Pith. sign in

REVIEW 4 major objections 6 minor 27 references

DQAOA-GPT: AI-Accelerated Distributed Quantum Optimization for Combinatorial Problems

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing DQAOA's per-sub-problem variational loop with a single GPT forward pass keeps solution quality while making runtime independent of sub-problem size.

desk verdict Sensible integration, but the central accuracy claim is in-sample: the GPT is trained on sub-problems of the same N=100 instance it then solves, so the ~0.78 accuracy at n=12 may be memorization rather than generalization. read the letter →

arxiv 2607.20225 v1 pith:6DBV5BT7 submitted 2026-07-22 quant-ph cs.AIcs.DCmath.OC

Seongmin Kim , Abhinav Rijal , Yuri Alexeev , Nora Bauer , Martin Roetteler , Mina Yoon , George Siopsis , In-Saeng Suh This is my paper · ORCID
classification quant-phcs.AIcs.DCmath.OC
keywords distributedquantumapproximateoptimizationgenerativecircuitsynthesisHUBOQAOAADAPT-QAOAGPTforcircuitscombinatorialgreedysub-problemaggregation
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

The paper tries to establish that the main bottleneck of distributed variational quantum optimization—the repeated quantum-circuit evaluations and classical parameter updates inside each sub-problem—can be removed entirely by training a transformer to generate the circuits directly. On dense third-order HUBO problems with 100 variables, the authors show that their DQAOA-GPT achieves roughly the same relative accuracy as standard DQAOA (about 0.78 for sub-problem size 12) while holding runtime at about 28 seconds, compared with about 684 seconds for DQAOA. The point of the result is that accuracy improves with larger sub-problems because they capture more interactions, and with GPT-based generation this improvement no longer carries a runaway computational cost. A sympathetic reader would take this as evidence that learned circuit synthesis, not incremental variational refinement, is a viable route to scaling quantum optimization.

What carries the argument

The load-bearing object is the trained GPT circuit generator, a decoder-only transformer that maps a tokenized sub-problem (one token per linear, quadratic, and cubic interaction with a shared coefficient vocabulary, conditioned by a graph embedding) to a sequence of ADAPT-QAOA operator and parameter tokens in one forward pass. Around it, the DQAOA loop supplies the other half: random sub-tensor extraction, generation of 10 candidate circuits, simulation, and greedy coordinate-wise acceptance that only applies bit flips lowering the global HUBO energy. The combination is what lets the method keep the expressiveness of larger sub-problems while removing the variational-loop cost that in DQAOA

What would settle it

Ablate the learned model: in the same DQAOA loop, replace GPT-generated circuits with randomly sampled circuits of matched depth from the same operator pool, keeping all other steps (10 candidates, greedy acceptance) identical. If relative accuracy stays near 0.78 at n=12, the result is due to decomposition and aggregation rather than learned circuit synthesis; if accuracy collapses, the generative mapping is load-bearing.

Watch

Extended reading notes

Core claim

DQAOA-GPT decomposes a large HUBO instance into random n-variable sub-problems, but instead of solving each with iterative variational optimization it feeds the sub-problem—tokenized as index/coefficient pairs plus a graph-structure embedding—to a GPT model trained on ADAPT-QAOA reference circuits. The model autoregressively emits a complete circuit; the authors sample 10 candidates per sub-problem, simulate all of them, and accept the lowest-energy bitstring only if it lowers the global objective. On 100-variable dense HUBO problems, increasing n from 4 to 12 raises relative accuracy from about 0.36 to about 0.78 for DQAOA-GPT, while runtime stays roughly constant at about 28 seconds; stand

Load-bearing premise

That a sub-problem solved in isolation—with all interactions to variables outside it held fixed—produces candidate local updates whose greedy acceptance improves the global solution; if isolated optima conflict with global coupling, no circuit generator can rescue the loop.

