acceptodds
Withdrawn from ICLR 2027

QuSSET: Shot Set Encoder for Quantum Property Estimation of Variable-Length States

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimating many observables from a small measurement record is difficult when each new quantum state can be sampled only a few times. We ask whether measurements from one variational-circuit family can support property estimation across both observables and system lengths. is a noncausal, variable-length shot-set encoder that preserves spatial order within each shot while aggregating repeated shots as an exchangeable set. Masked-outcome pretraining learns an ansatz-specific prior from unlabeled random-Pauli records without property labels or circuit parameters. Separate supervised heads then map the same property-agnostic records to either a symmetric isotropic spin-correlation matrix or an antisymmetric Dzyaloshinskii–Moriya (DM) correlation matrix. Pretraining and supervision use only lengths ; the resulting estimators are evaluated without further adaptation on new states at , using 64 shots per state. On a depth-4 long-range circuit family, QuSSET outperforms learned and calibrated statistical estimators using the same records, as well as deterministic estimators granted separately acquired, target-specific measurements, across both observables. Crucially, this advantage persists at the unseen lengths \(L=18,20\): without length-specific adaptation, each estimator processes longer shot records and predicts correlation matrices larger than any encountered during training. Matched controls show that masked pretraining provides its largest benefit with few labels and a persistent gain for long-range correlations. Together, these results show that one ansatz-level measurement corpus and pretrained initialization can support distinct property estimators and extrapolate to chain lengths absent from training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Property Estimation, Encoder, Quantum Measurements, Length-OOD

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hW9uOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces QuSSET, a framework for quantum property estimation across variable system sizes. It consists of masked-outcome pretraining on random-Pauli measurement records, followed by separate finetuning of the encoder and a property-specific head for each downstream task. The architecture preserves spatial order within measurement shots while treating repeated shots as an exchangeable set. Experiments on a long-range circuit ensemble show that models trained only on systems with up to 16 qubits can estimate isotropic and DM correlation matrices at unseen sizes of 18 and 20 qubits without further adaptation. At a fixed budget of 64 shots per test state, pretraining provides its largest improvement when labeled training states are scarce, with smaller gains as the labeled dataset grows.

Critical strengths

This paper proposes a well-motivated architecture that supports length extrapolation without additional adaptation. The distinction between spatial structure within shots and exchangeability across shots is appropriate for quantum measurement records. Combined with shared pairwise prediction parameters, this design allows the same estimator to process longer records and produce larger correlation matrices than those encountered during training. Although the demonstrated extrapolation is modest, this is a meaningful capability for quantum characterization, where accurate training labels may be accessible only for smaller systems.

Critical weaknesses
  1. The experimental scope is narrow, and the incremental contribution of pretraining remains limited. The evaluation uses one noiseless, depth-four ansatz with long-range interactions, and length extrapolation extends only from a maximum training size of 16 qubits to 18 and 20 qubits. The two main targets are distinct combinations of two-body Pauli correlations; the auxiliary energy analysis provides limited additional task breadth. Consequently, the experiments establish modest extrapolation within one circuit distribution, while applicability to substantially larger systems or other state families remains unverified. Moreover, Table 2 shows that the advantage over the architecture-matched scratch model generally diminishes as the number of labeled states increases. At (N_t=405), the OOD RMSE reductions are approximately 1.4% for isotropic correlations and 1.2% for DM correlations. Appendix E further indicates that the largest low-label gains depend partly on the optimization schedule.

  2. The positioning and comparisons against closely related learned approaches are insufficient. The paper includes useful statistical baselines and matched controls, but its principal external learned comparator is an adapted LLM4QPE-like model. The relationship to earlier quantum-state representation and property-estimation frameworks should be discussed more thoroughly. GQNQ by Zhu et al. (Nature Communications, 2022) learns measurement-based representations for predicting unperformed measurement statistics. Wu et al. (Nature Communications, 2024) study multi-task property estimation and representation transfer across system sizes. Du et al. (Nature Communications, 2025) develop property prediction for bounded-gate circuits using a different input setting. The manuscript should explicitly distinguish these supervision and inference settings, identify the resulting methodological advance, and provide a convincing comparison with compatible learned estimators under comparable information budgets. Without this positioning and evaluation, it is difficult to determine how much the proposed framework advances existing approaches beyond a tailored architecture for the selected benchmark.

Questions for authors
  1. Could the authors clarify the practical setting and the specific limitation of existing methods that motivate QuSSET’s reuse across properties and system sizes? Why should predicting stochastic measurement outcomes produce features that improve downstream property estimation, and what information is expected to remain reusable at unseen system sizes?
  2. What is the scaling behavior when generalizing the method to much larger systems, e.g., 30 - 50 qubits?
  3. Can the method be applied in other types of ansatz? Can the method efficiently characterize ground state properties when setting the parameters of the ansatz circuit to a ground state preparation circuit?
  4. Does the method support predicting non-linear properties like the R'enyi entropy?
  5. What is the contribution of single-qubit statistics? The calibrated baselines in Appendix A.3 use two-site measurement statistics, but omit single-qubit expectation values and their products. How does the method compare with a simple estimator that incorporates these additional features from the same 64-shot records?
Reviewer_vxRBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed QuSSET uses masked-outcome pretraining to learn a variable-length encoder from random-Pauli measurement records. The same pretrained backbone is then adapted to estimate two correlation matrices and generalizes from systems of up to 16 qubits to 18 and 20 qubits.

Critical strengths
  1. The architecture appropriately treats shots as an unordered set while preserving the spatial order of qubits. And the experiments include matched controls and evaluate two distinct observables using the same measurement corpus.

  2. The method performs well at unseen system lengths without length-specific adaptation.

Critical weaknesses
  1. The experiments use noiseless simulated measurements from one circuit family, so performance on real hardware and other state distributions remains unknown.
  2. The main benefit of pretraining is concentrated in the low-label setting and becomes small with more labels. The appendix also shows that part of the low-label gain depends on the training schedule.
  3. Length extrapolation is limited to 18 and 20 qubits, only slightly beyond the maximum training length of 16.
Questions for authors
  1. How robust is QuSSET to realistic measurement and gate noise? If access to real quantum hardware is unavailable, could you evaluate it using noisy simulations, such as those incorporating depolarizing noise?
  2. Does QuSSET generalize to different circuit families or substantially larger system sizes?
Reviewer_JMH2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a quantum shot-set encoder for transfer model for finite-shot property estimation. It is able to extrapolate length beyond the training range and uses nested labeled subsets for data modeling.

Critical strengths

The proof of concept experiments show that QuSSET is better than the models it compares with. The math and model development seem rigorous and what is expected of ICLR.

Critical weaknesses

The weakness is that it reads as though it were summarized by AI in parts.

Questions for authors

I do not have questions.

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

Related papers

Discussion (0)

Sign in to comment.