acceptodds
Under review as a conference paper at ICLR 2027

Unlearning-driven Quantum Architecture Search

Abstract

Quantum machine unlearning aims to remove the influence of specified training data from quantum models. However, existing quantum architecture search (QAS) primarily selects architectures based on predictive accuracy and does not account for potential post-deployment unlearning needs. To bridge this research gap, we conduct an empirical study and reveal for the first time that quantum model architecture itself can substantially affect unlearning capability, even among architectures with comparable predictive accuracy. Building on this observation, we propose Unlearning-driven Quantum Architecture Search (UQAS), a framework that incorporates unlearning capability into QAS to better prepare quantum models for post-deployment unlearning requests. In UQAS, we first introduce an Unlearning Capability Evaluator (UCE) to score the unlearning capability of a quantum architecture using synthetic unlearning requests. Then we design two complementary algorithms, UQAS-Select and UQAS-Design, to identify the quantum architecture with the best UCE score under two QAS settings (with and without an existing architecture pool) respectively. To evaluate the performance of the proposed UQAS, we conduct extensive experiments in both simulation and on IBM quantum computers, across sixteen evaluation scenarios spanning four diverse datasets and four accuracy-based QAS methods. The performance analysis reveals that UQAS effectively identifies quantum architectures with stronger unlearning capability while maintaining competitive predictive accuracy. Furthermore, UQAS-Select and UQAS-Design demonstrate complementary trade-offs between search cost and unlearning capabilities, enabling flexible selection under different computational budgets.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum machine unlearning, Quantum architecture search, Quantum machine learning, Variational quantum circuits

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_BKeGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing quantum architecture search (QAS) primarily selects architectures based on predictive accuracy and does not account for potential post-deployment unlearning needs. This paper conducts an empirical study and reveals for the first time that quantum model architecture itself can substantially affect unlearning capability, even among architectures with comparable predictive accuracy. Building on this observation, the paper proposes Unlearning-driven Quantum Architecture Search (UQAS), a framework that incorporates unlearning capability into QAS to better prepare quantum models for post-deployment unlearning requests.

Critical strengths
  1. The entry point of the paper is novel. This is the first paper to empirically reveal that quantum architectures with comparable predictive accuracy can exhibit substantially different unlearning capability.

  2. Empirical studies are meaningful. Experiments on Fashion-MNIST architectures attain the same validation accuracy of 98.5%, yet their mean TV distances are 0.0081 and 0.3894, this motivates the study of the method.

  3. Experiments are carried out across sixteen evaluation scenarios spanning four diverse datasets and four accuracy-based QAS methods.

Critical weaknesses
  1. Total variation (TV) distance may not guarantee data removal.

  2. The largest circuit width is ten qubits in the experiments, which is limited. And tasks like Fashion-MNIST, WDBC are restricted to simple classification. The experiments may be extended to more complex settings.

  3. The paper uses masked fine tuning of the trained parameters as the default unlearning operator and discusses full fine tuning, NegGrad+, adapted diagonal influence in the appendix. It is not clear whether UQAS just works on fine tuning-based approaches.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer BKeG (Part 1/2)

Thank you very much for your careful review and for your support of our work.

To W1: This is an insightful observation. We agree that a small TV distance does not by itself certify data removal, as explained in Appendix A.5. Retraining without the removed data is the gold standard for evaluating unlearning [1]. To assess closeness to retraining, we use TV, a widely used probability metric [2]. TV quantifies differences between the output distributions of the unlearned and retrained models, with smaller values indicating greater similarity [3].

Beyond output similarity, we examined whether models with lower TV also leak less membership information. To do so, we conducted a membership inference audit adapted from Sarmasarkar et al. [4]. In this audit, an attacker uses outputs of the unlearned model to predict whether the removed batch was part of the original training data. Table A reports the area under the receiver operating characteristic curve (AUC) for the attack under full fine tuning, averaged over the four tasks. UQAS achieves a lower mean attack AUC than the baseline.

Table A: Attack AUC of the batch audit (retraining: 0.5000).

