acceptodds
Under review as a conference paper at ICLR 2027

Quantum architecture search for syndrome extraction in fault-tolerant quantum computation

Abstract

Quantum error correction (QEC) is a key enabler of fault-tolerant quantum computing, yet the strong code-level properties of modern quantum low-density parity-check (qLDPC) codes do not automatically translate into equally strong circuit-level performance. A central bottleneck lies in syndrome extraction, where CNOT ordering, global scheduling, and ancilla faults jointly determine error propagation and logical reliability. We introduce syndrome-extraction architecture (SEA) search, which jointly designs local CNOT orderings, global scheduling, and selective flag deployment. Focusing on Abelian two-block group-algebra (2BGA) codes, we exploit their group structure to share ordering templates across stabilizer measurements while retaining flexible flag placement. We develop a probabilistic quantum architecture search framework that samples candidate architectures, compiles them into valid global circuits using CP-SAT, and updates the architecture distribution from decoder-in-the-loop circuit-level feedback. Across 20 Abelian 2BGA codes, SEA search consistently achieves the lowest logical error rate among all baselines, including solver-based scheduling, fully flagged scheduling, and AlphaSyndrome. Moreover, the resulting circuits recover the code-level distance in 16 of the 20 evaluated instances. These results show that syndrome extraction is an architecture-level design problem and establish quantum architecture search as a practical methodology for fault-tolerant QEC circuit design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Computing, AI for Quantum, Quantum Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_odsPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a syndrome-extraction architecture search framework for fault-tolerant quantum computation. It jointly optimizes CNOT ordering, subgroup-based template sharing, and selective flag-ancilla deployment, using CP-SAT for valid circuit compilation and decoder-based feedback for architecture optimization. Experiments on 20 Abelian 2BGA codes report lower logical error rates (LERs) than the compared baselines, with circuit-level distances matching the corresponding code distances in 16 instances.

Critical strengths

S1. The paper addresses an important problem in quantum error correction by jointly optimizing several interdependent circuit-design choices rather than focusing solely on gate scheduling. This formulation is technically meaningful and motivated by fault-propagation considerations. S2. The evaluation covers 20 quantum codes and reports consistent LER improvements over the compared baselines. The additional ablations on ordering optimization and flag deployment provide useful evidence for the effectiveness of the overall framework.

Critical weaknesses

W1. Although the authors use independent Monte Carlo samples for final evaluation, the sampling procedure stops at 30 decoding failures or a maximum of (10^6) shots per memory basis. Since some reported LER improvements are relatively small, clearer uncertainty estimates are needed to establish their statistical reliability. The manuscript also reports numerical circuit-distance bounds, and it would be helpful to distinguish exactly certified distances from numerical bounds. W2. The framework combines an expanded architecture search space with decoder-guided optimization, but the experimental comparisons do not fully isolate the contribution of subgroup-based architecture learning. In particular, more controlled comparisons with fixed subgroup-sharing strategies or random architecture search under matched evaluation budgets would help determine whether the improvements primarily arise from the proposed search mechanism or from increased search flexibility and computational effort.

Questions for authors

Q1. Could the authors provide uncertainty estimates or repeated independent evaluations to clarify the statistical reliability of the reported LER improvements, particularly for cases with relatively small gains? Additionally, could the authors clarify which circuit-distance results are supported by exact certificates? Q2. Could the authors further isolate the contribution of the learned subgroup-based architecture search through matched-budget comparisons with random architecture search or fixed subgroup-sharing strategies? Addressing these concerns with convincing evidence would improve my assessment.

8 comments
AuthorsOpenReview ↗

Response to Reviewer odsP — W1/Q1 (Part 1/2)

We appreciate the reviewer's constructive suggestions, which have helped us strengthen the experimental validation of our framework.

W1&Q1: Statistical reliability of LER and exact circuit-distance certification

We thank the reviewer for highlighting the importance of statistical reliability and rigorous circuit-distance verification. We agree that the original evaluation budget of 30 decoding failures per memory basis could introduce considerable statistical uncertainty, particularly when comparing circuits with relatively small LER differences. To address these concerns, we conducted additional high-precision evaluations and exact circuit-distance calculations for all 80 final circuits across the 20 benchmark codes and four methods.

