of syndrome extraction — to the estimated probability Pr(y=1|s) that s was sampled from a high-physical-error-rate ensemble. It is trained by minimizing cross-entropy on two syndrome ensembles generated at low (p_low) and high (p_high) physical error rates, with labels indicating only the ensemble, not logical success. The output is used as a post-selection score: runs with C(s) above a cutoff are aborted before decoding. This machinery is decoder-agnostic because the training data require no logical-error labels, correction operators, or decode

Load-bearing premise

The classifier is trained on simulated syndromes generated by a calibrated circuit model at rescaled error rates, and then applied to experimental syndrome data; if the simulation's noise correlations do not match the device, the high-noise score may not correspond to logical-failure likelihood in the experiment, invalidating the experimental claim.

What would settle it

Measure the actual logical-error rate of experimental runs as a function of the learned score (using the experiment's decoder to label outcomes) and check whether high-score runs are systematically more likely to fail than low-score runs within the same noise-regime ensemble; if the score's ranking shows no significant correlation with true logical failures on the device, the method's experimental utility is refuted. A simpler test: train on a deliberately mismatched noise model and observe that the score no longer separates logical successes from failures in real data.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Post-selection can happen before decoding: runs with high learned scores are discarded without invoking the decoder, so the classical decoding workload is reduced along with the conditional logical error rate of the accepted runs.
  • In the experimental magic-state distillation data, the learned score alone outperforms syndrome-weight post-selection, and a two-stage ML+logical-gap filter exceeds the logical-gap-only fidelity at a comparable acceptance rate, indicating the learned score captures syndrome structure beyond bit counts.
  • For the surface code under code-capacity noise, the classifier output exhibits a universal crossing at a post-selection threshold p_ML ≈ 0.0866, below the decoding threshold p_th ≈ 0.1037; near p_ML the improvement factor becomes almost independent of code distance for d ≥ 18, suggesting the method remains effective as codes scale.
  • The training procedure requires no logical-error labels, correction operators, or code-specific likelihood calculations, so syndrome ensembles can be generated efficiently from Clifford circuit simulation or from calibration data at deliberately varied error rates.
  • The method is modular: it can be inserted before any decoder, and combining it with decoder-based soft information (logical gap) gives further gains, so it can serve as a low-cost pre-filter in a layered post-selection pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the method only needs ensemble labels, it could be trained entirely on experimental data by deliberately varying the device's error rate, avoiding the simulation-to-experiment transfer assumption altogether; the paper trains on simulation in the experimental application, but direct experimental training is a natural next step.
  • The existence of a post-selection transition p_ML distinct from p_th suggests that syndrome distinguishability is a separate statistical property of the code; one might predict p_ML from the overlap of low- and high-noise syndrome distributions or from Tanner-graph parameters without training a classifier.
  • The success of combining the learned score with logical-gap filtering hints at a general pipeline: use the cheap score to discard the majority of runs, then apply expensive soft-information decoding only to the ambiguous tail, reducing decoding cost in fault-tolerant architectures where soft information is expensive.
  • The method's flexibility across memory and distillation settings suggests it may apply to other probabilistic protocols, such as logical gate preparation or repeat-until-success circuits, where the 'high-noise ensemble' could be defined operationally rather than by a single physical error rate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper introduces a machine-learning post-selection method for quantum error correction. The key idea is to train a binary classifier on syndromes from low- and high-physical-error-rate simulations, and then use the classifier's class-1 probability C(s) as an abort score: runs with C(s) above a cutoff are discarded before decoding. The classifier is trained without logical-error labels or decoder outputs. The paper validates the approach in three settings: (i) experimental logical magic-state distillation data from QuEra, where the authors report that ML post-selection outperforms syndrome-weight filtering and, combined with logical-gap post-selection, improves output fidelity beyond logical-gap alone; (ii) circuit-level simulations of the [[144,12,12]] Gross bivariate-bicycle code, where ML and syndrome-weight give comparable improvements; and (iii) code-capacity simulations of the rotated surface code, where the learned score shows a finite-size crossing at p_ML≈0.0866, distinct from the decoding threshold p_th≈0.1037.

