acceptodds
Under review as a conference paper at ICLR 2027

Variational Relevance-Informed Sparsification of Term Measurements for Variational Quantum Algorithms

Abstract

Repeated measurements are a major bottleneck in variational quantum algorithms (VQAs). We uncover and characterize parameter-specific sparsity in the matrix of qubit-wise commuting (QWC) group contributions to parameter-shift gradients: most absolute contribution mass concentrates in a subset of groups whose relevance varies across parameters and optimization states. Building on this structure, we introduce a learned controller that predicts group relevance from Hamiltonian, circuit, light-cone, and current-state features, then selects measurements through a tunable coverage threshold. Exact contributions provide offline supervision without being required at deployment. Across 15 molecular VQE settings, our method saves 49.6% of nominal gradient-measurement shots with a mean absolute energy gap of 5.67 mHa to Full-QWC. We then ask whether large language models (LLMs) can solve the same selection problem without fine-tuning. The results reveal both promise and a clear limitation: LLMs can identify useful measurement groups, yet without guidance, local models do not consistently prioritize the physical information that matters most. Our learned controller uses these cues more effectively. This contrast points toward a broader direction for AI-assisted science: combining the flexibility of general-purpose models with explicit physical representations, task-specific objectives, and domain-grounded validation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Variational Quantum Algorithms, Measurement Sparsification, Learning to Rank, Large Language Models

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

This paper aims to address the measurement bottleneck in Variational Quantum Algorithms (VQAs) by identifying and exploiting a structural property termed parameter-specific sparsity, which characterizes the contributions of qubit-wise commuting (QWC) measurement groups to parameter-shift gradients. Based on this observation, the authors propose a learned controller that predicts which measurement groups are most relevant to each parameter, thereby reducing the measurement overhead of gradient estimation. The authors further investigate whether LLMs, without task-specific training, can identify relevant measurement groups and achieve similar performance.

Critical strengths
  1. The paper addresses the measurement bottleneck in VQAs, which remains a significant challenge for the practical application and scalability of quantum machine learning and variational quantum algorithms. 2.The paper provides extensive numerical experiments to evaluate the proposed approach under different settings.
Critical weaknesses
  1. The paper lacks several necessary preliminary definitions and explanations, making it difficult for audiences who are not familiar with quantum computing or measurement optimization to follow the proposed methodology.
  2. The paper is difficult to follow due to numerous undefined and confused notations. Several important mathematical quantities and concepts are introduced without sufficient explanation, which obscures the core methodology and makes it challenging to understand.
  3. The paper would benefit from a more comprehensive discussion of existing approaches to reducing measurement overhead in VQAs. For example, arXiv:2306.03227v3 discusses related ideas involving the grouping of simultaneously measurable terms to reduce measurement costs.
  4. Several essential components required to understand and evaluate the proposed approach, including the loss functions, evaluation metrics, and experimental settings, are presented primarily in the appendix. These details should be sufficiently explained in the main text to improve readability and reproducibility.
  5. Rather than clearly presenting the motivation, key insights, methodology, and principal results of the proposed approach, the paper devotes considerable space to experiments involving LLMs. However, the motivation for introducing LLMs and their specific role in addressing the measurement bottleneck are not sufficiently justified.
Questions for authors
  1. In Fig.1, What is the precise definition of the "contribution-mass percentages"? Additionally, the meanings of "parameter bin" and "QWC-group bin" are unclear. Please provide formal definitions and explain how these quantities are computed.
  2. In Fig.2, How are the QWC groups constructed? Why are identical shift-angle values represented by different colors? What does the index KK in GKG_K represent? In the main text, GG denotes the size of G\mathcal{G}, which makes the notation potentially confusing. Why is CC represented as a vector in the figure, whereas it is defined as a contribution matrix in the main text? What is the definition of SS in the figure caption?
  3. In line 145, what is the definition of w^\hat{w}? What does the expression "groups in SjS_j" mean? Please clarify the notation SjS_j and SS because SS also refers to the number of measurement shots.
  4. In table 1, what are the specific experimental configurations corresponding to "Depth 1" and "Depth 2"? Additionally, please clarify the definition of "cell-wise shots" and how they are allocated.
  5. How are the top-τ\tau measurement groups selected? Please specify such ranking criterion.
  6. Is there any theoretical analysis of the bound of gradient estimation error introduced by exploiting parameter-specific sparsity in QWC groups?
  7. In line 161, what exactly is meaning by the "structural-OOD test"? Please clarify the definition of structural out-of-distribution generalization and specify which structural properties differ between the training and evaluation settings.
  8. In table 2 and line 236, what does the budget parameter bb represent? Please provide a precise definition, including how this budget is determined and allocated in the experiments.
Reviewer_XPqwOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes learning parameter-specific relevance scores for qubit-wise commuting measurement groups to reduce the measurement cost of variational quantum algorithms. A controller trained on exact gradient contributions selects a subset of groups for each circuit parameter using a coverage threshold. The authors evaluate the resulting accuracy-cost trade-off on small molecular VQE problems, with additional transfer, noise, and shot-allocation experiments. They also compare the controller with general-purpose LLMs and investigate how different models use the supplied physical features.

