Pith. sign in

REVIEW 5 major objections 6 minor 37 references

Leveraging Diffusion Models for Parameterized Quantum Circuit Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A text-conditioned diffusion model can generate complete parameterized quantum circuits—gate layout and rotation angles together—and demonstrates high-fidelity GHZ state preparation and accurate classification without iterative parameter…

desk verdict A plausible but under-evidenced extension of diffusion-based circuit synthesis to continuous gate parameters; the missing ablation on whether the angles actually matter is the load-bearing gap. read the letter →

arxiv 2505.20863 v3 pith:TS5AGASF submitted 2025-05-27 quant-ph cs.LG

Daniel Barta , Darya Martyniuk , Johannes Jung , Adrian Paschke This is my paper · ORCID
classification quant-phcs.LG MSC 81P6868T07 PACS 03.67.Lx
keywords parameterizedquantumcircuitsdiffusionmodelsarchitecturesearchgenerativecircuitsynthesisGHZstatepreparationmachinelearningclassifier-freeguidancetext-conditionedgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a single denoising diffusion model can generate a complete parameterized quantum circuit, including both the discrete choice of gates and the continuous rotation angles, directly from a short text prompt and without iterative parameter optimization. The payoff would be a way to design circuits for tasks like state preparation and quantum machine learning that avoids the costly simulation-in-the-loop search used by most quantum architecture search methods. The authors demonstrate the idea on 3-qubit GHZ state preparation with two gate alphabets and on a linear classification task, where the model's outputs track the requested fidelity or accuracy when guidance is increased. They also report that a model trained only on 3-qubit circuits can produce valid 4-qubit circuits, and that inference time stays roughly constant as qubit count grows, which supports the scalability claim. The central caveat is that the model's quality depends on the SQuASH training circuits and on a cosine-similarity decoding step that shows rising error rates at larger sizes.

What carries the argument

The load-bearing object is a joint tensor encoding: one axis indexes qubits, one indexes discrete circuit time steps, and the last stores a learned embedding of the gate type concatenated with its continuous parameter, currently at most one normalized parameter per gate. The same tensor is the diffusion model's data representation, its noising target, and its output, so the reverse process produces architectures and parameters together. The conditioning machinery is a pre-trained CLIP text encoder feeding cross-attention layers of a U-Net, with classifier-free guidance mixing conditional and unconditional predictions at inference; decoding then uses cosine similarity to snap each generated embedding to the nearest known gate identity. Together these pieces let a single forward pass answer what gates, in what order, and with what angles, for a given prompt.

What would settle it

A decisive test would be to train the same pipeline on SQuASH data with each circuit's fidelity and accuracy labels randomly permuted before the prompts are assigned: if the model still produces circuits whose measured fidelities track the requested targets, the conditioning signal is not driving the result, whereas if it does not, the method's power depends on labels that may not exist for new tasks.

Watch

Extended reading notes

Core claim

The central claim is that the two problems that quantum architecture search usually solves in stages, choosing where gates go and then tuning their parameters, can be solved jointly by a generative model. The pipeline encodes a parameterized quantum circuit as a tensor that stacks a discrete gate-identity embedding with a continuous parameter channel, diffuses noise into that tensor, and learns a U-Net to reverse the process under a CLIP-embedded text prompt. At inference, the denoised tensor is decoded by matching each position's embedding to the nearest known gate and reading off the parameter value. On the GHZ task with gate set gs1 and high guidance, nearly all synthesized circuits reach perfect fidelity while retaining structural diversity; on the hardware-native gate set gs2 the gains are more limited and higher guidance raises decoding errors. For the QML task, the model produces circuits clustered around requested accuracy targets such as 0.2 and 0.8. The authors interpret the emergence of layer-like ansatze, alternating single-qubit rotations with entangling gates, as evidence that the model has learned reusable circuit-design heuristics rather than only memorizing samples.