Editorial extensions

If this is right

  • Larger sub-problem sizes become affordable: n=12 yields about 0.78 relative accuracy at about 28 seconds, whereas standard DQAOA at n=12 costs about 684 seconds, so the accuracy-cost sweet spot shifts upward.
  • Runtime no longer scales with sub-problem size, so within the trained range the per-iteration cost is a constant overhead of generation plus a fixed number of circuit simulations.
  • Sub-problems are solved independently, so the loop parallelizes naturally across GPUs, promising further speedups on distributed systems.
  • The monotonic greedy acceptance guarantees the global energy never increases across iterations, so the decomposition cannot destroy the current solution.
  • The framework moves the expensive part of quantum optimization from online variational updates to offline training, making inference cost predictable and hardware-bound rather than optimization-bound.

Reading between the lines

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

  • If the flat-runtime result extends beyond n=12, the optimal strategy becomes to make sub-problems as large as the GPT model supports; the practical ceiling then becomes model accuracy and simulator memory, not optimization cost.
  • The method's accuracy ceiling may actually be set by the locality assumption the authors flag: on problems with dense long-range coupling, isolated sub-problem optima could systematically mislead the greedy acceptance, so a testable extension is to bias sub-problem sampling to include high-energy fixed neighbors.
  • Because runtime is dominated by a fixed number of simulations, the same framework could trade accuracy for speed by adjusting the number of candidate circuits sampled per sub-problem or the sampling temperature, without retraining the model.
  • A direct comparison against classical local-search heuristics with the same iteration budget would clarify whether the gains are due to the quantum circuit structure or to the greedy sub-problem sampling itself.
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

4 major / 6 minor

Summary. The paper introduces DQAOA-GPT, a hybrid framework that combines a distributed QAOA (DQAOA) decomposition strategy with a GPT-based generative model that directly synthesizes ADAPT-QAOA-style circuits for sub-problems, bypassing the variational optimization loop. The authors train separate GPT models for sub-problem sizes n ∈ {4,6,8,10,12} on sub-problems sampled from a single N=100 HUBO instance. At inference, Algorithm 1 iteratively samples variable subsets, generates candidate circuits with one forward pass each, simulates 10 candidate circuits per sub-problem, and accepts only greedy updates that lower the global energy. Experiments on one N=100 dense HUBO instance (Fig. 3) report relative accuracy improving from ~0.36 at n=4 to ~0.78 at n=12 for DQAOA-GPT, while runtime stays ~28 s versus DQAOA's ~684 s at n=12. The central claim is that DQAOA-GPT significantly reduces computational cost while maintaining competitive solution quality, with larger accelerations at larger n.

Significance. If the claims were validated on held-out instances with proper statistical rigor, the idea of replacing the variational loop with a single generative inference pass inside a DQAOA decomposition would be a meaningful contribution to the growing literature on AI-assisted quantum circuit design. The paper clearly describes the tokenization pipeline, the FEATHER conditioning (Eq. 8), and the greedy acceptance protocol ensuring monotonic global energy reduction. However, the current evidence is insufficient: the GPT is trained and evaluated on sub-problems of the same N=100 instance, creating an in-sample prediction loop; all reported numbers come from single runs without error bars or seeds; and the claimed 'approximately constant' runtime is in tension with the autoregressive token count growing as O(n^3) from the problem encoding. These issues are load-bearing for the abstract's quality and speedup claims, which are the paper's main contributions. The paper is therefore not yet ready for publication; the evaluation must be redone with held-out instances, multiple repetitions, and a runtime breakdown.

