acceptodds
Under review as a conference paper at ICLR 2027

QaQD: Learning to design circuits for scalable sample-based quantum diagonalization

Abstract

Ground-state energy estimation (GSE) is one of the most promising applications of quantum computers for understanding quantum systems across physics, chemistry, and materials science, yet remains challenging at large scales. Recent advances in sample-based quantum diagonalization (SQD) offer a promising route to this task. However, its scalability critically depends on the sampling circuit used to construct the reduced subspace. To address this bottleneck, here we first reformulate GSE tasks as a family-level learning problem across related Hamiltonians, a setting that naturally arises in many-body physics and quantum chemistry. Building on this formulation, we develop QaQD, a learning-guided quantum architecture search framework for automatically designing effective SQD sampling circuits. At its core is a multimodal predictor that jointly learns from circuit structure, Hamiltonian information, and low-shot observations to predict high-shot SQD performance, enabling circuit-performance knowledge to transfer across related Hamiltonians. We systematically evaluate QaQD on quantum systems with up to qubits. On TFIM and XXZ benchmarks, the multimodal predictor preserves strong circuit-performance rankings using only of the reference shots per circuit, enabling QaQD to reduce the error by up to over vanilla SQD. These results demonstrate the potential of QaQD for tackling challenging quantum systems and advancing quantum computing toward practical utility.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Architecture Search, Sample-Based Quantum Diagonalization, Ground state energy estimation

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_tviMOpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This is a timely and technically interesting paper. The central idea is to learning a surrogate that ranks SQD sampling circuits from circuit structure, Hamiltonian descriptors, and low-shot outcomes. The key step is a successive-halving/evolutionary search, which is clearly described. The experiments show consistent improvements over vanilla SQD and low-shot-guided QAS. On TFIM and XXZ benchmarks, it reports strong ranking correlations, generalization to nine unseen 20-qubit TFIM Hamiltonians, and lower SQD errors than several baselines. However, the strongest scalability claims rely on per-target retraining with very large offline high-shot datasets, so the practical amortized advantage is not yet established.

Critical strengths

The family-level formulation targets an important SQD bottleneck and is conceptually novel (but there are multiple surrogate based quantumn circuit design works as of 2026). The multimodal design is sensible; ablations support the contributions of measurements, circuit information, and Hamiltonian context.

Critical weaknesses

Offline cost is substantial and excluded from headline budget comparisons. Training labels use roughly one million shots per circuit, reaching billions of simulated shots; the paper does not quantify how many target problems are needed to amortize this cost. The 10 to 40-qubit scaling, VQE, noise, and molecular studies train separate predictors for each target Hamiltonian. Thus, the strongest scalability results demonstrate transfer across circuits, not the central claim of transfer across Hamiltonian. All evidence is simulation-based. The noise study retrains on the tested noise levels, leaving robustness to real quantunm device drift or unseen noise unclear.

Several comparisons use only five seeds and then three for ablations, all comes without statistical tests. Some improvements over low-shot QAS are modest.

Questions for authors

What is the full offline-plus-online shot and classical-compute cost, and when does QAS2QD become cheaper than instance-wise search?

Can one predictor trained across Hamiltonians reproduce the 30/40-qubit, molecular, and noisy results?

How does performance change under unseen noise or real-hardware drift?

Are the conclusions robust across more seeds and paired significance tests?

8 comments

I have done a more careful reading through this article given the conflicting review scores; I have to say that I tend to agree that point 3 of B3vd is a critical flaw, and I cannot see how reviewer B3vd's concern can be addressed convincingly without more extensive experiments.

In addition, I did a quick search in surrogate-based circuit design papers recently posted (e.g., https://arxiv.org/abs/2510.00145 and https://arxiv.org/abs/2507.12256). I would tend to suggest that authors can expand these literature, make a more comprehensive comparison to make the contribution more solid. I do hope my comment could encourage authors to make this interesting work more evidenced.

We thank the reviewer for the constructive feedback on resource costs, Hamiltonian transfer, noise robustness, and statistical reliability. In response, we clarify offline-cost amortization, evaluate shared predictors on held-out Hamiltonians, and provide additional evidence from unseen-noise tests and experiments on a real quantum processor. We also expand the multi-seed evaluation and add paired significance tests to assess statistical robustness. We will incorporate these clarifications and additional results into the revised manuscript.