Significance. The proposed method is simple, decoder-agnostic, and potentially useful as a pre-decoding filter for near-term QEC experiments. Its main strengths are that training does not require logical labels or expensive likelihood computations, and the numerical benchmarks use standard decoders (BP-OSD, MWPM) and a real experimental dataset. The surface-code scaling collapse is a falsifiable prediction and the Gross-code application demonstrates applicability beyond surface-code graphs. However, the experimental claims — specifically that ML outperforms syndrome-weight and that ML+LG beats LG — are not yet supported with adequate statistical rigor or a direct validation that the learned score tracks logical failure in the device. These gaps are central to the paper's headline result.

major comments (4)
  1. [III, Fig. 3] The central experimental claim that ML outperforms SW and that ML+LG exceeds LG alone is presented without confidence intervals for ML and SW (Fig. 3a,b). Only Fig. 3c shows 68% CIs, and only for ML+LG and LG. The differences between ML and SW, especially at R≈0.01, appear comparable to or smaller than the quoted uncertainties. To support the claim 'ML achieves higher fidelity than SW for all R', the authors should provide confidence intervals for all four curves (e.g., bootstrap over shots) and a statistical test at representative acceptance rates.
  2. [III, Eq. (5) and II.B] Sim-to-experiment transfer is load-bearing. The classifier is trained exclusively on simulated syndromes obtained by rescaling calibrated per-gate error rates by E={0.01,...,0.15} (low) and {1.75,...,2.5} (high), i.e., a one-parameter global rescaling. The training labels (Eq. 5) identify the noise ensemble, not logical failure. For the experimental fidelity gains to follow, C(s) must be monotonically related to the actual logical-failure probability under the device's noise. The paper does not provide a direct check of this monotonicity on experimental data. Since the experimental dataset includes logical tomography, the authors can compute C(s) for each run and compare binned C(s) against the empirical failure rate, or report the ROC-AUC of C(s) for actual logical failure. Without such a diagnostic, the experimental superiority claim is not established.
  3. [V, SM A] The post-selection transition p_ML≈0.0866(1) distinct from p_th≈0.1037(2) is obtained from a three-parameter polynomial fit (A1) with free p_ML and ν, but the paper does not report ν, the fit residuals, or a robustness check with different training p values. The collapse in Fig. 6 is only visual. Since this transition is advertised in the abstract and Sec. V, report the fitted ν with uncertainty, a goodness-of-fit measure, and demonstrate that p_ML is stable when the low/high training ensembles are varied. If this is not a central claim, it should be de-emphasized.
  4. [III, Fig. 3c] In ML+LG, the authors state that 'we use ML to discard 80% of the data' before applying LG. The choice of 80% is not justified and appears to be an operating point chosen post hoc. If the threshold is optimized on the same experimental data used to report the final fidelity, the ML+LG improvement over LG could be optimistically biased. Specify how the 80% discard fraction was selected (e.g., on a validation set) and show the sensitivity of the ML+LG curve to this fraction.
minor comments (5)
  1. [II.B] The phrase 'our decoder can be trained' should be 'our classifier can be trained' (or 'model').
  2. [III, Fig. 3 caption] Clarify whether the 68% intervals are over the posterior of the fidelity or over bootstrap resamplings, and why they are shown only for LG and ML+LG.
  3. [SM B] The statement that the best internal cross-validation score reached 1.0 should be replaced with a more informative metric (e.g., ROC-AUC, precision-recall) since a perfect CV score is surprising for overlapping syndrome distributions and may indicate leakage.
  4. [IV, Fig. 4 caption] The conversion of logical error rates from 12 cycles to per-cycle values (p_L = 1 - (1-p_L^{(12)})^{1/12}) should be defined in the main text, not just the caption.
  5. [References] Refs. [34,35] are scikit-learn user guide URLs; better to cite the scikit-learn library or a standard textbook. The paper would also benefit from a data/code availability statement.

