acceptodds
Under review as a conference paper at ICLR 2027

Agent-VQE: Learning Transferable Quantum Circuit Programs for Gradient-Free Deployment at Scale

Abstract

The high measurement overhead of gradient-based optimization is a bottleneck for deploying the variational quantum eigensolver at scale. We introduce Agent-VQE, an AI-assisted framework that learns transferable quantum circuit programs from small systems for gradient-free deployment. The agent analyzes local Hamiltonian features to propose circuit architectures and parameter-sharing rules, then iteratively refines these designs through small-system optimization and energy evaluation. The resulting circuit programs encode reusable circuit-building rules and shared circuit parameters. Larger circuits directly reuse these rules and coefficients without target-side gradient evaluations or parameter updates. Across six spin and fermionic many-body benchmarks spanning chain, ladder, and fully connected interaction structures, we fit circuit programs for ground-state energy estimation on systems of at most 16 qubits and evaluate the resulting circuits in noiseless simulations with 100–1000 qubits. Relative differences from reference energies remain below for five of the six benchmarks, while Fermi–Hubbard has the largest difference of . We further execute the learned one-dimensional transverse-field Ising model (TFIM) circuits on an OriginQ quantum computer. Measurements from 4–16-qubit circuits support measured-energy extrapolation up to 38 qubits. A separate comparison shows that the learned 32-qubit circuit achieves lower measured energy than randomized circuits of the same logical depth, despite hardware noise. These results support small-system circuit learning as a route to scalable ground-state energy estimation without costly target-side optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Agentic, Transfer Learning, Ground States

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BwDtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Agent-VQE, which leverages large language models to assist in designing quantum circuit programs capable of scaling with system size. Each program comprises a circuit layout, parameter assignment rules based on local Hamiltonians or geometric features, and shared coefficients obtained through joint optimization on small-scale systems. Following training on small systems, the program structure and coefficients are fixed to directly generate circuits for larger systems, eliminating the need for further structural search or parameter optimization at the target scale. Independent programs are learned for six benchmarks, with transfer occurring across different sizes within the same Hamiltonian family.

Critical strengths

For inhomogeneous systems such as Fibonacci and Aubry–André models, assigning gate angles based on local coupling or field information aligns better with the physical structure than using globally uniform parameters, while ensuring the number of learnable parameters does not scale with system size. The Fibonacci experiments with matched circuit depths and parameter counts demonstrate the efficacy of context-conditioned parameterization.

The role of local context-based parameter sharing in systems with inhomogeneous interactions offers valuable insight. However, the paper fails to provide sufficient evidence regarding the necessity of the agent, search success rates, or computational costs, nor does it clearly establish advantages over existing scalable circuit design methods.

Critical weaknesses

SC-ADAPT-VQE has already demonstrated the concept of deriving circuits from small systems and deploying them on large systems without further optimization; the authors also cite works involving the discovery of scalable ansätze from few-body instances and the iterative design of quantum circuits using LLMs. Given the name "Agent-VQE," the paper needs to explicitly demonstrate the incremental value provided by the agent whether in terms of circuit discovery, contextual rule design, resource efficiency, or the reduction of manual design effort.

The Fibonacci ablation study demonstrates the effectiveness of parameter allocation based on local coupling, but it fails to show that the agent-based approach is superior to manually specified rules or other search methods. The paper lacks comparisons under equivalent budgets against manual physical templates, LLMs without feedback loops, and non-LLM search methods; furthermore, it does not report success rates, failure cases, or resource consumption across multiple independent runs.

The Schwinger model comparison also fails to fill these evidentiary gaps. While the authors specify a circuit depth of 32, they do not provide the full circuit, gate counts, or comparable optimization budgets.

Figure 4 evaluates the size-predictability of noisy outputs by fitting a straight line to noisy energy measurements from small-scale systems and comparing this against measurements from large-scale systems. However, in my view, this experiment neither proves that the measured values ​​approach the true ground state nor demonstrates a resource advantage at equivalent precision. The necessary evidentiary links connecting numerical transferability, the benefits of agent-based design, and hardware practicality remain missing.

Questions for authors

What are the quantitative advantages of Agent-VQE compared to manually specified physical context templates, non-LLM search methods, and LLM-generated proposals lacking iterative feedback given the same source system and computational budget? While AI agents are powerful, there must be compelling reasons for their selection and a clear basis for comparison.

Please explicitly distinguish between design decisions proposed by the agent and those specified or revised by humans, and provide specific technical details for each model.

What specific gains are realized through the feedback and memory mechanisms? If iterative feedback or historical memory were removed while retaining the same LLM and numerical optimizer, would performance degrade significantly?