major comments (4)
  1. [Section III-C and Section IV] The GPT is trained on sub-problems 'generated by random sampling from the target problem' (III-C), and the evaluation in Section IV is run on 'an original optimization problem of size N=100' — the same target. Algorithm 1 samples sub-problems from that same instance during inference. There is no train/test split, held-out instance, or statement that evaluation sub-problems were excluded from training. Since the token sequence contains exact interaction coefficients of the sub-problem, the model can memorize instance-specific patterns. The reported relative accuracy of ~0.78 at n=12 therefore does not establish the generalization claim in the abstract ('maintaining competitive solution quality'). The authors must either train on a distinct set of instances and test on unseen ones, or report a train/test split and show per-instance transfer results.
  2. [Section IV, Fig. 3] All performance numbers appear to come from a single run: no error bars, no seeds, no repetitions, and no statistical significance testing are reported. The claims '~0.78' relative accuracy and '~28 s' runtime are point values from one trajectory. Given the stochasticity in random sub-problem sampling, GPT sampling temperature (0.8), and circuit simulation outcomes, the observed differences (e.g., DQAOA vs DQAOA-GPT at n=4) could fluctuate. The paper should report mean and standard deviation over multiple independent runs, including different initial assignments and random seeds, for both accuracy and runtime.
  3. [Section III-D and Section IV, Fig. 3(b)] The paper states that DQAOA-GPT runtime is 'approximately constant' as n increases and shows ~28 s for all n. However, the autoregressive GPT token generation cost scales with sequence length, and the problem tokenization alone yields n + C(n,2) + C(n,3) tokens (e.g., 12 + 66 + 220 = 298 at n=12 vs 4 + 6 + 4 = 14 at n=4). Even if circuit layers are fixed, total token count grows as O(n^3), so the inference pipeline cannot be truly n-independent unless additional assumptions hold (e.g., constant circuit length and negligible tokenization cost). The constant runtime curve is suspicious and needs a detailed breakdown: graph projection, FEATHER embedding, GPT generation, the 10 CUDA-Q simulations, and energy evaluation separately for each n. Without this, the speedup claim at larger n is not credible.
  4. [Section III-A] The authors themselves acknowledge in Section III-A that a sub-problem solved in isolation—with interactions to variables outside S fixed—does not fully determine the best local update for the global problem. This locality assumption is structural: even a perfect sub-problem solver can produce candidate updates that would not lower the global energy, and the greedy acceptance in Algorithm 1 may then lead to wasted cycles or suboptimal convergence. The paper does not analyze how often accepted updates actually occur, nor does it compare against simpler local-search baselines (e.g., random bit flips with same greedy acceptance). Such an ablation is needed to isolate the contribution of the GPT-generated circuits from the DQAOA decomposition/greedy update mechanism.
minor comments (6)
  1. [Abstract / Introduction] The term 'maintaining competitive solution quality' is too strong given that DQAOA at n=12 achieves ~0.78 relative accuracy and DQAOA-GPT achieves ~0.78 as reported; the actual difference (if any) should be stated quantitatively. Also, the phrase 'eliminates the variational loop' is imprecise: DQAOA-GPT still runs a fixed-number-of-inference loop and evaluates 10 candidates per sub-problem.
  2. [Eq. (2)] The gradient expression in Eq. (2) appears adapted from ADAPT-QAOA but the notation ε_j is introduced without definition; specify that ε_j is the parameter for the trial operator. Also, the global re-optimization step is not described precisely—clarify how γ and β are optimized.
  3. [Section III-B] The graph projection in Eqs. (5)-(6) is a coarse summary, but the statement 'the projection is not an exact reduction' could be strengthened by noting that the projected graph discards sign information of coefficients. This is relevant for FEATHER embeddings, which may then be ambiguous for terms with mixed signs.
  4. [Section III-C] The token vocabulary size for circuit tokens is given as 2n^2 - n + 1. For n=12 this is 277, but the operator pool size in ADAPT-QAOA typically grows with the number of possible operators; please confirm this formula and clarify what operator pool is used (e.g., all single-qubit and two-qubit Pauli strings on n qubits).
  5. [References] Several references share authors with the current manuscript (e.g., [13], [17], [22]). This is not problematic per se, but the paper should clearly state the relationship to [17] (QAOA-GPT) and [13] (DQAOA) and what new contribution beyond those works is being made. The current text in Sections II-E and III sometimes reads as a summary of [17] rather than a new development.
  6. [Fig. 3] The figures show relative accuracy and runtime but do not label axes with units or show error bars. The runtime axis for DQAOA should be distinguished from DQAOA-GPT (perhaps log scale) to make the constant-trend claim visible. Also, the text mentions '~0.38 for DQAOA and ~0.36 for DQAOA-GPT' at n=4, but Fig. 3(a) appears to show different values; please verify the numbers.