High-precision independent LER evaluation. We retain the originally selected circuits and reevaluate them using fresh Monte Carlo samples independent of the search process. The sampling target is increased from 30 to 1,000 decoding failures per memory basis, which was reached for all evaluated circuits. For failure-limited sampling, the approximate relative standard error scales as 1/F1/\sqrt{F}, where FF denotes the number of observed failures (Mendo and Hernando, 2010). Increasing FF from 30 to 1,000 therefore reduces the approximate relative standard error from 18.3% to 3.2% per memory basis. Under this substantially more reliable evaluation, our method continues to achieve the lowest observed LER among all four methods across all 20 benchmark codes. The updated average LER reductions relative to ASC, ASC+flag, and AlphaSyndrome are 31.31%, 9.82%, and 83.35%, respectively. While individual small differences should still be interpreted with their statistical uncertainty, the reevaluation confirms the overall performance advantage of our approach.

Table 1. High-precision LER results across 20 codes. All LER values in this response are in units of 10−310^{-3}.

CodeSEAASCASC+flagAlphaSyndromeReduction vs. ASCReduction vs. ASC+flagReduction vs. AlphaSyndrome
SB [[60,4,8]] w50.95081.0941.081.83413.09%11.96%48.16%
SB [[54,4,8]]1.1531.731.3698.38533.35%15.78%86.25%
SB [[60,4,9]] w70.65170.96650.662411.0632.57%1.62%94.11%
SB [[84,6,10]]0.2790.39310.34932.52929.03%20.13%88.97%
SB [[42,2,7]]2.8024.5553.24226.6738.49%13.57%89.49%
BB [[72,12,6]]1.2681.4411.3143.32912.01%3.50%61.91%
BB [[90,8,10]]0.25530.34290.281.15225.55%8.82%77.84%
BB [[108,8,10]]0.11490.1950.1271.07441.08%9.53%89.30%
BB [[144,12,12]]0.068570.1170.079630.84541.39%13.89%91.89%
GBB [[56,6,8]]0.96471.2411.0182.96922.26%5.24%67.51%
GBB [[48,4,7]]1.1851.5791.217.03624.95%2.07%83.16%
GBB [[72,4,8]]0.58750.81180.61474.42427.63%4.42%86.72%
GBB [[96,8,10]]0.12680.20520.15591.79438.21%18.67%92.93%
GBB [[32,2,6]]4.8457.3485.1521.8734.06%5.92%77.85%
GBB [[48,2,8]]1.6512.9061.85817.4843.19%11.14%90.55%
GBB [[64,2,10]]0.95881.8771.09424.1548.92%12.36%96.03%
TB [[60,4,8]]0.80370.98080.86924.1918.06%7.54%80.82%
TB [[84,6,10]]0.25570.38740.26492.00434.00%3.47%87.24%
TB [[90,8,9]]0.40780.55150.44042.54526.06%7.40%83.98%
TB [[128,20,8]]0.47810.82840.59346.25442.29%19.43%92.36%
AuthorsOpenReview ↗

Response to Reviewer odsP — W1/Q1 (Part 2/2)

Exact circuit-distance certification. We further clarify that the circuit-distance values in the original submission were obtained from heuristic searches and therefore constituted upper bounds rather than exact certificates (Pryadko et al., 2022). To strengthen this evaluation, we employ the random-window (RW) and connected-cluster (CC) search procedures implemented in dist-m4ri (official method documentation; Dumer et al., 2017). RW identifies explicit undetectable logical-fault configurations, providing constructive upper bounds, while completed CC searches exclude lower-weight logical faults and establish corresponding lower bounds. When these bounds coincide, the circuit distance is exactly certified (see also Cheng et al., 2026). The calculations use the detector error models of the same final circuits evaluated above, including syndrome and flag detectors, preparation faults, and terminal measurement faults. We evaluate both X- and Z-basis memory circuits and report the smaller distance.

The additional calculations establish exact circuit distances for all 80 final circuits. Importantly, all 20 circuit-distance values of our method are confirmed to be identical to their previously reported upper bounds, with 16 of the 20 architectures preserving the corresponding code distance. In contrast, the exact distances of ASC, ASC+flag, and AlphaSyndrome are lower than their previously reported upper bounds on 10, 4, and 8 instances, respectively. With exact distances, our method outperforms ASC and AlphaSyndrome on all 20 codes, while achieving higher circuit distance than ASC+flag on 8 codes and matching it on the remaining 12.

Table 2. Original circuit-distance upper bounds and newly certified exact distances.

