Pith. sign in

REVIEW 4 major objections 5 minor 18 references

An LLM-guided feedback loop can design quantum circuit Born machine ansätze that are much shallower and more noise-robust than a standard deep baseline, demonstrated on a 12-qubit financial modeling task.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 20:39 UTC pith:OQ47O65Q

load-bearing objection Real hardware numbers, real confound: the LLM ansatz wins only without post-selection and only against a much deeper baseline, so the paper's central attribution is not yet established. the 4 major comments →

arxiv 2509.08385 v1 pith:OQ47O65Q submitted 2025-09-10 quant-ph cs.LG

LLM-Guided Ans\"atze Design for Quantum Circuit Born Machines in Financial Generative Modeling

Yaswitha Gujju , Romain Harang , Tetsuo Shibuya This is my paper
classification quant-ph cs.LG MSC 81P68 PACS 03.67.-a03.67.Lx
keywords quantum circuit Born machinegenerative modelinglarge language modelsansatz designfinancial time seriesnoisy intermediate-scale quantum deviceshardware-aware circuitsKL divergence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a language-model-driven search loop can design hardware-aware quantum circuit Born machines (QCBMs) that are both compact and expressive enough to model a real financial distribution on noisy hardware. The empirical anchor is a 12-qubit generative task over daily Japanese government bond interest-rate changes: the LLM-designed ansatz has depth 28, while the standard layered baseline has depth 85. On a noiseless simulator the two perform comparably (KL 3.67 vs 3.10), but on real hardware the shallower LLM circuit clearly wins (KL 7.37 vs 9.32 without error mitigation, 6.91 vs 8.92 with it). The paper presents this as a practical route to deployable quantum generative models because hardware constraints are fed into the prompt and iterative feedback balances generative fidelity, depth, and validity.

Core claim

On the paper's own terms, the discovery is that conditioning ansatz search on hardware topology, qubit connectivity, basis gates, and gate error rates—and refining proposals with measured reverse KL divergence, depth, and validity—lets a language model find QCBM circuits that survive real-device noise better than a standard deep template. The supporting result is a 12-qubit experiment on Japanese government bond interest-rate changes in which the LLM-generated depth-28 ansatz reaches hardware KL divergence 7.37 without error mitigation and 6.91 with it, compared with 9.32 and 8.92 for the depth-85 TwoLocal baseline. On a noiseless simulator the gap reverses slightly (3.67 vs 3.10), so the pa

What carries the argument

The load-bearing mechanism is an iterative prompt-feedback loop rather than a new quantum gate or identity. An initial prompt encodes the hardware profile (qubit count, connectivity, basis gates, error rates); the LLM proposes a circuit; each round the system returns three numbers—reverse KL divergence after training, circuit depth, and a validity flag—plus the instruction to prioritize KL reduction over depth. The LLM responds by pruning or appending layers. Training itself uses maximum mean discrepancy loss with a Gaussian kernel and parameter-shift gradients, and the hardware-aware initialization lets the final circuit run without transpilation.

Load-bearing premise

The load-bearing premise is that the LLM's guided search, rather than simply the much shallower circuit depth, drives the hardware improvement; the paper tests only a depth-85 baseline against its depth-28 circuit.

What would settle it

Run a standard layered ansatz with depth matched to 28 (and comparable parameter count) on the same 12-qubit Japanese bond-rate task and the same hardware; if its KL divergence matches the LLM circuit's value of about 7.37, then circuit depth is the dominant factor and the LLM-specific contribution is not established.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, automated LLM-guided architecture search can replace manual ansatz design for quantum generative models on near-term hardware.
  • Shallow, hardware-native circuits become a concrete design target: the depth reduction from 85 to 28 is the clearest lever behind the observed noise resilience.
  • The same feedback loop—prompt with hardware constraints, evaluate, refine—can be transferred to other quantum machine learning tasks such as classifiers or feature maps, and to other device topologies.
  • The post-selection results, where KL drops to roughly 0.7 for both circuits, imply that most of the hardware degradation is correctable noise, so error mitigation and post-processing remain complementary to architectural search.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The comparison is between depth 28 and depth 85; without a depth-matched shallow baseline, the cleanest conclusion is that shallower hardware-native circuits are noise-robust, and the specific contribution of the LLM's search remains to be isolated.
  • A natural control experiment would fix the optimization loop but replace the LLM's proposals with random or evolutionary circuits constrained to depth 28; matching hardware KL would suggest the loop, not the LLM's prior knowledge, is doing the work.
  • The use of reverse KL on hardware may reward conservative distributions that avoid improbable outcomes; reporting forward KL or MMD on real-device samples would test whether the reported advantage reflects genuine distribution fidelity rather than mode-avoidance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a prompt-based LLM framework for generating hardware-aware QCBM ansätze and evaluates it on a 12-qubit financial generative modeling task (daily changes in JGB interest rates). The LLM is prompted with hardware topology, gate error rates, and basis gates, and receives iterative feedback on KL divergence, circuit depth, and validity. The LLM-generated circuit (depth 28) is compared to a Qiskit TwoLocal baseline (18 repetitions, depth ~85) on a noiseless simulator and on IBM fez hardware. Reported results: simulator KL 3.67 vs. 3.10; hardware KL 7.37±0.11 vs. 9.32±0.06 without error mitigation, and 6.91±0.13 vs. 8.92±0.07 with error mitigation. Post-selection KL values are approximately 0.78 (LLM) and 0.68 (TwoLocal). The paper claims the LLM-generated ansatz is significantly shallower and achieves superior generative performance on real hardware.