Critical strengths

The parameter-specific treatment of measurement relevance is interesting. Rather than selecting the same measurement groups for every parameter, the method accounts for differences in their contributions to individual gradient components.

Critical weaknesses
  1. The method only reduces one component of VQE measurement cost. It sparsifies the number of QWC groups measured for each parameter, but it does not remove the linear 2P2P parameter-shift overhead or the potentially large shot complexity required to estimate small gradients. While the approach can provide some measurement savings, it does not address these fundamental scalability limitations.
  2. The main scalability evidence is limited to small systems. The principal molecular experiments involve only 4-12 qubits and up to 179 QWC groups, which is not sufficient to establish whether the learned relevance structure remains useful when the number of parameters and measurement groups becomes much larger.
  3. The authors seem to use exact contributions from classical simulation to construct the training labels. While this is feasible for the small systems considered, generating such supervision may become computationally prohibitive for larger quantum systems. It remains unclear how the proposed approach could be trained efficiently in regimes where classical simulation is no longer feasible.
Questions for authors
  1. How does the method behave in regimes where VQE gradients exhibit exponential concentration or barren plateaus? Since the relevance labels are normalized relative contributions, could the controller still rank groups correctly while the absolute gradient signal becomes exponentially difficult to estimate?
  2. What is the total amortized cost, including exact-label generation and controller training, and how many downstream VQE runs would be required before the offline cost is recovered?
  3. The method assumes a fixed QWC partition. How sensitive are the results to the quality of this partition, and how would the overall approach scale when the QWC grouping itself becomes difficult for larger Hamiltonians?
Reviewer_HHAHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The paper studies the measurement cost of parameter-shift gradients in the variational quantum eigensolver (VQE). In the standard estimator, which the paper calls Full-QWC, each shifted circuit is measured in every qubit-wise commuting (QWC) group of the Hamiltonian. The authors write each gradient component as a sum of contributions from the groups. They observe that, for each parameter, most of the absolute contribution sits in a few groups, and that these groups change with the parameter and the optimization state. They call this parameter-specific sparsity and propose to measure, for each parameter, only the groups that matter for its gradient.

The contributions are not known during optimization, so the authors train a predictor, which they call a controller, on exact contributions from statevector simulation. It scores every pair of a parameter and a group from Hamiltonian, circuit, light-cone and current-state features, and each parameter then keeps the smallest set of groups that covers a chosen share of the predicted mass.

In 28-step runs on five small molecules, with exact expectations and shots that are counted rather than sampled, the method saves about half of the gradient shots at a mean energy gap of a few millihartree to Full-QWC. The baselines are random selection of the same number of groups, a ranking by Hamiltonian coefficients, parameter freezing (TITAN) and adaptive shot allocation methods. The appendices add transfer to new depths, molecules, Hamiltonians and ansatz families, a noise model, and a 400-step study with shot noise.

The second part asks whether LLMs can make the same selection without fine-tuning. Local open models, which receive the same features and take the number of groups from the controller, do much worse than the controller. Frontier GPT and Claude models that also see retrieved training examples reach millihartree-scale gaps, at lower savings. From a feature-deletion study and an attribution analysis, the authors conclude that the local models make little use of the current optimization state, which the controller does use. They argue that this motivates AI-assisted scientific workflows in which domain experts guide the problem representation, the objectives and the validation.

Critical strengths
  1. Per-parameter group selection Grouping reduces the number of measurement settings, shot allocation reduces the shots per setting, and parameter freezing reduces the number of gradient components. Choosing which groups enter each gradient component is a different way to save measurements, and the paper shows that it can be combined with freezing and with variance-aware allocation. The decomposition into group contributions is simple, and I expect it to be useful to VQA researchers independently of the particular predictor.

The best evidence that this works in practice is the 400-step study with shot noise in the appendix. There the final energy stays within 0.18 mHa of Full-QWC on all five molecules while saving between 25% and 92% of the gradient shots. On LiH, the largest system, random selection of the same number of groups fails. The appendices also describe the protocols and their caveats in enough detail to check most claims.

Critical weaknesses

have two main concerns. The experiments do not show what the learned selection adds over simpler rules, and the LLM part does not support its conclusions.

  1. The baselines do not show what the learned selection adds

The main results come from 28-step runs with exact expectations, in which shots are counted but never sampled, as Appendix N.1 states. In this setting, Full-QWC with fewer shots per group would give the same energies at any saving, so dropping groups can only help when there is shot noise. The relevant question is then whether a fixed budget is better spent on the selected groups or spread over all groups, and none of the finite-shot experiments makes this comparison.