CodeCode distanceSEA: original upper bound → exactASC: original upper bound → exactASC+flag: original upper bound → exactAlphaSyndrome: original upper bound → exact
SB [[60,4,8]] w588 → 86 → 68 → 85 → 5
SB [[54,4,8]]87 → 75 → 57 → 74 → 4
SB [[60,4,9]] w799 → 97 → 68 → 85 → 5
SB [[84,6,10]]1010 → 1010 → 710 → 108 → 6
SB [[42,2,7]]75 → 54 → 45 → 53 → 3
BB [[72,12,6]]66 → 64 → 46 → 64 → 4
BB [[90,8,10]]1010 → 107 → 610 → 97 → 6
BB [[108,8,10]]1010 → 108 → 710 → 109 → 7
BB [[144,12,12]]1212 → 1210 → 712 → 1010 → 6
GBB [[56,6,8]]88 → 85 → 57 → 75 → 5
GBB [[48,4,7]]77 → 76 → 67 → 74 → 4
GBB [[72,4,8]]88 → 86 → 68 → 84 → 4
GBB [[96,8,10]]1010 → 1010 → 710 → 98 → 7
GBB [[32,2,6]]66 → 64 → 45 → 53 → 3
GBB [[48,2,8]]88 → 87 → 68 → 84 → 4
GBB [[64,2,10]]1010 → 108 → 79 → 97 → 6
TB [[60,4,8]]88 → 87 → 77 → 76 → 4
TB [[84,6,10]]109 → 97 → 69 → 97 → 6
TB [[90,8,9]]99 → 99 → 89 → 95 → 5
TB [[128,20,8]]87 → 75 → 58 → 74 → 4

These additional evaluations address both concerns raised by the reviewer: the LER comparisons are now supported by substantially more precise independent measurements, and all reported circuit-distance results are supported by matching lower and upper bounds rather than heuristic estimates alone.

AuthorsOpenReview ↗

Response to Reviewer odsP — W2/Q2 (Part 1/2)

W2&Q2: Contributions of subgroup-based architecture learning and matched-budget comparisons

We thank the reviewer for this constructive suggestion. We agree that the advantages of an expanded architecture space should be distinguished from those of learning-based optimization. To address this concern, we further investigate the contributions of subgroup-based template sharing and selective flag deployment, and introduce matched-space, matched-budget random search experiments.

Search-stage ablations. We first clarify that all ablations in the original manuscript are performed during architecture search, rather than by modifying the selected circuits only at evaluation time. Specifically, the no-flag and full-flag variants fix flag deployment to zero and full coverage, respectively, while optimizing subgroup selection and CNOT ordering. The random-ordering variant samples ordering templates uniformly while retaining learned subgroup selection and flag deployment. These experiments demonstrate the importance of jointly optimizing architectural choices rather than fixing individual components.

Subgroup-based template sharing. To further examine the contribution of subgroup sharing, we conduct additional experiments with fixed numbers of sharing groups on three representative codes: SB ([[60,4,8]]), BB ([[72,12,6]]), and TB ([[84,6,10]]). Under the no-flag setting, we vary the sharing granularity from a single template per Pauli sector to independent templates for individual checks, while searching the subgroup partitions and CNOT orderings under each fixed group count. The initial screening identifies intermediate group counts of 3, 3, and 14, respectively, illustrating that neither complete sharing nor fully independent ordering is necessarily preferable. The rankings exhibit some variation under subsequent independent evaluation, indicating that the optimal sharing granularity can be sensitive to finite-shot estimation. Nevertheless, these experiments demonstrate the importance of treating template-sharing granularity as an architectural design variable.

Table 3. Fixed-group-count comparisons under the no-flag setting. LER values are from initial screening. Here, rr denotes the number of sharing groups in each X/Z check family.

CodeAvailable group counts rr, including endpointsNumber of choicesGroup count selected during screeningSingle-group LERSelected-group LERNo-sharing LER
SB [[60,4,8]] w51, 2, 3, 5, 6, 10, 15, 30830.7902000.7733520.836521
BB [[72,12,6]]1, 2, 3, 4, 6, 9, 12, 18, 36931.1378731.0066741.184824
TB [[84,6,10]]1, 2, 3, 6, 7, 14, 21, 428140.2481720.2474420.253347

Selective flag deployment. We additionally examine the effect of flag coverage while fixing the selected subgroup partitions. For each flag budget, we reoptimize flag placement and CNOT ordering before compiling the complete circuit. On the three representative codes, partial coverage uses 49/60, 63/72, and 80/84 available flags, respectively, and achieves lower independently evaluated LER than full coverage, with relative reductions of 6.76%, 4.38%, and 2.55%. These results further support the benefit of jointly considering flag overhead and ordering decisions rather than always deploying the maximum number of flags.

Table 4. Independent final evaluation of partial versus full flag coverage with fixed subgroup partitions.