Load-bearing premise

The load-bearing premise is that the SQuASH circuits, collected by other search algorithms and described in a manuscript under submission, are representative enough that a diffusion model trained on them learns a general recipe for turning a text prompt into a good circuit, and that the cosine-similarity decoding step turns the model's tensor outputs into valid circuits; the paper's own scaling tables show this decoding step failing more often as circuits grow.

Editorial extensions

If this is right

  • A user can steer the same trained model to different fidelity or accuracy targets by changing the prompt and the guidance scale, without retraining.
  • Because generation time stays nearly constant as qubit count increases, the approach could replace some search-based quantum architecture search runs whose optimization loops require repeated quantum simulation.
  • The model's zero-shot transfer from 3 to 4 qubits suggests that generative circuit priors can extrapolate to sizes not seen in training, although error counts rise.
  • High-fidelity outputs are not a single memorized circuit: the model produces structurally distinct architectures and distinct parameter sets, which is useful for noise-aware exploration.
  • The emergence of layer-like single-qubit-rotation and entangling-gate ansatze indicates the model captures a transferable design heuristic applicable beyond the exact training circuits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the scaling limit of this approach is currently set by the decoding step, not the diffusion model: the paper's own scaling tables show decoding errors rising sharply at 4 to 5 qubits and at deep gate counts, so replacing cosine-similarity decoding with a learned parameter decoder is a natural next step.
  • A direct test of whether the model generalizes rather than interpolates would be to train on circuits produced by one search algorithm and evaluate on circuits from a different search algorithm with the same gate set; if performance drops sharply, the model is capturing dataset-specific patterns rather than a universal circuit prior.
  • Conditioning prompts could be extended to include noise rates or device calibration parameters, making the generator output hardware-aware circuits; nothing in the current pipeline prevents adding such scalar conditions to the CLIP prompt, but the paper does not test it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes a text-conditioned denoising diffusion model that generates parameterized quantum circuits (PQCs) in a single pass, producing both the discrete gate architecture and the continuous gate parameters. The authors extend the discrete-only diffusion pipeline of Fürrutter et al. by concatenating a continuous parameter tensor to the discrete gate embedding, conditioning the U-Net with CLIP text embeddings, and using classifier-free guidance to steer generation toward user-specified fidelity or accuracy targets. They evaluate the method on 3-qubit GHZ state preparation with two gate sets (gs1 and gs2), on zero-shot generalization to 4- and 5-qubit GHZ circuits, and on a QML linear classification task, reporting fidelity/accuracy histograms, uniqueness metrics, and inference-latency scaling tables. The paper claims that the approach bypasses iterative parameter optimization and quantum simulation during inference and exhibits strong generalization across gate sets and qubit counts.

Significance. If the claims are substantiated, the paper would make a useful contribution to quantum architecture search by demonstrating that a generative model can propose complete, deployable PQCs without the usual inner-loop parameter optimization. The authors provide a concrete pipeline with two task demonstrations, a zero-shot scaling experiment, and an inference-efficiency analysis, which are valuable empirical starting points. The paper is also honest about limitations, including rising decoding errors at larger sizes and the reliance on a purpose-built dataset. However, the central novelty—simultaneous generation of continuous parameters—is not isolated in the evaluation, and the absence of baselines and statistical error bars leaves the main claims under-supported as they stand.

