acceptodds
Under review as a conference paper at ICLR 2027

QECFM: Quantum Error Correcting Code Discovery via Training Free Guidance

Abstract

Designing quantum error-correcting codes is difficult: the search space is discrete and highly constrained, while evaluating a candidate often requires an expensive Monte Carlo estimate of its logical error rate. We introduce QECFM, a two-stage conditional flow matching model with a deterministic group-algebra readout layer, which maps each continuous sample to discrete parity-check matrices and enforces CSS commutativity. We then use D-Flow with a differentiable noise-aware critic to optimise the source variables toward codes with lower logical error rates. We evaluate the resulting method across six code families, code-capacity and circuit-level noise, and block lengths up to . Our method finds codes with a lower median logical error rate than both learned and classical search baselines within the Bivariate Bicycle Code family. Under code-capacity depolarising noise, our best code also outperforms the Gross code at different physical error rates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
QEC, Flow Matching, Optimisation, Code Discovery

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f79fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces QECFM, a method for finding quantum error-correcting codes with a low logical error rate (LER) under a specified noise model and decoder. It learns a two-stage conditional flow over X/Z Tanner-Laplacian spectra, retains eigenvector bases from training examples, and uses a learned LER critic for D-Flow source optimization. A deterministic group-algebra readout enforces CSS commutativity. The experiments cover several code families, block lengths, and noise regimes. The authors report positive search results within the bivariate-bicycle (BB) family and a selected code that outperforms the Gross code under the reported code-capacity BP-OSD evaluator.

Critical strengths
  1. The framework provides a practical way to search a constrained code space when objective evaluations are expensive. It combines continuous search with exact algebraic feasibility: the group-algebra readout guarantees CSS commutativity. Final candidates are evaluated through Monte Carlo decoding rather than critic predictions alone. Appendix B.2 also gives an explicit code representation whose structural properties can be checked independently.

  2. The paper reports encouraging results within the BB family. Table 1 compares relevant learned and classical baselines across ten search runs under a stated online Monte Carlo budget. The median winner LER is 0.0310 for QECFM+DFlow, compared with 0.0495 for random search. Table 5 also reports non-overlapping count-based confidence intervals for the selected BB codes found by these methods at a low physical error rate. These results suggest that the learned search framework can be useful after pretraining.

Critical weaknesses
  1. Several details are missing from the generation and evaluation procedures, making it difficult to trace how the reported results were obtained. Equation (13) specifies the same support size ww for both blocks, while the selected code in Appendix B.2 has support sizes 2 and 6. Producing the displayed parameterization from the stated rule therefore requires an additional step that is not described. This is a gap in the method specification, not evidence that the code is invalid.

    Training is described only for n=144n = 144, with eigenbases retained from training examples. Figure 5 evaluates other sizes without explaining how their bases are obtained. The evaluation also needs clarification: for the BB result at p=0.002p = 0.002, the count ratio reported in Table 6 is 19/40,000=4.75×10−419/40{,}000 = 4.75 \times 10^{-4}, while the reported LER is 9.50×10−49.50 \times 10^{-4}. The reference to "channel-derived intervals" may account for the difference, but the paper does not define the estimator. These omissions make it difficult to verify the discovery pipeline, scaling results, and circuit-level performance.

  2. The experiments do not yet establish how much each component contributes or whether the framework offers an end-to-end computational advantage. Table 1 supports improved online BB search after pretraining under a matched online evaluation budget. However, the critic uses 50,000 labeled codes, including 20,000 examples from the low-LER tail. The authors acknowledge the precomputation cost, but give too little detail on how much label acquisition costs or what reuse would be needed to amortize those costs. The readout also enumerates candidates and prioritizes the requested kk, while the denominator used to calculate yield and the treatment of filtering costs across methods remain unclear. These gaps limit both the comparison of total computational costs and the attribution of gains to the spectral proposal, critic-guided optimization, and candidate selection by the readout.

    The architecture also needs a more targeted comparison. Within the Abelian BB training family, the X/Z Tanner graphs are isomorphic and have identical Laplacian spectra. The experiments do not isolate the benefit of the second conditional flow over a tied-spectrum alternative. Table 4 provides evidence that D-Flow guidance helps, but the independent-search and candidate-selection protocols need clarification to judge how robust this benefit is.