CodeFixed group count rrPartial flag count (X/Z)Full flag count (X/Z)Partial-coverage LERFull-coverage LERASC+flag LERReduction vs. full coverageReduction vs. ASC+flag
SB [[60,4,8]] w5349 (26/23)60 (30/30)0.94451.0131.0806.76%12.55%
BB [[72,12,6]]363 (33/30)72 (36/36)1.2231.2791.3144.38%6.93%
TB [[84,6,10]]1480 (40/40)84 (42/42)0.24850.25500.26492.55%6.19%
Reviewer_xrK8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the improvement in circuit level performance of the quantum error correction for quantum low-density paritycheck (qLDPC) codes by proposing a syndrome extraction search algorithm that jointly addresses the cnot ordering, global X/Z scheduling and selective flag deployment. It focuses on 2BGA codes and exploits its group structure to share the ordering template between the stabilizer measurements and uses probabilistic architecture search framework to sample the architectures into valid structures using CP-SAT optimization.

Critical strengths
  1. This paper addresses the critical problem of the circuit level performance of quantum error correction codes and identifies the bottlenecks.
  2. It presents sufficient background and preliminary material required to make the ideas clear.
  3. The results show considerable improvement in the logical error rates and circult-level performance
Critical weaknesses
  1. Some of the choices in the architecture search such as flag check placement and noise model used were fixed, but there is no proper explanation provided for this choice, was there any offline experiments conducted to determine these choices. Although, ablation study shows different error rates, it is not clear if it is just for the testing or conducting architecture search. Also, how would the performance adapt if a completely different noise model is used, can the selected architecture still work well?
  2. The search algorithm exploits the group structure of 2BGA codes to to share the ordering template and reduce the search space, what is the scalability of the algorithm for the codes that do not have this group structure?
  3. How does the scalability of the algorithm change with increase in number of qubits?
  4. What is the transferrability of the idea to different types of error correction codes?
Questions for authors

Please see weaknesses

8 comments
AuthorsOpenReview ↗

Response to Reviewer xrK8 — W1 (Part 1/3)

We appreciate the reviewer's constructive suggestions, which have helped us strengthen the experimental validation of our framework.

W1: design choices, ablation settings, and noise-model robustness

choice of flag gadgets

We choose flag gadgets because they can detect harmful fault propagation with a small number of auxiliary qubits and can be incorporated into individual stabilizer measurements. Typical Shor-style syndrome extraction uses verified multiqubit cat states, while Steane- and Knill-style schemes rely on encoded ancillary blocks and their preparation. Flag circuits instead provide additional measurement information that identifies faults capable of propagating into dangerous multiqubit data errors, with relatively low ancillary-qubit overhead. This motivation follows the flag protocols of Chao and Reichardt and Chamberland and Beverland. 1, 2

This protection at the level of individual checks is well suited to our architecture-search problem. The internal form of the flag gadget is fixed, while its deployment count, placement, and associated CNOT ordering are optimized. The search can therefore balance protection against the additional operations introduced by flags using the final circuit-level LER. Related flag-sharing work also studies how deployment and sharing affect resource use, and experiments on diamond and superconducting platforms have demonstrated the role of flag information in fault identification and error correction. 3, 4, 5

choice of noise model

We use SI1000 because it assigns different relative error strengths to gates, preparation, measurement, and idle operations, reflecting the differing noise costs of these operations in superconducting circuits. These distinctions are directly relevant to our optimization: deploying flags adds gates and measurements, while changing the schedule affects idle exposure. SI1000 allows these costs to enter the LER evaluation of the complete circuit, and all methods use the same noise model for a consistent comparison. 6

SI1000 has also been used in several QEC studies, including the simulated pretraining of AlphaQubit, circuit-level evaluation of Yoked surface codes, and comparisons of BB-code decoders in Decoding correlated errors in quantum LDPC codes. 7, 8, 9 We further evaluate the selected circuits under a uniform noise model below to examine whether their performance advantage persists under different relative operation error rates.

AuthorsOpenReview ↗

Response to Reviewer xrK8 — W1 (Part 2/3)

search-stage ablations

All ablations in the original manuscript are imposed during architecture search. Each condition is searched separately, followed by candidate selection and independent evaluation of the selected circuit. The no-flag variant fixes all flag decisions to zero while optimizing subgroup partitioning and CNOT ordering. The full-flag variant deploys flags on all checks while optimizing the same remaining choices. The random-ordering variant samples ordering templates randomly while retaining learned subgroup selection and flag deployment. The full method jointly optimizes these choices. In every condition, CP-SAT generates a valid global schedule for the resulting architecture. 10

We conducted additional experiments to examine the contribution of flag coverage. We first searched the available group counts under the no-flag setting, optimizing the subgroup partition and ordering templates for each count. Here, rr denotes the number of sharing groups within each X/Z check family. The two endpoints are one shared template per Pauli sector and independent templates for individual checks. Table 1 reports the initial screening results used to select the parent circuits for the flag-count experiments.