W1-1: Offline cost is substantial and excluded from headline budget comparisons. Training labels use roughly one million shots per circuit, reaching billions of simulated shots; the paper does not quantify how many target problems are needed to amortize this cost. Q1: What is the full offline-plus-online shot and classical-compute cost, and when does QAS2QD become cheaper than instance-wise search?

A: We thank the reviewer for highlighting the need to account for the offline investment. Below, we quantify this cost and clarify the conditions under which it can be amortized by reusing the same trained predictor across target problems.

(1) Measurement cost and amortization. Our family-level formulation separates a one-time offline investment from the measurement costs incurred for each target Hamiltonian. We compare against instance-wise high-shot SQD-guided search, which evaluates candidate circuits with high-shot measurements separately for each target. The offline dataset contains 2,000 source circuits measured with 10610^6 shots each. The same records can support multiple Hamiltonians and low-shot training inputs without additional measurements. During online search, predictor guidance scores active candidates using substantially fewer shots than the high-shot reference, which uses the same per-circuit shot budget as training-label estimation (10610^6 shots). The table below summarizes the measurement costs for the 20-qubit TFIM example.

ComponentPredictor-guided methodInstance-wise high-shot search
Offline acquisition shots, once per predictor2,000×1062,000\times 10^600
Search-guidance shots per target4.242×1064.242\times10^61.344×1091.344\times10^9
Final-evaluation shots per target5×1065\times10^65×1065\times10^6
Total measurements per target (search + final evaluation)9.242×1069.242\times10^61.349×1091.349\times10^9

For MM target Hamiltonians, the total costs are Ctotalpred=Coffline+MConlinepredC_{\mathrm{total}}^{\mathrm{pred}}=C_{\mathrm{offline}}+M C_{\mathrm{online}}^{\mathrm{pred}} and Ctotalhigh=MConlinehighC_{\mathrm{total}}^{\mathrm{high}}=M C_{\mathrm{online}}^{\mathrm{high}}. Here, ConlineC_{\mathrm{online}} denotes the per-target measurement cost, including circuit search and final SQD evaluation, while CofflineC_{\mathrm{offline}} denotes the one-time offline acquisition cost. Both methods independently evaluate five final circuits with 10610^6 shots each; their final errors are 0.198 and 0.486, respectively. Using the selected-dataset acquisition cost lower bound gives $ M>\frac{C_{\mathrm{offline}}}{C_{\mathrm{online}}^{\mathrm{high}}-C_{\mathrm{online}}^{\mathrm{pred}}}=\frac{2\times10^9}{1.349\times10^9-9.242\times10^6}\approx1.49. $ This yields an optimistic conditional projection of two target Hamiltonians, assuming the same frozen predictor continues to attain the required accuracy under the same search budget. Additional offline acquisition costs can increase the required number of targets.

(2) Classical computation. The classical costs and their frequency are summarized below.

ComponentClassical computationFrequency
Offline labelingSubspace construction and SQD diagonalization for circuit–Hamiltonian pairsOnce per offline dataset
Predictor trainingNeural-network optimizationOnce per predictor
Circuit searchEvolutionary operations and candidate selectionPer target search
Candidate evaluationLow-shot SQD diagonalization and predictor inferencePer active candidate at each search stage
Final SQD calculationSubspace construction and diagonalizationPer final evaluation

Relative to direct SQD-guided search, predictor guidance adds offline label construction, neural-network training, and recurring inference costs. Offline labeling and training are shared across target Hamiltonians, while search and final evaluation recur for each target.

W1-2: The 10 to 40-qubit scaling, VQE, noise, and molecular studies train separate predictors for each target Hamiltonian. Thus, the strongest scalability results demonstrate transfer across circuits, not the central claim of transfer across Hamiltonian. Q2: Can one predictor trained across Hamiltonians reproduce the 30/40-qubit, molecular, and noisy results?

A: We thank the reviewer for raising this distinction. We revisited the manuscript’s scaling, molecular, and noisy experiments by changing the predictor-training regime from target-specific training to joint training across Hamiltonians. The original model architecture, circuit space, target test datasets, initial populations, KK, search and evaluation procedures, measurement budgets, and paired search seeds remain unchanged. Shared predictors are frozen before target evaluation. Target-Hamiltonian labels are excluded from fitting, normalization, and checkpoint selection.

