Pith. sign in

REVIEW 4 major objections 6 minor 28 references

A Symmetry-Integrated Approach to Surface Code Decoding

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a trained neural decoder for the toric surface code can be improved after the fact by reoptimizing it with a smooth surrogate of the syndrome map, so that the syndrome of its residual error is driven to zero.

desk verdict A plausible fine-tuning trick for neural surface-code decoders, but the missing same-epoch control means the 'universal effectiveness' claim is not yet earned. read the letter →

arxiv 2509.10164 v1 pith:YLN5XNFR submitted 2025-09-12 cs.LG quant-ph

Hoshitaro Ohnishi , Hideo Mukai This is my paper
classification cs.LGquant-ph MSC 81P7068T07 PACS 03.67.Pp
keywords surfacecodetoricquantumerrorcorrectionneuraldecodersyndromemeasurementreoptimizationdeeplearninglogicalrate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that the accuracy of a deep-learning surface-code decoder is not fixed at the end of its main training run. It proposes a fine-tuning step in which the decoder's own error guess e′ is compared with the true error e through a differentiable approximation of the syndrome measurement, and the decoder is updated to make the syndrome of |e−e′| zero. Across MLP, CNN, RNN, and transformer decoders, and for MLP at two code distances, this reoptimization reduced logical error rates in every case; the paper reports a 60.7% improvement for the transformer at distance 5 and a roughly 5-fold reduction in the training data needed to match the reoptimized MLP's accuracy. The general claim is that the benefit is independent of network architecture and code distance, so the method is a drop-in improvement for neural toric-code decoders.

What carries the argument

The load-bearing object is a continuous surrogate for the syndrome measurement: for each stabilizer, a smooth trigonometric function of the sum of the four surrounding qubit-error values, which reproduces the binary parity syndrome while remaining differentiable. This surrogate is approximated by a small neural network, then fixed during reoptimization. The decoder is reoptimized by minimizing a binary-cross-entropy loss between the surrogate's output on the residual |e−e′| and the zero vector, so gradients can flow from syndrome consistency back into the decoder's weights. The surrogate is what makes the discrete, non-differentiable syndrome map usable as a training signal.

What would settle it

Take a trained decoder and restrict the reoptimization dataset to syndromes generated by errors in one fixed logical sector. If the reoptimization converges to syndrome-consistent predictions that land in the wrong logical sector on a measurable fraction of test cases, the logical error rate will fail to drop or will rise, showing that syndrome-only consistency is not a safe objective.

Watch

Extended reading notes

Core claim

The central claim is that a decoder trained in the usual way—predicting an error vector from a syndrome using a classification-style loss—has only learned a probability distribution over possible errors, not the code's stabilizer symmetry. The paper's reoptimization adds a second training phase: fix a neural approximation f of the syndrome function, feed the residual |e−e′| into f, and update the decoder's weights to push f(|e−e′|) toward zero. Since the true syndrome of e and the syndrome of e′ will then match, the decoder is forced to be consistent with the stabilizer structure that the original training ignored. In all tested configurations—MLP at code distances 5 and 7, and CNN, RNN, and

Load-bearing premise

The method assumes that making the predicted error produce the same syndrome as the true error is a good proxy for lowering logical error rate, even though syndrome equality cannot distinguish errors that differ by a logical operator.