Circularity Check

0 steps flagged

No significant circularity: all claims are evaluated on held-out or experimental data, and the training labels are ensemble membership rather than the target post-selection outcome.

full rationale

The paper's derivation chain is not circular. The classifier is trained only to distinguish syndromes drawn from low- and high-noise ensembles (Eq. 5), and the score C(s)=Pr(y=1|s) is then applied to held-out simulated data or to independent experimental data; the reported logical-error rates and fidelities are measured outcomes, not re-used training targets. The surface-code 'post-selection transition' p_ML is a fitted characterization of the trained classifier's crossing, explicitly described as 'a property of distinguishability between two syndrome ensembles as represented by the learned classifier,' and the distinct decoding threshold p_th is independently fitted from logical-error data. Neither quantity is an input that forces the claimed prediction. The experimental application trains on Stim simulations calibrated to the QuEra device, but the subsequent fidelity comparison on experimental data is an external evaluation; the sim-to-experiment transferability assumption is a correctness/validity concern, not a definitional circularity. No load-bearing self-citation is present: Ref. [33] includes an author but is a background tutorial, not the basis of the method. The scaling ansatz in SM A is an explicit fitting form, not a hidden premise, and the syndrome-weight transition is credited to prior work. Therefore no step reduces to its own inputs, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. The main free choices are the training ensemble noise rates and the experimental rescaling factors, both hand-selected. The transition parameters p_ML and ν are fitted outputs used to support the transition claim. The load-bearing assumptions are the faithfulness of the noise models and the transferability of the learned score from training ensembles to the target data.

free parameters (3)
  • Training ensemble physical error rates p_low, p_high = varies by application (e.g., 0.05/0.06 and 0.14/0.15 for surface code)
    Chosen by hand to represent easy and hard noise regimes; the classifier's behavior and the resulting post-selection curve depend on this choice.
  • Experimental noise-rescaling factors E = E_low = {0.01, 0.05, 0.10, 0.15}, E_high = {1.75, 2.00, 2.25, 2.50}
    Used to generate simulated training data for the QuEra experiment; these factors are calibrated guesses, not derived from first principles.
  • Scaling-fit parameters p_ML, ν = p_ML ≈ 0.0866(1), ν from collapse fit
    Obtained by fitting the classifier output C(p,d) with a polynomial; the fit is used to claim a distinct post-selection transition.
axioms (4)
  • standard math Gottesman-Knill theorem: Clifford circuits can be simulated efficiently.
    Used in Sec. II B to argue that training data can be generated efficiently for stabilizer codes; the method relies on this for scalable data generation.
  • domain assumption The chosen noise models (circuit-level depolarizing for the Gross code, code-capacity bit-flip for the surface code, calibrated Stim circuit for the experiment) are faithful representations of the relevant physics.
    The method's performance on simulations and the transfer to experiment depend on these models producing syndrome distributions that reflect real noise; any mismatch weakens the claims.
  • domain assumption A classifier trained to separate low- and high-noise syndrome ensembles yields a score whose ranking transfers to the target data (experimental or other noise rates).
    The entire method rests on this ordering assumption; the paper explicitly states perfect calibration is not required, but the ranking must transfer, which is an unproven assumption, especially for sim-to-experiment transfer.
  • domain assumption Decoding is performed by an unchanged decoder (MWPM or BP-OSD) and the logical error rate is computed on the accepted shots.
    Standard practice; the comparison of post-selection methods assumes the decoder is fixed.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine-learned syndrome post-selection for reliable quantum error correction." pith.science (2026). https://pith.science/paper/Q4DDY7GC