major comments (5)
  1. [Section IV] The results contain no comparison against any baseline, including random search, evolutionary search, reinforcement learning, the generative quantum eigensolver (GQE, Ref. [30]), or even the discrete-only diffusion pipeline of Ref. [1]. The abstract and Section V claim computational efficiency and advantages over search-based methods, but without baseline quality/cost comparisons the reader cannot judge whether the proposed method improves on existing approaches. This is load-bearing because the paper's stated motivation is to overcome the bottlenecks of search-based QAS.
  2. [Section III-A and Section IV-B] No metric isolates whether the generated continuous parameters are actually meaningful. All reported success measures (fidelity distributions, accuracy histograms, 'unique hash') are end-to-end circuit properties. Since a GHZ state on 3 qubits can be prepared with parameter-free gates (e.g., H followed by CNOT/CX chains) achieving fidelity exactly 1, the high-fidelity counts in Fig. 2 and Tables I-II could be driven entirely by the discrete topology, with the rotation angles being irrelevant or even wrong. I request a parameter-level evaluation: for example, compare the fidelity of the generated circuit against the same architecture with random or zero-filled parameters, or report the error between decoded rotation angles and the angles of the nearest valid training circuit. Without this, the claimed advance over discrete-only diffusion is not established.
  3. [Section IV-A] The training data come from SQuASH, cited as Ref. [33], which is the authors' own manuscript under submission. The post-hoc balancing ratios (75/25 for gs1, 90/10 for gs2, 50/50 for the ML task) are introduced without justification or sensitivity analysis, and the exact dataset split is not released. Because the benchmark construction is not independent of the method, and the balancing choices could inflate apparent performance, the paper should either release the exact datasets and splits or evaluate on an independently constructed benchmark, and should justify the chosen ratios or show that results are robust to them.
  4. [Section IV (Figs. 2-3, Tables I-II)] All quantitative results appear to be single realizations with no error bars, confidence intervals, or repeated-seed analysis. For instance, Tables I and II report single counts (e.g., High Fid. Count = 82) without variance across diffusion sampling seeds, and the histograms in Figs. 2-3 do not state the number of seeds or runs used to produce the 500 or 250 samples. Given that diffusion sampling is stochastic, I cannot assess the statistical reliability of the guidance-tradeoff and scaling conclusions. Please report means and standard deviations over multiple independent sampling runs, or explicitly state that the shown histograms aggregate all samples from a single generative pass and justify that choice.
  5. [Abstract and Table I] The abstract claims 'strong generalization across varying gate sets and scaling qubit counts,' but Table I shows that for 5 qubits (gs1, guidance 10) only 2 out of a large sample achieve high fidelity, with 88 decoding errors; the 4-qubit case yields 23 high-fidelity circuits and 51 errors. The paper's own Section V acknowledges these limitations, but the abstract and parts of Section V overstate the generalization results. The claims should be qualified to match the empirical evidence, e.g., by reporting that generalization to unseen qubit counts is partially successful and mostly limited to 4 qubits with high error rates at larger sizes.
minor comments (6)
  1. [Section III-A] The normalization range of the continuous parameters is never stated. The paper says 'normalized parameters' and 'dp ≤ 1', but the reader cannot tell whether decoded rotation angles are calibrated (e.g., mapped to [0, 2π)) or arbitrary, which affects the interpretability of all downstream results.
  2. [Section IV-B] The metric 'unique hash' is used to report parameter uniqueness, but its precise definition is missing. Does it hash the full parameter vector, and if so, with what quantization? Please define it explicitly in Section IV.
  3. [Section IV-B] The decoding procedure uses cosine similarity to the closest gate embedding, but the text does not state what happens on ties or whether a similarity threshold is applied to reject ambiguous decodings. This is relevant because decoding errors increase sharply at larger sizes (Tables I-II).
  4. [Section II] In the related work, Ref. [1] is described as using 'fixed-dimensional vectors drawn from a finite embedding set'; please clarify how your continuous parameter tensor differs from simply having a larger, still discrete set of angle values, since the paper does not describe how angles are embedded and denoised in practice.
  5. [Section IV-D and Table II] The rows of Table II are labeled 'Max. Gates' but the text refers to 'gate count'; it would be clearer to call this 'maximum number of gates' and to specify how circuits of variable length are padded to this maximum during encoding.
  6. [References] There are minor reference formatting issues: 'Nat. Mac. Intell.' should be 'Nat. Mach. Intell.' and 'F¨urrutter' should use the proper umlaut encoding. Also, Ref. [33] is a manuscript under submission; it should be marked as such in the reference list to avoid implying it is a published dataset descriptor.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the generated circuits are evaluated on external fidelity/accuracy metrics, and the SQuASH self-citation is a data dependency rather than an imported conclusion.