Significance. If the attribution were established, this would be a practical demonstration of LLM-guided quantum architecture search on real hardware, with relevance to QCBM-based financial modeling and to the broader goal of automating hardware-aware circuit design. The paper has genuine strengths: it reports real 12-qubit hardware measurements, uses a standard baseline, applies measurement error mitigation, and the direct numeric comparison in Table I does show a hardware advantage for the shallower LLM-generated circuit. However, the central interpretation—that LLM guidance, rather than circuit depth, produces the hardware advantage—is not supported by the current experimental design. The post-selection result also reverses the ranking. As a proof-of-concept the paper is suggestive, but additional control experiments are needed before the stated claims can be accepted.

major comments (4)
  1. [§V.B, §IV.C, Table I] The headline hardware comparison is between the LLM-generated circuit (depth 28) and TwoLocal with 18 repetitions (depth ~85). No depth-matched control is run: there is no shallow TwoLocal, no random shallow ansatz, and no noise-simulation study that varies depth independently of the design method. On the noiseless simulator the LLM circuit is worse (KL 3.67 vs. 3.10, §V.A), so the hardware reversal is fully consistent with the LLM circuit simply having less gate noise. The abstract's attribution of the improvement to LLM-guided search is therefore not established. Please add depth-matched controls and/or a noise-sweep experiment that isolates depth from the design method.
  2. [§V.B.a] After post-selection on valid bitstrings, the KL divergence is reported as approximately 0.78 for the LLM-generated ansatz and 0.68 for TwoLocal, reversing the ranking from Table I. The abstract's unqualified statement of 'superior generative performance' is not robust to a standard post-processing step. The paper should either temper the claim, report post-selected metrics in the main comparison, or explain why the ranking reversal does not affect the conclusion.
  3. [§III.A, §IV.D] The LLM search is a single run: GPT-4.1 at default temperature, with one prompt sequence and one resulting circuit. There are no independent LLM search runs, no training random seeds, and the reported hardware statistics are shot noise only. The stochasticity of LLM sampling and of QCBM training is therefore uncharacterized. Please report results over multiple independent LLM search runs and training initializations to show that the presented circuit is representative rather than a lucky draw.
  4. [§II.A] The KL divergence formula is written as D_KL(P∥Q) = Σ_x P(x) log(P(x)/Q(x)) with the sum over P(x)>0, but the text says the paper computes the reverse KL D_KL(Q∥P) with Q the model distribution. For the reverse KL the formula and support condition should be over Q, not P. Since all reported quantitative claims depend on this quantity, please clarify the exact formula actually used in the evaluation.
minor comments (5)
  1. [Title, Authors] The title contains an encoding artifact 'Ans ¨atze'; it should read 'Ansätze'. The third affiliation line lists '3nd' instead of '3rd'.
  2. [§IV.D, §V.B] The text states that GPT-4.1 and GPT-4o were used and then says GPT-4.1 was selected for all experimental results. Please make the final choice explicit earlier and avoid the impression that both models contributed results. Also, 'Mthreepackage' should be 'Mthree package'.
  3. [§IV.B] There is an inconsistency in the training budget: §IV.B says 'stochastic gradient descent over 30 epochs', but later in the same section it says training typically spans 50 to 100 epochs. Please harmonize these statements.
  4. [§III.A] The description of how the LLM 'prunes' or 'appends' layers is qualitative. A small pseudocode listing or explicit prompt template would improve reproducibility and clarify what information is fed back to the model.
  5. [§V.A, Fig. 2] The simulator comparison reports KL 3.67 vs. 3.10 and MMD convergence, but the text says 'both achieve similar performance.' The KL difference is roughly 18% relative to the baseline; please state whether this difference is within expected training variability, especially since no seeds are reported.