@misc{pith2026260719563,
  author       = {Pith},
  title        = {Pith review of: Machine-learned syndrome post-selection for reliable quantum error correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4DDY7GC}},
  note         = {Machine review of arXiv:2607.19563}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Quantum error correction can be enhanced by post-selecting out runs that are likely to produce a logical failure, but the most accurate measures for that require costly decoder-level information. We introduce a practical, decoder-agnostic post-selection method that learns directly from syndrome data. The method trains a supervised classifier to distinguish between syndromes from low- and high-noise regimes, and then uses the classifier's output as an abort score for new runs, without requiring logical-error labels, correction operators, or code-specific likelihood calculations. We validate the approach in three complementary settings: circuit-level simulations of the Gross bivariate-bicycle code, code-capacity simulations of the surface code, and experimental logical magic-state distillation data from the QuEra neutral-atom processor. In the Gross and surface codes, learned syndrome post-selection reduces the conditional logical error rate at a fixed acceptance rate, with performance comparable to syndrome-weight filtering. For the surface code, the learned classifier reveals a post-selection transition distinct from the conventional decoding threshold. In the experimental data, the machine-learning score outperforms syndrome-weight post-selection and, when combined with logical-gap filtering, improves the output fidelity beyond using the logical gap alone. These results show that syndrome-only learning provides a scalable and hardware-compatible route to improving the reliability of quantum error correction.

Figures

Figures reproduced from arXiv: 2607.19563 by Askery Canabarro, Leandro Aolita, Tobias Haug.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: compares four post-selection scores: logical gap (LG), machine-learning (ML), syndrome-weight (SW), and machine-learning post-selection followed by logical gap (ML+LG). We also show the fidelity FSI ≈ 0.951 for direct state-injection (SI) into the QEC code without distillation and post-selection as reported in Ref. [15]. We show fidelity F against post-selection rate R in [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Surface-code classifier threshold. The machine-learning predictor [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. As reference, we show the (non-post-selected) logical error rate [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 1 canonical work pages

  1. [1]

    B. M. Terhal, Quantum error correction for quantum memories, Reviews of Modern Physics87, 307 (2015)

    2015

  2. [2]

    H.Chen, T.J

    E. H.Chen, T.J. Yoder, Y. Kim, N.Sundaresan, S.Srini- vasan, M.Li, A.D.Córcoles, A.W.Cross,andM.Takita, Calibrated decoders for experimental quantum error cor- rection, Physical Review Letters128, 110504 (2022)

    2022

  3. [3]

    Knill, Quantum computing with realistically noisy de- vices, Nature434, 39 (2005)

    E. Knill, Quantum computing with realistically noisy de- vices, Nature434, 39 (2005)

    2005

  4. [4]

    Aliferis, D

    P. Aliferis, D. Gottesman, and J. Preskill, Accuracy threshold for postselected quantum computation, arXiv preprint quant-ph/0703264 (2007)

    Pith/arXiv arXiv 2007

  5. [5]

    Bravyi and A

    S. Bravyi and A. Kitaev, Universal quantum computa- tion with ideal clifford gates and noisy ancillas, Physical Review A—Atomic, Molecular, and Optical Physics71, 022316 (2005)

    2005

  6. [6]

    Bravyi and J

    S. Bravyi and J. Haah, Magic-state distillation with low overhead, Physical Review A—Atomic, Molecular, and Optical Physics86, 052329 (2012)

    2012

  7. [7]

    Bombín, M

    H. Bombín, M. Pant, S. Roberts, and K. I. Seetharam, Fault-tolerant postselection for low-overhead magic state preparation, PRX Quantum5, 010302 (2024)

    2024

  8. [8]

    Gidney, N

    C. Gidney, N. Shutty, and C. Jones, Magic state culti- vation: growing t states as cheap as cnot gates, arXiv preprint arXiv:2409.17595 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Akahoshi, K

