acceptodds
Under review as a conference paper at ICLR 2027

FlowQC: Conditional Flow Matching for Efficient and Robust Quantum Circuit Synthesis

Abstract

Quantum circuit synthesis is crucial for near-term quantum computing, yet existing generative models built on diffusion models suffer from slow iterative sampling, limited conditional controllability, and high sensitivity to sampling steps and the guidance scales. To address these challenges, we propose flowQC, a conditional flow matching framework for efficient quantum circuit generation. To improve conditional generation quality, we introduce the Prompt-Aware Quantum Transformer (PAQ-former), which employs adaLN-Zero modulation and sigmoid residual gating to control the conditional signal dynamically in the generation trajectory. For dataset generation and verification, We firstly replace Qiskit's DensityMatrix representation with its Statevector implementation, enabling dataset generation to scale up to 12 qubits for large-scale training and evaluation. Experiments on conditional circuit generation tasks demonstrate that flowQC improves sampling efficiency and circuit quality over diffusion-based baselines. For SRV-conditioned circuit generation, flowQC achieves 4.74 inference acceleration and reduces generated circuit gate operations by 35.5% on average. Moreover, across varying numbers of sampling steps from 20 to 5 and guidance scales from 5 to 1, flowQC maintains more stable accuracy than the diffusion-based baseline. For unitary-conditioned generation, flowQC produces diverse candidate circuits validated by phase-invariant unitary equivalence evaluation. These results establish flowQC as an efficient and robust conditional generative framework for quantum circuit synthesis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Computing, flowmatching, synthesis of quantum circuit

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_ef2XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the conditional flow matching framework for quantum circuit generation and prompt-aware quantum transformer to improve conditional generation quality. The paper improves the generation speed and accuracy compared to prior state of the art.

Critical strengths

The paper is the first to apply conditional flow matching for quantum circuit generation, and shows strong performance improvement over other ML-based works.

Critical weaknesses

The evaluation performs only on small circuits. The experimental settings for unitary compilation is not clear. What unitary is tested for performance should be clarified. Although the paper shows good results on entanglement generation, unitary synthesis remains a more important task for quantum circuit synthesis. It is not clear the evaluation settings for other non-ML works. For example, the performance of Qiskit may depend on the random seed and different optimization passes.In addition to Qiskit, BQSKit, MQT, and Synthetic are all advanced quantum circuit synthesis tools that should be included into comparison.

Questions for authors
  1. Can you provide some background introduction to entanglement generation and explain why it is relevant or important to the quantum circuit generation?
  2. Why for unitary synthesis, the circuit size is limited to 3 qubits based on Table 7?
  3. One motivation for using ML-based method for quantum circuit synthesis is the classical method does not scale to large circuits. How do we argue that the ML-based method can have better scalability?
  4. How to justify the use of SRV for accuracy? SRV only characterize the entanglement level of a circuit but two circuits with the same entanglement level can have complete different functionality.
  5. How do we ensure the generated circuits are feasible? If the generated circuit is not correct, do we have post processing procedure to revise the circuit? For the large circuit, how does one verify the circuit correctness?
9 comments
AuthorsOpenReview ↗

Response to Reviewer ef2X (1/6)

Weakness

Reviewer concern 1: The experimental settings for unitary compilation is not clear. It is not clear the evaluation settings for other non-ML works.

Response.

Thanks for pointing out this distinction that can help our manuscript more reliable. We will add experimental setting for unitary compilation in the rebuttal version.

MethodImplementation and synthesis settings
Qiskit QSQiskit 1.4.5. We applied Quantum Shannon decomposition directly to the target matrix using qs_decomposition(U) with its default settings. The evaluator enforced the time budget.
QSearchBQSKit 1.2.1, using QSearchSynthesisPass with SimpleLayerGenerator, a native {U3,CX}\{\mathrm{U3},\mathrm{CX}\} gate set, an internal synthesis threshold of 10−1010^{-10}.
LEAPBQSKit 1.2.1, using LEAPSynthesisPass with the same layer generator, native gate set, internal threshold.
SynthetiqOfficial C++ paper implementation, commit 59a0cc6. We used the six-gate pool listed above, one search thread per target, the implementation’s fixed search seed of 0, and a first-solution stopping rule.
flowQC (ours)For each of the 272 three-qubit targets, flowQC encodes the real and imaginary parts of the full 8 × 8 unitary matrix as two input channels and conditions generation on the allowed subset of {H, CX, X, Z, CCX, SWAP}. It samples 1,024 candidate circuits per target using 20 Euler steps, a classifier-free guidance scale of 7.5, and at most 12 gate slots per circuit. A target counts as solved if at least one candidate achieves a global-phase-invariant process infidelity of at most 10⁻⁶.