Table 1. Initial screening of subgroup-sharing granularity under the no-flag setting. All LER values in this response are in units of 10−310^{-3}.

CodeAvailable group counts rr, including both endpointsNumber of choicesScreening-selected rrOne-group LERSelected-grouping LERNo-sharing LER
SB [[60,4,8]] w51, 2, 3, 5, 6, 10, 15, 30830.7902000.7733520.836521
BB [[72,12,6]]1, 2, 3, 4, 6, 9, 12, 18, 36931.1378731.0066741.184824
TB [[84,6,10]]1, 2, 3, 6, 7, 14, 21, 428140.2481720.2474420.253347

The grouping rankings changed under subsequent independent reevaluation, so the screening-selected counts should not be interpreted as independently established optimal group counts. The flag-count experiments retain the three parent circuits selected at this screening stage.

We then fix the subgroup partitions and template-sharing relationships and reoptimize flag placement and CNOT ordering for each flag budget, followed by global scheduling. The selected partial- and full-coverage circuits are frozen and evaluated independently alongside ASC+flag with a higher sampling budget. For each X/Z memory basis, sampling stops at either 1,000 decoding failures or 10710^7 shots, whichever occurs first. The counts in parentheses in Table 2 give the numbers of flags on X/Z checks.

Table 2. Independent final evaluation of partial and full flag coverage with fixed subgroup partitions.

CodeFixed group count rrPartial-coverage flags (X/Z)Full-coverage flags (X/Z)Partial-coverage LERFull-coverage LERASC+flag LERReduction vs. full coverageReduction vs. ASC+flag
SB [[60,4,8]] w5349 (26/23)60 (30/30)0.94451.0131.0806.76%12.55%
BB [[72,12,6]]363 (33/30)72 (36/36)1.2231.2791.3144.38%6.93%
TB [[84,6,10]]1480 (40/40)84 (42/42)0.24850.25500.26492.55%6.19%

Partial coverage achieves lower independently evaluated LER than both full coverage and ASC+flag on all three codes, with reductions of 2.55%–6.76% relative to full coverage. These results support treating flag count and placement as architecture-search variables: even after the sharing partition has been selected, optimizing flag coverage together with ordering can improve circuit performance.

AuthorsOpenReview ↗

Response to Reviewer xrK8 — W1 (Part 3/3)

robustness across noise models

Figure 7 of the original manuscript already evaluates five codes over multiple physical error rates under SI1000, with the full method achieving the lowest LER at every tested point. We now extend this evaluation to two noise models on SB [[54,4,8]], TB [[128,20,8]], and NA-2BGA [[42,2,7]]. This gives three codes, two noise models, and five physical error rates, for a total of 30 conditions and 90 LER results across SEA, ASC, and ASC+flag.

For every method, we freeze the circuit selected under SI1000 at p=0.003p=0.003. Subgroup partitioning, CNOT ordering, flag deployment, and scheduling remain fixed throughout this evaluation. We rebuild the detector error model and decoder priors for each new noise condition before estimating LER. The uniform model uses the four settings in Table 3, applied consistently to data, syndrome, and flag qubits. Idle noise includes waiting locations during measurements or resets of other qubits.

Table 3. Uniform circuit-level noise model.

Noise locationUniform setting
Single-qubit gates and idle locationsSingle-qubit depolarizing noise with total probability pp; each of X,Y,ZX,Y,Z occurs with probability p/3p/3.
Two-qubit gatesTwo-qubit depolarizing noise with total probability pp; each of the 15 nonidentity two-qubit Pauli errors occurs with probability p/15p/15.
Preparation/resetError probability pp.
MeasurementMeasurement-outcome flip probability pp.

SEA achieves lower LER than both baselines in all 30 conditions, including every tested point under the uniform model. Relative to ASC+flag under uniform noise, the LER reductions are 24.97%–38.66% for SB [[54,4,8]], 58.38%–64.55% for TB [[128,20,8]], and 4.01%–44.17% for NA-2BGA [[42,2,7]]. These results show that the advantage of the SI1000-selected architectures persists across the tested physical error rates and the change to uniform operation error strengths.

Table 4. Evaluation of fixed SI1000-selected circuits under SI1000 and uniform noise.