Unlearning updateBaseline AUCUQAS AUCBaseline accuracy (%)UQAS accuracy (%)
Full (4 steps)0.88450.814094.5294.99
Full (16 steps)0.85350.760494.4994.91
Full (64 steps)0.70110.589394.9295.00

The AUC gap is statistically significant for every update in Table A. UQAS already has a lower mean attack AUC than the baseline at 4 steps. The AUC reduction from 4 to 64 steps is also larger for UQAS. With masked fine tuning, the attack AUC is also lower for the UQAS architectures. Across architectures, lower TV corresponds to lower attack AUC after unlearning. These results link lower TV to reduced membership leakage and support its use for evaluating unlearning. We will add these results and the complete membership inference audit to the revised version.

To W2: Thank you for this suggestion. Our circuit widths follow common practice in QML, where circuits of up to 10 qubits are a standard setting. For example, QuantumDARTS uses 8 qubits for classification [5] and QuACK uses 10 qubits [6]. QDEQ uses 4 and 10 qubits [7]. Following your suggestion, we have extended our experiments to 12 and 14 qubits (Table B). UQAS-Select lowers the mean TV over the four tasks from 0.4174 to 0.1299 at 12 qubits and from 0.3038 to 0.1007 at 14 qubits. At 12 qubits, UQAS-Design reduces the mean TV further to 0.0786.

Table B: Mean unlearning discrepancy of UQAS-Select at 12 and 14 qubits.

QubitsTaskBaselineUQAS-Select
12Digits0.45590.3546
Fashion-MNIST0.49150.0291
WDBC0.49470.1196
TFIM0.22740.0165
14Digits0.35560.2284
Fashion-MNIST0.31410.0160
WDBC0.51430.1490
TFIM0.03130.0094

Regarding task complexity, the paper includes a task on quantum data (TFIM) and a multiclass task on Digits. Select lowers the mean TV on both under all three request types (Table 2 and Table 9 of Appendix C.3). We have also extended Digits to four and five classes, where Select lowers the mean TV from 0.0879 to 0.0723 and from 0.2392 to 0.1691. In future work we will extend UQAS to larger widths and more complex tasks. We will add the complete experimental results of both extensions to the revised version.

[1] Guo et al. Certified Data Removal from Machine Learning Models. ICML 2020.

[2] Gibbs and Su. On Choosing and Bounding Probability Metrics. International Statistical Review, 2002.

[3] Hangleiter and Eisert. Computational Advantage of Quantum Random Sampling. Reviews of Modern Physics, 2023.

[4] Sarmasarkar et al. Auditing of Unlearning Algorithms. arXiv:2607.05898, 2026.

[5] Wu et al. QuantumDARTS: Differentiable Quantum Architecture Search for Variational Quantum Algorithms. ICML 2023.

[6] Luo et al. QuACK: Accelerating Gradient-Based Quantum Optimization with Koopman Operator Learning. NeurIPS 2023.

[7] Schleich et al. Quantum Deep Equilibrium Models. NeurIPS 2024.

AuthorsOpenReview ↗

Response to Reviewer BKeG (Part 2/2)

To W3: We apologize for causing this confusion. The unlearning methods evaluated in the paper are mainly based on fine tuning, which gave the impression that UQAS only works with such approaches. However, UQAS is in fact not limited to methods based on fine tuning. To demonstrate this, we have added two unlearning methods that perform no fine tuning (Table C). The first adapts the trajectory correction of the Hessian free method HF [1] and the second adapts the Fisher correction of Golatkar et al. [2]. Both only combine derivative statistics cached beforehand and bring the model closer to retraining than leaving it unchanged. UCE scores candidates using the same unlearning method as in evaluation.

Table C: Mean unlearning discrepancy without fine tuning.

Unlearning methodTaskBaselineUQAS-SelectUQAS-Design
HFDigits0.0057300.0048600.000517
Fashion-MNIST0.0010730.0000870.000348
WDBC0.0168310.0032300.003425
TFIM0.0022930.0000060.000008
FisherDigits0.0269160.0197090.003085
Fashion-MNIST0.0074180.0023060.001799
WDBC0.0493370.0171770.025045
TFIM0.0202610.0018500.001740

Select and Design both lower the mean TV on every task under both methods, by 15.2% to 99.7%. At the same time, the absolute decrease in mean test accuracy compared with the baseline is at most 1.00%. Together with the four unlearning methods of Table 13 (Appendix C.5), UQAS thus lowers the discrepancy under six methods. We will add the results for these two methods to the revised version.

Thank you again for your constructive comments. We would be glad to answer any further questions.

[1] Qiao et al. Hessian-Free Online Certified Unlearning. ICLR 2025.

[2] Golatkar et al. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. CVPR 2020.