full rationale

The paper's claimed contribution is an empirical conditional generative model: a diffusion U-Net is trained to denoise noisy tensor encodings of parameterized quantum circuits, with CLIP-embedded prompts specifying target fidelity or accuracy, and is then sampled to produce new circuits. The training objective is the standard DDPM noise-prediction loss, which is a reconstruction loss over the training distribution; the conditioning labels (fidelity and accuracy) are external scalar properties of the training circuits and are not defined in terms of the model's outputs. At inference, generated circuits are decoded by cosine similarity to a fixed gate-embedding table and then evaluated by reported fidelity and accuracy distributions on the newly synthesized circuits. There is no equation in the paper in which the predicted quantity is defined in terms of the fitted parameters, and no parameter is fitted to the evaluation outcome. The SQuASH reference [33] is authored by the same group and serves as the training-data source; this is a data dependency and a reproducibility consideration, but it is not a circular step because the benchmark supplies input examples and target labels rather than an imported theorem or uniqueness claim that forces the paper's conclusions. The layer-like ansatz observation is presented as an empirical pattern, not as a first-principles derivation. No 'prediction' reduces to a fit by construction, and no load-bearing argument depends on an unverified self-citation. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on the reliability of the authors' own unpublished SQuASH dataset and on the fidelity of the tensor encoding and decoding scheme. The dataset balancing ratios are hand-set and not justified by a principled criterion, and the decoding assumption is known to fail at larger scales.

free parameters (3)
  • Dataset balance ratio for gs1 GHZ task = 75% high fidelity (greater than 0.9), 25% low fidelity (less than or equal to 0.9)
    Chosen post hoc because it substantially improved model performance. This hand-set ratio affects the model's ability to hit high-fidelity targets.
  • Dataset balance ratio for gs2 GHZ task = 90% high fidelity, 10% low fidelity
    Chosen post hoc to improve training stability and output quality.
  • Dataset balance ratio for ML classification task = 50% accuracy below 0.7, 50% accuracy at least 0.7
    Hand-set balance that notably enhances model performance, affecting the reported accuracy distributions.
assumptions (4)
  • standard math The standard DDPM training objective with epsilon-prediction learns the circuit distribution correctly.
    The paper adopts the DDPM framework from [31] without modification; this is standard background.
  • domain assumption The concatenated gate-embedding and parameter tensor representation faithfully encodes PQCs, and cosine-similarity decoding recovers valid circuits with high probability.
    Section III-A. The model assumes that the continuous embedding space separates discrete gate types well enough for nearest-neighbor decoding. This assumption is contradicted by the observed decoding errors at larger sizes.
  • domain assumption The SQuASH dataset [33] contains a representative, correctly labeled distribution of PQCs for the GHZ and classification tasks.
    Section IV-A. The dataset is the authors' own unpublished benchmark; its construction and labeling procedure are not independently described.
  • ad hoc to paper The post-hoc balancing of the dataset does not introduce a bias that inflates the apparent performance.
    Section IV-A states that strategic balancing yielded substantial improvements. The hand-selected ratios are not justified by a principled criterion and affect the central results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Diffusion Models for Parameterized Quantum Circuit Generation." pith.science (2026). https://pith.science/paper/TS5AGASF

@misc{pith2026250520863,
  author       = {Pith},
  title        = {Pith review of: Leveraging Diffusion Models for Parameterized Quantum Circuit Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TS5AGASF}},
  note         = {Machine review of arXiv:2505.20863}
}
read the original abstract

Quantum computing holds immense potential, yet its practical success depends on multiple factors, including advances in quantum circuit design. In this paper, we introduce a generative approach based on denoising diffusion models (DMs) to synthesize parameterized quantum circuits (PQCs). Extending the recent diffusion model pipeline of F\"urrutter et al. [1], our model effectively conditions the synthesis process, enabling the simultaneous generation of circuit architectures and their continuous gate parameters. We demonstrate our approach in synthesizing PQCs optimized for generating high-fidelity Greenberger-Horne-Zeilinger (GHZ) states and achieving high accuracy in quantum machine learning (QML) classification tasks. Our results indicate a strong generalization across varying gate sets and scaling qubit counts, highlighting the versatility and computational efficiency of diffusion-based methods. This work illustrates the potential of generative models as a powerful tool for accelerating and optimizing the design of PQCs, supporting the development of more practical and scalable quantum applications.