Reviewer concern 2: For example, the performance of Qiskit may depend on the random seed and different optimization passes.In addition to Qiskit, BQSKit, MQT, and Synthetic are all advanced quantum circuit synthesis tools that should be included into comparison.

Response.

Then we added advanced quantum circuit synthesis tools such as BQSKit, MQT and Synthetic. The results help establish flowQC's balanced performance across reliability, sampling speed, and cir Cuit compactness to diffusion baselines and competitive QCS performance more broadly.

MethodSuccess Rate ↑Time / Target ↓CX Count ↓Circuit Depth ↓
Qiskit QS100.00%0.030 s1939
LEAP100.00%2.839 s613
Synthetiq100.00%0.132 s70106
QSearch93.38%2.680 s511
MQT69.12%0.128 s35
genQC98.16%1.573 s813
flowQC (ours)99.28%1.232 s512
AuthorsOpenReview ↗

Response to Reviewer ef2X (2/6)

Q1: Can you provide some background introduction to entanglement generation and explain why it is relevant or important to the quantum circuit generation?

Response.

We thank the reviewer for pointing out this puzzle. We will add this part to our appendix to help other readers who are confused about the field. Your confusion will greatly help us improve our manuscript understanding.

Entanglement generation is a natural task for quantum circuit design because applications often require a specific pattern of correlations. In the recent work, Roh et al. (2025) experimentally prepared a photonic cluster state with three-dimensional connectivity, a resource for fault-tolerant measurement-based quantum computation; Arlt et al. (2026) illustrated this connection by automatically generating circuits for families of entangled states, including GHZ states, and verifying their outputs at a system size beyond those used for training.

What we want to emphasize is that a target entanglement pattern can often be realized by many different circuits, yet their depths and two-qubit gate counts can vary substantially. Identifying resource-efficient circuits for a prescribed entanglement pattern is therefore an important objective in quantum circuit synthesis. A dedicated optimizer may find a particularly compact implementation, but a trained generative model can amortize the search cost across targets and sample many candidates in parallel. Within a fixed time budget, this gives us a pool of circuits to screen: we can verify which candidates realize the requested entanglement pattern and select among them for low depth or few two-qubit gates. That's the reason why we propose the flowQC method that can fasten the path of circuit synthesis.

[1] Roh, Chan, Geunhee Gwak, Young-Do Yoon, and Young-Sik Ra. “Generation of three-dimensional cluster entangled state.” Nature Photonics, vol. 19, 2025, pp. 526–532.

[2] Arlt, Sören, Haonan Duan, Felix Li, Sang Michael Xie, Yuhuai Wu, and Mario Krenn. “Meta-designing quantum experiments with language models.” Nature Machine Intelligence, vol. 8, 2026, pp. 148–157.

AuthorsOpenReview ↗

Response to Reviewer ef2X (3/6)

Q2: Why for unitary synthesis, the circuit size is limited to 3 qubits based on Table 7?

Response.

We chose three qubits to enable a direct comparison with the unitary-synthesis benchmark of Fürrutter et al. (2024), using the same gate set and 12-gate limit. Unitary synthesis imposes a stricter correctness criterion than SRV-conditioned generation: a circuit must reproduce the target transformation on every input state, up to global phase, whereas the entanglement task constrains only the output prepared from ∣0⟩⊗n|0\rangle^{\otimes n}. Scaling our current implementation is also more demanding because the unitary encoder takes a dense matrix with 4n4^n complex entries per target. Our 12-qubit results therefore apply to entanglement generation, and the scalability of unitary synthesis beyond three qubits remains to be evaluated.

We thank the reviewer for this suggestion. In future work, we will investigate how well our approach scales to unitary synthesis beyond three qubits.

[1] Fürrutter, F., Muñoz-Gil, G. & Briegel, H.J. Quantum circuit synthesis with diffusion models.Nat Mach Intell 6, 515–524 (2024). https://doi.org/10.1038/s42256-024-00831-9

Reviewer_wDW5OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces flowQC, a conditional flow-matching framework for quantum circuit synthesis, together with PAQ-former, which uses timestep-dependent adaptive normalization and residual gating for conditional generation. A statevector-based dataset pipeline enables scaling experiments to 12 qubits. The method is evaluated on SRV-conditioned entanglement generation and unitary synthesis, reporting improved generation accuracy, sampling efficiency, circuit compactness, and robustness to sampling steps and guidance scales compared with diffusion-based baselines, alongside comparisons to conventional unitary-synthesis methods and an architectural ablation.

Critical strengths

Novel application of flow matching to quantum circuit synthesis.