The three studies are reported separately below. Entries give mean ±\pm sample standard deviation of final SQD energy errors with five paired search seeds in units of 10−310^{-3}.

1. Scaling from 10 to 40 qubits. At each system size, one predictor is trained across Δ∈{3.5,4.5,6.5,7.5}\Delta\in\{3.5,4.5,6.5,7.5\}, validated at Δ∈{4.0,7.0}\Delta\in\{4.0,7.0\}, and evaluated at the held-out Δ=5.6\Delta=5.6, with K=1024K=1024.

QubitsTarget-specific predictorShared predictor across Hamiltonians
101.040±0.5211.040\pm0.5211.205±0.6131.205\pm0.613
200.757±0.0430.757\pm0.0430.928±0.1640.928\pm0.164
3010.037±0.64110.037\pm0.64110.763±1.64110.763\pm1.641
4023.566±0.64623.566\pm0.64624.064±1.96024.064\pm1.960

2. Molecular benchmark. For 16-qubit N$_2$ with K=256K=256, one predictor is trained across six source geometries, R∈{1.0,1.2,1.6,2.0,2.4,2.6}R\in\{1.0,1.2,1.6,2.0,2.4,2.6\} Å, validated at R∈{1.4,2.2}R\in\{1.4,2.2\} Å, and reused at the three held-out bond lengths below.

Held-out bond lengthTarget-specific predictorShared predictor across Hamiltonians
R=1.1R=1.1 Å9.400±1.5999.400\pm1.5999.013±2.3689.013\pm2.368
R=1.8R=1.8 Å19.437±15.07919.437\pm15.07911.658±3.68911.658\pm3.689
R=2.5R=2.5 Å12.013±9.77812.013\pm9.7789.085±5.7759.085\pm5.775

3. Noisy benchmark. For the noisy 20-qubit XXZ experiment, we use the same Hamiltonian split as in the scaling study and retain the original noise levels and K=1024K=1024. One predictor is trained jointly across the source Hamiltonians and noise levels, then frozen for evaluation at the held-out Δ=5.6\Delta=5.6.

Noise strength ppTarget-specific predictorShared predictor across Hamiltonians
001.326±0.3901.326\pm0.3901.286±0.2841.286\pm0.284
10−310^{-3}1.388±0.3191.388\pm0.3191.124±0.2911.124\pm0.291
3×10−33\times10^{-3}1.579±0.1151.579\pm0.1151.181±0.6421.181\pm0.642
10−210^{-2}2.082±0.2382.082\pm0.2381.767±0.5201.767\pm0.520

Yes: within each setting, a predictor trained across source Hamiltonians can be reused for the held-out targets without target-specific retraining, with one predictor per system size in the scaling study. These results extend the evidence for Hamiltonian transfer to the full 10–40-qubit scaling study, molecular geometries, and noisy circuits. We will replace the manuscript’s target-specific results in these three studies with the shared-predictor results, update the corresponding VQE comparison, and document the joint-training protocol and Hamiltonian splits.

Reviewer_B3vdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Qas$^2$QD, a learning-guided quantum architecture search framework for improving sampling circuits in sample-based quantum diagonalization (SQD). A multimodal neural predictor combines circuit structure, Hamiltonian information, and low-shot measurement outcomes to estimate high-shot SQD performance and guide circuit selection. Numerical experiments on TFIM, XXZ, and molecular systems demonstrate improvements over the evaluated baselines, with results extending to 40 qubits.

Critical strengths

The paper addresses the practical challenge of identifying informative sampling circuits under limited quantum measurement budgets. By learning to predict high-shot SQD performance from low-shot observations, the proposed framework improves circuit selection compared with conventional architecture search. The experiments demonstrate promising improvements across multiple quantum systems and measurement budgets.