CodeppSI1000: SEASI1000: ASCSI1000: ASC+flagUniform: SEAUniform: ASCUniform: ASC+flag
SB [[54,4,8]]0.0010.0073690.020840.0096430.0046060.011950.007439
SB [[54,4,8]]0.0020.16580.3310.20950.12490.17180.1703
SB [[54,4,8]]0.0031.1531.731.3690.88530.92521.18
SB [[54,4,8]]0.0044.4956.0065.0982.9223.5114.413
SB [[54,4,8]]0.00511.8214.3513.967.1589.88811.67
TB [[128,20,8]]0.0010.0020770.0061820.0029210.0014980.0053650.004117
TB [[128,20,8]]0.0020.055250.11220.071140.038680.091180.1091
TB [[128,20,8]]0.0030.47810.82840.59340.34290.67650.9322
TB [[128,20,8]]0.0042.4143.9032.9991.8042.8754.334
TB [[128,20,8]]0.0057.90610.639.6235.7668.08613.95
NA-2BGA [[42,2,7]]0.0010.017230.029040.018050.01030.010490.01073
NA-2BGA [[42,2,7]]0.0020.30890.46740.3260.15270.18380.2037
NA-2BGA [[42,2,7]]0.0031.7852.4071.9570.79931.1021.188
NA-2BGA [[42,2,7]]0.0046.0027.5726.642.5983.9674.215
NA-2BGA [[42,2,7]]0.00515.3216.6816.626.0469.97610.83
Reviewer_FF5XOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a probabilistic quantum architecture search framework for optimizing syndrome-extraction circuits in fault-tolerant quantum computing. The approach jointly optimizes CNOT ordering, subgroup-based template sharing, and selective flag-qubit placement for Abelian two-block group-algebra (2BGA) quantum codes. It uses a policy-gradient method to learn a probability distribution over candidate architectures, while a CP-SAT solver constructs valid global circuit schedules. Candidate architectures are evaluated using circuit-level simulations and decoder feedback.

Experiments on 20 quantum codes show that the proposed method achieves lower logical error rates than the considered baselines, including ASC, fully flagged ASC, and AlphaSyndrome. The results also demonstrate improvements in circuit-level distance and reductions in the number of flag ancillas.

Critical strengths

The paper addresses an important implementation challenge in quantum error correction. Optimizing syndrome-extraction circuits can improve logical reliability by reducing the impact of error propagation during stabilizer measurements.

The integration of CNOT ordering, scheduling, and selective flag placement is interesting. The experimental results demonstrate consistent improvements over the considered baselines, suggesting that jointly considering these design decisions is beneficial.

Critical weaknesses
  1. Insufficient motivation for using machine learning instead of conventional combinatorial optimization.

The problem considered in this paper is fundamentally a discrete combinatorial optimization problem involving CNOT ordering, subgroup partitioning, and flag-qubit placement. The authors employ a probabilistic architecture search method based on policy gradients, but it is unclear why machine learning is necessary or particularly advantageous for solving this problem. Conventional combinatorial optimization methods, such as simulated annealing, genetic algorithms, or local search, could potentially explore the same architecture space.

The paper does not compare its learning-based approach against these alternatives under equivalent computational budgets. Consequently, it is difficult to determine whether the reported improvements result from the proposed learning framework or simply from exploring a richer architecture space. A systematic comparison with conventional combinatorial optimization methods would substantially strengthen the paper.

  1. Limited novelty in the machine learning methodology.

The machine learning component largely relies on established techniques, including categorical sampling, Bernoulli sampling, and REINFORCE-style policy-gradient optimization. The paper does not introduce a new learning algorithm, model architecture, or optimization principle. Its main methodological contribution appears to be applying existing learning techniques to a specialized quantum circuit optimization problem.

Furthermore, the model is trained separately for each quantum code, without demonstrating generalization or transfer learning across different codes. This raises questions about the broader machine learning contribution, particularly for a conference such as ICLR.

  1. Narrow problem scope and limited broader significance.

The paper focuses on syndrome-extraction circuit optimization for Abelian 2BGA quantum codes, with fixed stabilizer structures and a prescribed flag gadget. Although this is a useful practical optimization problem, its scope is relatively specialized.

More fundamental challenges in quantum error correction include designing new quantum codes with improved distance, encoding rate, fault tolerance, and resource efficiency. The proposed approach does not address these code-design problems, nor does it demonstrate that its learning framework can be readily extended to broader quantum code families or more general fault-tolerant circuit architectures.

As a result, the broader significance and general applicability of the proposed framework remain insufficiently established.

