acceptodds
Under review as a conference paper at ICLR 2027

Auditing Claimed Speedups of Learned Surrogates for Variational Quantum Optimization

Abstract

A growing line of work accelerates variational quantum algorithms by learning a classical surrogate of the optimization trajectory and rolling it forward, so that quantum measurements are replaced by classical inference. We observe that the two headline results in this line—a physics-informed speedup and a Koopman speedup in an over-parameterised regime—are both measured against small-step gradient-descent reference trajectories, at and . We show this makes the reported quantity uninformative: for quadratics the reference needs iterations in the worst case, so a ratio against a fixed step diverges as . We also show that a rule predicting the next iterate from the step index and the parameter vector alone cannot represent the transition law of an optimizer with internal state, which covers one of the two families we audit. Running both released implementations unmodified through one counter that charges energy and gradient evaluations at their true circuit cost, Adam is more query-efficient in both families; for the physics-informed surrogate it also reaches accuracies the surrogate never attains, at , and qubits, on two Hamiltonians, and under finite sampling. On ibm_kawasaki we verify our sampling model to within and find a device error that, at the configurations tested, is two orders of magnitude larger than the accuracy at which those ratios are defined. Asked of ourselves, the same sensitivity analysis shows our ordering holds over a range of step size—and reverses below it, at exactly the reference value the audited work uses.

Primary area:
datasets and benchmarks
Keywords:
benchmarking, evaluation methodology, reproducibility, variational quantum algorithms

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_715zOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
1
Soundness
1
Read review
Paper summary

This paper examines PALQO and QuACK, two methods that reduce quantum measurements by predicting optimized trajectories. The authors argue that the reported speedups approximately 30-fold and over 200-fold rely on gradient descent baselines using small learning rates; consequently, these results do not directly demonstrate that learned surrogate models outperform conventional optimizers with well-tuned hyperparameters.

Critical strengths

In Table 2, the authors distinguish between the costs of energy evaluation and the full parameter-shift gradient calculation, while also tallying the number of inner optimization calls. For a 20-qubit system, PALQO requires approximately 48,000 circuit evaluations to reach a target energy difference of 0.200, whereas Adam achieves this goal within about 38,000 evaluations.

Figure 4 shows that Adam outperforms the tested PALQO variants at learning rates between 0.1 and 0.5, though this trend reverses at learning rates of 0.01 or lower. The authors also acknowledge that L-BFGS fails under limited sampling, Proposition 2 is inapplicable to QuACK, and the incremental benefit of QuACK over Adam with tuned hyperparameters has not been verified.

Critical weaknesses

Both Appendix F and Table 7 state that the target energy E⋆E^\star is defined using 6,000 gradient steps with the reference optimizer; however, the reference query count in Table 3 corresponds to 25,000 steps. The authors should clarify this inconsistency.

While the authors selected specific learning rates for the added GD and Adam optimizers, they kept the surrogate model configurations and runtime durations fixed at public defaults. This supports the conclusion that "these default implementations underperform the selected standard optimizers," but it fails to adequately distinguish whether the performance gap stems from a weak reference baseline, the surrogate model configuration, the stopping criteria, or discrepancies in reproduction.

Figure 3 lacks a legend entirely; overall, the paper is very difficult to read.

Questions for authors

See weaknesses.

PALQO failed to reach the target within the fixed horizon, whereas QuACK's multipliers were calculated based only on successful seeds. Please report costs and success rates across multiple precision thresholds.

Does QuACK still yield incremental gains compared to Adam after the latter has been tuned? If this comparison is not included, the conclusions should be explicitly qualified as applying only to the small-step-size GD setting examined.

Reviewer_h4jnOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on the question of whether existing algorithms that learn surrogates of optimization trajectories in VQAs truly achieve the acceleration claimed in the original papers. In particular, the work compares well-tuned classical optimizers (GD, momentum, Adam, and L-BFGS) with two surrogate methods, QuACK and PALQO, and reveals that classical optimizers with carefully selected learning rates can outperform the surrogate methods in terms of the quantum resources required to reach a target energy.

Critical strengths
  • The work rethinks the progress of learned surrogates for optimization of VQAs, providing useful insights for the research community.