Questions for authors
  1. Does the implementation use independent support sizes for the two blocks, or apply an additional transformation that reconciles Equation (13) with the support sizes 2 and 6 in Appendix B.2? Please describe the actual candidate-generation procedure. How are eigenbases obtained for n=36,72,288n = 36, 72, 288 in Figure 5, and does this require additional data or training?

  2. What counts toward the 600-evaluation budget, and at what stage is raw target-kk yield measured? Please specify each method's search domain, target-kk screening rules, and treatment of rejected or intermediate candidates. How were the 50,000 critic labels, including the examples from the low-LER tail, obtained? Please report offline and online costs separately and explain the intended reuse conditions.

  3. Given the identical X/Z Tanner-Laplacian spectra in the BB training family, what benefit does the second conditional flow provide over a tied-spectrum alternative? A focused control retaining the eigenbases, critic, readout, and evaluation budget would help isolate its contribution. Can the existing run-level results quantify the gains and their uncertainty for QECFM+DFlow relative to unguided QECFM and Diffusion+DFlow?

  4. Which events are counted in Table 6, and what formulas convert those counts into the reported LER and confidence intervals? Please state any assumptions about joint X/Z failures. For Table 4, please clarify the noise and decoder settings, number of independent searches, and candidate-selection protocol. Do the final evaluations use fresh Monte Carlo samples independent of selection? Please also provide the existing event counts and uncertainty estimates for Table 2.

4 comments
AuthorsOpenReview ↗

Rebuttal response

We thank the reviewer for the detailed technical assessment. We address implementation details, computational costs, component attribution, and statistical uncertainty.

W1 / Q1: Generation and evaluation details

Support sizes. We updated Eq. 13 to use separate wA,wBw_A,w_B, since AA and BB supports are selected independently. This yields ∣A∣=2|A|=2 and ∣B∣=6|B|=6.

Eigenbases and scaling. QECFM generates spectral coefficients, not eigenvectors. For each block length, Tanner-Laplacian eigenbases are obtained from valid codes and fixed during generation. Experiments at n=36,72,144,288n=36,72,144,288 use separate datasets and independently trained models.

Circuit-level estimator. Table 7 now reports channel-specific failures for the QECFM+DFlow BB code:

Physical error rateX failures / trialsZ failures / trialsCombined LER
0.00212 / 20,0007 / 20,0009.50×10−49.50\times10^{-4}
0.00350 / 4,09654 / 5,1202.26×10−22.26\times10^{-2}

We report X/Z failures for all families and define

p^L=1−(1−p^X)(1−p^Z),p^b=eb/sb.\widehat p_L=1-(1-\widehat p_X)(1-\widehat p_Z),\qquad \widehat p_b=e_b/s_b.

W2 / Q2: Component attribution, yield, and budget

Search budget and yield. Table 1 uses 600 evaluated candidates and 20,000 Monte Carlo trials per candidate, i.e. up to 12 million trials per search. Invalid or wrong-kk proposals are rejected before LER evaluation, with

k=n−rank⁡(HX)−rank⁡(HZ),k = n - \operatorname{rank}(H_X) - \operatorname{rank}(H_Z),

where the ranks are computed over F2\mathbb{F}_2. Search continues until 600 target-kk codes are found; yield is therefore 600 divided by the number of raw proposals.

Baseline distinction. QECFM, Diffusion+DFlow, and QECFM+Diffusion search in Tanner-Laplacian spectral space, whereas simulated annealing, Bayesian optimisation, and random search operate directly on the discrete polynomial code parameterisation.

Training costs. The n=144n=144 critic dataset contains 50,067 examples, including 20,000 low-LER samples, with k=12k=12 codes drawn uniformly from the BB parameter space. Label construction required approximately 1 billion Monte Carlo trials. We therefore report training separately and restrict search-efficiency claims to post-training discovery.

Amortisation. This cost corresponds to roughly 83 searches and can be amortised when constructing code libraries or algorithm-specific codes.