Editorial extensions

If this is right

  • Post-training reoptimization on the same dataset lowers logical error rate for MLP, CNN, RNN, and transformer decoders at code distance 5, and for MLP at code distance 7.
  • The reoptimized MLP at distance 5 matches the accuracy of an unreoptimized decoder trained on a five-fold larger dataset, so the method directly reduces the data required for a given accuracy.
  • The improvement persists under a biased noise model in which bit-flip and phase-flip errors are correlated, indicating the method is not tied to depolarizing noise.
  • The added cost is small: the syndrome surrogate is trained once and then held fixed, and the same reoptimization hyperparameters are used across different decoder architectures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the same syndrome-surrogate reoptimization should carry over to other stabilizer codes, since any code whose syndrome is a parity check admits a smooth surrogate that can be trained once and reused.
  • Editorial extension: a natural next test is to compare the reoptimized neural decoders directly against classical decoders at larger code distances; the paper compares neural architectures with one another, not with classical baselines.
  • Editorial extension: because the loss enforces only syndrome consistency, adding a homology-aware penalty—one that detects whether the residual error is a logical operator rather than a trivial cycle—could make the method more robust when a decoder's initial predictions are concentrated in the wrong logical sector.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a fine-tuning step for neural decoders of the toric surface code. After standard supervised training to predict an error vector e from a syndrome s, the decoder is reoptimized for 75 additional epochs by passing the absolute residual |e-e'| through a neural-network approximation f of the syndrome map and minimizing the binary cross-entropy of f(|e-e'|) against the zero vector. The authors test the method on an MLP decoder at code distances 5 and 7, on CNN, RNN, and transformer decoders at distance 5, on training-set size reduction, and on a biased (category) noise model. They report consistent improvements in logical error rate after reoptimization, with the transformer showing the largest relative gain, and claim 'universal effectiveness' independent of code distance and architecture.

Significance. If the claimed effect is real and attributable to the proposed loss, the paper would offer a simple, generic drop-in fine-tuning step for neural surface-code decoders, with possible dataset-size savings. The study has real strengths: results are averaged over 20 training runs with standard deviations, several architectures are compared under the same protocol, and the dataset-reduction experiment is a useful practical benchmark. The paper also provides a falsifiable empirical claim: reoptimization improves logical error rate across architectures and distances. However, the evidence as presented does not yet establish the central attribution claim: there is no control for the additional 75 epochs of training, and no comparison with a standard decoder such as MWPM. The proposed loss is also syndrome-only, so it cannot by itself distinguish residuals that differ by a logical operator, making the claimed symmetry-based mechanism undersupported.