Circularity Check

1 steps flagged · score 6.0 of 10

Quality 'prediction' is in-sample: GPT trained on sub-problems of the same N=100 target used for Fig. 3 evaluation.

  1. fitted input called prediction [Section III-C (Training the GPT model) and Section IV (Results)]
    "A large set of sub-problems is generated by random sampling from the target problem. ... Figure 3 shows relative accuracy and runtime as a function of n for an original optimization problem of size N=100"

    The GPT is trained on sub-problems extracted from the N=100 target instance (Section III-C), and the evaluation of DQAOA-GPT in Figure 3 is run on the same N=100 instance, with inference sub-problems sampled from it via Algorithm 1. No train/test split, held-out instance, or exclusion of training sub-problems is reported. The relative accuracy ~0.78 at n=12 therefore measures how well the model reproduces circuits for the very instance distribution it was fitted on; it is a memorization/in-sample fit result, not an independent prediction. The abstract's 'maintaining competitive solution quality' is thus supported only by an in-sample loop, not by a genuine predictive test.

full rationale

The central quality claim of DQAOA-GPT rests on Figure 3, but the GPT model was trained on sub-problems sampled from the same N=100 target problem on which Figure 3 evaluates performance. The paper reports no held-out instance, no train/test split, and no exclusion of training sub-problems from the evaluation set. As a result, the measured relative accuracy is an in-sample fit to the evaluation distribution: the model has seen the exact interaction coefficients and index patterns of the target instance during training, so the 'competitive solution quality' result is not an independent prediction. The runtime comparison is less affected by this issue because the reported DQAOA-GPT runtime covers per-cycle inference and circuit simulation rather than training; however, the abstract's overall claim of 'maintaining competitive solution quality' is load-bearing and is exactly what is contaminated. Self-citations to Refs. [13], [17], [22] are present and share authors with this paper, but the paper also provides direct experimental evidence for its main speed/quality claims, so self-citation alone is not scored as circular here. The identified in-sample evaluation loop is the main circularity, warranting a 6 rather than a higher score because the runtime claim retains independent content.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The framework introduces no new physical entity. Its numerical claims rest on many hand-set hyperparameters, the most consequential being the unreported ADAPT-QAOA quality threshold and the training setup. The GPT model is a fitted function whose training cost is excluded from the reported runtime, and the evaluation is in-sample with respect to the target problem instance.