Table 4 settings. Table 4 uses depolarising code-capacity noise at p=0.003p=0.003, BP-OSD decoding, and 10 independent searches per method. Final codes are selected using Monte Carlo LER estimates.

W2 / Q3: Second-field generation and D-Flow

We added the following ablation:

MethodTarget-kk yield
Tied X/Z spectrum80.9%
Independent HX,HZH_X,H_Z FM models93.1%
Conditional HZ∣HXH_Z\mid H_X (QECFM)95.9%

Independent generation improves yield by 12.2 percentage points over tied spectra, and conditioning adds a further 2.8 points. All methods share the same readout layer, so separate field representations provide greater flexibility for projection to valid CSS codes.

D-Flow and diffusion run-level advantage.

ComparisonMedian QECFM+DFlow vs baselineRelative median gain ± s.d.
vs QECFM0.0310 vs 0.0361+14.1% ± 8.7
vs Diffusion+DFlow0.0310 vs 0.0333+6.8% ± 7.9

Thus D-Flow improves median final-code LER by 14.1%, while QECFM+DFlow improves by 6.8% over Diffusion+DFlow. Table 4 additionally shows a 1.92× lower selected-code LER with D-Flow guidance.

Q4: Selection protocol and uncertainty

We distinguish search-to-search variability from Monte Carlo uncertainty. Table 1 summarises independent 600-step searches using medians and standard deviations. Table 4 re-evaluates selected codes with independent Monte Carlo simulations, so its confidence intervals quantify LER-estimation uncertainty.

For Table 2, we added failure counts and 95% confidence intervals:

pMethodFailures / samplesLER [95% CI]
0.05QECFM+DFlow1,556 / 60,0002.593e-02 [2.468e-02, 2.724e-02]
Gross [[144,12,12]]3,561 / 60,0005.935e-02 [5.747e-02, 6.127e-02]
0.01QECFM+DFlow20 / 400,0005.000e-05 [3.054e-05, 7.722e-05]
Gross [[144,12,12]]25 / 40,0006.250e-04 [4.045e-04, 9.225e-04]

The full table is now in the appendix of the manuscript. The intervals do not overlap, confirming the advantage under the code-capacity noise model used for optimisation.

Table 7 instead uses channel-specific circuit-level noise. We report X and Z counts separately and combined 95% confidence intervals using

L=1−(1−LX)(1−LZ).L=1-(1-L_X)(1-L_Z).

These additions separate Monte Carlo uncertainty from variability across independent searches.

We thank the reviewer for these questions and have incorporated the missing details into the revised manuscript. We are happy to clarify any remaing concerns and questions.

Thank you for the response and the additional experiments. The clarifications about Eq. (13), the LER estimator, and the counts in Table 2 address my presentation concerns.

Labelling data for the critic takes about as many MC trials as 83 searches. In your reply to Reviewer ogbE, random search matches your best code at about 12× the search budget and finds a better code at 16×. The case for amortising the labelling cost still needs evidence that the critic can be reused across settings.

Proposals with the wrong k do not count towards the MC budget, so the yield ablations alone do not show whether the learned components improve search. In your reply to Reviewer J7AK, QECFM’s median LER at n = 288 is 12.5% lower than that of Gaussian sampling (0.00630 vs. 0.00720). The comparison gives neither the number of runs nor a measure of uncertainty. The ablation comparing tied and independent spectra reports yield only.

The gain over Diffusion+DFlow is within one standard deviation. The difference between the best codes found with D-Flow and unguided QECFM is not statistically significant (reply to Reviewer ogbE).

I’m keeping my original score.

Thank you for acknowledging that the implementation and evaluation clarifications address your presentation concerns. We provide additional uncertainty estimates below.

Learned proposals versus Gaussian sampling. The results reported in our response to Reviewer J7AK summarise five independent searches per method. We now additionally report the standard deviations across searches and 95% confidence intervals for the selected codes.