  • The work adopts a unified resource metric to fairly compare the quantum costs of conventional VQA optimization and surrogate-assisted VQA optimization.

Critical weaknesses
  • Although the work provides numerical evidence that classical optimizers with well-tuned learning rates can outperform the baselines employed in existing papers, these results do not directly establish whether learned optimization-trajectory surrogates can provide additional benefits when combined with well-tuned optimizers.

  • The work audits only two surrogate methods. More comprehensive benchmarks and experimental settings would make the audit more convincing.

Questions for authors
  • If we further tune the surrogate models for every task, will the acceleration re-appear again?

  • Could the authors further compare QuACK applied to a well-tuned Adam optimizer against the same Adam optimizer without QuACK? This would help clarify whether the surrogate provides additional acceleration beyond optimizer tuning.

Reviewer_rwYsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper audits two “learned trajectory surrogate” methods for variational quantum algorithms: the physics-informed PALQO model and the Koopman/DMD-based QuACK. Both report large speedups (about 30× and over 200×) over the gradient-descent optimizer they accelerate. The authors argue that these numbers are mostly a property of the baseline. Both are measured against gradient descent with a very small fixed step. Overall, the paper’s key recommendation is that quantum-resource-reduction claims be reported against a tuned classical optimizer on an explicit query-count axis. The authors release an instrumented harness.

Critical strengths
  • The paper studies an important question, investigating speedup relative to what? As benchmarking hygiene in quantum ML is a real problem, and makes a worthwhile contribution.

  • The matched-budget comparison is a sound design because it does not depend on an accuracy threshold.

  • The paper is upfront about its own limitations. The authors don’t try to sweep anything under the rug: they readily admit that Proposition 2 is narrow and doesn't apply to Koopman, and that they didn't test the surrogate on top of a tuned Adam optimizer. They also openly share that L-BFGS falls apart under finite sampling and that SPSA misses the target entirely.

  • I like the fact that authors reported discarded results and error history over in Appendix F.3 which helps reproducibility.

Critical weaknesses
  • The tone feels a bit more aggressive than the data actually warrants. In fact, part of the comparison feels like it's taking down a point the surrogates weren't even claiming in the first place. The surrogate papers claim to accelerate a given base optimizer. They do not claim to beat the best optimizer in the field. The paper concludes that the Koopman work also reports gains on top of Adam (4.63× on LiH, 3.24× on Ising). It also concludes that it did not run the experiment that matters: a surrogate driving a tuned base optimizer. The abstract and conclusion (“standard optimizers exceed it on the same device-cost axis”) read as a stronger rebuttal than the experiments deliver. The title and abstract should be scoped to match.

  • The theory part is thin and weakly connected to the empirics (e.g., Proposition 1 is a textbook fact about gradient descent on quadratics, Proposition 2 is a one-line observation about state-free maps. The authors correctly say it does not explain any failure and does not apply to the Koopman family. It also does not preclude fitting a realised trajectory, which is what these surrogates do.

  • Proposition 1 is for quadratics. The VQA landscapes tested are not quadratic. The argument is plausible near a minimum, but no evidence beyond the synthetic Fig. 2(a) is given that the 1/η scaling holds in the actual experiments (for example, GD at several η on the TFIM).

  • The sampling experiments contain a significant error that undermines their interpretability such as the injected noise used a misread Hamiltonian, the sampling regime is covered only for n = 8 and 20, two shot counts, and Gaussian noise)

Questions for authors

In addition to concerns above, I was left wondering 1) Does the surrogate’s own cost depend on \eta?, 2) Can you show the 1/η scaling of Proposition 1 on the actual VQA instances (for example, GD with η ∈ {0.003, …, 0.3} on the TFIM at n = 8 or 20), and not only on the synthetic quadratic of Fig. 2(a)?, 3) The classical cost of the surrogate (73.5M to 294M parameters per outer iteration) is excluded from the axis. Could you add a wall-clock or FLOP comparison, even in an appendix, so that the “favourable under some cost models” statement is quantified?, and 4) For the hardware section, what is the specific conclusion you want the reader to draw, given that the surrogates were not run and the instance is too easy for the main effect to appear? Is the main point the device-error measurement? If so, how does it change the interpretation of the paper’s own 10⁻³ target?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.