    Y. Akahoshi, K. Maruyama, H. Oshima, S. Sato, and K. Fujii, Partially fault-tolerant quantum computing ar- chitecture with error-corrected clifford gates and space- time efficient analog rotations, PRX quantum5, 010337 (2024)

    2024

  10. [10]

    J. W. Staples, W. Fu, and J. D. Thompson, Scal- able postselection of quantum resources, arXiv preprint arXiv:2603.08697 (2026)

    arXiv 2026

  11. [11]

    Meister, C

    N. Meister, C. A. Pattison, and J. Preskill, Efficient soft-output decoders for the surface code, arXiv preprint arXiv:2405.07433 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    H. Chen, D. Xu, G. M. Sommers, D. A. Huse, J. D. Thompson, and S. Gopalakrishnan, Scalable accuracy gains from postselection in quantum error correcting codes, arXiv preprint arXiv:2510.05222 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    S. C. Smith, B. J. Brown, and S. D. Bartlett, Mitigating errors in logical qubits, Communications Physics7, 386 (2024)

    2024

  14. [14]

    Gidney, M

    C. Gidney, M. Newman, P. Brooks, and C. Jones, Yoked surface codes, Nature Communications16, 4498 (2025)

    2025

  15. [15]

    Sales Rodriguez, J

    P. Sales Rodriguez, J. M. Robinson, P. N. Jepsen, Z. He, C. Duckering, C. Zhao, K.-H. Wu, J. Campo, K. Bagnall, M. Kwon,et al., Experimental demonstration of logical magic state distillation, Nature645, 620 (2025)

    2025

  16. [16]

    S.-H. Lee, L. H. English, and S. D. Bartlett, Efficient post-selection for general quantum ldpc codes, npj Quan- tum Information (2026)

    2026

  17. [17]

    Hsieh and F

    M.-H. Hsieh and F. Le Gall, Np-hardness of decoding quantum error-correction codes, Physical Review A83, 052331 (2011)

    2011

  18. [18]

    P.Iyerand D.Poulin, Hardnessofdecoding quantum sta- bilizer codes, IEEE Transactions on Information Theory 61, 5209 (2015)

    2015

  19. [19]

    Li, A magic state’s fidelity can be superior to the operations that created it, New Journal of Physics17, 023037 (2015)

    Y. Li, A magic state’s fidelity can be superior to the operations that created it, New Journal of Physics17, 023037 (2015)

    2015

  20. [20]

    Singh, A

    S. Singh, A. S. Darmawan, B. J. Brown, and S. Puri, High-fidelity magic-state preparation with a biased-noise architecture, Physical Review A105, 052410 (2022)

    2022

  21. [21]

    L. H. English, D. J. Williamson, and S. D. Bartlett, Thresholds for postselected quantum error correction from statistical mechanics, Physical Review Letters135, 120603 (2025)

    2025

  22. [22]

    Torlaiand R.G.Melko,Neuraldecoderfortopological codes, Physical review letters119, 030501 (2017)

    G. Torlaiand R.G.Melko,Neuraldecoderfortopological codes, Physical review letters119, 030501 (2017)

    2017

  23. [23]

    Krastanov and L

    S. Krastanov and L. Jiang, Deep neural network proba- bilistic decoder for stabilizer codes, Scientific reports7, 11003 (2017)

    2017

  24. [24]

    Baireuther, M

    P. Baireuther, M. D. Caio, B. Criger, C. W. Beenakker, and T. E. O’Brien, Neural network decoder for topolog- ical color codes with circuit level noise, New Journal of Physics21, 013003 (2019)

    2019

  25. [25]

    Meinerz, C.-Y

    K. Meinerz, C.-Y. Park, and S. Trebst, Scalable neural decoder for topological surface codes, Physical Review Letters128, 080505 (2022)

    2022

  26. [26]

    Bausch, A

    J. Bausch, A. W. Senior, F. J. H. Heras, T. Edlich, A. Davies, M. Newman, C. Jones, K. Satzinger, M. Y. Niu, S. Blackwell,et al., Learning high-accuracy error de- coding for quantum processors, Nature635, 834 (2024)