I thank the authors for the detailed response. I will keep my score.

Thank you for your time and continued support. We deeply value your recognition of the novelty of our work and the breadth of our experiments. Your feedback has been instrumental in strengthening our paper. Thank you again for your constructive engagement throughout the review process.

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

The paper proposes UQAS, a framework incorporating unlearning capability into quantum architecture search. It introduces a TV-distance-based evaluator that compares approximately unlearned and retrained quantum models, with two approaches: architecture reranking (UQAS-Select) and evolutionary search (UQAS-Design). Results on four datasets and IBM hardware show reduced unlearning discrepancy with limited accuracy degradation.

Critical strengths
  • Relevant research direction by considering future unlearning requirements during quantum circuit design.
  • The combination of unlearning evaluation and QAS is a potentially novel contribution.
  • Broad experimental coverage, including four QAS baselines, different unlearning operators, and three IBM devices.
Critical weaknesses
  1. Novelty is mainly at the level of the optimization objective. UQAS-Select is accuracy-constrained reranking, while UQAS-Design adapts regularized evolution. Chen and Zhang (2026) already investigate quantum model amenability to unlearning and its relationship to optimization geometry. The architecture-search contribution is distinct, but the claimed novelty needs more precise positioning.

  2. The proposed metric is not sufficient to verify unlearning. Small TV distance between unlearned and retrained Born distributions does not establish removal of training-data information. Privacy leakage and membership inference should be evaluated, especially considering Su et al. (2026).

  3. Insufficient explanation of architecture-dependent behavior. The experiments demonstrate correlations between circuit architecture and unlearning discrepancy, but do not isolate the influence of depth, entanglement, parameter count, and optimization geometry. This limits the physical interpretation of the results.

  4. Scalability is a major concern. Experiments cover only 4–10 qubits. Full Born-distribution estimation requires exponentially many outcomes, while UCE repeatedly trains and retrains candidate circuits. Appendix A.4 acknowledges the measurement-scaling issue, but no scalable alternative is demonstrated.

  5. Hardware claims should be qualified. Training and unlearning are performed in simulation; IBM hardware is used to measure the resulting circuits. In Table 7, UQAS-Select worsens WDBC (0.199 to 0.217) and shows no improvement on TFIM (0.096 to 0.096).

  6. Cost and comparison fairness remain unclear. UQAS-Design uses 1.79–6.67x the shots of UQAS-Select at 2048 shots per distribution (Table 5). Consequently, the superior discrepancy does not establish superiority under equal total measurement budgets.

Questions for authors
  1. What is the specific technical novelty beyond applying existing QAS algorithms to an additional unlearning objective?
  2. Can the authors demonstrate actual forgetting using membership-inference attacks, instead of relying on TV distance?
  3. Which circuit properties determine unlearning performance, and do the improvements remain after matching architecture complexity?
  4. How does UQAS compare with standard QAS under identical total computational and shot budgets?
  5. How can the method scale beyond 10 qubits without full Born-distribution estimation and repeated retraining?
  6. Why is unlearning not performed directly on hardware, and are the small hardware improvements statistically significant?
7 comments
AuthorsOpenReview ↗

Response to Reviewer SSZS (Part 1/5)

We appreciate your time and your detailed and constructive feedback.

To W1: Thank you for this thoughtful comment on our novelty. The contribution of UQAS goes beyond adding an unlearning objective to existing QAS algorithms. We show that quantum architectures with comparable predictive accuracy can differ substantially in unlearning capability. The technical challenge is to use this finding for architecture selection before real unlearning requests are known. UCE addresses this challenge by evaluating synthetic requests, with theoretical bounds on the estimation error (Appendix A.4). Select and Design use this evaluator to guide pool selection and architecture search under accuracy constraints.

Chen and Zhang [1] compare quantum and classical models and relate the observed quantum advantage in unlearning to loss landscape geometry. They evaluate unlearning with known forget sets within a prescribed quantum circuit template and identify broader architecture studies as future work. UQAS instead optimizes the architecture before deployment, when future forget sets are unknown. Our work thus addresses the question that they leave open: how can architecture choice improve future unlearning capability? We provide a method for selecting and designing architectures that better approximate retraining under a limited unlearning budget. On independent evaluation requests, Select reduces the mean TV from 0.140 for the baseline to 0.028 (Appendix C.2). We will clarify this positioning in the revised version.