Figures

Figures reproduced from arXiv: 2505.20863 by the authors.

Figure 1
Figure 1. Overview of our diffusion-based PQC synthesis pipeline, inspired by [1] (a) Circuit Encoding: A circuit is converted into a tensor by concatenating the embeddings of discrete gates and continuous parameters (e.g., rotation angles). (b) Forward Diffusion: Gaussian noise is progressively injected into the tensor according to a predefined scheduler, emulating a corruption process. (c) Textual Conditioning: A pretrained… view at source ↗
Figure 2
Figure 2. Comparison of fidelity distributions and evaluation metrics for two different gate sets: gs1 (top) and gs2 (middle) and generalization ability of our DM on 4 qubits for gs1 (bottom). Shown are the fidelity scores of synthesized PQCs under varying guidance scales (left) and their respective evaluation results, e.g. error counts (invalid circuits) and uniqueness (right). high fidelity without length balancing, and the… view at source ↗
Figure 3
Figure 3. Output distribution to match different accuracy targets demon￾strating the model’s adaptability and controlled circuit generation: synthesized PQCs with target accuracy of 0.2 (left) and synthesized PQCs with target 0.8 (right). able to generate valid 4-qubit circuits, despite never having encountered such configurations during training. Moreover, increasing the guidance scale enables the model to shift the fidelity… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 19 canonical work pages

  1. [30]

    The generative quantum eigensolver (gqe) and its application for ground state search,

    K. Nakaji, L. B. Kristensen, J. A. Campos-Gonzalez-Angulo, M. G. Vak- ili, H. Huang, M. Bagherimehrab, C. Gorgulla, F. Wong, A. McCaskey, J.-S. Kim et al. , “The generative quantum eigensolver (gqe) and its application for ground state search,” arXiv preprint arXiv:2401.09253 , 2024

    arXiv 2024

  2. [1]

    Quantum circuit synthesis with diffusion models,

    F. F ¨urrutter, G. Mu ˜noz-Gil, and H. J. Briegel, “Quantum circuit synthesis with diffusion models,” Nat. Mac. Intell., vol. 6, pp. 515–524,

    work page

  3. [33]

    Benchmarking quantum architecture search with surrogate assistance,

    D. Martyniuk, J. Jung, D. Barta, and A. Paschke, “Benchmarking quantum architecture search with surrogate assistance,” 2025, manuscript under submission

    work page 2025

  4. [2]

    Quantum computing in the nisq era and beyond,

    J. Preskill, “Quantum computing in the nisq era and beyond,” Quantum,

    work page

  5. [3]

    ´Eliv´agar: Efficient quantum circuit search for classification,

    S. Anagolum, N. Alavisamani, P. Das, M. Qureshi, and Y . Shi, “´Eliv´agar: Efficient quantum circuit search for classification,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 , 2024, pp. 336–353

    work page 2024

  6. [4]

    Curriculum reinforcement learning for quantum archi- tecture search under hardware errors,

    Y . J. Patel, A. Kundu, M. Ostaszewski, X. Bonet-Monroig, V . Dunjko, and O. Danaci, “Curriculum reinforcement learning for quantum archi- tecture search under hardware errors,” arXiv preprint arXiv:2402.03500, 2024

    arXiv 2024

  7. [5]

    Quantumnas: Noise-adaptive search for robust quantum circuits,

    H. Wang, Y . Ding, J. Gu, Y . Lin, D. Z. Pan, F. T. Chong, and S. Han, “Quantumnas: Noise-adaptive search for robust quantum circuits,” in 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 2022, pp. 692–708

    2022

  8. [6]

    Mog-vqe: Multiobjective genetic variational quantum eigensolver,

    D. Chivilikhin, A. Samarin, V . Ulyantsev, I. Iorsh, A. R. Oganov, and O. Kyriienko, “Mog-vqe: Multiobjective genetic variational quantum eigensolver,” arXiv preprint arXiv:2007.04424 , 2020

    arXiv 2007

Show all 37 references
  1. [7]