major comments (4)
  1. [§3.2–3.3, Algorithm 1, Tables 2 and 4] The reoptimization procedure adds 75 epochs of training on the same dataset and with a BCE loss, while the original training used 55 epochs. The paper concludes that the observed logical-error improvements (e.g., 3.3% for L=5 and 6.5% for L=7 at p=5%) demonstrate 'universal effectiveness of the proposed method.' This attribution is not valid without an ablation that continues the original supervised BCE training for the same additional 75 epochs (or controls for total gradient steps / data passes) at a comparable learning rate. If continued standard training gives the same improvement, the proposed loss is not the causal factor. This control is needed to support the central claim.
  2. [§2.3, Algorithm 1] The reoptimization objective drives f(|e-e'|) to zero, i.e., it enforces syndrome consistency of the residual. But the syndrome map is blind to logical operators: if e-e' is a logical operator, f(|e-e'|)=0 exactly, so the loss is minimized without reducing the logical error rate. The paper asserts that the method 'consider[s] the symmetry unable to be accomplished by the original decoder,' but the proposed loss actually operates in the syndrome quotient and cannot detect logical equivalence classes. The paper measures logical error rate directly, so this is not circularity, but it is a gap between the stated mechanism and the objective. The authors should either analyze the residual distribution (e.g., fraction of residuals that are logical operators before/after reoptimization) or modify the objective to include logical-operator information, and temper the symmetry-based interpretation.
  3. [§3, Figs. 1–6] All comparisons are relative to the authors' own originally trained neural decoders. There is no baseline such as MWPM, which is the standard decoder for toric/surface codes and the reference point used in most prior neural-decoder papers. Without such a baseline, the reader cannot judge whether the proposed reoptimization closes the gap to practical decoding performance or merely improves a weak starting model. The claim in the conclusion that the method is 'useful in a wide range of noise models' is also only supported by one family of biased noise models with no circuit-level or time-dependent noise. Please add MWPM comparisons and clearly scope the claims to the tested noise model.
  4. [§4, Conclusions] The phrase 'universal effectiveness ... independent of code distance or network architecture' is not supported by the experiments. Only two code distances (L=5,7) and four architectures at L=5 are tested, all with the same noise model family and the same data-generation procedure. Moreover, because the missing epoch control (above) is not addressed, the claim of independence from architecture and distance is at best a description of the observed trend, not a demonstrated property of the method. Please rephrase the conclusion to match the evidence.
minor comments (6)
  1. [§2.2, Eq. (3)–(4)] The definition of f uses cos(pi x'/2), which for integer x' gives 1 when x' is even and 0 when x' is odd, matching parity. Please state explicitly that x' is the sum of the four neighboring error bits and clarify the domain [0,1]^{4L^2} is the continuous relaxation of the binary error vector.
  2. [Table 2] The column header 'renormalization' appears to be a typo for 'reoptimization.' Also, the 'First Training' column lists 'Input dimension 4L^2' and 'Output dimension 2L^2', but the Reoptimization column leaves these blank; since the decoder architecture is unchanged, it would be clearer to state 'same as first training.'
  3. [§3.2, Fig. 2 caption] The caption says 'Solid line: mean of difference in logical error rate before and after reoptimization,' but the text and figure appear to plot the logical error rate itself, not the difference. Please correct the caption to avoid ambiguity.
  4. [§3.2, first paragraph] The phrase 'we wanted to confirm that no room was left for improving the optimization method' is not operational. State the criterion used to decide convergence (e.g., loss plateau) and how this was verified.
  5. [References] Some relevant prior work on neural surface-code decoders is cited, but the paper would benefit from a comparison with more recent neural decoders that incorporate syndrome information or iterative refinement (e.g., Bellman-Ford style or recurrent decoders). This is a presentation issue, not a correctness issue.
  6. [§2.1] The notation 'e ∈ {0,1}^{4L^2}' is described as concatenating bit-flip and phase-flip vectors, but the sentence 'Here, we prepared two 2L^2 vectors ... and concatenated them' is grammatically unclear. Please rewrite for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: reoptimization uses a fixed syndrome-consistency loss; accuracy gains are measured on held-out logical-error data.

full rationale

The paper's central claim is empirical: a fixed syndrome-consistency fine-tuning loss improves logical-error rate across decoders. The loss in Algorithm 1 is constructed from the forward syndrome map f, which is trained once on random (error, syndrome) pairs and then frozen; it is not fitted to the decoder's outputs or to logical-error rates. The reported improvements (3.3% for L=5, 6.5% for L=7 at p=5%) are evaluated on held-out logical-error measurements, not on the reoptimization loss itself. No fitted parameter is renamed as a prediction: the only fitted quantities are network weights, and the quantity claimed to improve (logical error rate) is not used as a training target. There are no self-citations that carry load-bearing justifications, no imported uniqueness theorems, and no ansatz smuggled in via citation. The absence of a control for additional training epochs is a potential experimental confound but does not make any step circular by the paper's own equations. Therefore the derivation chain is self-contained; no circularity is found.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are invented. The free parameters are hyperparameters and noise-model settings; the central claim depends on them, but the paper does not fit them to the target logical error rates. The important axioms are the validity of the continuous syndrome surrogate and the adequacy of the neural approximation.