To W2: Thank you for the suggestion to evaluate privacy leakage. We agree that output similarity alone does not verify data removal, as noted in Appendix A.5. Retraining without the removed data is the gold standard for evaluating unlearning [2]. We use TV to compare the output distributions of the unlearned and retrained models. A smaller value indicates a closer match to retraining.

Following your suggestion, we have also evaluated membership inference to examine leakage after unlearning. Our unlearning requests remove a batch of training data. We therefore audit the batch as a whole and adapt the audit of Sarmasarkar et al. [3]. An auditor observes measurement outcomes of the unlearned model and must decide whether the removed batch was part of its original training data. The auditor also knows the candidate datasets and the initialization. This is more access than the gray box setting of Su et al. [4]. Table A reports the area under the receiver operating characteristic curve (AUC), averaged over the four tasks. For the baseline, it is still 0.8845 after 4 steps of full fine tuning. Attacks on single records as in Su et al. [4] were less sensitive in this setting.

Table A: Attack AUC of the batch audit (retraining: 0.5000).

Unlearning updateBaselineUQAS
Full (4 steps)0.88450.8140
Full (16 steps)0.85350.7604
Full (64 steps)0.70110.5893

Averaged over the four tasks, the UQAS architectures leak less about the removed data than the baseline architectures. The AUC gap is significant at every step count. From 4 to 64 steps, the mean attack AUC decreases more for UQAS, widening the gap from 0.0705 to 0.1118. UQAS also lowers the AUC under masked fine tuning. Across architectures, lower TV is associated with lower attack AUC, linking closer approximation to retraining with reduced membership leakage. We will add these results and the relation of this audit to Su et al. [4] to the revised version.

[1] Chen and Zhang. Superior Resilience to Poisoning and Amenability to Unlearning in Quantum Machine Learning. Nature Communications, 2026.

[2] Guo et al. Certified Data Removal from Machine Learning Models. ICML 2020.

[3] Sarmasarkar et al. Auditing of Unlearning Algorithms. arXiv:2607.05898, 2026.

[4] Su et al. From Membership-Privacy Leakage to Quantum Machine Unlearning. Physical Review Applied, 2026.

AuthorsOpenReview ↗

Response to Reviewer SSZS (Part 2/5)

To W3: We appreciate this comment. Following your suggestion, we have isolated the factors you list in a factorial experiment on 288 trained circuits (Table B). The effects of layer count and trainable parameter count change sign across tasks. The entangling pattern has a consistent effect and the distance between the unlearned and the retrained parameters tracks the discrepancy on every task.

The experiment varies one factor at a time. Depth is varied through the number of layers. Parameter count is varied by training all rotations or only the RY rotations. The entangling pattern is varied through the CNOT arrangement (a line, a reversed line or a star). The three patterns have identical resource counts (qubits, parameters, depth and CNOTs).

Table B: Mean change in unlearning discrepancy when one circuit factor is varied.

Factor variedDigitsFashion-MNISTWDBCTFIM
Layers (per added layer)−0.0438−0.0009+0.0569+0.0581
Trainable rotations (RY only to all)+0.0534−0.0610+0.0090+0.0055
Pattern (line to reversed line)+0.0488+0.2045+0.1780+0.0427
Pattern (line to star)+0.0877+0.1605+0.0623+0.0056

Depth and parameter count have no consistent effect since their effects change sign across tasks. The entangling pattern has a consistent effect. The line has a lower mean discrepancy than the reversed line and the star on all four tasks although the resource counts are identical. The amount of entanglement measured by the entanglement entropy does not rank TV consistently. For optimization geometry, we examined the Hessian spectral norm and the distance between the unlearned and the retrained parameters. The Hessian norm does not rank TV consistently across tasks. The parameter distance tracks TV on every task (rank correlations 0.904 to 0.978).

These results are consistent with the analysis in Appendix A.5. The discrepancy has two sources. The first is how far the unlearning update ends from the retrained parameters. The second is how strongly the circuit turns this parameter gap into a change of the outcomes. A change of a rotation angle reaches the outcomes through the gates that follow it. The entangling pattern therefore shapes the discrepancy even at identical resource counts. We will add the factorial experiment and this interpretation to the revised version.