Critical weaknesses
  1. The proposed framework primarily combines existing techniques, including Transformer-based multimodal prediction, quantum architecture search, evolutionary optimization, and successive halving. Although applying these techniques to SQD is useful, the methodological contribution beyond adapting established approaches appears relatively limited.
  2. The demonstrated Hamiltonian generalization is restricted to unseen parameter settings within the same Hamiltonian family, without experiments on cross-family transfer or controlled generalization across quantum phase transitions. Furthermore, the system-size scaling experiments train separate predictors for each target Hamiltonian and qubit count. Thus, these results do not establish that the learned predictor can generalize to larger systems or different Hamiltonians without retraining.
  3. The proposed method improves sampling-circuit selection but does not address the fundamental dependence of SQD on the ground state being well approximated within a manageable computational-basis subspace. When the ground state amplitudes are distributed across exponentially many significant configurations, optimizing the sampling circuit alone may not overcome this limitation. Moreover, the main experiments focus on 1D TFIM and XXZ systems, which can be studied efficiently using classical methods. In particular, the 40-qubit XXZ experiments use a strongly anisotropic regime (Δ=5.6\Delta=5.6) and physically motivated circuits, which may be particularly favorable for SQD. Consequently, the experiments do not convincingly establish whether the proposed approach remains effective for more challenging, strongly correlated or frustrated systems.
Questions for authors
  1. Could the authors provide an end-to-end resource comparison that includes offline training measurements, neural network training, architecture search, and repeated classical SQD diagonalization? In particular, how many subsequent SQD calculations would be needed to amortize the cost of generating high-shot training labels?
  2. How well does the predictor generalize when trained in one quantum phase and evaluated across a quantum phase transition? Is there any evidence of cross-family transfer, for example from TFIM to XXZ, without retraining?
  3. Since separate predictors are trained for different system sizes, can a predictor trained on smaller systems be applied to larger systems without retraining? Furthermore, Appendix C.2.2 evaluates selected circuits at different subspace dimensions KK, but does the predictor itself maintain accurate circuit rankings when KK differs substantially from its training value?
  4. How does Qas$^2$QD perform when the ground state has broad computational-basis support and requires a much larger SQD subspace?Could the authors evaluate the method on more challenging regimes (such as XXZ near Δ=1\Delta=1), TFIM near its critical point, or frustrated and nonintegrable Hamiltonians? How does its performance depend on the effective computational-basis sparsity of the ground state?
  5. How does the amount of high-shot training data required for reliable circuit-performance prediction scale with the number of qubits, circuit depth, and desired accuracy? Can the authors provide evidence that these training-data requirements remain manageable as the problem size increases?
11 comments

We thank the reviewer for the detailed comments and questions. In the point-by-point responses below, we clarify our methodological contribution and resource accounting, and present additional experiments on generalization across Hamiltonian families and quantum phases, performance on challenging Hamiltonians, and training-data requirements. We will incorporate these clarifications and additional results into the revised manuscript.

W1: The proposed framework primarily combines existing techniques ... relatively limited.

A: Our contribution is a framework for actively designing SQD sampling circuits under limited measurement budgets, together with a family-level learning formulation that enables reusable circuit evaluation across related Hamiltonians. The established tools used in our experiments constitute one concrete implementation of this framework. For example, the learned evaluator can be instantiated with predictive models other than Transformers while retaining the same learning objective.

A key practical limitation of SQD is its dependence on the prepared sampling state. If this state assigns insufficient probability to configurations important for describing the target ground state, finite-shot sampling may miss them, producing a less informative subspace and limiting the accuracy of the resulting energy estimate¹. This motivates actively identifying sampling circuits that expose informative configurations efficiently.

Adapting QAS to SQD introduces a further methodological challenge. In the supervised-classification settings of representative QAS methods, candidates are evaluated using prediction loss or validation accuracy on a prescribed dataset²˒³. In SQD, evaluation requires constructing a subspace from measured configurations and diagonalizing the projected Hamiltonian.¹ Finite-shot sampling therefore determines which configurations enter the reduced eigenvalue problem, and their usefulness depends on their joint contribution under the target Hamiltonian. Consequently, changing the sampling budget can change the evaluated subspace and need not preserve candidate rankings. Obtaining useful search feedback thus requires accounting for this budget dependence while balancing candidate exploration against the measurement cost of reliable evaluation.

Our framework addresses these challenges by coupling QAS with SQD through low-shot subspace-quality feedback. It actively searches circuit structures and parameters, supports both direct low-shot SQD-guided QAS and predictor-guided QAS, and progressively allocates measurements to promising candidates. The learned evaluator combines circuit structure, Hamiltonian information, and low-shot observations, and is trained to predict reference-shot energies and preserve their ordering across candidate circuits. The framework has two key features:

(i) Family-level learning and efficiency. The evaluator is trained across a family of related Hamiltonians and reused for new target instances within that family, without training a separate predictor for each target. This enables shared evaluation knowledge to guide target-specific circuit searches and allows offline data-acquisition and training costs to be amortized across multiple SQD tasks.