free parameters (7)
  • subproblem_size_n = 4, 6, 8, 10, 12
    Sub-problem sizes chosen by the authors; a separate GPT model is trained for each n, and Figure 3 sweeps n to demonstrate the trend.
  • dqa_iterations_T = 100
    Number of DQAOA iterations is fixed in all experiments; not swept.
  • candidate_samples_per_subproblem = 10
    Each GPT inference cycle samples 10 candidate circuits at temperature 0.8 and selects the minimum-energy bitstring.
  • sampling_temperature = 0.8
    Sampling temperature for autoregressive circuit generation; chosen without reported sensitivity analysis.
  • adapt_quality_threshold = unreported
    Only ADAPT-QAOA reference circuits above a predefined quality threshold are retained for training; the threshold value is not stated.
  • coefficient_grid = 201 values in [-10,10], step 0.1
    Shared vocabulary quantizes HUBO coefficients and circuit parameters; the discretization is a modeling choice that affects circuit quality.
  • feather_embedding_dimension = 500
    Fixed embedding dimension for the graph-structure conditioning signal, projected into the transformer input.
assumptions (6)
  • standard math Third-order HUBO over binary variables maps to an Ising Hamiltonian via x_i=(1-z_i)/2, with cubic terms expanding to Z, ZZ, and ZZZ Pauli terms.
    Used in Section III-C to convert sub-problems into cost Hamiltonians for ADAPT-QAOA reference generation.
  • domain assumption ADAPT-QAOA, filtered by a quality threshold, produces near-optimal reference circuits that are a valid training target.
    Section III-C: 'Only reference circuits meeting a predefined quality threshold are retained for training.' No threshold or convergence analysis is given; training quality is inherited from ADAPT-QAOA.
  • domain assumption Greedy acceptance of locally solved sub-problems (each bit changed only if global energy decreases) monotonically improves the DQAOA-GPT solution.
    Section III-D and Algorithm 1: candidate updates are accepted only if they lower the original HUBO energy; this relies on the DQAOA decomposition strategy of Ref. [13].
  • domain assumption A GPT decoder trained with next-token prediction can autoregressively generate quantum circuits that are competitive with ADAPT-QAOA circuits.
    Section II-E and III-C: the method assumes the trained model's output circuits are good enough for local updates; no in-paper comparison against ADAPT-QAOA is provided.
  • domain assumption FEATHER embedding of the HUBO-to-graph projection provides a useful structural conditioning signal despite being an approximate projection.
    Section III-B explicitly states the projection is not exact; the authors assume it still helps circuit generation, but provide no ablation.
  • domain assumption CUDA-Q with cuQuantum accurately simulates the generated circuits so that measured runtime and bitstrings reflect real quantum execution.
    Section III-E: all experiments use GPU simulation; no quantum-hardware validation is performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DQAOA-GPT: AI-Accelerated Distributed Quantum Optimization for Combinatorial Problems." pith.science (2026). https://pith.science/paper/6DBV5BT7

@misc{pith2026260720225,
  author       = {Pith},
  title        = {Pith review of: DQAOA-GPT: AI-Accelerated Distributed Quantum Optimization for Combinatorial Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DBV5BT7}},
  note         = {Machine review of arXiv:2607.20225}
}
read the original abstract

While combinatorial optimization problems are central to many scientific and engineering applications, their solution remains challenging due to exponentially large search spaces. Variational quantum algorithms offer a promising route for tackling such problems, yet their practical performance is limited by repeated quantum circuit evaluations and classical parameter updates. In this work, we introduce DQAOA-GPT, a hybrid framework that integrates the distributed quantum approximate optimization algorithm (DQAOA), which decomposes a large optimization problem into smaller sub-problems, with GPT-based quantum circuit generation for solving those sub-problems. Rather than relying on iterative variational optimization, the proposed approach uses a trained generative model to directly generate high-quality quantum circuits for the decomposed sub-problems. As a benchmark, we evaluate DQAOA-GPT against conventional DQAOA on dense HUBO optimization problems with up to 100 decision variables. The results demonstrate that DQAOA-GPT significantly reduces computational cost while maintaining competitive solution quality, with larger acceleration observed for larger sub-problem sizes. Although this work focuses on benchmark-scale validation, the framework provides a promising foundation for larger-scale combinatorial optimization in hybrid HPC-QC environments through increased GPU resources and parallel computing capability.

Figures

Figures reproduced from arXiv: 2607.20225 by the authors.

