acceptodds
Under review as a conference paper at ICLR 2027

Commit or Defer: A Checkable Neural Pre-Decoder for Quantum LDPC Codes

Abstract

Quantum low-density parity-check (qLDPC) codes offer a route to lower qubit overhead in fault-tolerant quantum computing. However, algorithmic decoders can require costly iterative inference and candidate search, while standalone neural decoders often rely on large models and extensive training data. Accordingly, we propose CoD (Commit-or-Defer), a novel unified neural pre-decoding framework. Its lightweight predictor estimates fault posteriors by combining physical priors with gated recurrent unit (GRU) message passing on -hop active subgraphs around fired detectors in a detector error model. A commit-or-defer strategy bypasses the classical backend when a candidate passes a zero-residual syndrome check and otherwise uses the same posteriors to guide backend inference. We bound the logical risk of committed shots in terms of posterior uncertainty and the DEM logical distance, and give a sufficient condition under which CoD lowers overall logical risk below the unassisted backend. A common circuit-level benchmark spanning four code families and three backends shows up to % lower mean logical error rates (LERs) and –% fewer backend calls than unassisted decoding. On BB72 and BB144 under code-capacity noise, CoD achieves lower mean LERs than a standard graph neural decoder using roughly three million training shots and – fewer neural inference FLOPs

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Error Correction, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jefqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Commit-or-Defer (CoD), a new lightweight pre-decoder framework for quantum error correction. CoD can either commit and make a correction or defer and otherwise use its predicted posteriors to guide a classical decoder. The network uses K-hop subgraphs so that its cost will grow with active detectors rather than with the code. The method is evaluated over several code families: BB codes, surface codes, color codes and random quantum Tanner codes using MWPM, BP-OSD, and Relay-BP. The evaluation is performed under code-capacity and SI1000 noise models.

Critical strengths
  • The idea of using K-hop subgraphs is well motivated.
  • ⁠The model's ability to either commit or defer is an interesting design and the logical risk analysis is a compelling feature.
Critical weaknesses
  • ⁠In terms of experiments, a prominent candidate for quantum computing is the BB144 code, and figure 5 shows that the method matches Relay-BP under SI1000. In addition, there is no comparison to other methods that augment classical decoders (for example for surface codes: Belief Matching or Correlated Matching).
  • Framework Effectiveness: For larger codes, such as the BB144, surface d >=7 and color 11, the number of backend calls reaches around 70% and above for the upper end of physical error rates. In addition, neither wall-clock time for the network or backend nor parameter count is reported, it is therefore hard to sufficiently assess the portion of the network compared to solvers in terms of runtime.
Questions for authors
  • ⁠Could the authors provide detailed wall-clock times for each stage in the process (backend and network) to assess the cost of each stage in the architecture? Also, could the authors provide a parameter count?
  • ⁠In the code-capacity noise, could the authors provide the comparison for the p > 0.1?
  • To better evaluate the proposed method, could the authors provide a comparison with correlation aware decoders?
  • Could the authors provide the percentage of backend calls for the all tested physical error rate in figure 2 and 5? For example surface d=9 decoded in the range of $ 3 \times 10 {-3}$ to 7×10−37 \times 10 {-3} and the backend calls for surface code is only up to 4×10−34 \times 10 {-3}.
Reviewer_JoypOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoD presents a neural pre-decoding framework for quantum error correction. A lightweight network estimates fault probabilities on active subgraphs of a detector error model. Candidates that satisfy the syndrome constraints are committed directly; otherwise, the same probabilities guide a classical decoder. The paper provides conditional logical-risk bounds and evaluates decoding accuracy, backend utilization, and neural computation across several code families.

Critical strengths
  1. The mechanism-level probability output provides a useful interface between neural prediction and classical decoding. It supports both checking a candidate and guiding subsequent inference when that candidate fails the check. This gives the learned model a practical role even when it cannot independently resolve a shot.
  2. The evaluation covers multiple code families and classical backends, with comparisons on identical test shots. The reported mean logical error rate improves in 18 of 20 configurations, and the routing and deferred-shot analyses help explain these improvements. These results provide useful evidence that the approach can improve decoding accuracy beyond a single code–backend combination.
Critical weaknesses
  1. The contribution relative to the closest hybrid decoding methods needs clearer positioning. The existing ablations support several design choices, but the paper could more directly explain what the proposed combination adds beyond prior approaches. A representative comparison would help establish the significance of these differences.
  2. The reported efficiency results characterize neural computation and backend utilization, but leave the complete decoding cost unclear. These are useful measurements within their stated scope. Representative runtime results accounting for the added neural processing and changes in backend workload would help assess the practical benefit.
Questions for authors
  1. What is the main methodological advance over the closest hybrid decoding approaches? Could the authors explain which differences are responsible for the benefits, drawing on the existing ablations and, where applicable, a representative comparison?
  2. Could the authors provide representative end-to-end runtimes, or a cost breakdown covering the added neural processing and changed backend workload? Costs that are identical between the compared methods may be excluded consistently.
Reviewer_HbxvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CoD, a neural pre-decoder for qLDPC codes that sits in front of a classical decoder. A lightweight network runs bidirectional GRU message passing on the K-hop subgraph around fired detectors in the DEM bipartite graph, and outputs a posterior for each fault mechanism as a correction to the physical prior log-odds. The posteriors are thresholded into a candidate fault vector, and the syndrome residual is computed on the full DEM. If the residual is zero, the candidate is committed, and the backend is skipped. Otherwise, the shot is deferred to the backend, which is guided by the neural posteriors.