free parameters (4)
  • Reoptimization learning rate = 3e-8
    Hand-selected; no sensitivity analysis or ablation reported.
  • Reoptimization epochs = 75
    Hand-selected; no control separates the benefit of extra epochs from the benefit of the new loss.
  • Noise bias parameter eta = 0.5, 3, 5 (and 1 for unbiased)
    Sets the category noise distribution; chosen by hand, not fitted.
  • f-approximator hidden scale = 1000 (L=5), 750 (L=7)
    Hand-selected architecture size for the syndrome approximation network.
assumptions (4)
  • standard math Universal approximation theorem (Hornik 1991) guarantees a neural network can approximate the continuous syndrome interpolation f on a compact set.
    Invoked in Section 2.2 to justify using a neural network as the syndrome model; not machine-checked, but a standard theorem.
  • standard math The continuous function f, with g=1-cos(pi x/2), reproduces the discrete parity syndrome on integer error vectors.
    Needed for the reoptimization loss to be a valid surrogate for syndrome consistency; true for integer sums, approximate for continuous decoder outputs.
  • domain assumption Errors on physical qubits are independent and drawn from the discrete uniform or category noise model.
    Defines training and test distributions in Section 2.1; not validated against physical hardware or circuit-level noise.
  • domain assumption The trained syndrome-approximation network is accurate enough that gradients through it guide the decoder correctly.
    Table 5 reports high cosine similarity, but the paper does not test the effect of approximation error on reoptimization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Symmetry-Integrated Approach to Surface Code Decoding." pith.science (2026). https://pith.science/paper/YLN5XNFR

@misc{pith2026250910164,
  author       = {Pith},
  title        = {Pith review of: A Symmetry-Integrated Approach to Surface Code Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLN5XNFR}},
  note         = {Machine review of arXiv:2509.10164}
}
read the original abstract

Quantum error correction, which utilizes logical qubits that are encoded as redundant multiple physical qubits to find and correct errors in physical qubits, is indispensable for practical quantum computing. Surface code is considered to be a promising encoding method with a high error threshold that is defined by stabilizer generators. However, previous methods have suffered from the problem that the decoder acquires solely the error probability distribution because of the non-uniqueness of correct prediction obtained from the input. To circumvent this problem, we propose a technique to reoptimize the decoder model by approximating syndrome measurements with a continuous function that is mathematically interpolated by neural network. We evaluated the improvement in accuracy of a multilayer perceptron based decoder for code distances of 5 and 7 as well as for decoders based on convolutional and recurrent neural networks and transformers for a code distance of 5. In all cases, the reoptimized decoder gave better accuracy than the original models, demonstrating the universal effectiveness of the proposed method that is independent of code distance or network architecture. These results suggest that re-framing the problem of surface code decoding into a regression problem that can be tackled by deep learning is a useful strategy.

Figures

Figures reproduced from arXiv: 2509.10164 by the authors.

Figure 1
Figure 1. Horizontal axis: error rate of physical qubits. Vertical axis: logical error [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Logical error rate for L = 5, 7 before and after reoptimization by MLP decoder. Horizontal axis: error rate of physical qubit. Vertical axis: logical error rate of the decoder. Solid line: mean of difference in logical error rate before and after reoptimization. Blue shadow: Area within standard deviation. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Mean logical error rate of decoders before and after reoptimization. Horizontal [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Means and standard deviations of logical error rates of CNN, RNN, and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Means of logical errors of the decoder trained on data from x1 to x5. Horizontal [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Effect of the skewness of noise model. Horizontal axis: error rate of physical [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 linked inside Pith

  1. [1]

    35th Annu

    Shor P W 1994Proc. 35th Annu. Symp. Found. Comput. Sci. (FOCS)124–134

  2. [2]

    Broadbent A and Islam R 2020Theory of Cryptography (TCC 2020), Lecture Notes in Computer Science1255092–122 (Springer, Cham)

    2020

  3. [3]

    Zhou L, Wang S-T, Choi S, Pichler H and Lukin M D 2020Phys. Rev. X10021067

  4. [4]

    Intell.631

    Suzuki T, Hasebe T and Miyazaki T 2024Quantum Mach. Intell.631

  5. [5]

    15th Int

    Kim G S, Park S and Kim J 2024Proc. 15th Int. Conf. Inf. Commun. Technol. Converg. (ICTC)1145–1150

  6. [6]

    Amin M H, Andriyash E, Rolfe J, Kulchytskyy B and Melko R 2018Phys. Rev. X8 021050

  7. [7]

    Process.17154

    He Z, Li L, Huang Z and Situ H 2018Quantum Inf. Process.17154

  8. [8]

    Ibe Y, Nakagawa Y, Earnest N, Yamamoto T, Mitarai K, Gao Q and Kobayashi T 2022Phys. Rev. Research4013173