(ii) Generality beyond SQD. The framework separates circuit search and resource allocation from the QD-specific evaluation procedure, providing a design principle that can be adapted to other quantum diagonalization settings. For example, an extension to Krylov quantum diagonalization could optimize reference-state preparation circuits, with measurements and evaluation targets adapted to the projected Hamiltonian and overlap matrices of the resulting Krylov subspaces⁴.

W2: The demonstrated Hamiltonian ... without retraining.

A: We thank the reviewer for this distinction. As described in Section 3.1, our goal is to learn transferable circuit-performance knowledge within a specified Hamiltonian family, enabling predictor reuse across parameter settings. This setting naturally arises when studying ground-state properties across physical parameters and has also been explored in property prediction⁵˒⁶. We initially consider Hamiltonians with the same interaction topology but varying parameters, such as TFIM coupling and transverse-field strengths. We do not assume generalization to arbitrary families or system sizes without retraining. Cross-family or cross-phase transfer is a stronger extrapolation setting, where reliable performance cannot be assumed. Our training aims to cover the intended physical regimes, while the experiments below examine extrapolation beyond that coverage.

Generalization across Hamiltonian families. Since the original predictor already conditions on Hamiltonian parameters, our framework can be extended to accommodate different interaction families. Following the reviewer's suggestion, we extended the predictor's input encoding and conducted additional family-transfer experiments between TFIM and XXZ. Specifically, we represent both families on a shared 10-site graph and encode their Pauli coefficients in a fixed-length vector. The extended predictor also explicitly encodes continuous circuit angles.

We evaluate both transfer directions: training exclusively on TFIM and testing on XXZ, and vice versa. Each direction uses 2,048 source-family training circuits and 256 source-family validation circuits. Training, normalization, and checkpoint selection use source-family data only. The models and search protocol are frozen before generating the target Hamiltonians. During search, the target Hamiltonian is known, but its high-shot reference labels are not used for training or search decisions. The table below presents the family transfer results, evaluated with 20 search seeds. We use n=10n=10, K=64K=64, Slow=26S_{\rm low}=2^6, and Seval=215S_{\rm eval}=2^{15} shots per independent final SQD evaluation.

Transfer directionGround-state energy E0E_0Vanilla SQD: ϵ\epsilonPredictor-guided QAS: ϵ\epsilon
TFIM → XXZ−23.849.40 ± 0.983.24 ± 0.22
XXZ → TFIM−13.514.06 ± 0.700.15 ± 0.04

Here, ϵ\epsilon is the mean absolute high-shot SQD error defined in Eqn. (10) in the manuscript. For these two instances, our method reduces the mean absolute error relative to Vanilla SQD by approximately 65.5% and 96.4%, respectively. The transfer performance is asymmetric because TFIM and XXZ differ in their interactions, symmetries, and low-energy subspace structure. Consequently, circuit-performance relationships learned from one family need not transfer equally well to another. Differences in the tested target instances also contribute to the observed accuracy gap. These results highlight that successful transfer is possible, but comparable accuracy on an unseen family should not be assumed. To support multiple Hamiltonian families with a single predictor, the training set should therefore include diverse samples from the intended families and adequately cover their relevant parameter regimes.

Generalization across multiple quantum phases. Figure 4(a) evaluates a frozen TFIM predictor on unseen parameter combinations spanning both sides of the thermodynamic critical ratio, including J/h=1J/h=1, without retraining. To further examine transfer to a phase excluded from training, we additionally trained predictors exclusively in the ordered regime and evaluated them on six fresh paramagnetic TFIM Hamiltonians with N=20N=20, K=2048K=2048.

Training → evaluationSpearman correlation
Multiple phase regions → multiple phase regions0.816 ± 0.057
Ordered → paramagnetic0.402 ± 0.130

The results indicate limited ranking performance of predictor in this direct-transfer setting. Single-phase training can introduce distribution bias, so performance in an unrepresented phase cannot be guaranteed. This motivates including diverse samples from the intended physical regimes during training, as in our main experiments.

Generalization across multiple system scales. We agree that training separate predictors at each qubit count does not demonstrate transfer of a frozen predictor to larger, unseen systems. The current implementation also contains size-dependent input components, so such transfer requires a separately designed and validated extension.