Circularity Check

0 steps flagged

No circularity: LLM-ansatz evaluation is an empirical benchmark; the depth/KL confound is a validity threat, not a circular derivation.

full rationale

The manuscript is an empirical architecture-search study rather than a derivation. The claimed chain is: hardware-conditioned prompts -> LLM circuit proposals -> iterative feedback (KL divergence, depth, validity) -> MMD-based training -> simulator and hardware evaluation against a TwoLocal baseline. None of these steps defines its reported performance metric in terms of the method's own inputs. The KL divergence appears both as a feedback signal to the LLM and as the headline evaluation metric, but this is an optimization loop, not a definitional equivalence; the competing baseline is trained and measured with the same protocol. The only self-citation is [13], used in the introduction as a general QML reference, and it is not load-bearing for the ansatz generation or for the hardware result. The MMD bandwidth sigma=3 is adopted from prior work [2], not from the authors' own work. The depth difference (28 vs 85) is a possible confound for the hardware comparison, but confounding is a correctness risk, not circular reasoning. Therefore no specific circular step can be exhibited and the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim relies on standard QCBM training assumptions and one hand-chosen hyperparameter (σ=3). No new physical entities are introduced. The biggest implicit assumption is that simulator-trained parameters transfer to real hardware and that the chosen evaluation metric captures generative quality.

free parameters (2)
  • MMD Gaussian kernel bandwidth σ = 3
    Hand-chosen hyperparameter for the Gaussian kernel in the MMD loss, taken 'consistent with prior work [2]'. It affects the loss during training and therefore the learned circuit parameters.
  • Qubits per JGB maturity = 4
    Data encoding choice allocating 4 qubits per maturity, yielding 12 qubits. Not fitted to data, but a modeling choice that determines the target distribution dimension.
axioms (4)
  • domain assumption MMD loss with a Gaussian kernel is a suitable training objective for QCBMs
    Used throughout Section IV.B; adopted from prior work [2] without independent validation on this task.
  • domain assumption Reverse KL divergence D_KL(Q||P) is a reliable evaluation metric on noisy hardware
    Stated in Section III as a choice for numerical stability, but no comparison with forward KL or other metrics is given.
  • domain assumption Training on a noiseless simulator transfers to real hardware without retraining
    Implied in Section IV.B where circuits are trained on aer_simulator and then executed on ibm_fez. No noise-aware training or fine-tuning is performed.
  • standard math The parameter-shift rule yields unbiased gradients for the generated circuits
    Invoked in Section IV.B citing [18]; standard result but not verified for the LLM-generated circuits specifically.

pith-pipeline@v1.3.0-alltime-deepseek · 5083 in / 12469 out tokens · 124688 ms · 2026-08-04T20:39:09.021060+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LLM-Guided Ans\"atze Design for Quantum Circuit Born Machines in Financial Generative Modeling." pith.science (2026). https://pith.science/paper/OQ47O65Q

@misc{pith2026250908385,
  author       = {Pith},
  title        = {Pith review of: LLM-Guided Ans\"atze Design for Quantum Circuit Born Machines in Financial Generative Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQ47O65Q}},
  note         = {Machine review of arXiv:2509.08385}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Quantum generative modeling using quantum circuit Born machines (QCBMs) shows promising potential for practical quantum advantage. However, discovering ans\"atze that are both expressive and hardware-efficient remains a key challenge, particularly on noisy intermediate-scale quantum (NISQ) devices. In this work, we introduce a prompt-based framework that leverages large language models (LLMs) to generate hardware-aware QCBM architectures. Prompts are conditioned on qubit connectivity, gate error rates, and hardware topology, while iterative feedback, including Kullback-Leibler (KL) divergence, circuit depth, and validity, is used to refine the circuits. We evaluate our method on a financial modeling task involving daily changes in Japanese government bond (JGB) interest rates. Our results show that the LLM-generated ans\"atze are significantly shallower and achieve superior generative performance compared to the standard baseline when executed on real IBM quantum hardware using 12 qubits. These findings demonstrate the practical utility of LLM-driven quantum architecture search and highlight a promising path toward robust, deployable generative models for near-term quantum devices.

Figures

Figures reproduced from arXiv: 2509.08385 by Romain Harang, Tetsuo Shibuya, Yaswitha Gujju.

Figure 1
Figure 1. Figure 1: Overview of our experimental setup B. Previous Works The use of LLMs for quantum circuit design is a recent and rapidly developing research direction. Typically, the LLM is treated as a black box that takes a set of instructions (a prompt) and generates textual outputs conditioned on that prompt. For example, the work in [4] introduced a supervised learning approach to ground-state problems, generating tas… view at source ↗
Figure 2
Figure 2. Figure 2: Probability distributions of the true data (green) and those generated by the QCBM using both the LLM-generated and standard (naive) ansatz, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MMD loss during training for the ansatz generated using the LLM [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 13 canonical work pages · 1 internal anchor

  1. [1]