The paper provides, to my knowledge, the first conditional flow-matching approach for generating quantum circuits, offering a conceptually well-motivated alternative to existing diffusion-based approaches.

Strong empirical performance and improved scalability.

flowQC consistently improves conditional generation accuracy over genQC and UDiTQC and demonstrates substantially better time-to-valid-sample efficiency. The statevector-based dataset pipeline further enables evaluation up to 12 qubits, beyond the scale considered in the main prior generative approaches.

Critical weaknesses

Insufficient positioning and unclear presentation.

The related-work discussion focuses narrowly on diffusion-based QCS while omitting recent generative work [1-3], flow-matching approaches in quantum computing [4-6], and alternative learning/optimization paradigms for QCS [7-11]. More concerningly, several citations do not support the associated claims e.g., an RNA-focused CLIP paper cited for the ML CLIP method (Hafner et al. 2021), geospatial XAI for timestep-dependent conditioning (Sing et al. 2025), and Earth-observation diffusion for quantum-circuit generation (Mauro et al. 2025). This makes it difficult to reliably assess the claimed novelty and positioning. More generally, the paper is difficult to follow: important definitions and methodological/experimental details are deferred to the appendix without clear references from the main text. For example, SRV conditioning is introduced without adequately explaining what an SRV represents. The main paper should be substantially more self-contained and provide a coherent flow sufficient to understand and assess the method and experiments.

Several experimental claims are insufficiently supported or overstated.

The reported 31× “inference acceleration” combines runtime with generation success rather than measuring inference latency alone, without reporting model size or per-step computational cost. Circuit-depth improvements are claimed across test conditions despite being reported only up to 6 qubits; conventional synthesis success rates are presented without adequately accounting for documented backend failures; and the small unitary-success improvement is based on only 272 targets without repeated runs. Moreover, the diversity metric is confounded by the number of correct samples, and an important ablation of the proposed CLIP-based SRV conditioning is missing.

[1] Beaudoin, Collin, and Swaroop Ghosh. "Q-fusion: Diffusing quantum circuits." 2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI). Vol. 1. IEEE, 2025.

[2] Barta, Daniel, et al. "Leveraging diffusion models for parameterized quantum circuit generation." 2025 IEEE International Conference on Quantum Computing and Engineering (QCE). Vol. 1. IEEE, 2025.

[3] Fürrutter, Florian, et al. "Synthesis of discrete–continuous quantum circuits with multimodal diffusion models." Machine Learning: Science and Technology 7.2 (2026): 025065.

[4] Cui, Zidong, Pan Zhang, and Ying Tang. "Quantum flow matching." PRX Intelligence 1.1 (2026): 013009.

[5] Hahm, Jaehoon, et al. "Generative modeling of quantum distribution with functional flow matching." arXiv preprint arXiv:2607.00301 (2026).

[6] Dai, Jun, Michael Rizvi-Martel, and Guillaume Rabusseau. "FlowQ-Net: A Generative Framework for Automated Quantum Circuit Design." arXiv preprint arXiv:2510.26688 (2025).

[7] Kremer, David, Ali Javadi-Abhari, and Priyanka Mukhopadhyay. "Optimizing the non-clifford-count in unitary synthesis using reinforcement learning." arXiv preprint arXiv:2509.21709 (2025).

[8] Rietsch, Sebastian, et al. "Unitary synthesis of clifford+ t circuits with reinforcement learning." 2024 IEEE international conference on Quantum Computing and Engineering (QCE). Vol. 1. IEEE, 2024.

[9] Theißinger, Lukas, et al. "Towards Foundation Models for Quantum Unitary Synthesis via Zero-Shot MDL." ICLR 2026 Workshop on Foundation Models for Science: Real-World Impact and Science-First Design.

[10] Paradis, Anouk, et al. "Synthetiq: Fast and versatile quantum circuit synthesis." Proceedings of the ACM on Programming Languages 8.OOPSLA1 (2024): 55-82.

[11] Zak, Dekel, et al. "Reducing quantum circuit synthesis to# sat." arXiv preprint arXiv:2508.00416 (2025).

Questions for authors

1. Entanglement generation — efficiency and circuit compactness (clarification).

The reported up-to-31.44× inference acceleration is based on TjointT_{\mathrm{joint}}, which combines runtime with the fraction of valid/correct samples. Could you disentangle these effects by reporting raw inference latency/throughput and, ideally, model size or per-step computational cost? Furthermore, the manuscript claims reduced circuit depth across all test conditions, while Table 3 only reports results for 3–6 qubits despite evaluation up to 12 qubits. Does the reported compactness advantage persist for 7–12 qubits?