Instead, our scaling experiments examine whether the framework remains effective when applied with size-specific training. Table 2 of our manuscript reports lower mean absolute SQD errors than Vanilla SQD at all tested sizes, from 10 to 40 qubits, under the stated evaluation protocols. These results support the applicability of our method to larger systems.

W3: The proposed method improves ... frustrated systems.

A: We thank the reviewer for highlighting the dependence of SQD on computational-basis support. We agree that broad ground-state support can limit the accuracy achievable with a fixed number of configurations KK. This is an inherent representation limitation of fixed-KK SQD, which also constrains our method. As discussed in our Limitations section, overcoming this limitation remains a direction for future work, potentially through joint adaptation of the sampling circuits and configuration selection.

As described in Appendix A.1, SQD involves two stages: configuration sampling and configuration selection. The first encompasses state preparation and measurement, which determine how efficiently useful configurations are observed. The second determines the configurations retained in the reduced subspace, whose size KK constrains the best achievable accuracy. Our work focuses on the first stage: using a learned predictor to guide sampling-circuit design under a limited measurement budget. With KK and the configuration-selection procedure held fixed, our aim is to improve the sampled subspace relative to vanilla SQD.

In this context, “challenging” refers to the representation difficulty faced by fixed-KK SQD, rather than a limitation specific to our circuit-design method. To characterize computational-basis support, we report reference coverage: the ground-state probability mass captured by the KK most frequent configurations in multiple reference samples. Lower coverage indicates broader support under this diagnostic.

Following the reviewer’s advice, we evaluated the following four regimes. We use the same Pauli-operator convention as the manuscript, with open boundaries.

RegimeHamiltonianParametersQubits
XXZ near Δ=1\Delta=1J∑i=1N−1(XiXi+1+YiYi+1+ΔZiZi+1)J\sum_{i=1}^{N-1}(X_iX_{i+1}+Y_iY_{i+1}+\Delta Z_iZ_{i+1})J=0.0625J=0.0625, Δ=0.80\Delta=0.8040
TFIM at its critical point−J∑i=1N−1ZiZi+1−h∑i=1NXi-J\sum_{i=1}^{N-1}Z_iZ_{i+1}-h\sum_{i=1}^{N}X_iJ=h=0.50J=h=0.5020
Frustrated J1J_1–J2J_2 modelJ1∑i=1N−1(XiXi+1+YiYi+1+ZiZi+1)+J2∑i=1N−2(XiXi+2+YiYi+2+ZiZi+2)J_1\sum_{i=1}^{N-1}(X_iX_{i+1}+Y_iY_{i+1}+Z_iZ_{i+1})+J_2\sum_{i=1}^{N-2}(X_iX_{i+2}+Y_iY_{i+2}+Z_iZ_{i+2})J1=0.0625J_1=0.0625, J2/J1=0.30J_2/J_1=0.3040
Nonintegrable field XXZJ∑i=1N−1(XiXi+1+YiYi+1+ZiZi+1)+∑i=1NgiZiJ\sum_{i=1}^{N-1}(X_iX_{i+1}+Y_iY_{i+1}+Z_iZ_{i+1})+\sum_{i=1}^{N}g_iZ_iJ=0.0625J=0.0625, nonuniform fields with max⁡i∥gi∥=0.025\max_i\|g_i\|=0.02540

For the near-critical XXZ experiment, we changed only Δ\Delta from the original value of 5.605.60 to 0.800.80. Across the additional 40-qubit experiments, we retained the original circuit pool, configuration-selection procedure, and sampling budgets, using training labels corresponding to each Hamiltonian. The field XXZ experiment uses one fixed field realization. For critical TFIM, we used the frozen predictor in Figure 2 of the manuscript. In each comparison, our method and vanilla SQD use the same KK and final measurement budget. The results are shown in following table.