MethodMedian LERSD across runsSelected-code LER, 20k shots [95% CI]Selected-code LER, 1M shots [95% CI]
QECFM0.006300.003340.00520 [0.00425, 0.00630]0.00478 [0.00464, 0.00492]
Gaussian0.007170.001110.00645 [0.00539, 0.00766]0.00672 [0.00656, 0.00688]

The 20k-shot and 1M-shot columns evaluate the same selected code from each method. QECFM achieves a 12.1% lower median LER across searches and a 28.9% lower selected-code LER at 1M shots, with non-overlapping confidence intervals.

Diffusion+DFlow versus QECFM+DFlow. We constructed Diffusion+DFlow by replacing the flow-matching formulation with a diffusion formulation while retaining the remaining components. Its competitive performance demonstrates that our framework also supports diffusion. Table 1 numerically favours flow matching over diffusion. The comparison with classical search algorithms supports the benefit of the shared learned framework under the same search budget after pretraining.

Statistical evidence for the framework and for D-Flow. Under the same search budget in Table 1, a two-sided Mann–Whitney U test comparing per-search winner LERs from unguided QECFM and random search gives p < 0.005. We compare the worst-performing method in our framework with the best-performing classical search method. This provides statistical evidence for the benefit of the learned code-discovery model over random search in the evaluated setting.

The incremental contribution of D-Flow is a separate comparison. QECFM+D-Flow reduces the observed median winner LER by 14.1% relative to unguided QECFM. Furthermore, Table 4 provides complementary evidence: an independent Monte Carlo evaluation finds approximately 48% lower LER for the code selected with D-Flow than for the code selected via unguided generation.

Thank you for the additional uncertainty estimates and the comparison with random search. The comparison supports a benefit from the learned framework once it has been pretrained.

For QECFM versus Gaussian sampling, the median LER difference (0.00087) is small relative to QECFM’s standard deviation across runs (0.00334). I’m still unsure how consistently that gain holds across searches. The non-overlapping confidence intervals quantify only the Monte Carlo uncertainty for a single selected code, the same type of uncertainty you identified for the intervals in Table 4.

The 48% figure for D-Flow repeats the result in Table 4, and the gain over Diffusion+DFlow remains within one standard deviation. I still haven’t seen evidence that the critic can be reused across settings to amortise its cost, and the tied-spectrum ablation reports only yield.

I’m keeping my original score.

Reviewer_ogbEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes QECFM, a flow matching-based framework for discovering quantum error-correcting codes. The method generates Tanner-graph spectra and uses a structured readout layer to guarantee CSS commutativity. The authors further incorporate D-Flow with a learned LER critic to guide code generation toward lower logical error rates. Experiments across multiple code families demonstrate promising improvements over classical and learning-based search methods.

Critical strengths
  • The combination of flow matching and structured code generation is an interesting approach that guarantees CSS commutativity while exploring a continuous search space.

  • The method achieves promising logical error rates compared with several classical and learning-based baselines, including improvements over the Gross code under the evaluated noise and decoding settings.

Critical weaknesses
  • The computational fairness of the comparison is unclear. The proposed method requires substantial offline training data and LER evaluations, but these costs are not fully reflected in the matched-search comparisons.

  • The additional contribution of D-Flow is not entirely convincing. In Table 1, QECFM+DFlow achieves only modest improvements over QECFM and Diffusion+D-Flow relative to the reported variability.

Questions for authors
  • How does the proposed method compare with classical search methods when the cost of generating training data and training the flow model and LER critic is included? Would classical methods perform better under the same total computational budget?

  • In Table 1, the improvements from QECFM+DFlow over QECFM and Diffusion+D-Flow are relatively small. Are these improvements statistically significant, and what is the main advantage of the proposed combination?

I would be willing to increase my score if these concerns are sufficiently addressed.

2 comments
AuthorsOpenReview ↗

Rebuttal response

We thank the reviewer for their assessment and address the computational costs and D-Flow's contribution.

W1 / Q1: Computational fairness

We agree that Table 1 compares methods under a fixed search budget rather than equal total computation. Learned methods incur training costs, but the proposal distribution and critic can be reused across subsequent searches.