    Ga4qco: genetic algorithm for quantum circuit optimization,

    L. S ¨unkel, D. Martyniuk, D. Mattern, J. Jung, and A. Paschke, “Ga4qco: genetic algorithm for quantum circuit optimization,” arXiv preprint arXiv:2302.01303, 2023

    2023 arXiv

  2. [8]

    Gasp–a genetic algorithm for state preparation,

    F. M. Creevey, C. D. Hill, and L. C. Hollenberg, “Gasp–a genetic algorithm for state preparation,” arXiv preprint arXiv:2302.11141, 2023

    2023 arXiv

  3. [9]

    Quantum circuit compilation by genetic algorithm for quantum approximate optimization algorithm applied to maxcut problem,

    L. Arufe, M. A. Gonz ´alez, A. Oddi, R. Rasconi, and R. Varela, “Quantum circuit compilation by genetic algorithm for quantum approximate optimization algorithm applied to maxcut problem,” Swarm and Evolutionary Computation , vol. 69, p. 101030, 2022. [Online]. Available: http...

    2022

  4. [10]

    Quantum circuit structure learning,

    M. Ostaszewski, E. Grant, and M. Benedetti, “Quantum circuit structure learning,” arXiv: Quantum Physics , 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:162184021

    2019

  5. [11]

    Quantum compiling by deep reinforcement learning,

    L. Moro, M. G. A. Paris, M. Restelli, and E. Prati, “Quantum compiling by deep reinforcement learning,” Communications Physics , vol. 4, no. 1, Aug. 2021. [Online]. Available: http://dx.doi.org/10.1038/ s42005-021-00684-3

    2021

  6. [12]

    Reinforcement learning for optimization of variational quantum circuit architectures,

    M. Ostaszewski, L. M. Trenkwalder, W. Masarczyk, E. Scerri, and V . Dunjko, “Reinforcement learning for optimization of variational quantum circuit architectures,” 2021. [Online]. Available: https: //arxiv.org/abs/2103.16089

    2021 arXiv

  7. [13]

    Quantum neural architecture search with quantum circuits metric and bayesian optimization,

    T. Duong, S. T. Truong, M. Tam, B. Bach, J.-Y . Ryu, and J.-K. K. Rhee, “Quantum neural architecture search with quantum circuits metric and bayesian optimization,” arXiv preprint arXiv:2206.14115 , 2022

    2022 arXiv

  8. [14]

    An adaptive variational algorithm for exact molecular simulations on a quantum computer,

    H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, “An adaptive variational algorithm for exact molecular simulations on a quantum computer,” Nature communications , vol. 10, no. 1, p. 3007, 2019

    2019

  9. [15]

    Adaptive quantum approximate op- timization algorithm for solving combinatorial problems on a quantum computer,

    L. Zhu, H. L. Tang, G. S. Barron, F. Calderon-Vargas, N. J. Mayhall, E. Barnes, and S. E. Economou, “Adaptive quantum approximate op- timization algorithm for solving combinatorial problems on a quantum computer,” Physical Review Research, vol. 4, no. 3, p. 033029, 2022

    2022

  10. [16]

    Machine learning method for state preparation and gate synthesis on photonic quantum computers,

    J. M. Arrazola, T. R. Bromley, J. A. Izaac, C. R. Myers, K. Br’adler, and N. Killoran, “Machine learning method for state preparation and gate synthesis on photonic quantum computers,” Quantum Science and Technology , vol. 4, 2018. [Online]. Available: https://api.semanticscho...