    2024

  27. [27]

    Lange, P

    M. Lange, P. Havström, B. Srivastava, I. Bengtsson, V. Bergentall, K. Hammar, O. Heuts, E. van Nieuwen- burg, and M. Granath, Data-driven decoding of quantum error correcting codes using graph neural networks, Phys- ical Review Research7, 023181 (2025)

    2025

  28. [28]

    A. S. Maan and A. Paler, Machine learning message- passing for the scalable decoding of qldpc codes, npj Quantum Information11, 10.1038/s41534-025-01033-w (2025)

    work page doi:10.1038/s41534-025-01033-w 2025

  29. [29]

    Gicev, L

    S. Gicev, L. C. L. Hollenberg, and M. Usman, Fully con- volutional 3d neural network decoders for surface codes with syndrome circuit noise, arXiv preprint (2025), arXiv:2506.16113 [quant-ph]

    Pith/arXiv arXiv 2025

  30. [30]

    Dentelski, Neural network decoder confidence as a learned proxy for the logical gap, arXiv:2606.08758 (2026)

    D. Dentelski, Neural network decoder confidence as a learned proxy for the logical gap, arXiv:2606.08758 (2026)

    Pith/arXiv arXiv 2026

  31. [31]

    A. Gu, J. Ataides, M. D. Lukin, and S. F. Yelin, Scal- able neural decoders for practical fault-tolerant quantum computation, arXiv preprint arXiv:2604.08358 (2026)

    Pith/arXiv arXiv 2026

  32. [32]

    D. Seip, L. Colmenarez, M. Schmitt, and M. Müller, Machine Learning Optimal Quantum Error Correction Thresholds, arXiv:2606.22194v1 (2026)

    Pith/arXiv arXiv 2026

  33. [33]

    D. J. Spencer, S. P. Jain, A. Tanggara, Z. Sun, T. Haug, D. Khu, and K. Bharti, Quantum error correction and fault tolerance: A comprehensive tutorial, arXiv preprint arXiv:2605.29137 (2026)

    Pith/arXiv arXiv 2026

  34. [34]

    scikit-learn developers, Tuning the decision threshold for class prediction,https://scikit-learn.org/stable/ modules/classification_threshold.html(2026), scikit-learn User Guide, accessed June 29, 2026

    2026

  35. [35]

    scikit-learn developers, Probability calibration,https:// scikit-learn.org/stable/modules/calibration.html (2026), scikit-learn User Guide, accessed June 29, 2026

    2026

  36. [36]

    Gottesman, The heisenberg representation of quan- tum computers, arXiv preprint quant-ph/9807006 (1998)

    D. Gottesman, The heisenberg representation of quan- tum computers, arXiv preprint quant-ph/9807006 (1998)

    Pith/arXiv arXiv 1998

  37. [37]

    C.Gidney,Stim: afaststabilizercircuitsimulator,Quan- tum5, 497 (2021)

    2021

  38. [38]

    R. S. Olson, N. Bartley, R. J. Urbanowicz, and J. H. 9 Moore, Evaluation of a tree-based pipeline optimiza- tion tool for automating data science, inProceedings of the Genetic and Evolutionary Computation Conference 2016, GECCO ’16 (ACM, New York, NY, USA, 2016) pp. 485–492

    2016

  39. [39]

    Ribeiro, A

    P. Ribeiro, A. Saini, J. Moran, N. Matsumoto, H. Choi, M. Hernandez, and J. H. Moore, Tpot2: A new graph- based implementation of the tree-based pipeline opti- mization tool for automated machine learning, inGenetic programming theory and practice XX(Springer, 2024) pp. 1–17

    2024

  40. [40]

    experi- mental demonstration of logical magic state distillation

    QuEra Computing and collaborators, Data for “experi- mental demonstration of logical magic state distillation”, 10.1038/s41586-025-09367-3 (2025)

    doi:10.1038/s41586-025-09367-3 2025

  41. [41]