AuthorsOpenReview ↗

Response to Reviewer SSZS (Part 3/5)

To W4: Thank you for raising this concern. The widths of 4 to 10 qubits in our paper are the standard setting of current studies on quantum unlearning and QAS. For example, Su et al. [1] use 10 qubits and QuantumDARTS uses 8 qubits [2]. To address your concern, we have extended UQAS-Select to 12 and 14 qubits on all four tasks (Table C).

Table C: Mean unlearning discrepancy of UQAS-Select at 12 and 14 qubits.

QubitsTaskBaselineUQAS-SelectUQAS-Select (pair marginals)
12Digits0.45590.35460.3624
Fashion-MNIST0.49150.02910.0927
WDBC0.49470.11960.0966
TFIM0.22740.01650.0165
14Digits0.35560.22840.3327
Fashion-MNIST0.31410.01600.0160
WDBC0.51430.14900.1802
TFIM0.03130.00940.0088

Select lowers the TV on every task at both widths. Averaged over the four tasks, the mean TV falls from 0.4174 to 0.1299 at 12 qubits and from 0.3038 to 0.1007 at 14 qubits. UQAS-Design lowers it further to 0.0786 at 12 qubits. The absolute decrease in mean validation accuracy stays below 1%.

We also apologize that Appendix A.4 only acknowledged the scaling of the measurement cost and did not demonstrate an alternative. We extend UCE to score architectures using the marginal distributions of neighbouring qubit pairs. Each marginal distribution describes how often the four outcomes 00, 01, 10 and 11 occur for one pair and can be estimated from measured bit strings. We compute the TV between the unlearned and retrained models for each pair and average it across pairs. At 14 qubits, each model has 13 such distributions with four probabilities each, giving 52 in total, compared with 16,384 for the full distribution. This substantially reduces the measurement cost of UCE scoring as the number of qubits increases. Averaged over the four tasks, selection using pair marginals lowers TV from 0.4174 to 0.1420 at 12 qubits and from 0.3038 to 0.1344 at 14 qubits (last column of Table C). On the 219 pools of Table 2, it keeps 96% of the gain of the full score.

The second cost you raise is the repeated retraining of the candidates. These retraining runs allow UCE to assess unlearning with synthetic requests before real removal requests are known. The number of retrainings is set by the numbers of candidates and synthetic requests and does not depend on the width. Each retraining is one more training of the same circuit, and its cost grows with the width like the training in the base search. In our cost measurements at 4 to 10 qubits, the time added by Select stays between 0.88 and 1.26 times the base search time and shows no upward trend. These retraining runs are performed during architecture selection. Unlearning requests after deployment do not repeat them. We will add the results at 12 and 14 qubits, the marginal score, and this cost analysis to the revised version.

[1] Su et al. From Membership-Privacy Leakage to Quantum Machine Unlearning. Physical Review Applied, 2026.

[2] Wu et al. QuantumDARTS: Differentiable Quantum Architecture Search for Variational Quantum Algorithms. ICML 2023.

Reviewer_6q7jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper identifies the unlearning ability as the overlooked factor for existing quantum architecture search. By introducing a quantified evaluator, this paper adopt two methods to select or derive quantum circuit architectures with stronger unlearning capability. The experiments show that UQAS can effectively identify the promising architectures across datasets.

Critical strengths
  1. The paper clearly presents the unlearning ability as the overlooked factor in previous work and showcases the variation. This could serve as as interesting topic and motivate relevant work.

  2. The final experimental results show substantial improvement in the unlearning ability for the selected architectures, which proves the effectiveness of this unlearning-aware search. The evaluation is also extended to real hardware experiments, which make the effectiveness more convincible.

Critical weaknesses
  1. The methodology and framework design are very unclear. (1) how the proposed UCE is integrated into search flow is not shown. (2) the methodology of UQAS-Select and -Design is largely missing, even in the appendix description. The current -Select algorithm seems just plain and brute-force comparison among the candidate pool; the -Design algorithm also only applies mutation operations, which lacks of justification in terms of both efficiency and effectiveness. Besides, the abstract describe -Select and -Design as trade-off, but the later detailed contents show they serve for different scenarios or purpose, which makes it confusing to justify and correspond to experiment design.