    2018

  11. [17]

    Neural predictor based quantum architecture search,

    S.-X. Zhang, C.-Y . Hsieh, S. Zhang, and H. Yao, “Neural predictor based quantum architecture search,” Machine Learning: Science and Technology , vol. 2, 2021. [Online]. Available: https://api. semanticscholar.org/CorpusID:232185636

    2021

  12. [18]

    Discovering quantum circuit components with program synthesis,

    L. Sarra, K. Ellis, and F. Marquardt, “Discovering quantum circuit components with program synthesis,” 2023. [Online]. Available: https://arxiv.org/abs/2305.01707

    2023 arXiv

  13. [19]

    Hybrid discrete-continuous compilation of trapped-ion quantum circuits with deep reinforcement learning,

    F. Preti, M. Schilling, S. Jerbi, L. M. Trenkwalder, H. P. Nautrup, F. Motzoi, and H. J. Briegel, “Hybrid discrete-continuous compilation of trapped-ion quantum circuits with deep reinforcement learning,” Quantum, vol. 8, p. 1343, May 2024. [Online]. Available: http://dx.doi.o...

    2024 doi

  14. [20]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,”

  15. [21]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  16. [22]

    Video diffusion models,

    J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion models,” Advances in Neural Information Processing Systems, vol. 35, pp. 8633–8646, 2022

    2022

  17. [23]

    Dif- fwave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Dif- fwave: A versatile diffusion model for audio synthesis,” arXiv preprint arXiv:2009.09761, 2020

    2009 arXiv

  18. [24]

    Eigenfold: Generative protein structure prediction with diffusion mod- els,

    B. Jing, E. Erives, P. Pao-Huang, G. Corso, B. Berger, and T. Jaakkola, “Eigenfold: Generative protein structure prediction with diffusion mod- els,” arXiv preprint arXiv:2304.02198 , 2023

    2023 arXiv

  19. [25]

    Uditqc: U-net-style diffusion transformer for quantum circuit synthesis,

    Z. Chen and H. Tang, “Uditqc: U-net-style diffusion transformer for quantum circuit synthesis,” arXiv preprint arXiv:2501.16380 , 2025

    2025 arXiv

  20. [26]

    Prepare ansatz for vqe with diffusion model,

    Y . Shen, “Prepare ansatz for vqe with diffusion model,” 2023. [Online]. Available: https://arxiv.org/abs/2310.02511

    2023 arXiv

  21. [27]

    Quantum circuit optimization with deep reinforcement learning,

    T. F ¨osel, M. Y . Niu, F. Marquardt, and L. Li, “Quantum circuit optimization with deep reinforcement learning,” 2021. [Online]. Available: https://arxiv.org/abs/2103.07585

    2021 arXiv

  22. [28]

    Quantum architecture search: a survey,

    D. Martyniuk, J. Jung, and A. Paschke, “Quantum architecture search: a survey,” in2024 IEEE International Conference on Quantum Computing and Engineering (QCE) , vol. 1. IEEE, 2024, pp. 1695–1706

    2024

  23. [29]

    Challenges for reinforcement learning in quantum circuit design,

    P. Altmann, J. Stein, M. K ¨olle, A. B ¨arligea, M. Zorn, T. Gabor, T. Phan, S. Feld, and C. Linnhoff-Popien, “Challenges for reinforcement learning in quantum circuit design,” in 2024 IEEE International Conference on Quantum Computing and Engineering (QCE) , vol. 1. IEEE, 202...

    2024

  24. [31]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https://arxiv.org/abs/2006.11239

    2020 arXiv

  25. [32]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” 2022. [Online]. Available: https://arxiv.org/abs/2207.12598

    2022 arXiv

  26. [34]

    Better than classical? the subtle art of benchmarking quantum machine learning models,

    J. Bowles, S. Ahmed, and M. Schuld, “Better than classical? the subtle art of benchmarking quantum machine learning models,” arXiv preprint arXiv:2403.07059, 2024

    2024 arXiv

  27. [2015]

    Available: https://arxiv.org/abs/1503.03585

    [Online]. Available: https://arxiv.org/abs/1503.03585

    arXiv

  28. [2018]

    Available: https://api.semanticscholar.org/CorpusID: 44098998

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 44098998

  29. [2023]

    Available: https://api.semanticscholar.org/CorpusID: 265018897

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 265018897

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.