We therefore report training costs explicitly: approximately 1 billion MC trials for the n=144n=144 critic landscape, 2.7 GPU-hours for flow-matching training at n=144n=144, and 3.4 GPU-hours for QECFM at n=288n=288. The critic trains in 1.5 GPU-hours. A 600-step search uses up to 12 million MC trials, so critic construction requires approximately 83× the MC trials of one fully evaluated search. This cost can be amortised across compatible searches, e.g., algorithm-specific codes, code libraries, or changing hardware.

Our claim is therefore improved search effectiveness under a fixed search budget, with training costs amortisable across searches.

To assess search fairness, we also ran random search for longer:

BudgetEvaluationsMedianBest
QECFM+D-Flow6000.03100.0247
1×6000.05090.0313
2×1,2000.04420.0313
3×1,8000.04420.0313
4×2,4000.04330.0313
6×3,6000.03780.0313
8×4,8000.03740.0295
12×7,2000.03310.0244
16×9,6000.02970.0216
24×14,4000.02920.0216
32×19,2000.02770.0216
42×25,2000.02770.0204

Random search requires approximately 12× the search budget to match or slightly improve upon the best QECFM+D-Flow code.

W2 / Q2: Contribution of D-Flow

D-Flow reduces median winner LER from 0.0361 to 0.0310 across independent searches, a 14.1% relative improvement. A one-sided Mann–Whitney U test gives p=0.06p=0.06, providing suggestive, though not conventionally significant at the 0.05 level, evidence that D-Flow improves winner LER.

The controlled guidance ablation in Table 4 further shows that D-Flow achieves 1.92× lower selected-code LER than unguided generation and 2.30× lower LER than random- or zero-gradient controls.

Compared with Diffusion+DFlow, QECFM+DFlow achieves a lower median winner LER (0.0310 versus 0.0333). Across five searches per method under the same budget:

MetricQECFM+DFlowDiffusion+DFlow
Median winner LER ↓0.030800.03325
Best winner LER ↓0.024700.02925

Flow matching directly learns a continuous velocity field through which D-Flow optimises source variables. The guidance experiments demonstrate the benefit of optimisation within this representation, while the diffusion comparison numerically favours QECFM.

Beyond the final winner, D-Flow also shifts the full search distribution toward lower LER:

nnmethodpool median5th pctbest
144QECFM0.4120.1150.0250
144QECFM + D-Flow0.3370.1250.0247
pp1.1×10−51.1\times10^{-5}0.0120.53
288QECFM0.2460.0370.0033
288QECFM + D-Flow0.2050.0320.0029
pp0.00160.00160.47

The Mann–Whitney U tests show that D-Flow consistently improves the search distribution over unguided QECFM, particularly its median, with the advantage becoming more pronounced at n=288n=288. The best-code differences are not significant, as expected for an extreme statistic, but the broader evidence is consistent: QECFM+DFlow outperforms unguided QECFM and classical baselines in Table 1, scales to n=288n=288 (Fig. 5), improves under circuit-level noise (Fig. 4b–c), and generalises across five unseen code families (Table 3). At p=0.003p=0.003, using 20M Monte Carlo shots, it also achieves lower LER than unguided and random guidance with non-overlapping 95% confidence intervals (Table 4), with a similar advantage over Diffusion+DFlow for the best code in Table 1.

This consistency matters because code-capacity performance alone does not determine practical utility: decoder, hardware, and fault-tolerant circuit constraints must also be considered. Especially at n=288n=288 and beyond, the goal is therefore to reliably generate promising candidates for subsequent hardware-level testing. Across Tables 1, 3, and 4 and Figs. 4–5, D-Flow systematically biases QECFM toward better-performing regions of the valid-code manifold rather than producing isolated improvements.

We thank the reviewer for identifying these missing technical details and hope the revision resolves their concerns.

Reviewer_ogbEOpenReview ↗

Response to Authors

Thank you for the detailed rebuttal and additional experiments. The responses help clarify the computational costs and the contribution of D-Flow. However, the comparison under equal total computational budgets remains unclear, and the improvements from D-Flow are still relatively modest. Therefore, I will maintain my original rating.