  2. The experiments are not sufficient and missing critical metrics or insights. (1) The abstract and introduction mention that UQAS maintain the high accuracy among the architecture search, but this critical accuracy metrics are consistently missing in experimental results. Relevant analysis is also missing. (2) For Table 2, how different search algorithm works similarly or differently with UQAS is not clear. Right now the results are just put plain without relevant analysis and insights into the search algorithm. (3) For the computational cost, it would be more informative to see the ablation or progressively additional overhead compared to baseline search method. (4) The finite shot experiments are confusing. It's not clear what the purpose to list different shot results and how it shows the robustness. The intuition is greater number of shots would be closer to the theoretical performance, so the paper should how the meaning for showing lower shots results, especially for 128 shots. (5) The experimental setting for the IBM hardware are weird. It's not clear why the two different methods are using different Devices, which hinders the direct and easy comparison.

Questions for authors

(1) Can the authors provide relevant background and method comparison to classical machine unlearning field? This area has been studied for some time and shares conceptual similarities, making it worthwhile to see how the proposed method stands in contrast.

(2) Can the accuracy be presented along with the unlearning metrics in the experiment for comparison? How is the accuracy tolerance? What's the gap between the selected ones by UQAS and the baseline?

(3) How is UQAS integrated to base search methods? How is the variance or preference among all these methods to apply UQAS? What is the computation overhead to the original base search? Especially for the UCE design, how is the overhead for enumerating comparison and retrain among all the candidates?

(4) Can the authors justify the purpose of -Select and -Design? If -Design always achieve better architecture and somehow has the ability to find new ones, why we do not apply on the existing pools? Finding new architectures with same accuracy but much improved unlearning capability seems more impactful. If it's not feasible, does that indicate there are critical scaling issues?

(5) Can the authors explain more about the experimental choices of the finite shots and real hardware deployment?

7 comments
AuthorsOpenReview ↗

Response to Reviewer 6q7j (Part 1/5)

We sincerely thank you for your time and for the careful review. Your comments have helped us make the method and the experiments clearer, and we respond to each of them below.

To W1(1): Thank you for raising this point about how UCE is integrated into the search flow. UCE evaluates the unlearning capability of candidate architectures using synthetic requests before deployment. For each request, it compares the output distributions of the unlearned and retrained models with the same architecture. The mean discrepancy across these requests gives the UCE score (Section 5.1).

Select applies UCE after the base search returns a candidate pool. UCE scores the architectures within the accuracy tolerance, and Select chooses the one with the lowest score (Section 5.2).

Design applies UCE during the search, after each candidate model is trained. The UCE score and a penalty for falling below the required accuracy guide the choice of parents. After the search, Design selects the architecture with the lowest UCE score among those within the accuracy tolerance (Section 5.3). Algorithms 1 and 2 in Appendix A.3 detail both procedures. We will clarify this integration in the revised version.

To W1(2): We appreciate these detailed comments on the methodology and the relationship between Select and Design.

Select. Select compares all candidates within the accuracy tolerance and finds the global minimum of the UCE score in this eligible pool. For a fixed number of synthetic requests, the number of unlearning and retraining runs grows linearly with the number of eligible candidates. This scoring cost is incurred once during architecture selection. Appendix A.4 further analyzes UCE estimation error and its effect on selection within a fixed pool. On independent evaluation requests, Select reduces mean discrepancy from 0.140 for the baseline to 0.028 (Appendix C.2). These results support using unlearning responses to select architectures before deployment.

Design. Evolutionary methods are widely used in architecture search, including quantum architecture search [1, 2]. Design adopts regularized evolution because it can directly use UCE scores to guide parent selection in a discrete architecture space. Starting from a random population, it selects parents using the UCE score and an accuracy penalty. Mutation generates a new architecture by inserting or removing circuit blocks from the search space (Appendix B.1). The resulting model is trained and scored, and the new candidate replaces the oldest population member. This process continues within the search budget.

Design uses the same mutation rules as the baseline. To examine whether additional search alone explains its gains, we triple the number of candidates evaluated by the baseline. The largest reduction in mean discrepancy from this expansion is 0.002. Design achieves mean discrepancy values 0.033 to 0.291 below those of the expanded baseline (Table 10, Appendix C.3). We also compare the methods from the same starting pool under equal limits on simulator work and total shots. Design achieves lower mean discrepancy on all four tasks. Averaged over these tasks, the discrepancy is 0.0475 for Design and 0.1135 for the baseline.

Scenarios and cost tradeoff. Select applies when architecture selection is restricted to an existing candidate pool. Design supports further architecture generation and can also start from that pool. When resources permit further search, both methods are available. We then compare their mean TV, runtime, and total measurement shots under the same validation accuracy tolerance. This comparison assesses the additional reduction in TV alongside the extra computational and measurement costs of Design. Tables 2 and 3 evaluate each method separately. Tables 4 and 5 compare their unlearning performance, while Figure 3 and Table 5 report computational and measurement costs.