2. Unitary synthesis — generalization and conventional baselines (clarification).

How are the 272 evaluation unitaries separated from the 500k training circuits, and can you confirm that the evaluation targets (or equivalent implementations) are unseen during training? Furthermore, how should the comparison to QSearch/LEAP in Table 5 be interpreted given the appendix statement that their success rates are affected by BQSKit backend failures and should not be viewed as intrinsic algorithmic limitations? What would the comparison look like when these failures are accounted for?

3. Positioning with respect to related work (clarification).

The related-work discussion is largely restricted to genQC and UDiTQC, while recent generative QCS work, flow-matching work in quantum computing, and alternative QCS paradigms such as reinforcement learning, supervised learning, simulated annealing, and automated reasoning are largely absent. Could you clarify the relationship of flowQC to these approaches and more precisely identify which aspects of the proposed method are novel relative to the broader QCS literature?

4. Additional evidence for conditioning and comparison to established QCS methods.

Could you provide an ablation establishing the benefit of representing SRVs as text strings and encoding them with OpenCLIP, compared with a simpler categorical/learned SRV embedding? Relatedly, comparison against an established non-generative learning-based QCS approach (e.g., Synthetiq) would help establish whether the reported improvements are specific to diffusion baselines or translate into competitive QCS performance more broadly.

8 comments
AuthorsOpenReview ↗

Response to Reviewer wDW5 (1/4)

Reviewer concern. The reported up-to-31.44× inference acceleration is based on TjointT_{\mathrm{joint}}, which combines runtime with the fraction of valid/correct samples. Could you disentangle these effects by reporting raw inference latency/throughput and, ideally, model size or per-step computational cost? Furthermore, the manuscript claims reduced circuit depth across all test conditions, while Table 3 only reports results for 3–6 qubits despite evaluation up to 12 qubits. Does the reported compactness advantage persist for 7–12 qubits?

Respose.

We thank the reviewer for pointing out this distinction. The 31.44× improvement in Table 1 measures time-to-correct-samples speedup, combining sampling r Untime and the probability of producing a valid, condition-correct circuit. Tables 2 and 6 report average and per-condition accuracy, respectively. In the paper, we originally intended to compare the advantages of flowQC compared with other previous diffusion-based models through the combination of Table1,2 and 6. We report parameter counts under the same hardware, classifier-free guidance scale and sampling step that help solve your questions.

We are very grateful to review for paying attention to the issue of reduction about circuit depth, we clarify under the manuscript that with the number of qubits increases, the difference will become less obvious.At present, it only has obvious effects on quantum circuits with low qubits (3-6).

AuthorsOpenReview ↗

Response to Reviewer wDW5 (2/4)

Unitary synthesis — generalization and conventional baselines (clarification).

Reviewer concern. How are the 272 evaluation unitaries separated from the 500k training circuits, and can you confirm that the evaluation targets (or equivalent implementations) are unseen during training? Furthermore, how should the comparison to QSearch/LEAP in Table 5 be interpreted given the appendix statement that their success rates are affected by BQSKit backend failures and should not be viewed as intrinsic algorithmic limitations? What would the comparison look like when these failures are accounted for?

Respose

Thank you for pointing out this confound. Several of its suggestions are now in the revision, including description of evaluation unitaries and QSearch/LEAP results under BQSKit backend success.

Firstly, We need to clarify the 272-target benchmark. Each accepted target was exhaustively compared against the entire corpus using float64, global-phase-invariant process fidelity. All accepted targets have minimum process infidelity greater than 10−610^{-6}, with an observed minimum of approximately 0.38965. We verified that the targets do not overlap with the available 500,000-unitary corpus. This guarantee is limited to the checked corpus, and the filtered benchmark represents a different target distribution.

Second, We agree that the original QSearch/LEAP success rates should not be interpreted as intrinsic algorithmic limitations. We reran all 272 original Table 5 targets using isolated BQSKit runtimes, a 120-second online budget per target, and a process-infidelity threshold of 10−610^{-6}. QSearch achieved 254/272 successes 93.38%93.38\%, with 18 timeouts, while LEAP achieved 272/272 successes 100100%. There were no backend or evaluation errors, and all successful circuits passed independent semantic verification.

MethodSuccess Rate ↑Time / Target ↓CX Count ↓Circuit Depth ↓
Qiskit QS100.00%0.030 s1939
LEAP100.00%2.839 s613
QSearch93.38%2.680 s511
Synthetiq100.00%0.132 s70106
genQC98.16%1.573 s813
flowQC (ours)99.28%1.232 s512

↑ Higher is better; ↓ Lower is better.

AuthorsOpenReview ↗

Response to Reviewer wDW5 (3/4)

Positioning with respect to related work (clarification).