Figure 1
Figure 1. Schematic illustration of the QAOA-GPT workflow. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Schematic illustration of the DQAOA-GPT workflow. Inspired by Refs. [13], [17] [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. DQAOA-GPT performance for different sub-problem sizes [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 4 linked inside Pith

  1. [1]

    Benchmarking the performance of portfolio optimization with qaoa: S. brandhofer et al.,

    S. Brandhofer et al., “Benchmarking the performance of portfolio optimization with qaoa: S. brandhofer et al.,” Quantum Information Processing, vol. 22, no. 1, p. 25, 2022

    2022

  2. [2]

    Multi- level combinatorial optimization across quantum archi- tectures,

    H. Ushijima-Mwesigwa, R. Shaydulin, C. F. Negre, S. M. Mniszewski, Y . Alexeev, and I. Safro, “Multi- level combinatorial optimization across quantum archi- tectures,”ACM Transactions on Quantum Computing, vol. 2, no. 1, pp. 1–29, 2021

    2021

  3. [3]

    Sampling frequency thresholds for the quantum advantage of the quantum approximate opti- mization algorithm,

    D. Lykov, J. Wurtz, C. Poole, M. Saffman, T. Noel, and Y . Alexeev, “Sampling frequency thresholds for the quantum advantage of the quantum approximate opti- mization algorithm,”npj Quantum Information, vol. 9, no. 1, p. 73, 2023

    2023

  4. [4]

    Harnessing quantum computing for energy materials: Opportunities and challenges,

    S. Kim, I.-S. Suh, T. S. Humble, T. Beck, E. Lee, and T. Luo, “Harnessing quantum computing for energy materials: Opportunities and challenges,”ACS Energy Letters, 2026

    2026

  5. [5]

    Designing metamaterials with quantum annealing and factorization machines,

    K. Kitai et al., “Designing metamaterials with quantum annealing and factorization machines,”Physical Review Research, vol. 2, no. 1, p. 013 319, 2020

    2020

  6. [6]

    High-performance transparent radiative cooler designed by quantum computing,

    S. Kim, W. Shang, S. Moon, T. Pastega, E. Lee, and T. Luo, “High-performance transparent radiative cooler designed by quantum computing,”ACS Energy Letters, vol. 7, no. 12, pp. 4134–4141, 2022

    2022

  7. [7]

    Quantum annealing for combinatorial optimization: A benchmarking study,

    S. Kim, S.-W. Ahn, I.-S. Suh, A. W. Dowling, E. Lee, and T. Luo, “Quantum annealing for combinatorial optimization: A benchmarking study,”npj Quantum Information, vol. 11, no. 1, p. 77, 2025

    2025

  8. [8]

    Variational quantum algorithms,

    M. Cerezo et al., “Variational quantum algorithms,” Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

    2021

Show all 27 references
  1. [9]

    A benchmarking study of quantum al- gorithms for combinatorial optimization,

    K. Sankar et al., “A benchmarking study of quantum al- gorithms for combinatorial optimization,”npj Quantum Information, vol. 10, no. 1, p. 64, 2024

    2024

  2. [10]

    Combinatorial optimization with quantum imaginary time evolution,

    N. M. Bauer, R. Alam, G. Siopsis, and J. Ostrowski, “Combinatorial optimization with quantum imaginary time evolution,”Physical Review A, vol. 109, no. 5, p. 052 430, 2024

    2024

  3. [11]

    A quantum approximate optimization algorithm,

    E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,”arXiv preprint arXiv:1411.4028, 2014

    2014 arXiv

  4. [12]

    Evidence of scaling advantage for the quantum approximate optimization algorithm on a classically intractable problem,

    R. Shaydulin et al., “Evidence of scaling advantage for the quantum approximate optimization algorithm on a classically intractable problem,”Science Advances, vol. 10, no. 22, eadm6761, 2024

    2024

  5. [13]

    Distributed quantum approximate optimiza- tion algorithm on a quantum-centric supercomputing architecture,

    S. Kim, V . R. Pascuzzi, Z. Xu, T. Luo, E. Lee, and I.-S. Suh, “Distributed quantum approximate optimiza- tion algorithm on a quantum-centric supercomputing architecture,”npj Quantum Information, 2026

    2026

  6. [14]