Questions for authors

Why is policy-gradient-based machine learning preferable to conventional combinatorial optimization for this problem? Can the authors compare against simulated annealing, genetic algorithms, or local search using the same architecture space and evaluation budget?

What is the fundamental machine learning novelty beyond applying established probabilistic sampling and REINFORCE optimization techniques? Can the learned model generalize or transfer to previously unseen quantum codes?

How readily can the proposed approach be extended beyond Abelian 2BGA codes and the specific flag-measurement architecture considered in this work?

8 comments
AuthorsOpenReview ↗

Response to Reviewer FF5X — W1/Q1 (Part 1/2)

W1&Q1: motivation for learning-based search and comparison with conventional optimization

We thank the reviewer for this important question. We agree that the motivation for adopting learning-based optimization should be supported by both the characteristics of the underlying problem and direct comparisons with conventional search methods. We address these two aspects below.

Challenges of conventional optimization. Syndrome-extraction architecture search is a tightly coupled, multi-stage combinatorial optimization problem. Subgroup partitioning determines the number of shared CNOT ordering templates and their associated constraints, while CNOT ordering and flag deployment jointly influence fault propagation and global scheduling. Consequently, optimizing these components independently does not necessarily produce the best overall architecture. More importantly, the final logical error rate (LER) is obtained through circuit simulation and decoding, making it an expensive black-box objective that cannot be directly expressed as a conventional constraint-programming objective. While solvers such as CP-SAT are effective for enforcing circuit validity and optimizing scheduling depth, directly optimizing the complete architecture for LER remains challenging.

Motivation for learning-based optimization. Learning-based approaches have demonstrated considerable potential in complex combinatorial design problems, particularly when design decisions are interdependent and the final objective is difficult to optimize analytically. Representative examples include machine-learning-based electronic design automation (Huang et al., 2021) and chip placement (Mirhoseini et al., Nature, 2021). Motivated by these developments, we adopt policy-gradient optimization to learn a probability distribution over syndrome-extraction architectures. This approach allows us to directly incorporate circuit-level LER as optimization feedback and jointly update subgroup selection, CNOT ordering, and flag deployment without differentiating through the CP-SAT compiler, circuit simulator, or decoder. Our framework therefore combines learning-based architecture exploration with conventional constraint optimization for valid circuit compilation, rather than replacing the latter.

AuthorsOpenReview ↗

Response to Reviewer FF5X — W1/Q1 (Part 2/2)

Matched-space and matched-budget comparisons. To empirically validate this choice, we conducted additional experiments with three conventional search methods: random search (RS), simulated annealing (SA), and genetic algorithms (GA). All methods explore the same architecture space, including subgroup partitioning, CNOT ordering, and selective flag deployment. Each method evaluates the same budget of 5,120 candidates per code, using identical circuit compilation, screening, and final evaluation protocols. Across the original 20 benchmark codes, these comparisons involve 60 additional search experiments. The selected circuits are independently evaluated using fresh Monte Carlo samples, targeting 1,000 decoding failures per memory basis.

Table 1. Matched-space and matched-budget comparisons across 20 benchmark codes. LER values in both tables are reported in units of 10−310^{-3}. Relative reductions are calculated as 100(1−SEA/baseline)100(1-\mathrm{SEA}/\mathrm{baseline}).

CodeOur methodRSSAGAReduction vs. RSReduction vs. SAReduction vs. GA
SB [[60,4,8]] w50.95081.0600.99181.04310.30%4.13%8.84%
SB [[54,4,8]]1.1531.2801.2901.2519.92%10.62%7.83%
SB [[60,4,9]] w70.65170.75700.72810.686213.91%10.49%5.03%
SB [[84,6,10]]0.27900.30460.33670.32608.40%17.14%14.42%
SB [[42,2,7]]2.8023.7563.8493.29425.40%27.20%14.94%
BB [[72,12,6]]1.2681.4191.4291.37510.64%11.27%7.78%
BB [[90,8,10]]0.25530.33400.28010.261123.56%8.85%2.22%
BB [[108,8,10]]0.11490.14370.12890.122920.04%10.86%6.51%
BB [[144,12,12]]0.068570.085060.077890.0784119.39%11.97%12.55%
GBB [[56,6,8]]0.96471.3071.0240.976626.19%5.79%1.22%
GBB [[48,4,7]]1.1851.2051.3961.3991.66%15.11%15.30%
GBB [[72,4,8]]0.58750.68220.62470.625713.88%5.95%6.11%
GBB [[96,8,10]]0.12680.16810.15090.136024.57%15.97%6.76%
GBB [[32,2,6]]4.8455.6735.2795.02514.60%8.22%3.58%
GBB [[48,2,8]]1.6512.4151.9672.04331.64%16.07%19.19%
GBB [[64,2,10]]0.95881.2291.2981.25221.99%26.13%23.42%
TB [[60,4,8]]0.80370.88590.85700.82259.28%6.22%2.29%
TB [[84,6,10]]0.25570.28200.26640.27999.33%4.02%8.65%
TB [[90,8,9]]0.40780.53970.51130.426224.44%20.24%4.32%
TB [[128,20,8]]0.47810.66100.55800.599827.67%14.32%20.29%