Reviewer concern. The related-work discussion is largely restricted to genQC and UDiTQC, while recent generative QCS work, flow-matching work in quantum computing, and alternative QCS paradigms such as reinforcement learning, supervised learning, simulated annealing, and automated reasoning are largely absent. Could you clarify the relationship of flowQC to these approaches and more precisely identify which aspects of the proposed method are novel relative to the broader QCS literature?

Response.

Thank you for this detailed comments. flowQC sits in the conditional-generative branch of QCS. It is not a search or optimization algorithm: one trained model amortizes candidate generation across conditions. Compared with diffusion-based circuit generators, it replaces iterative denoising with a continuous-time conditional velocity field integrated by an ODE, which makes the number of function evaluations and the guidance strength explicit inference-time controls. Compared with GFlowNet-based generators, it performs continuous vector-field regression rather than discrete trajectory-balance training. Compared with quantum flow matching, it generates a classical gate sequence rather than transporting a quantum state. Its distinctive conditioning is the Schmidt Rank Vector, an entanglement-structure condition complementary to unitary conditioning. Its advantages are therefore amortized conditional generation, inference-time controllability, and a statevector labeling pipeline that scales SRV data to 12 qubits. It provides no optimality guarantee and does not dominate per-target search, annealing, RL, or exact synthesis.

[1] Beaudoin, Collin, and Swaroop Ghosh. "Q-fusion: Diffusing quantum circuits." 2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI). Vol. 1. IEEE, 2025.

[2] Barta, Daniel, et al. "Leveraging diffusion models for parameterized quantum circuit generation." 2025 IEEE International Conference on Quantum Computing and Engineering (QCE). Vol. 1. IEEE, 2025.

[3] Fürrutter, Florian, et al. "Synthesis of discrete–continuous quantum circuits with multimodal diffusion models." Machine Learning: Science and Technology 7.2 (2026): 025065.

[4] Cui, Zidong, Pan Zhang, and Ying Tang. "Quantum flow matching." PRX Intelligence 1.1 (2026): 013009.

[5] Hahm, Jaehoon, et al. "Generative modeling of quantum distribution with functional flow matching." arXiv preprint arXiv:2607.00301 (2026).

[6] Dai, Jun, Michael Rizvi-Martel, and Guillaume Rabusseau. "FlowQ-Net: A Generative Framework for Automated Quantum Circuit Design." arXiv preprint arXiv:2510.26688 (2025).

[7] Kremer, David, Ali Javadi-Abhari, and Priyanka Mukhopadhyay. "Optimizing the non-clifford-count in unitary synthesis using reinforcement learning." arXiv preprint arXiv:2509.21709 (2025).

[8] Rietsch, Sebastian, et al. "Unitary synthesis of clifford+ t circuits with reinforcement learning." 2024 IEEE international conference on Quantum Computing and Engineering (QCE). Vol. 1. IEEE, 2024.

[9] Theißinger, Lukas, et al. "Towards Foundation Models for Quantum Unitary Synthesis via Zero-Shot MDL." ICLR 2026 Workshop on Foundation Models for Science: Real-World Impact and Science-First Design.

[10] Paradis, Anouk, et al. "Synthetiq: Fast and versatile quantum circuit synthesis." Proceedings of the ACM on Programming Languages 8.OOPSLA1 (2024): 55-82.

[11] Zak, Dekel, et al. "Reducing quantum circuit synthesis to# sat." arXiv preprint arXiv:2508.00416 (2025).

Reviewer_uv4yOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper replaces the DDPM/DDIM denoising objective used by genQC (Fürrutter et al., 2024) and UDiTQC (Chen & Tang, 2025) with independent conditional flow matching (I-CFM, a linear interpolation path with Euler sampling), while keeping the same continuous embedding of tokenized gate-level circuits. It also proposes a backbone variant called PAQ-former: a U-Net whose transformer blocks use time-conditioned adaLN modulation and apply a per-channel sigmoid gate on the self-attention and cross-attention residuals (Eqs. 4-6). The third contribution is more of an engineering one: the authors label training circuits with Schmidt rank vectors (SRV) using Qiskit Statevector rather than DensityMatrix, and use this to build balanced H/CX datasets from 3 to 12 qubits (Table 7). There are three groups of experiments. (i) SRV-conditioned entanglement generation: using a joint metric, "time per k valid circuits" (the time needed to obtain a given number of valid circuits), the speedup reaches up to 31.44× (Table 1); mean exact SRV accuracy goes from 67.67 (genQC) and 71.63 (UDiTQC) to 75.95 (Table 2); circuit depth is lower for 3 to 6 qubits (Table 3). (ii) A sensitivity sweep over 5 to 20 sampling steps and CFG scales from 1 to 5 (Table 4). (iii) 3-qubit unitary compilation with a pool of 6 gate types, comparing the number of distinct correct circuits per target (Fig. 3), and comparing success rate, runtime, CX count and depth against genQC, Qiskit QS, QSearch and LEAP (Table 5). The ablation shows a +2.4 pp gain from PAQ-former (Fig. 4).