    Bravyi, A

    S. Bravyi, A. W. Cross, J. M. Gambetta, D. Maslov, P. Rall, and T. J. Yoder, High-threshold and low- overhead fault-tolerant quantum memory, Nature627, 778 (2024)

    2024

  42. [42]

    Rosenfeld, C

    E. Rosenfeld, C. Gidney, G. Roberts, A. Morvan, N. Lacroix, D. Kafri, J. Marshall, M. Li, V. Sivak, D. Abanin,et al., Magic state cultivation on a superconducting quantum processor, arXiv preprint arXiv:2512.13908 (2025)

    arXiv 2025

  43. [43]

    R. S. Gupta, N. Sundaresan, T. Alexander, C. J. Wood, S. T. Merkel, M. B. Healy, M. Hillenbrand, T. Jochym- O’Connor, J. R. Wootton, T. J. Yoder,et al., Encoding a magic state with beyond break-even fidelity, Nature625, 259 (2024)

    2024

  44. [44]

    Bluvstein, S

    D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter,et al., Logical quantum processor based on reconfigurable atom arrays, Nature626, 58 (2024)

    2024

  45. [45]

    Google Quantum AI and Collaborators, Quantum error correction below the surface code threshold, Nature638, 920 (2025)

    2025

  46. [46]

    F. H. Watson and S. D. Barrett, Logical error rate scaling of the toric code, New Journal of Physics16, 093045 (2014). 10 Supplemental Information In the Supplemental Material, we provide additional analyses and methodological details. CONTENTS A. Post-selection and error correction threshold 10 B. Machine-learning methodology 10

    2014

  47. [47]

    Dataset construction 10

  48. [48]

    Automated pipeline search 11

  49. [49]

    Probabilistic inference 12

  50. [50]

    Post-selection rule 12 SM A: Post-selection and error correction threshold In this section, we show for the surface code the explicit collapse of machine-learning predictorConto a single curve by appropriate rescaling. To getpML, we fit the data with the polynomial [46] C=A+Bx+Dx 2 withx= (p−p ML)d1/ν.(A1) For the fitting, we remove data with very small a...

  51. [51]

    Dataset construction As an example, we describe the dataset generation for our Gross bivariate-bicycle code application. For the Gross bivariate-bicycle code, each input sample was constructed from the syndrome information associated with a single 11 0.050 0.075 0.100 0.125 0.150 p 0.0 0.2 0.4 pL d = 12 d = 18 d = 24 d = 30 b FIG. 7. As reference, we show...

  52. [52]

    Automated pipeline search The classifier was selected using the Tree-based Pipeline Optimization Tool, TPOT, version0.12.1[38, 39]. TPOT formulates model selection and hyperparameter optimization as a genetic-programming problem in which candidate machine-learning pipelines are generated, evaluated through cross-validation, and iteratively evolved accordi...

  53. [53]

    We define the machine-learning score as the inferred probability of class1, C(s) =Pr(y= 1|s),(B4) where class1denotes the high-error-rate ensemble

    Probabilistic inference For each new syndrome records, the selected classifier returns the estimated probabilities of the two training classes. We define the machine-learning score as the inferred probability of class1, C(s) =Pr(y= 1|s),(B4) where class1denotes the high-error-rate ensemble. Larger values ofC(s)indicate that the observed syndrome is more c...

  54. [54]

    Syndromes assigned a sufficiently large probability of belonging to class1are therefore discarded before decoding

    Post-selection rule For a chosen cutoffC0 ∈[0,1], a QEC run with syndromesis accepted when C(s)< C0,(B5) and rejected otherwise. Syndromes assigned a sufficiently large probability of belonging to class1are therefore discarded before decoding. By varyingC 0, we control the trade-off between the fractionRof accepted runs and the conditional logical error r...

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.