Reviewer_J7AKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for generating quantum error correcting codes, QECFM, which use two stages of conditional flow matching model in order to find CSS quantum error correcting codes. The method represent the tanner graph with their Laplacian eigenvalues with a fixed eigenvectors from valid dataset. The proposed method generate the spectrum of Hx and then used as conditioned for the generation of Hz. The proposed method also use deterministic group algebra readout layer to get get a valid binary CSS.The proposed method trained separate transformer critic network in order to find codes with good logical error rate. The proposed method tests on 6 code families up to block length 288. The proposed method generate 144, 12, 8 BB code.

Critical strengths
  1. The experiment section is good and use large n which is equal to 288 physical qubits. The generated code 144, 12, 8 is very interesting since it better than the well known BB 144, 12, 12 Gross code.

  2. The separation of the continuous optimization from the algebraic constraint  with the group algebra readout layer is novel and useful. The proposed method also get valid CSS codes and also right k parameter. Table 1 shows 95.9 percent target k yield  which is much higher than baselines.

Critical weaknesses
  1. It is not well clear how the zero shot transfer to other code families comes from the deterministic readout layer itself or the flow matching model.

  2. The proposed model trained on 30,000 BB codes, the LER critic is trained on 50,000 labeled BB codes. Therefore the method use 50,000 Monte Carlo LER evaluations. So when you compare your method with simulated annealing and Bayesian optimization and random search it ignore the compute spent on generating this dataset.

Questions for authors
  1. report the wall clock time and the number of Monte Carlo running to create the 50k lables codes.

  2. The generate 144, 12, 8 code have check weight 8, wheras the Gross code has check weight 6 and distance 12. Can you compare you code with the Gross code under SI1000 circuit level noise ?

  3. what hppen to the results when you remove the  flow matching model and give  random Gaussian noise matrices to the readout?

3 comments
AuthorsOpenReview ↗

Rebuttal Answer

Rebuttal — Reviewer J7AK

We thank the reviewer for recognising the novelty of separating continuous optimisation from exact algebraic constraints. We address the contribution of learned generation, computational costs, and the Gross-code comparison with additional experiments.

W1 / Q3: Learned flow matching model vs. Gaussian sampling of the Tanner graph spectrum

To isolate the contribution of flow matching, we replace our learned model with sampling the Tanner graph spectrum from a Gaussian distribution and keep the remaining setup unchanged, using the same eigenbases, structured readout layer, and evaluation protocol.

MetricQECFMGaussian distribution samples
Target-kk yield95.90%91.26%
Median LER, n=288n=288 ↓0.006300.00720
Best LER, n=288n=288 ↓0.005200.00645

QECFM improves target-kk yield by 4.64 percentage points and reduces median LER by 12.5%. Both approaches share the same structured readout layer, which guarantees algebraic validity independently of the learned model. The results show that learning the spectral proposal distribution improves the quality of the discovered codes when compared with Gaussian sampling.

To support our zero-shot transfer claim, we further investigate the yield across different code families. Our Flow Matching model is only trained on the unlabeled BB code dataset. We use this model to compare its sampling yield with a random spectrum sampled from a Gaussian distribution.

FamilyQECFMGaussian
Z12×Z6Z_{12}\times Z_697.06%96.76%
A4×Z6A_4\times Z_699.98%100.00%
S4×Z3S_4\times Z_399.86%99.88%
D36D_{36}96.55%95.99%
D4×Z9D_4\times Z_985.84%80.13%
Q8×Z9Q_8\times Z_983.83%78.51%
Z6×Z6×Z2Z_6\times Z_6\times Z_2100.00%100.00%

We notice that the advantage of our learned flow matching model is bigger the more the code distributions differ from Gaussian noise.

We thank the reviewer for bringing this up. We have added this experiment to the paper in Appendix B, Table 5 and clarified the contribution of the readout layer.

W2 / Q1: Training data and computational costs

We agree that training and search/code-discovery costs should be distinguished. Flow matching is trained without LER labels, whereas the D-Flow critic requires Monte Carlo-labelled data.

Training the flow matching model takes approximately 2.7 GPU-hours at n=144n=144 and 3.4 GPU-hours at n=288n=288. Critic dataset generation at n=144n=144 required 51 single-core CPU-hours, comprising approximately 1 billion Monte Carlo trials. Training the critic model for the DFlow addition took 1.5 h on a GPU, but it is independent of the QECFM training.