Critical strengths
  1. The motivation is practical and the method fits the problem well. Sampling latency and sensitivity to step count and guidance scale are real problems for diffusion-based circuit generators. Switching to a straight-path flow matching objective is a reasonable, low-risk change that matches the continuous embedding genQC already uses (Sec. 3.3, Eq. 2), and it would be easy to adopt on top of the existing genQC code.
  2. The entanglement experiments cover every SRV class from 3 to 12 qubits (Table 6, with 4084 labels at 12 qubits; Table 7), which is a wider range than the at most 8 qubits handled by earlier generative work. The per-entanglement-level breakdown in Table 6 shows where the gains come from, mainly at smaller qubit counts and intermediate entanglement levels.
  3. The step-count and guidance sensitivity study (Table 4) is the most convincing result. At the same number of steps and CFG scale, flowQC loses much less accuracy among successfully decoded circuits. At a guidance scale of 5, going from 20 steps to 5 steps drops its accuracy from 0.785 to 0.675 (a relative drop of about 14%), while the diffusion baseline drops from 0.716 to 0.505 (about 29%), and flowQC is ahead in all 40 cells. This directly supports the "robust" part of the title.
Critical weaknesses
  1. The 31× speedup is not explained, and it is hard to reconcile with the claim of an "identical computational budget". Sec. 4.1 says all models use 20 sampling steps "under an identical computational budget" with the same U-Net. If so, per-sample wall-clock time should be similar, and differences in the joint time-per-valid-circuit metric should come only from the fraction of valid and correct samples (Sec. 4.2). Table 2 shows that this fraction differs by at most about 1.4× (for example 63.75 vs. 72.92 at 12 qubits), not 18× to 31×. Something else must be driving Table 1, for example:

    • a different number of network evaluations (the doubling from CFG, or overhead differences between DDIM and Euler);
    • different model sizes;
    • a much higher invalid-decoding rate for the baseline;
    • verification time being counted in the total time.

    Table 1 also does not say which diffusion baseline is used (genQC or UDiTQC). Please break the joint time metric down per method into per-sample sampling time, NFE, parameter count, invalid rate, error rate and verification time, and report the speedup against genQC and UDiTQC separately.

  2. The descriptions of the backbone, training and baselines are inconsistent, which makes it hard to tell where the gains come from. Sec. 4.1 says all models use the "same U-Net", but the same section says genQC uses "the original Spatial Transformer", and UDiTQC is by definition a different (U-DiT) architecture. Sec. 4.4 says the ablation variant uses "the original U-Net backbone", while Sec. 4.1 calls it a Spatial Transformer. Appendix A.6 gives the training configuration only for flowQC and its ablation variant (100 epochs, lr 3e-4, batch 256); parameter counts, epoch counts and tuning for genQC and UDiTQC are not reported. The variant that isolates the effect of the objective, flowQC without PAQ-former, is defined in Sec. 4.1, but its results against genQC are never shown. A single table with genQC, flowQC without PAQ (objective change only) and full flowQC (objective plus backbone), at matched parameter count, epochs and NFE, plus a diffusion model with the PAQ-former backbone, would separate the effect of flow matching from the effect of a better-tuned backbone.

  3. Standard few-step diffusion baselines are missing. The robustness claim in Table 4 compares flowQC only against DDIM. More natural comparisons are stronger few-step samplers or objectives (DPM-Solver(++), v-prediction or EDM-style samplers, rectified-flow reflow, consistency or distilled models), and, since circuits are discrete tokens, discrete-state methods such as masked or discrete diffusion and discrete flow matching. Without these, I cannot tell whether the gain comes from flow matching itself or from a weak DDIM setup. Please at least add DPM-Solver++ on the genQC and UDiTQC models, ideally also a discrete flow matching or masked diffusion baseline, and compare at 5, 10 and 20 steps.