Show all 28 references
  1. [9]

    Ohgoe T, Iwakiri H, Kohda M, Ichikawa K, Nakagawa Y O, Valencia H O and Koh S 2024Phys. Rev. Res.6L022022 10

  2. [10]

    Scr.99065106

    Weber T, Borras K, Jansen K, Kr¨ ucker D and Riebisch M 2024Phys. Scr.99065106

  3. [11]

    disserta- tion, California Institute of Technology)

    Gottesman D 1997Stabilizer Codes and Quantum Error Correction(Ph.D. disserta- tion, California Institute of Technology)

  4. [12]

    Phys.303 2–30

    Kitaev A Y 2003 Fault-tolerant quantum computation by anyonsAnn. Phys.303 2–30

    2003

  5. [13]

    Raussendorf R and Harrington J 2007 Fault-tolerant quantum computation with high threshold in two dimensionsPhys. Rev. Lett.98190504

    2007

  6. [14]

    deMarti iOlius A, Etxezarreta Martinez J, Fuentes P and Crespo P M 2023 Perfor- mance enhancement of surface codes via recursive minimum-weight perfect-match decodingPhys. Rev. A108022401

    2023

  7. [15]

    Rep.711003

    Krastanov S and Jiang L 2017 Deep neural network probabilistic decoder for stabilizer codesSci. Rep.711003

    2017

  8. [16]

    Wang H-W, Xue Y-J, Ma Y-L, Hua N and Ma H-Y 2022 Determination of quantum toric error correction code threshold using convolutional neural network decoders Chin. Phys. B31010303

    2022

  9. [17]

    Advisor: Qun Li

    Wu K Y 2024Improving the Scalability of Neural Network Surface Code Decoders Bachelor’s Thesis, Department of Computer Science, William & Mary. Advisor: Qun Li. Committee Member: Chi-Kwong Li

  10. [18]

    Lange M 2023Decoding the Surface Code Using Graph Neural Networks(Master’s thesis, University of Gothenburg)

  11. [19]

    Fjelddahl F and Bengtsson I 2024Data-driven Decoding of the Surface Code Using a Neural Matching Decoder(Master’s thesis, Chalmers University of Technology, MSc in Complex Adaptive Systems)

  12. [20]

    2023 IEEE/ACM Int

    Wang H, Liu P, Shao K, Li D, Gu J, Pan D Z, Ding Y and Han S 2023Transformer- QEC: Quantum Error Correction Code Decoding with Transferable Transformers Proc. 2023 IEEE/ACM Int. Conf. on Computer Aided Design (ICCAD)

    2023

  13. [21]

    Fitzek D, Eliasson M, Kockum A F and Granath M 2020Phys. Rev. Res.2023230

  14. [22]

    Technol.3015004

    Varsamopoulos S, Criger B and Bertels K 2017Quantum Sci. Technol.3015004

  15. [23]

    Varbanov B M, Serra-Peralta M, Byfield D and Terhal B M 2025Phys. Rev. Res.7 013029

  16. [24]

    and Zhong H.-S

    Hu G., Ouyang W., Lu C.-Y., Lin C. and Zhong H.-S. 2025 Efficient and Uni- versal Neural-Network Decoder for Stabilizer-Based Quantum Error Correction arXiv:2502.19971 [quant-ph]

    2025 arXiv

  17. [25]

    Tan X, Zhang F, Chao R, Shi Y and Chen J 2023PRX Quantum4040344

  18. [26]

    Hornik K 1991 Approximation capabilities of multilayer feedforward networksNeural Netw.4251–257 11

    1991

  19. [27]

    IEEE Conf

    He K, Zhang X, Ren S and Sun J 2016 Deep residual learning for image recognition Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR)770–778

    2016

  20. [28]

    Neural Inf

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, Kaiser L and Polosukhin I 2017 Attention is all you needAdv. Neural Inf. Process. Syst.30 12

    2017

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.