These costs are for dataset preparation and training, while Table 1 compares search performance under a fixed budget of 600 search steps, requiring up to 12 million MC trials. The learned proposal model and critic can be reused across compatible searches, amortising their initial computational cost, whereas MC cannot benefit from such amortisation. This additional information is now added in Appendix A.1.

Q2: Gross code under SI1000 circuit-level noise

Importantly, our search optimised code-capacity LER at p=0.05p=0.05, not circuit-level performance. Under the actual optimisation objective and evaluated BP-OSD decoder, our selected code outperforms Gross (Table 2). These results demonstrate that code rankings depend on the noise model and decoder, motivating future extensions to circuit-level optimisation.

We then follow the reviewer’s suggestion and evaluated our selected [[144,12,8]][[144,12,8]] BB code against the Gross [[144,12,12]][[144,12,12]] code under SI1000 circuit-level noise, using 12 syndrome rounds and Relay-BP decoding.

Physical error rateOur selected BB codeGross code
0.0029.5×10−49.5\times10^{-4}0 failures / 305,152 trials
0.0032.26×10−22.26\times10^{-2}9.0×10−49.0\times10^{-4}
0.0045.4×10−15.4\times10^{-1}1.75×10−21.75\times10^{-2}

Gross achieves approximately 25× and 31× lower LER at p=0.003p=0.003 and p=0.004p=0.004, respectively.

The Gross code was specifically optimised for circuit-level performance, whereas our search targets code-capacity noise. Nevertheless, under SI1000 noise, our discovered codes outperform those obtained by the other search methods, which also optimised over code-capacity noise. We omitted training the critic model with SI1000 noise measurements because simulating a full dataset was too costly. As our method is foundational, we can in the future expand to different noise models, decoders, and code families. This therefore suggests merely that considering different noise models is important when training the critic neural network.

We thank the reviewer for suggesting those additional experiments. We hope that the Gaussian ablation, additional compute information, and SI1000 comparison clarify our contributions and the method's limitations, and we are happy to answer any remaining questions.

Thank you for the answer and the additional expmirents that you conduct.

I still think that the paper has issues:

  1. sampling from the with Gaussian distribution samples has very strong results (±91% compare to 95%) meaning that the Flow matching model has modest improvemnt.
  2. The LER is much smaller in Gross code comapre to the proposed method on SI1000 circuit-level noise

Therefore, I would like to keep my score.

AuthorsOpenReview ↗

Response to reviewers comments

Dear Reviewer J7AK,

Thank you for your time and for reviewing our additional experiments. We would like to clarify the scope of our additional results and claims.

Contribution of flow matching . We see that Gaussian sampling is already a strong baseline. Nonetheless, we would like to point out that with our flow matching model, we can increase the sampling yield to 95.90%. Furthermore, the median LER of codes generated by QECFM is lower than the lowest LER from the codes generated directly from Gaussian noise. This shows a clear advantage of using a flow matching model over sampling the code spectrum from a Gaussian distribution.

Gross code comparison. Our code was optimised for code-capacity LER. Gross performs better under SI1000, showing that our advantage in the code-capacity setting does not carry over to circuit-level noise. We acknowledge this limitation. However, when comparing against classical search methods that also optimise code-capacity LER, QECFM+DFlow produces codes with lower circuit-level LER in the reported experiments (Fig. 4b–c and Table 7). This provides evidence for the benefit of our search method, despite the gap to Gross under SI1000.

Therefore, we see our contribution as a framework combining a learned flow matching model with a code constraint readout layer, which demonstrates good results in the search domain in which it was trained, which, in our case, is code capacity noise. Adopting the critic neural network to other noise regimes leaves room for future studies. Thank you again for suggesting experiments that helped establish this distinction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewers,

Thank you for your time and constructive feedback, which has helped us improve the manuscript. Your suggestions for additional experiments have also helped clarify the contributions of the individual components. As the discussion period draws to a close, we would appreciate hearing whether you have any remaining concerns. We remain available to answer further questions and provide clarification.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.