  4. PAQ-former has limited novelty, parts of its description are technically inaccurate, and the ablation is thin.

    • The module is the adaLN modulation from DiT (Peebles & Xie, 2023), with a sigmoid of the gate parameter in place of the gate parameter itself. This is not adaLN-Zero in the DiT sense: zero-initializing the gate parameter gives a gate value of 0.5, not an identity block.
    • Zero-initializing the projection weights and bias makes the adaLN scale and shift zero at initialization for every timestep, not only at time zero as stated on p. 5.
    • The "probabilistic interpretation" of the sigmoid gate, and the argument that sign-flipped attention "lacks physical meaning" (p. 6), are both unsupported.
    • The claim that the cross-attention gate "converges to a sigmoidal decay" over time (p. 6) is not backed by evidence anywhere in the paper.
    • The ablation (Fig. 4) gives an aggregate gain of +2.4 pp with no seeds or variance reported, and it does not separate the effect of the adaLN modulation from that of the sigmoid gating.

    Please plot the learned cross-attention and self-attention gate values (after the sigmoid) over time for each layer, add separate ablations (adaLN only, gating only, standard adaLN-Zero gating), and report mean and standard deviation over at least 3 seeds.

  5. The dataset contribution is overstated. Switching SRV labeling from density matrices to statevectors (Sec. 3.5, A.5) is a routine implementation choice, and computing single-qubit reduced density matrices from a statevector is textbook material. At 12 qubits a dense density matrix (4096 by 4096 complex entries) is also not prohibitive, so the claim that this is "the key factor behind the 8-qubit ceiling" needs evidence. Also, with a gate pool of only H and CX, every training circuit is a Clifford circuit, so SRVs can be computed in polynomial time with a stabilizer simulator. This weakens both the bottleneck argument and the scalability claim. I suggest presenting this as an implementation detail, giving a wall-clock comparison of the labeling pipelines, and discussing the Clifford structure. The scalability claim would mean more with results on a non-Clifford gate pool (for example adding T gates or parameterized rotations).

  6. The unitary compilation experiment is small, may be an in-distribution test, and the classical baselines seem to be at a disadvantage.

    • The experiment uses only 3 qubits and 272 targets, drawn from the same random circuit generator as the training data (A.5.1, A.6). There is no description of deduplicating training and test targets up to unitary equivalence.
    • QSearch and LEAP should in principle be able to solve 3-qubit unitaries exactly, yet they succeed on only 38% and 47% of targets. The Fig. 5 caption attributes part of this to "backend execution failures", so Table 5 partly reflects tooling problems rather than limits of the algorithms.
    • The diversity count (Fig. 3) measures syntactically distinct sequences. It can be inflated by reordering commuting gates or inserting identity pairs (for example two consecutive H gates), which runs against the goal of producing compact circuits.

    Please deduplicate targets against the training unitaries, fix or explain the BQSKit failures (or report success rates only on runs that executed), report the diversity count after canonicalization (for example after Qiskit optimization or using a DAG canonical form), and add at least one 4- or 5-qubit setting.

  7. Statistical reporting is missing, and some claims go beyond what the data support.

    • None of Tables 1-6 or Fig. 4 give seeds, confidence intervals or the number of prompts evaluated per SRV class. At 12 qubits the balanced dataset has only 180 circuits per label, so the test set is roughly 18 per label.
    • The text says flowQC has higher exact SRV accuracy "across all evaluated system sizes". This holds for the averages, but Table 6 has cells where UDiTQC is higher (6 qubits, level 6: 0.70 vs. 0.66; 9 qubits, level 9: 0.59 vs. 0.57; 10 qubits, level 2: 0.98 vs. 0.97).
    • The 35.5% depth reduction is computed only over 3 to 6 qubits (Table 3), but the abstract states it without that qualification.
    • The standard deviation of flowQC's depth is very small (for example 1.26 vs. 11.21), which may indicate low structural diversity in entanglement generation and should be checked.

    Please add confidence intervals or multi-seed results, give the depth table for all qubit counts in the appendix, qualify the claim in the abstract, and also report a diversity metric for the SRV task.

Questions for authors
  1. In Table 1, how many network evaluations per sample (including CFG) does each method use, and what are the parameter counts? Which baseline is the denominator of the speedup? Please give the per-sample sampling time and the invalid and error rates behind each speedup at 12 qubits.
  2. How does flowQC without PAQ-former (objective change only) compare to genQC in Tables 1, 2 and 4? This is the cleanest test of whether the gain comes from flow matching itself rather than the architecture.
  3. How were the 272 unitary targets generated? Did you check that none of them is equivalent (up to global phase) to a unitary from one of the 500k training circuits?
  4. Why do QSearch and LEAP fail on more than half of the 3-qubit targets? What exactly are the "backend execution failures"? What is the success rate when the tools run correctly?
  5. Can you show the learned gate trajectories over time to support the "sigmoidal decay" claim on p. 6? Do the self-attention and cross-attention gates behave differently?
  6. Since circuits built from H and CX are Clifford circuits, did you consider stabilizer-based labeling? Do your speed and accuracy conclusions still hold for a universal gate set (for example one including T gates or parameterized Z rotations)?
  7. Table 4 reports accuracy among successfully decoded circuits. What is the decoding failure rate of each method at 5 steps? A few-step sampler could be trading validity for accuracy.
  8. Why do max_gates and the guidance scale differ between experiments (32 vs. 52; a guidance scale of 5 vs. 7.5, A.6)? Were the baselines tuned for each setting?
8 comments
AuthorsOpenReview ↗

