An LLM System for Autonomous Variational Quantum Circuit Design

Source: arXiv:2606.13380 · Published 2026-06-11 · By Kenya Sakka, Wataru Mizukami, Kosuke Mitarai

TL;DR

This paper addresses the challenge of automating the design of high-performing quantum circuits, a process historically reliant on human expert intuition and iterative manual refinement. The authors propose an autonomous, large language model (LLM)-driven agentic framework that integrates seven components—Exploration, Generation, Discussion, Validation, Storage, Evaluation, and Review—to perform iterative quantum circuit design under explicit constraints. Unlike prior works that focus on isolated sub-tasks or fixed templates, this system implements a closed-loop workflow combining web-based research, multi-expert discussion grounded in literature, automated code synthesis, and experimental feedback for active refinement.

They evaluate the framework on two distinct tasks: construction of quantum feature maps for quantum machine learning, and ansatz design for variational quantum eigensolvers (VQE) in quantum chemistry. The best-generated quantum feature maps outperform standard quantum kernels and, when scaled to larger qubit counts, surpass classical radial basis function kernels on image classification datasets like MNIST and CIFAR-10. For the molecular ground-state estimation task across seven molecules, the generated ansatz circuits achieve competitive accuracy compared to widely used chemisty-inspired and hardware-efficient ansatzes, while also respecting critical scaling constraints for gate count and parameters. These results highlight the system’s ability to generate scalable, interpretable, and high-performing quantum circuits autonomously, establishing LLM-driven agentic systems as a viable paradigm for scientific workflow automation in quantum algorithm design.

Key findings

• The system’s best quantum feature map outperforms representative quantum baseline feature maps and the classical radial basis function (RBF) kernel on MNIST, Fashion-MNIST, and CIFAR-10 datasets.
• Generated quantum feature map circuits maintain logical consistency and scalability across varying numbers of qubits, demonstrating design principles that generalize beyond fixed circuit sizes.
• For VQE molecular ground-state energy estimation, the generated ansatz circuits achieve accuracy comparable to unitary coupled cluster and hardware-efficient ansatzes across seven molecules, while having significantly fewer gates and parameters.
• Iterative agentic refinement guided by LLM-powered multi-expert discussion reduces evaluation of low-potential designs, enhancing computational efficiency compared to brute-force search.
• The use of external domain knowledge from 2,785 academic papers and 385 quantum library classes in a vector database enhances idea critique and code correctness, addressing LLM internal knowledge limitations.
• Validation with static and dynamic code analysis incorporating quantum library documentation and web-based code error search reduces runtime errors and ensures executable quantum circuits.
• Incorporation of domain-specific review criteria (e.g., particle number preservation, symmetry constraints) guides ansatz designs toward physical interpretability and hardware feasibility.

Threat model

n/a — The paper focuses on autonomous scientific workflow and quantum circuit design automation rather than adversarial or security threats.

Methodology — deep read

1. Threat Model & Assumptions: The system assumes an environment where adversarial interference is not central; the challenge is automating quantum circuit design under explicit task constraints using LLMs with external domain knowledge. It assumes access to scientific literature and quantum software libraries but does not model adversarial manipulation of outputs or data.

2. Data: The framework leverages a VectorDB containing 2,785 academic papers related to quantum machine learning, variational quantum eigensolvers, and quantum chemistry, as well as 385 classes from quantum programming libraries (e.g., PennyLane). These documents provide the external knowledge base for critique and reference. Benchmark datasets used for evaluation include MNIST, Fashion-MNIST, CIFAR-10 (image classification), and seven molecules relevant for ground state energy estimation.

3. Architecture & Algorithm: The core system contains seven components:

• Exploration: Uses a research agent combining web search and code interpreter to generate a comprehensive research report and seed diverse candidate designs.
• Generation: An LLM generates multiple candidate circuit design ideas and corresponding Python code implementing quantum feature maps or ansatz circuits, structured with overview, mathematical formulas, and literature references.
• Discussion: An automated multi-agent system assigns expert roles (e.g., quantum algorithm expert, quantum chemistry theorist). The critic LLM generates pointed critique questions using the literature vector DB; expert LLMs answer; the advocate LLM evaluates critiques and refines circuit ideas iteratively for up to three rounds.
• Storage: Maintains the vector DB of papers and local JSON storage of programming library documentation for reference during critique and code validation.
• Validation: Enforces static (syntax, API usage matching library docs) and dynamic (test runs on dummy data) validity of generated code. Errors feed back to LLM with web search assistance for correction, up to three retry attempts.
• Evaluation: Executes the validated circuit code on target tasks, measuring accuracy (QML classification accuracy, VQE energy estimation error), resource metrics (gate, parameter count), and optimization difficulty.
• Review: Analyzes recent trial results with LLM guidance, highlighting success factors and areas for improvement. Provides rich feedback (e.g., performance metrics, computational cost, domain-specific criteria) that informs the next generation iteration.

4. Training Regime: Not a traditional ML training setup; the LLM is prompted with carefully designed templates to produce candidate ideas and code. Iterative prompting with feedback from Evaluation and Review drives progressive refinement. The maximum number of retries for code validation is three; discussion component involves up to three rounds of critique and refinement.

5. Evaluation Protocol: The system is evaluated on two tasks. For QML quantum feature maps, classification accuracy on MNIST, Fashion-MNIST, and CIFAR-10 is measured versus classical RBF kernels and established quantum maps. For VQE ansatz generation, ground-state energy estimation error is compared against unitary coupled cluster and hardware-efficient ansatzes across seven molecules. Resource metrics (gate/parameter count scaling) and training iteration statistics are also assessed. Ablation on impact of Discussion component and scaling tests are described. Cross-validation details are not explicit; datasets are standard benchmarks.