Critical strengths
  1. The design is clean and has engineering value.
  2. The data-efficiency and inference-cost arguments are fairly convincing, and the claim of generality across codes and backends has some empirical support.
Critical weaknesses
  1. The theoretical guarantee is essentially decorative and is never actually validated. For example, Proposition 1 relies on Assumption 1, yet the quantities in Assumption 1 are not measured anywhere in the paper. Moreover, the "strictly lower risk" conclusion is a tautology once the assumptions hold, and the results contain counterexamples: RQT60 (LER up 1.70%) and BB144 + Relay-BP (up 0.98%). These indicate the conditions fail in those cases, but the paper never uses the theory to explain them.
  2. The statistics and comparison setup weaken the conclusions.
  • 2.1 The gains on Relay-BP, the strongest current backend, are noticeably smaller, and on BB144 there is no improvement at all. The benefit from the algorithmic backends is concentrated on BP-OSD (especially the weak order-0 configuration), so the baselines look weak. The configurations are also inconsistent: the Tanner codes use order-60 OSD while all other codes use OSD-0.
  • 2.2 I could not find end-to-end latency or wall-clock time reported anywhere in the manuscript. Fewer backend calls do not necessarily mean less total compute. For example, the neural forward pass and graph construction are excluded from the cost accounting, and the bottleneck of a GNN is often indexing and aggregation. The core "cheap" claim is therefore never actually measured.
  • 2.3 The comparison against neural baselines covers only Astra, and only under code-capacity noise on two codes. The circuit-level experiments include no neural decoder at all (e.g., Blue et al., GraphQEC). The claim of outperforming a "standard neural decoder" is therefore very narrow in scope.
Questions for authors

Please address the above weaknesses.

Comments on the paper

1 comment
Area_Chair_yu6rOpenReview ↗

Follow-up questions from the AC

Thank you for the detailed and diligent responses during the rebuttal period. I have a few additional questions. Given the limited time left, answers without new experiments are sufficient.

1. Positioning

The zero-residual commit gate is conceptually similar to the convergence check in BP+OSD [1, 2]. More broadly, committing or deferring based on the output of a first stage is a natural design in two-stage (pre-/post-) decoders, and well-known decoders such as Belief Matching [3] and the scalable neural decoder of Meinerz et al. [4] already use a two-stage structure in which a first stage informs or reduces the work of a classical decoder. Neural models that guide classical decoders have also been explored, e.g., [4, 5]. Although these methods may not use an explicit commit/defer rule, they are similar in pairing a first-stage (neural) decoder with a classical one. Please clarify what CoD adds beyond these.

2. Comparison with BP+OSD

  • (a) Thank you for the OSD-order sweep suggested by Reviewer Hbxv, which addresses the concern about the OSD-0 baseline. If convenient, a log-scale version or values at p = 0.002 would make the low-error regime easier to read.
  • (b) When comparing GNN+BP+OSD (CoD) with BP+OSD, the OSD call reduction alone does not make the role of the GNN clear. Is the reduction mainly due to the GNN committing shots before BP is invoked, or to the GNN posteriors strengthening BP so that BP converges more often?

3. Relay-BP hyperparameters

Hyperparameters of Relay-BP [6], such as the number of legs, iterations per leg, memory strengths, and the stopping criterion, are not reported. Please report them for each code.

4. GNN overhead in FLOPs

If available, please report the additional overhead of the GNN in FLOPs by comparing GNN+BP+OSD with BP+OSD per shot, giving both the average and the worst case.

References

[1] P. Panteleev, G. Kalachev. Degenerate Quantum LDPC Codes With Good Finite Length Performance. Quantum, 2021.

[2] J. Roffe, D. R. White, S. Burton, E. Campbell. Decoding across the quantum low-density parity-check code landscape. Physical Review Research, 2020.

[3] O. Higgott, T. C. Bohdanowicz, A. Kubica, S. T. Flammia, E. T. Campbell. Improved Decoding of Circuit Noise and Fragile Boundaries of Tailored Surface Codes. Physical Review X, 2023.

[4] K. Meinerz, C.-Y. Park, S. Trebst. Scalable Neural Decoder for Topological Surface Codes. Physical Review Letters, 2022.

[5] Y. Peled, D. Zenati, E. Nachmani. Neural Minimum Weight Perfect Matching for Quantum Error Codes. ICML 2026.

[6] T. Müller, T. Alexander, M. E. Beverland, M. Bühler, B. R. Johnson, T. Maurer, D. Vandeth. Improved belief propagation is sufficient for real-time decoding of quantum memory. arXiv, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.