We will clarify these design choices in the algorithm descriptions and add the equal-budget comparison to the revised version.

[1] Real et al. Regularized Evolution for Image Classifier Architecture Search. AAAI 2019.

[2] Wang et al. QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits. HPCA 2022.

AuthorsOpenReview ↗

Response to Reviewer 6q7j (Part 2/5)

To W2(1): Thank you for pointing this out. We have added the validation accuracy results for Select to the main text to complement the unlearning results in Table 2. Table A summarizes the results for random requests, with each row averaged over the four QAS methods in Table 2. Select lowers mean unlearning discrepancy on all four tasks. Averaged over these tasks, the absolute decrease in validation accuracy relative to the baseline is 0.78%, within the 3% tolerance. Table 3 reports the validation accuracy of Design, and Table 12 (Appendix C.4) reports the test accuracy after unlearning for Select and Design.

Table A: Discrepancy and validation accuracy of UQAS-Select on random requests.

TaskBaselineUQAS-SelectBaseline accuracy (%)UQAS-Select accuracy (%)
Digits0.12100.043397.4596.41
Fashion-MNIST0.13930.011899.6798.76
WDBC0.24680.043489.8988.74
TFIM0.08740.0188100.0099.98

To W2(2): We appreciate the request for a deeper analysis of Table 2. We chose Random, Evolution, MCTS and DQAS to cover four distinct search strategies: random sampling, evolutionary mutation, tree search and gradient-based optimization. This comparison tests whether Select improves unlearning across different candidate-generation strategies. For each method, the baseline selects by validation accuracy, and Select uses UCE to choose among candidates within the accuracy tolerance.

For Design, we adopt regularized evolution to use costly UCE evaluations to guide subsequent search. Each evaluation involves unlearning and retraining. Existing scores guide parent selection, each iteration evaluates a new mutated candidate, and replacing the oldest population member encourages continued exploration [1]. Under equal limits on simulator work and shots, Design achieves lower discrepancy than the baseline on all four tasks (Appendix C.3).

In Table 2, we average the results over the four tasks. Select reduces mean discrepancy by 85%, 66%, 88% and 77% for Random, Evolution, MCTS and DQAS, respectively. The larger gain for MCTS is consistent with its wider discrepancy range among eligible architectures (0.49 versus 0.30 for Evolution), suggesting more opportunity for UCE to improve selection. These results show that Select benefits all four search strategies, while the gains vary with the candidate pools they produce. We will add this analysis to the discussion of Table 2 in the revised manuscript.

To W2(3): This suggestion on the computational cost is very helpful. We have measured the overhead stage by stage relative to the baseline search. For Select, the baseline cost covers QAS preparation from candidate generation to selection by validation accuracy. For each candidate within the accuracy tolerance and each synthetic request, UCE applies the unlearning method, retrains a model on retained data, and compares the output distributions. Across the sixteen combinations of dataset and QAS method, Select increases mean preparation time by 36.6% to 147.0% relative to the baseline. Retraining accounts for 93.5% to 93.7% of the added time. For Design, the new ablation compares total preparation work with the work needed to train and validate the baseline candidate pool. Measured in simulator work, the mean cost ratio ranges from 4.22 to 4.52 across the four tasks. Retraining on retained data accounts for 64.6% to 65.9% of the additional work. These architecture selection costs are incurred once before deployment. Subsequent unlearning requests use the selected architecture without repeating this preparation. We will add both overhead breakdowns to the revised version.

[1] Real et al. Regularized Evolution for Image Classifier Architecture Search. AAAI 2019.

AuthorsOpenReview ↗

Response to Reviewer 6q7j (Part 3/5)

To W2(4): Thank you for this comment on the finite shot experiment. The purpose is to test whether UCE can still select architectures with lower unlearning discrepancy when measurements are limited. The shot count specifies how many measurements are used to estimate the output distribution of each model at each input during UCE scoring. The reported discrepancy is evaluated on separate requests using exact output distributions from simulation (Appendix C.7). As you note, more shots generally improve the accuracy of the UCE score estimate. However, each additional shot requires a circuit execution and consumes QPU time. Low shot counts therefore test whether UCE retains an unlearning advantage despite sampling noise, with 128 shots as the lowest tested setting. At 128 shots, Select still lowers mean discrepancy relative to the baseline in 15 of the 16 dataset and QAS method combinations. Averaging discrepancies across the four QAS methods within each dataset gives reductions of 40.6% to 76.7%. At 1024, 2048, and 4096 shots, Select lowers mean discrepancy in all 16 combinations (Table 15, Appendix C.7). We will state the purpose of the finite shot experiment in the revised version.