Our method achieves lower independently evaluated LER than all three conventional baselines across all 20 benchmark codes. The relative LER reductions range from 1.66% to 31.64% against RS, 4.02% to 27.20% against SA, and 1.22% to 23.42% against GA. Since all methods operate within the same architecture space and candidate-evaluation budget, these results provide direct empirical evidence that the performance gains arise not merely from increased design flexibility, but also from the effectiveness of learning-guided optimization within that space.

AuthorsOpenReview ↗

Response to Reviewer FF5X — W2/Q2

W2&Q2: scientific and methodological contributions and cross-code transferability

We appreciate the reviewer's comments on the machine learning novelty and broader significance of our work. We acknowledge that our approach does not introduce a new general-purpose learning algorithm. However, we believe that the scientific contribution of this work lies in formulating and addressing an important problem in fault-tolerant quantum computing, rather than modifying established policy-gradient algorithms.

Problem formulation and scientific contribution. A quantum error-correcting code with strong algebraic properties does not necessarily achieve comparable protection at the circuit level, because syndrome-extraction operations introduce additional faults and error propagation. Addressing this gap requires designing the extraction architecture rather than merely optimizing its scheduling. Our work formulates this as a joint architecture-search problem involving subgroup-based template sharing, CNOT ordering, and selective flag deployment. These choices are interdependent: template sharing determines the ordering parameterization, while ordering and flag placement jointly affect fault propagation, scheduling constraints, and logical reliability. By integrating a code-structure-aware architecture representation, probabilistic search, constraint-aware compilation, and decoder-in-the-loop evaluation, we provide an automated solution to this problem. The contribution therefore lies in the problem formulation, architecture representation, and experimentally validated solution framework, rather than a new generic optimization rule.

Relevance to ICLR. We would also like to emphasize that ICLR explicitly welcomes applications to physical sciences, including physics, in its 2027 Call for Papers. More importantly, the ICLR 2027 Reviewer Guidelines recognize improving solutions to established application problems and identifying new application problems as legitimate research objectives, alongside theoretical and algorithmic advances. We believe our work fits this scope by introducing a meaningful architecture-design formulation for fault-tolerant QEC and demonstrating that learning-based optimization can address it effectively. The significance of such application-driven research should be assessed through its scientific contribution, methodological design, and experimental evidence, rather than solely through the novelty of its underlying general-purpose learning algorithm.

Cross-code transferability. Our framework currently optimizes a separate architecture distribution for each target code, and we do not claim that a trained policy transfers directly to previously unseen codes. This design reflects the nature of the task: different quantum codes have different stabilizer supports, measurement structures, and legal circuit configurations, leading to different architecture spaces. Our objective is to discover a high-quality, fixed syndrome-extraction circuit for a given code, rather than to train a universal policy that must generalize across codes. We therefore do not view cross-code policy transfer as a prerequisite for demonstrating the value of code-specific architecture optimization.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses and Additional Experiments

Dear Area Chair,

We thank you and the reviewers for the constructive feedback. We appreciate that Reviewer odsP recognized the importance of the problem, describing it as “important,” and Reviewer xrK8 described it as “critical.” Reviewer FF5X also found the integration of CNOT ordering, scheduling, and selective flag placement “interesting.”

We have provided point-by-point responses to the nine numbered concerns raised by the three reviewers and supplemented our study with seven sets of additional experiments and evaluations. These cover matched-budget optimizer comparisons, subgroup sharing, flag coverage, additional code families, noise robustness, high-sampling LER reevaluation, and exact circuit-distance certification. Specifically, the additions include 60 matched-budget optimizer searches, evaluations on eight additional code instances and across 30 noise conditions, and high-sampling LER reevaluation and exact distance certification for all 80 original final circuits. We have also incorporated the corresponding results and clarifications into the revised manuscript and uploaded the updated PDF.

With approximately 50 hours remaining in the discussion period, we are eager to hear the reviewers’ feedback on whether our responses have addressed their concerns. If any questions remain, we are happy and ready to provide further answers and clarification.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.