HamiltonianKKReference coverage (%)Vanilla SQD: ϵ\epsilonOur method: ϵ\epsilonReduction in mean ϵ\epsilon (%)
Original XXZ, Δ=5.60\Delta=5.601,02499.07–99.140.06±0.010.06\pm0.010.02±0.000.02\pm0.0059.21
Near-critical XXZ, Δ=0.80\Delta=0.801,0240.29–0.311.15±0.021.15\pm0.021.05±0.011.05\pm0.018.64
Critical TFIM2,04847.21–47.305.24±0.405.24\pm0.403.44±0.313.44\pm0.3134.30
Frustrated Hamiltonian1,0240.41–0.431.15±0.021.15\pm0.021.07±0.001.07\pm0.007.32
Nonintegrable Hamiltonian1,0241.89–2.001.05±0.021.05\pm0.020.92±0.010.92\pm0.0112.57

The additional regimes exhibit substantially lower reference coverage than the original Δ=5.60\Delta=5.60 instance, supporting the reviewer’s observation that the original instance has relatively concentrated support. The three additional 40-qubit instances also yield substantially larger energy errors under the same KK and energy convention. Nevertheless, our method achieves a lower mean ϵ\epsilon than vanilla SQD in all four additional regimes, with reductions ranging from 7.32% to 34.30%. The remaining errors are substantial, and we do not attribute them solely to the fixed-KK representation limit, since the circuit pool and finite sampling budget also affect performance. These results support our intended contribution: improving configuration sampling within SQD while remaining subject to the representation constraints of the underlying framework.

Reviewer_2NYKOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Qas²QD, a learning-guided framework for designing sampling circuits for sample-based quantum diagonalization (SQD). A multimodal predictor combines circuit structure, Hamiltonian descriptors, and low-shot measurement statistics to predict high-shot SQD performance. This predictor guides evolutionary circuit search with adaptive measurement allocation. Experiments demonstrate improved circuit selection on spin systems up to 40 qubits, with additional molecular and simulated-noise evaluations. The 20-qubit experiments also demonstrate transfer to unseen Hamiltonians. Overall, the results support learned circuit evaluation as an effective way to improve SQD search under limited online measurement budgets.

Critical strengths
  1. A valuable formulation that addresses a consequential bottleneck in SQD. The paper treats sampling-circuit design as a learning problem and uses inexpensive observations to predict the outcome of more expensive evaluations. This is a compelling application of surrogate learning because circuit quality directly determines the usefulness of the retained subspace. The experiments on unseen 20-qubit Hamiltonians provide evidence that circuit-performance knowledge can transfer across related problems.

  2. Convincing evaluation of the predictor’s downstream value. The paper evaluates the quality of circuits found through search, rather than relying solely on prediction accuracy or rank correlation. Comparisons with direct low-shot-guided search use matched online measurement budgets. At 40 qubits, the mean SQD error decreases from approximately 0.0299 to 0.0234. Modality ablations, independent final evaluations, and paired search seeds provide complementary support for the usefulness of the learned search signal.

  3. Validation across varied settings. The molecular and simulated-noise experiments extend the evidence beyond noiseless spin-system benchmarks. Sweeps over measurement budgets and retained-subspace sizes further clarify where the method is effective, strengthening its relevance to practical SQD workflows.

Critical weaknesses

The main limitation is that total measurement savings depend on amortizing the offline training cost. A clearer accounting of this cost would strengthen the efficiency claim. The reported improvements under matched online search budgets remain convincing.

Questions for authors
  1. Given access to the same offline dataset, how do selecting the best stored circuit and warm-starting direct low-shot search compare with predictor-guided search?

  2. Can the authors state explicitly which results demonstrate transfer to unseen Hamiltonians and which use target-specific training, and align the abstract and conclusion with that distinction?

3 comments

We thank the reviewer for the positive assessment and for recognizing the value of our approach to SQD sampling-circuit design and the results across different settings. We appreciate your helpful suggestions and address each point below.

W1: The main limitation is ... convincing.

A: We thank the reviewer for recognizing the improvements under matched online search budgets and for highlighting the need to account for the offline investment. Our family-level formulation separates this investment from the measurement costs incurred for each target Hamiltonian.

The following table summarizes the completed-search accounting for the 20-qubit TFIM example, comparing predictor guidance with instance-wise high-shot SQD-guided search. The offline dataset contains 2,000 source circuits measured with 10610^6 shots each. The same records can support multiple Hamiltonians and low-shot training inputs without additional measurements.