Reporting results solely for "GPT-5.6 Sol," without comparing different LLMs or agent configurations, makes it impossible to assess the importance of model selection or to validate the advantages of the agent-based mechanism.

Overall, I do not believe that transferring a model trained on a small system to a large system inherently requires an agent; the SC-ADAPT-VQE already employs a similar approach. Therefore, the introduction of an agent must demonstrate tangible benefits: more efficient circuit design, reduced expert involvement, or the discovery of superior solutions within the same budget. Unfortunately, the paper lacks sufficient experimental evidence to substantiate these claims: there is no quantification of human effort, no fair comparison against expert-designed templates, and no budget-matched comparison with alternative high-level design methods such as feedback-free LLM generation or random/evolutionary search.

Reviewer_wsNjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes an agent for VQE to automatically generate quantum circuits for larger-scale VQE problems without further optimization after training on smaller-scale instances. Numerical results on several quantum spin models demonstrate that the generated circuits achieve relative energy estimation differences below 1% on most of the benchmarks.The authors also deploy the generated circuit to OriginQ quantum computer. It achieves lower energy than random circuits with the same depth.

Critical strengths
  • The research topic is valuable for improving the scalability of VQE for quantum many-body systems.

  • The experiments are conducted on different quantum spin models, covering numerical simulations with size up to 1000 qubits and evaluations on physical hardware.

Critical weaknesses
  • The concept of the proposed framework is clear, but the implementation is not sufficiently specified. Besides, the necessity of AI agent is unclear.

  • The approximately linear energy scaling in Figure 2 is largely expected from the extensivity of the local quantum many-body Hamiltonians. Therefore, these scaling trends provide limited evidence to support the effectiveness of the proposed method.

  • The manuscript does not sufficiently evaluate the variability of the generated circuits and their performance across independent agent runs.

Questions for authors
  • Since the circuit generated by the AI agent may vary across independent runs, how sensitive are the resulting circuit architectures and achievable energies to this randomness?

  • When deploying the generated circuits on quantum hardware, device noise can substantially affect their performance. Does the proposed agent account for these hardware constraints during circuit design?

Reviewer_r13AOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the cost of optimizing variational quantum circuits at large system sizes. Specifically, the authors propose Agent-VQE, which is an LLM-assisted workflow that designs circuit layouts and context-dependent parameter-sharing rules, while conventional numerical optimizers fit shared coefficients on small systems. The resulting programs generate larger circuits without target-side optimization. The main result is that small-system circuit design and fitting can produce accurate, transferable ground-state energy estimators while reducing model-specific design effort. The authors conduct comprehensive experiments with six many-body benchmarks at 100~1000 qubits. The author also compares with SC-ADAPT-VQE on the Schwinger model and conducts TFIM experiments on quantum hardware.

Critical strengths
  1. The paper demonstrates a novel and feasible way to represent and express size transfer. The author specifically separates circuit layouts, parameter-assignment rules, and shared coefficients makes the frozen-deployment procedure clear.

  2. The experiments are comprehensive and cover several distinct Hamiltonian structures. The experiments include uniform and quasiperiodic chains, a ladder, fermionic interactions, and a fully connected model. The authors reported relative energy differences remain below 1% across the tested large sizes for five of the six benchmarks.

  3. The paper distinguishes simulation from hardware results and recognize the hardware noise limit. These details help assess the scope of the results.

Critical weaknesses
  1. The specific contribution of the agent is not very clear to me. The author describes the general propose-optimize-evaluate workflow, but the experiments do not compare it against a fixed physics-informed ansatz, non-agent search, or a one-shot LLM design under matched optimization budgets. The Fibonacci experiment isolates the benefit of parameter conditioning, rather than the benefit of agent-guided discovery.
  2. The SC-ADAPT-VQE comparison does not establish a clear advantage. The authors report lower Schwinger-model energies using a depth-32 program trained at 12, 16, 20 qubits. However, Table 1 does not compare elementary gate counts, compiled depth, parameter counts, or search costs against the published SC-ADAPT-VQE circuits. The result therefore establishes better energy for the reported configurations, but not superior accuracy at comparable resources.
  3. The hardware results demonstrate a benefit over randomized circuits, but not useful ground-state accuracy. The hardware result support a narrower conclusion than practical ground-state preparation, so a more appropriate address of this would be beneficial.
Questions for authors
  1. What improvement comes specifically from the iterative agent? Can you conduct ablations to demonstrate such advantage?
  2. How does the learned hardware circuit compare with a simple product-state preparation on the same device? This would clarify whether the observed improvement over randomized circuits translate into useful ground-state estimation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.