acceptodds
Withdrawn from ICLR 2027

MAQCal: A Bounded Sequential Decision Workflow for Quantum Readout Calibration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum readout calibration repeatedly requires experimenters to locate a resonator frequency and select a readout power, with sweep ranges, resolutions, and stopping decisions often configured from device-specific experience. We present MAQCal, a bounded sequential decision workflow for these two coupled stages. Stage-specific planners propose typed sweep actions from measurement feedback. A deterministic runtime enforces stage order, numerical constraints, and iteration limits, and selects the returned estimates. We provide the planner interfaces and execution rules to support reproducibility. We also study a weighted -nearest-neighbor (kNN) prior for frequency-search initialization using simulated chip records. On a simulation-driven dataset of 2006 chip designs and three synthetic noise families, the mean of MAQCal's relative SNR gains over the strongest fixed preset in each of 15 reported settings is 13.7%, with setting-specific gains from 4.4% to 23.4%. Stage-wise and scan-budget experiments show consistent improvements in the evaluated frequency and power decisions. MAQCal automates repeated sweep choices and improves readout quality over the tested fixed configurations in simulation under matched downstream processing.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum Readout Calibration, Automated Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gMLfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MAQCal, a bounded sequential workflow for quantum readout calibration that coordinates resonator frequency localization and readout power selection. It combines stage-specific planners with deterministic execution constraints, feedback-driven sweep selection, and a weighted kNN initialization prior. Simulation experiments across 2,006 chip designs and three synthetic noise families report improved readout SNR and calibration accuracy relative to fixed sweep configurations.

Critical strengths
  1. Practical and reproducible workflow design. The separation of planning and deterministic execution, including numerical constraints, stopping conditions, and typed actions, makes the system auditable and potentially useful for laboratory automation.
  2. Consistent simulation improvements. Evaluation across multiple noise conditions and stage-specific analyses provides evidence that feedback-driven calibration can outperform predefined sweep schedules.
Critical weaknesses
  1. Insufficient comparison with competitive adaptive optimization methods. The main baselines are fixed sweep presets. Comparisons against Bayesian optimization, adaptive experimental design, or deterministic feedback-based controllers are needed to establish the advantage of the proposed planning approach.
  2. Limited evidence of practical generalization. All principal results are simulation-based, without validation on physical quantum hardware. Moreover, the documented deterministic fallback controller is not separately evaluated end to end, making the contribution of the agent-based planning component difficult to isolate.
Questions for authors
  1. How does MAQCal compare against budget-matched Bayesian optimization or other adaptive calibration algorithms under the same simulation and measurement constraints?
  2. What fraction of the observed gains remains when the stage-specific planners are replaced by the deterministic fallback controllers described in the appendix?
Reviewer_SJaCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents MAQCal, a bounded sequential decision workflow for the two coupled stages of superconducting readout calibration: localizing the resonator frequency (Stage I) and choosing the readout power at a stable operating point (Stage II). GPT-4o mini planners receive structured JSON summaries of fit diagnostics and propose typed sweep actions; a deterministic Python runtime enforces stage order, clamps proposals through numerical guards, falls back to a deterministic action on request or parsing failure, and selects the returned estimates by fixed scoring rules. A weighted kNN retrieval prior over 200 simulated chip records seeds the frequency search. On 2006 Qiskit-Metal/HFSS chip designs under three synthetic noise families and eleven noise levels, MAQCal improves readout SNR over the strongest of six fixed scan presets in all 15 settings, mean relative gain 13.7% (range 4.4-23.4%), with stage-wise ablations, a matched scan-budget study, a 20-round drift study and a retrieval-size study.

Critical strengths
  1. A rigorously bounded, fully specified agent interface. Constrained action spaces, schema validation, numeric guards, bounded iteration and a deterministic fallback make the system auditable and reproducible, and the interface documentation is a model for the subfield.

  2. Breadth and honesty of the evaluation. 2006 designs, three noise families, eleven levels, six presets, per-stage ablations, a matched-budget study, a drift study and a retrieval-size study, with the limitations and the unresolved attribution stated plainly rather than buried.

Critical weaknesses
  1. The paper never runs the control experiment it exists to make, and the comparison gives the proposed method oracle information. The rule-based fallback controller is already implemented and would settle the attribution question on the existing benchmark; running it is a few GPU-hours, and the paper leaves it to "future evaluation". The Stage II oracle compounds this. Together, no claim in the paper can currently be attributed to LLM planning.