Response to Reviewer Reviewer uv4y (1/9)

Weakness

reviewer main concern. The 31× speedup is not explained, and it is hard to reconcile with the claim of an "identical computational budget". Sec. 4.1 says all models use 20 sampling steps "under an identical computational budget" with the same U-Net. If so, per-sample wall-clock time should be similar, and differences in the joint time-per-valid-circuit metric should come only from the fraction of valid and correct samples (Sec. 4.2). Table 2 shows that this fraction differs by at most about 1.4× (for example 63.75 vs. 72.92 at 12 qubits), not 18× to 31×. Something else must be driving Table .

Response.

We thank the reviewer for drawing our attention to the reported 31× speedup. On re-examining the comparison, we found that we had evaluated flowQC, genQC, and UDiTQC under different parameter settings. We also described a metric that combines sampling time and generation success as an inference speedup. We will withdraw this claim from the manuscript and report raw sampling latency and generation success rates separately under matched inference settings.

AuthorsOpenReview ↗

Response to Reviewer Reviewer uv4y (2/9)

Q1: In Table 1, how many network evaluations per sample (including CFG) does each method use, and what are the parameter counts? Which baseline is the denominator of the speedup? Please give the per-sample sampling time and the invalid and error rates behind each speedup at 12 qubits.

Response.

Thanks for pointing out the constructive comment. In order to answer this question, we re-measure the 12-bit and entanglement level 11 where the peak is located. We fixed 32 gate slots, 20 sampling steps and CFG intensity 5 on the same RTX PRO 6000, covering all 12 ordered SRVs of this level. Each target uses 3 sampling seeds, and 64 seeds are generated each time, resulting in a total of 2,304 candidate circuits for each method. The four configurations are measured to be 20 network forward calls per batch. Each call will combine and calculate conditional and unconditional input, so each candidate corresponds to 40 single-branch equivalent network evaluations. The same number of calls does not guarantee the same running time.

MethodGenerator parametersMedian sampling latency, B = 1 (ms)Sampling time per candidate, B = 64 (ms)Invalid (%)Valid but incorrect SRV (%)Correct SRV (%)
diffusion-based model5,393,556554.7519.2018.4037.3744.23
flowQC6,972,564293.1615.2515.7231.6853.87
AuthorsOpenReview ↗

Response to Reviewer Reviewer uv4y (3/9)

Q2: How does flowQC without PAQ-former (objective change only) compare to genQC in Tables 1, 2 and 4? This is the cleanest test of whether the gain comes from flow matching itself rather than the architecture.

Thanks to the reviewer for suggesting that we directly compare the flowQC and genQC of PAQ-former. In order to distinguish the contribution of the backbone architecture and the generation method, we fixed the backbone, training data and training budget, and compared flowQC with I-CFM training and Euler sampling with genQC with diffusion training and DDIM sampling under the same number of network calls. The following table reports the proportion of all candidates that meet the target SRV and the proportion of successfully decoding circuits that meet the target SRV.

Network callsgenQC: correct / allflowQC w/o PAQ: correct / allgenQC: correct / decodedflowQC w/o PAQ: correct / decoded
553.84% ± 0.2067.82% ± 0.3356.83%69.93%
1058.01% ± 0.9174.54% ± 0.5361.52%76.64%
2061.34% ± 1.6477.32% ± 0.5764.02%79.35%

Under 5, 10 and 20 network calls, the correctness of all candidates increased by 13.98, 16.53 and 15.98 percentage points respectively, compared with the flowQC of PAQ-former compared with genQC. When only the circuit of successful decoding is examined, the corresponding improvement is still 13.10, 15.12 and 15.33 percentage points. This shows that in this evaluation, even if the same backbone is used, I-CFM/Euler can generate a higher proportion of target matching circuits under three sampling budgets.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer Reviewer uv4y (9/9)

Q8: Why do max_gates and the guidance scale differ between experiments (32 vs. 52; a guidance scale of 5 vs. 7.5, A.6)? Were the baselines tuned for each setting?

Response.

This is because the original intention of this experiment is to ensure that flowQC has a certain acceleration effect on the previous diffusion-based model under different parameters.

Yes, in each setting, baselines and flowQC have the same parameter settings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.