    Gpu-accelerated distributed qaoa on large-scale hpc ecosystems,

    Z. Xu et al., “Gpu-accelerated distributed qaoa on large-scale hpc ecosystems,”arXiv preprint arXiv:2506.10531, 2025

    2025 arXiv

  7. [15]

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

    K. Nakaji et al., “The generative quantum eigensolver (gqe) and its application for ground state search,”arXiv preprint arXiv:2401.09253, 2024

    2024

  8. [16]

    Artificial intelligence for quantum computing,

    Y . Alexeev et al., “Artificial intelligence for quantum computing,”Nature Communications, vol. 16, no. 1, p. 10 829, 2025

    2025

  9. [17]

    Qaoa-gpt: Efficient generation of adaptive and regular quantum approximate optimization algorithm circuits,

    I. Tyagin, M. H. Farag, K. Sherbert, K. Shirali, Y . Alexeev, and I. Safro, “Qaoa-gpt: Efficient generation of adaptive and regular quantum approximate optimization algorithm circuits,” in2025 IEEE International Confer- ence on Quantum Computing and Engineering (QCE), IEEE, vo...

    2025

  10. [18]

    Extending qaoa-gpt to higher-order quantum optimization problems,

    L. Sunny, A. Rijal, and G. Siopsis, “Extending qaoa-gpt to higher-order quantum optimization problems,”arXiv preprint arXiv:2511.07391, 2025

    2025

  11. [19]

    Higher- order factorization machine for accurate surrogate mod- eling in material design,

    S. Hwang, S. Kim, Z. Xu, T. Luo, and E. Lee, “Higher- order factorization machine for accurate surrogate mod- eling in material design,”Scientific Reports, vol. 15, no. 1, p. 35 392, 2025

    2025

  12. [20]

    Compressed quadratization of higher or- der binary optimization problems,

    A. Mandal, A. Roy, S. Upadhyay, and H. Ushijima- Mwesigwa, “Compressed quadratization of higher or- der binary optimization problems,” inProceedings of the 17th ACM International Conference on Computing Frontiers, 2020, pp. 126–131

    2020

  13. [21]

    Space-efficient binary optimization for variational quantum comput- ing,

    A. Glos, A. Krawiec, and Z. Zimbor ´as, “Space-efficient binary optimization for variational quantum comput- ing,”npj Quantum Information, vol. 8, no. 1, p. 39, 2022

    2022

  14. [22]

    Distributed quantum optimization for large-scale higher-order problems with dense interac- tions,

    S. Kim et al., “Distributed quantum optimization for large-scale higher-order problems with dense interac- tions,”arXiv preprint arXiv:2604.20599, 2026

    2026 arXiv

  15. [23]

    Adaptive quantum approximate optimiza- tion algorithm for solving combinatorial problems on a quantum computer,

    L. Zhu et al., “Adaptive quantum approximate optimiza- tion algorithm for solving combinatorial problems on a quantum computer,”Physical Review Research, vol. 4, no. 3, p. 033 029, 2022

    2022

  16. [24]

    Characteristic func- tions on graphs: Birds of a feather, from statistical descriptors to parametric models,

    B. Rozemberczki and R. Sarkar, “Characteristic func- tions on graphs: Birds of a feather, from statistical descriptors to parametric models,” inProceedings of the 29th ACM international conference on information & knowledge management, 2020, pp. 1325–1334

    2020

  17. [25]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  18. [26]

    nvidia.com/cuda-q, Accessed: 2026-05-12, 2026

    NVIDIA Corporation,CUDA-Q, https : / / developer . nvidia.com/cuda-q, Accessed: 2026-05-12, 2026

    2026

  19. [27]

    Cuquantum SDK: A high- performance library for accelerating quantum science,

    H. Bayraktar et al., “Cuquantum SDK: A high- performance library for accelerating quantum science,” arXiv preprint arXiv:2308.01999, 2023.DOI: 10.48550/ arXiv.2308.01999 [Online]. Available: https://doi.org/ 10.48550/arXiv.2308.01999

    2023 doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.