To W2(5): We apologize for the confusion. In Table 6, Select and Design were each compared with their own baseline on the same device. Gate errors, readout errors, and coherence times vary across quantum devices, so the same circuit can produce different measurement results across devices. To cover this variation, we evaluated UQAS on several of these devices. Table 7 compares the methods directly on one device (Kingston), and we have now extended it to Marrakesh and Fez (Table B). On each device, Select, Design, and the baseline use the same unlearning requests, physical qubits, evaluation inputs, and shot count. A shot is one circuit execution followed by a measurement. For each evaluation input, the unlearned and retrained circuits are each measured 2048 times to estimate their output distributions. Design has the lowest mean discrepancy in all twelve device and task combinations, with reductions of up to 66.4% from the baseline. On all three devices, Select lowers mean discrepancy on Digits and Fashion-MNIST. On WDBC and TFIM, its mean discrepancy is higher than the baseline in Table B. Gate errors, readout errors, and the limited number of shots may also affect the measured discrepancy. Across all pools in the simulation study (Table 4), Select lowers mean discrepancy on both tasks. The revised version will report this comparison on the three devices.

Table B: Mean unlearning discrepancy on three IBM devices at 2048 shots.

DeviceTaskBaselineUQAS-SelectUQAS-Design
MarrakeshDigits0.12840.06350.0443
Fashion-MNIST0.08840.04650.0408
WDBC0.19220.20990.0988
TFIM0.09570.09860.0910
FezDigits0.11530.06020.0464
Fashion-MNIST0.08360.04380.0423
WDBC0.20210.22060.1039
TFIM0.09410.09610.0929
KingstonDigits0.13630.05870.0458
Fashion-MNIST0.09150.04550.0411
WDBC0.19850.21720.1014
TFIM0.09600.09610.0891

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their valuable feedback. We have uploaded the revised manuscript with changes marked in blue. Section and table numbers below refer to this new version.

Main text. We have made the following additions and clarifications:

  • The introduction clarifies that UQAS-Select selects an architecture from an existing pool, while UQAS-Design generates new architectures.
  • Section 2 adds background on classical machine unlearning and clarifies the limitations of prior work on quantum machine unlearning.
  • Section 4.2 clarifies the interpretation of TV distance, and Section 5 explains how UCE guides UQAS-Select and UQAS-Design (Appendix A.3).
  • Section 6.2 adds validation accuracy results for UQAS-Select (Table 3) and clarifies UCE preparation costs and finite-shot scoring.
  • Section 6.3 compares UQAS-Select and UQAS-Design with a common baseline on all three IBM devices (Table 7).

Appendix. We have added the following experiments and analyses:

  • A membership inference audit of the unlearned batch (Appendix C.9, Table 25).
  • UQAS-Select at 12 and 14 qubits with full or pair-marginal scoring, and UQAS-Design at 12 qubits (Appendices A.4 and C.8, Table 24).
  • Evaluation with two unlearning operators that require no fine tuning (Appendix C.5, Table 20).
  • Selection under exact complexity matching and experiments varying one circuit factor at a time (Appendix C.2, Tables 9 and 10).
  • UCE guidance during architecture generation, preparation time, initialization of UQAS-Design from an existing pool, and comparisons under identical budget limits (Appendix C.3, Tables 13 to 16).
  • UQAS-Select on Digits with four and five classes (Appendix C.3).
  • Analysis of gains across QAS methods (Appendix C.6) and comparisons under different shot budgets (Appendix C.7).
  • Unlearning performed on IBM Fez and bootstrap confidence intervals for the hardware comparisons (Appendix B.4).

We thank the reviewers for the time they have spent on our paper. Further suggestions on the revision are welcome.

Update. Following Reviewer SSZS’s valuable additional feedback, we have discussed our limitations in Section 7 (p. 10) of the revised paper.

We thank the reviewers again for their constructive insights and the AC for considering our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.