The finite-shot comparisons that are made do not separate the method from simple controls. At 28 steps, every method differs from Full-QWC by tens to hundreds of millihartree, including variance-aware allocation with the same number of shots, so these comparisons mostly reflect variation between runs.

In the 400-step study, random selection also ends within about a millihartree of Full-QWC on the four smaller molecules. TITAN at matched cost, which ranks parameters with exact gradients, ends closer to the exact reference than the method on four of the five molecules. The large gaps to these controls in Table 1 therefore seem to come partly from the short, noiseless runs.

It is also unclear how much of the selection needs learning. As plotted, Figure 1 does not show dominant groups that change with the parameter, since in every molecule the first group bin holds most of the mass for nearly every parameter. The coefficient ranking is close to the learned ranking in most cells of Table 1 and better under the joint molecule and depth shift. Its ordering needs no training, but it borrows its per-parameter counts from the learned model, so no fully training-free selector is evaluated.

The input features suggest one. A group with no term in the light cone of a parameter contributes exactly zero, and the light-cone overlap of each group is one of the few explicit pair features. It would be useful to know how much of the saving such a rule recovers before crediting the learned model, especially since the pairwise MLP without attention does about as well as the Transformer.

Finally, the labels require statevector simulation of the target system, so the method is useful in practice only if it transfers. Under transfer to unseen depths, ansatz families and spin Hamiltonians, the savings drop to roughly 6% to 20%. Leaving one molecule out is more encouraging, but all five molecules are small.

  1. The LLM experiments do not support the conclusions

Close to half of the main text and the end of the abstract are about LLMs. The stated finding is that local models "do not consistently prioritize the physical information that matters most", while the controller "uses these cues more effectively". I do not think the experiments show this. The local models get their group counts from the controller, and missing or invalid IDs are filled in from the coefficient ranking, at a rate that is logged but not reported. Without this repair (Table 50), all four models end hundreds of millihartree from Full-QWC, and GPT-OSS beats random selection in 7 of 15 cells instead of all 15.

The frontier models see the training relevance labels of the same parameter at its two nearest training snapshots, so their millihartree-scale gaps mainly show that nearby labels are informative. A nearest-neighbor rule without an LLM is the obvious control, and it is missing.

The mechanism claim of Section 6 rests on integrated gradients for a single snapshot at iteration 0. There the step fraction is zero and the angle is small, so most trajectory features sit at or near the numeric-zero baseline used for the LLMs, while the controller is attributed against its training mean. Part of the difference between the controller and the LLMs may therefore come from the baselines, and Appendix N.5 says that the cross-model percentages are not on a common scale. The LLM gradients also come from 4-bit checkpoints, not from the runtime that produced the actions, and most fail the completeness check of the paper.

Nothing in the paper shows that trajectory features improve selection either. Removing them slightly raises the captured mass of the LLMs (the share of the true contribution mass in the selected groups, Table 56), and the controller is never run without them. This part carries the second contribution of the paper and the end of the abstract, so these are not minor problems.

  1. (Minor) Some results are presented too favorably

The 1.6 mHa "no worse" criterion is one-sided and applied to four-seed means, and several standard deviations are larger than the differences being ranked. Appendix M.4 calls these comparisons descriptive, and the main text should present them that way. The negative shot gains of iCANS and Rosalin come from one untuned configuration inside a common Adam loop, so Table 2 says little about these methods as published. The main text also reports that the trajectories in Figure 7 reach parity with Full-QWC for H2 and H3+, but not that they never do for H4 and LiH. Finally, the best LLM prompt was selected on a panel that includes an evaluation seed.

Overall, I think the per-parameter decomposition is a sensible idea, and the long-horizon study suggests that it can save gradient shots in practice. What VQA researchers would want to know is how much of this sparsity follows from light cones and coefficients, and whether selecting groups beats spending fewer shots on every group. The paper does not answer either question yet, and the LLM part does not support its conclusions, so I lean towards rejection in its current form.

Questions for authors

Clarification requests

  1. How are the groups ordered into bins in Figure 1, and how often does the top group of a parameter differ from the top group of the coefficient ranking?
  2. What fraction of the parameter and group pairs dropped by the method have zero contribution because the group has no term in the light cone of the parameter?
  3. How often were the outputs of the local LLMs in Table 3 filled in from the coefficient ranking?

Suggestions for future work

  1. It would be useful to add two baselines to the 400-step study with shot noise: Full-QWC with fewer shots per group at the same total number of shots, and the coefficient ranking.
  2. It would also help to evaluate a training-free selector that drops groups outside the light cone of a parameter and ranks the rest by the coefficient-weighted light-cone overlap, with its own threshold. A nearest-neighbor rule that copies the selection of the closest training snapshot of the same parameter would be a second useful control.
  3. For the LLM part, the controller could be trained without current-state features and compared in closed loop, and the attribution could be repeated at later iterations with the same kind of baseline for all models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.