  2. The closest published system is neither cited nor differentiated. LLM-agent calibration of superconducting processors already exists in peer-reviewed form before the contemporaneous-work cutoff: "Agents for self-driving laboratories applied to quantum computing" (arXiv:2412.07978), published as "Automating quantum computing laboratory experiments with an agent-based AI framework" in Patterns 2025 (DOI 10.1016/j.patter.2025.101372), which drives closed-loop calibration of a superconducting processor with LLM agents organised as state machines. The submission does not cite it. What remains specific here - the bounded typed-action interface, the two-stage frequency-to-power coupling, and the quantitative comparison against fixed presets - should be stated as the contribution.

Questions for authors
  1. What happens when the fallback controller runs end to end (no planner at every iteration) under the protocol of Table 1? How much of the SNR gain survives without the LLM?

  2. Does Stage II keep its advantage if the simulator's ground-truth noise-level metadata is removed from the planner payload and from the fallback rule? If not, how would a hardware deployment obtain an equivalent estimate?

  3. From the recorded traces: what fraction of planner proposals were altered by the numerical guards, how often did the fallback fire, and how many model calls occur per episode?

  4. What are the total frequency-sample and power-sample counts per chip for each method in Table 1, and can the matched-budget analysis be extended to end-to-end SNR across the full noise grid? Wall-clock time and API cost are also unmeasured.

  5. (Clarification.) What exactly is aggregated in Figures 3-7 (mean or median over the 1806 evaluation chips, number of simulator repetitions), and what are the uncertainties on the headline numbers?

  6. (Suggestion for future work.) All results come from a bespoke simulator whose fidelity to real readout physics is never validated against measured data, and no standard errors are reported. A comparison against measured data from a real device, or a validation of the noise families against hardware spectroscopies, would substantially strengthen the study.

Reviewer_r2WzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents MAQCal, an LLM-based workflow for superconducting quantum readout calibration. It uses two sequential planners to adjust frequency and power sweeps based on measurement feedback, while a deterministic Python runtime handles execution, fitting, constraints, and final parameter selection. The paper also evaluates a weighted kNN prior for frequency-search initialization. Experiments on simulated chip designs under three synthetic noise families show an average relative SNR improvement of 13.7% over the best fixed preset in each of 15 reported settings. Additional experiments examine frequency localization, power selection, measurement budgets, drift tracking, and retrieval-based initialization.

Critical strengths

The paper presents a well-structured two-stage calibration workflow. The planner interfaces, numerical constraints, stopping rules, and deterministic fallback are described in detail, making the implementation easy to follow and inspect.

Critical weaknesses
  1. The research gap and methodological novelty are not clearly established. Previous studies have explored feedback-driven quantum calibration [1] and LLM-based agents for closed-loop quantum experiments [2]. Although these works address different tasks, the paper does not sufficiently explain what gap remains for MAQCal to address. The proposed method mainly combines LLM-based planning, finite-state coordination, and deterministic execution. The two-stage design also follows the natural dependency between frequency localization and power selection. While the resulting workflow is well structured and auditable, its methodological novelty beyond integrating existing techniques for this specific calibration task remains limited.
  2. The experiments show that MAQCal outperforms six fixed sweep presets, but do not establish whether the LLM planner offers additional benefits over conventional adaptive control. Appendix C.3 already provides a deterministic fallback that can be enabled using planner=None, while retaining the same fitting routines, numerical guards, and candidate selection. However, this baseline is not evaluated. Without this comparison, it remains unclear whether the LLM planner provides any advantage over a rule-based adaptive controller. The comparison also does not fully account for calibration costs. Although the paper includes a matched-budget experiment, it only covers frequency localization rather than the complete frequency and power calibration process. The total measurement cost and calibration time are not reported either. Therefore, while MAQCal improves upon the tested fixed presets, its additional benefit over conventional adaptive control and its performance under comparable overall calibration costs remain unclear.

[1] 2025 – Efficient Qubit Calibration by Binary-Search Hamiltonian Tracking

[2] 2025 – Automating Quantum Computing Laboratory Experiments with an Agent-Based AI Framework

Questions for authors
  1. How does MAQCal differ from previous work on feedback-driven quantum calibration [1] and LLM-based quantum experiment automation [2]? What specific research gap does it address, and which parts of the proposed workflow are methodologically new?
  2. Could the authors evaluate the deterministic fallback described in Appendix C.3 (planner=None) under the same experimental settings? How does MAQCal compare with this rule-based adaptive controller?
  3. The matched-budget experiment only covers frequency localization. Could the authors compare the complete two-stage workflow under matched measurement budgets and report the total measurement cost and calibration time?

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

Related papers

Discussion (0)

Sign in to comment.