ComponentPredictor-guided methodInstance-wise high-shot search
Offline acquisition, selected dataset2,000×1062,000\times10^600
Search-guidance shots per target4.242×1064.242\times10^61.344×1091.344\times10^9
Final-evaluation shots per target5×1065\times10^65×1065\times10^6
Search + final evaluation per target9.242×1069.242\times10^61.349×1091.349\times10^9

For MM target Hamiltonians, the total costs are Ctotalpred=Coffline+MConlinepredC_{\mathrm{total}}^{\mathrm{pred}}=C_{\mathrm{offline}}+M C_{\mathrm{online}}^{\mathrm{pred}} and Ctotalhigh=MConlinehighC_{\mathrm{total}}^{\mathrm{high}}=M C_{\mathrm{online}}^{\mathrm{high}}. Here, ConlineC_{\mathrm{online}} denotes the per-target measurement cost, including circuit search and final SQD evaluation, while CofflineC_{\mathrm{offline}} denotes the one-time offline acquisition cost. Both methods independently evaluate five final circuits with 10610^6 shots each; their final errors are 0.198 and 0.486, respectively. Using the selected-dataset acquisition cost lower bound gives $ M>\frac{C_{\mathrm{offline}}}{C_{\mathrm{online}}^{\mathrm{high}}-C_{\mathrm{online}}^{\mathrm{pred}}}=\frac{2\times10^9}{1.349\times10^9-9.242\times10^6}\approx1.49.$ This yields an optimistic conditional projection of two target Hamiltonians, assuming the same frozen predictor continues to attain the required accuracy under the same search budget. Additional offline acquisition costs can increase the required number of targets.

Q1: Given access to the ... search?

A: We conducted a supplementary comparison using the same 4,096-circuit offline archive on two 10-qubit TFIM Hamiltonians, with 20 paired search seeds per Hamiltonian. We compare four strategies:

  • Best stored circuit: select the circuit with the lowest archived high-shot mean energy for the target Hamiltonian, without further search.
  • Low-shot SQD-guided QAS: start from 64 randomly generated circuits and use direct 64-shot SQD scores to guide search.
  • Warm-start low-shot SQD-guided QAS: initialize all 64 candidates from the dataset’s top-ranked circuits and use direct 64-shot SQD scores.
  • Warm-start predictor-guided QAS: use exactly the same archive initialization as warm-start direct search, replacing its candidate scoring rule with the frozen predictor.

We thank the reviewer for this constructive suggestion. The table below reports the error of the single circuit selected before final evaluation, averaged over independent high-shot measurements. The results show a numerical benefit from warm-starting search with the offline dataset.

MethodMean energy error ϵ\epsilonMean relative energy error
Best stored circuit0.337092.470%
Low-shot SQD-guided QAS0.118310.866%
Warm-start low-shot SQD-guided QAS0.098270.720%
Warm-start predictor-guided QAS0.089940.656%

Q2: Can the authors state explicitly ... that distinction?

A: We thank the reviewer for requesting this clarification. We have replaced the target-specific results in the scaling, simulated-noise, and molecular experiments with evaluations of predictors trained across Hamiltonians and kept frozen on held-out targets. The evaluation scope of Figs. 2–4 is summarized below.

ResultsEvaluation scope
Fig. 2Generalization on unseen circuits
Fig. 3Generalization on unseen circuits
Fig. 4Generalization unseen Hamiltonians within the same family and generalization across measurement budgets

Note that, in the scaling experiments, a separate predictor is trained at each system size and shared across Hamiltonians. These results therefore evaluate within-size Hamiltonian transfer. We will make these training and evaluation scopes explicit in the revised manuscript and align the abstract and conclusion accordingly.

Thank you for the constructive discussion. After considering the concerns raised by the other reviewers and the new evidence in the rebuttal, I maintain my original rating of clear acceptance.

The central contribution is to use quantum architecture search to design sampling circuits for the subspaces they produce in quantum diagonalization. Under matched measurement budgets and subspace sizes, the new experiments show that circuit search improves SQD even on near-critical, frustrated, and nonintegrable systems. The Krylov proof of concept shows that the same design principle can guide reference-state preparation beyond SQD.

The rebuttal also demonstrates predictor reuse on held-out Hamiltonians at each tested system size. In an end-to-end hardware experiment, a simulation-trained predictor improved on vanilla SQD without device-specific retraining.

Dear Reviewer 2NYK,

Thank you for your thorough review and continued support. We appreciate your constructive feedback and are glad that our responses and new experiments have helped address your concerns and questions.

Best,

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.