6. Reproducibility: The authors provide open access to prompt configurations, code, and implementation details at a public repository. The use of publicly available datasets and open quantum libraries supports reproducibility. However, the exact vector DB content and internal LLM weights are not disclosed.

Concrete example: The process begins by querying the Exploration component to generate a research report and diverse seed ideas for a quantum feature map for MNIST classification. The top 10 seed ideas are experimentally evaluated; 5 highest performers are selected and passed iteration to the Generation phase. The LLM then refines designs based on Review comments, and Discussion experts critique and suggest improvements, with advocate refining ideas. Validated code is executed and evaluated for accuracy and resource use. The loop iterates until the generated circuit exceeds classical and quantum baselines, producing scalable Python code compatible with different qubit sizes.

Technical innovations

• Integration of a seven-component closed-loop agentic framework combining web-based exploration, multi-perspective LLM critique, code generation, validation, and experimental feedback for autonomous quantum circuit design.
• Use of natural language multi-agent discussion between critic, multiple domain-specific expert roles, and advocate LLMs informed by academic literature vector searches to iteratively critique and refine quantum circuit design ideas.
• Implementation of scalable quantum circuit generation producing executable Python code independent of qubit count, enhancing adaptability across quantum system sizes.
• A dynamic code validation loop incorporating both static API checks against local quantum library documentation and dynamic test runs, augmented by LLM-guided web search for code error rectification.

Datasets

• MNIST — 70,000 images — public benchmark
• Fashion-MNIST — 70,000 images — public benchmark
• CIFAR-10 — 60,000 images — public benchmark
• Quantum chemistry molecules dataset — 7 molecules — public quantum chemistry benchmarks
• Academic literature vector DB — 2,785 quantum ML, VQE, quantum chemistry papers — curated corpus
• Quantum programming library classes — 385 classes from PennyLane and related libraries — curated local storage

Baselines vs proposed

• Classical RBF kernel (MNIST): accuracy ~baseline; Proposed LLM-generated quantum feature map: accuracy surpasses RBF kernel by margin (exact number not specified).
• Representative quantum feature maps (MNIST): baseline accuracy X%; Proposed feature map: outperforms by several percentage points (Fig. 3 referenced but exact delta unclear).
• Unitary coupled cluster ansatz (VQE): baseline ground state estimation error E; Hardware-efficient ansatz: error E+delta; Proposed ansatz: error comparable or lower than both, with significantly fewer gates and parameters.
• Seed ideas count reduction: from 10 initial seed feature maps to 2 final candidates after evaluation, improving overall efficiency.

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2606.13380.

Fig 1: Overview of the proposed agentic system for automatic quantum circuit generation. The system

Fig 14: Kernel matrices of representative classical kernel methods, the worst-performing seed feature map, and the

Fig 15: Circuit diagrams of the five seed ans¨atze.

Fig 4 (page 45).

Fig 5 (page 45).

Fig 6 (page 45).

Fig 7 (page 45).

Limitations

• The system validation is limited to simulation and benchmark datasets; no demonstration on live quantum hardware subject to noise or device-specific errors.
• The multi-expert discussion uses LLM-based surrogate experts rather than true human experts, which may limit the depth or originality of critique.
• The impact of hyperparameters in prompting strategy, number of discussion rounds, and retry limits on final performance is not thoroughly ablated.
• Scaling tests assess logical scalability by code parameterization but do not explicitly analyze execution cost or noise sensitivity on large-scale quantum devices.
• The framework’s dependency on external web resources during validation may introduce variability or reproducibility challenges.
• Adversarial settings or robustness of generated circuits under worst-case noise or attack scenarios are not considered.

Open questions / follow-ons

• How well does the framework generalize to circuit design tasks involving noisy intermediate-scale quantum (NISQ) hardware with device-specific constraints?
• Can human-in-the-loop integration enhance the multi-expert discussion component to improve circuit quality beyond LLM-generated critique?
• What are the limits of interpretability and insight that purely LLM-driven systems can provide for complex quantum algorithm design?
• How might this agentic framework be extended to automate parameter optimization and noise mitigation techniques alongside circuit structure generation?

Why it matters for bot defense

While not directly related to bot defense or CAPTCHA, this paper exemplifies how large language models can be integrated into closed-loop agentic workflows to autonomously generate, evaluate, and refine complex computational artifacts—in this case, quantum circuits. For bot-defense practitioners, this reinforces the potential of LLMs beyond static code generation toward iterative, multi-agent refinement cycles that incorporate external knowledge and multi-perspective critique. Similar principles could inspire advanced CAPTCHA generation and validation schemes where challenges and defenses evolve adaptively based on automated multi-step reasoning and exploitation of domain-specific knowledge. The validation and feedback components highlight effective uses of dynamic testing and domain-aware error correction that could inform robustness evaluations in CAPTCHAs or bot detection models.

Cite

@article{arxiv2606_13380,
  title={ An LLM System for Autonomous Variational Quantum Circuit Design },
  author={ Kenya Sakka and Wataru Mizukami and Kosuke Mitarai },
  journal={arXiv preprint arXiv:2606.13380},
  year={ 2026 },
  url={https://arxiv.org/abs/2606.13380}
}

Read the full paper

• arXiv:2606.13380 (abstract)
• arXiv:2606.13380 (PDF)