    Circuit Design based on Feature Similarity for Quantum Generative Modeling

    Makarski, Mathis, et al. ”Circuit Design based on Feature Similarity for Quantum Generative Modeling.” arXiv preprint arXiv:2503.11983 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    ”Trainability barriers and opportunities in quantum generative modeling.” npj Quantum Information 10.1 (2024): 116

    Rudolph, Manuel S., et al. ”Trainability barriers and opportunities in quantum generative modeling.” npj Quantum Information 10.1 (2024): 116

    work page 2024

  3. [3]

    ”Optimizing Ansatz Design in Quan- tum Generative Adversarial Networks Using Large Language Models.” arXiv preprint arXiv:2503.12884 (2025)

    Ueda, Kento, and Atsushi Matsuo. ”Optimizing Ansatz Design in Quan- tum Generative Adversarial Networks Using Large Language Models.” arXiv preprint arXiv:2503.12884 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    ”The generative quantum eigensolver (GQE) and its application for ground state search.” arXiv preprint arXiv:2401.09253 (2024)

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

    arXiv 2024

  5. [5]

    ”Automating quan- tum feature map design via large language models.” arXiv preprint arXiv:2504.07396 (2025)

    Sakka, Kenya, Kosuke Mitarai, and Keisuke Fujii. ”Automating quan- tum feature map design via large language models.” arXiv preprint arXiv:2504.07396 (2025)

    arXiv 2025

  6. [6]

    Quantum Computing with Qiskit

    Javadi-Abhari, Ali, et al. Quantum Computing with Qiskit. arXiv:2405.08810, arXiv, 15 May 2024. arXiv.org, arxiv.org/abs/2405.08810

    Pith/arXiv arXiv 2024

  7. [7]

    ”Scalable mitigation of measurement errors on quantum computers.” PRX Quantum 2.4 (2021): 040326

    Nation, Paul D., et al. ”Scalable mitigation of measurement errors on quantum computers.” PRX Quantum 2.4 (2021): 040326

    work page 2021

  8. [8]

    ”Differentiable learning of quantum circuit born machines.” Physical Review A 98.6 (2018): 062324

    Liu, Jin-Guo, and Lei Wang. ”Differentiable learning of quantum circuit born machines.” Physical Review A 98.6 (2018): 062324

    work page 2018

  9. [9]

    ”A generative modeling approach for bench- marking and training shallow quantum circuits.” npj Quantum informa- tion 5.1 (2019): 45

    Benedetti, Marcello, et al. ”A generative modeling approach for bench- marking and training shallow quantum circuits.” npj Quantum informa- tion 5.1 (2019): 45

    work page 2019

  10. [10]

    Hinton, and Terrence J

    Ackley, David H., Geoffrey E. Hinton, and Terrence J. Sejnowski. ”A learning algorithm for Boltzmann machines.” Cognitive science 9.1 (1985): 147-169

    work page 1985

  11. [11]

    Kullback, Solomon, and Richard A. Leibler. ”On information and sufficiency.” The annals of mathematical statistics 22.1 (1951): 79-86

    work page 1951

  12. [12]

    ”Challenges and opportunities in quantum machine learning.” Nature computational science 2.9 (2022): 567-576

    Cerezo, Marco, et al. ”Challenges and opportunities in quantum machine learning.” Nature computational science 2.9 (2022): 567-576

    work page 2022

  13. [13]

    Gujju, Yaswitha, Atsushi Matsuo, and Rudy Raymond. ”Quantum ma- chine learning on near-term quantum devices: Current state of supervised and unsupervised techniques for real-world applications.” Physical Re- view Applied 21.6 (2024): 067001

    work page 2024

  14. [14]

    ”Quantum machine learning.” Nature 549.7671 (2017): 195-202

    Biamonte, Jacob, et al. ”Quantum machine learning.” Nature 549.7671 (2017): 195-202

    work page 2017

  15. [15]

    ”Gpt-4o: The cutting-edge advancement in multimodal LLM.” Authorea Preprints (2024)

    Islam, Raisa, and Owana Marzia Moushi. ”Gpt-4o: The cutting-edge advancement in multimodal LLM.” Authorea Preprints (2024)

    work page 2024

  16. [16]

    ”A kernel two-sample test.” The Journal of Machine Learning Research 13.1 (2012): 723-773

    Gretton, Arthur, et al. ”A kernel two-sample test.” The Journal of Machine Learning Research 13.1 (2012): 723-773

    work page 2012

  17. [17]

    ”Adam: A method for stochastic optimization.” arXiv preprint arXiv:1412.6980 (2014)

    Kingma, Diederik P., and Jimmy Ba. ”Adam: A method for stochastic optimization.” arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  18. [18]

    ”General parameter-shift rules for quantum gradients.” Quantum 6 (2022): 677

    Wierichs, David, et al. ”General parameter-shift rules for quantum gradients.” Quantum 6 (2022): 677

    work page 2022