acceptodds
Under review as a conference paper at ICLR 2027

ConQuER: Modular Conditional Control of Frozen IQP Generators

Abstract

Pretrained instantaneous quantum polynomial (IQP) generators learn data distributions but lack direct control over the generated samples when trained unconditionally. Practical applications often require generation conditioned on a specific mode or class. We study how to add this capability without updating the pretrained backbone or departing from the IQP circuit family. We propose ConQuER, a modular post-training framework that equips frozen IQP generators with a reusable conditional generation interface. Its controller adopts the same IQP structure and composes directly with the backbone to steer generation toward requested conditions. On Binary Blobs, a 784-support controller achieves mean conditional squared maximum mean discrepancy () of and a mean requested-mode hit rate of . On binarized MNIST-3/5, a controller with 1,380 trainable parameters serves both classes and reduces mean conditional by relative to the frozen unconditional backbone. Further evaluations on these tasks and Shapes20 characterize the trade-off between conditional distribution quality and parameter budget and demonstrate parameter savings through shared control across conditions.

Primary area:
generative models
Keywords:
Quantum Neural Network, Quantum Machine Learning

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_qhS5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ConQuER adds class conditioning to a pretrained IQP generator by learning per-condition phase offsets on K chosen supports. The offsets add to the backbone phases, so the result is one IQP circuit. Training uses conditional MMD^2. A shared/private split of offsets cuts parameters and lets a new condition be added with q new values.

Critical strengths
  1. The composition is exact. Zero offsets recover the backbone. Adding a condition leaves existing conditional distributions unchanged.
  2. Table 1 reports postselection, full fine-tuning and a classical AR model next to the method, with sampling cost per returned sample.
Critical weaknesses
  1. Simple baselines beat the method. Postselection on the frozen backbone gives lower MMD^2 on Blobs (0.0042 vs 0.0105) and Shapes20 (0.0073 vs 0.0242). A 290-parameter classical AR model beats the 3,136-parameter controller on Shapes20 (MMD^2 0.0123 vs 0.0242, hit 89.8% vs 64.5%). Zeroing the backbone phases gives a higher mode hit (81.7% vs 79.2%) at modestly worse MMD^2, so the pretrained backbone adds little.

  2. The method is per-condition sparse phase fine-tuning, as the paper notes for M ⊆\subseteq G_base. The stated reason to stay in the IQP family is the scalable classical MMD estimator. Yet all experiments are at 16 to 20 qubits with full probability vectors. Blobs uses all 65,535 supports and a Hellinger projection that needs the full state. Nothing is shown at a size where IQP structure matters.

Questions for authors

Clarifications:

  • Table 8: the 1,232-parameter shared controller beats the 6,272-parameter independent table that contains it (0.0098 vs 0.0147). Is optimizer variance as large as the reported effects? Please give multi-seed results for Tables 2 and 8.
  • Appendix E.1 says the Blobs test split was accessed during development. Which numbers does this affect?
  • Can you show one experiment beyond exact simulation?
4 comments
AuthorsOpenReview ↗

rebuttal for Q1

We thank the reviewer for the detailed feedback. We provide multi-seed results and optimization diagnostics for Tables 2 and 8, an independent test-set re-evaluation, 64-qubit experiments beyond exact simulation, and baseline trade-off comparisons.

For Q1 No—the reported effects are not artifacts of optimizer variance. Because Blobs uses a full-batch, exact-PMF objective, controller seeds vary initialization rather than sampling noise. Repeating both studies across five initialization seeds on each of the three backbones (Table 1) shows that the Table 8 sharing advantage (0.0049 at seed 0; 0.0044 across five seeds) is 3.0–7.3× the independent s.d. at seed 0 and 2.7–6.5× for the five-seed mean, while the shared controller itself has an initialization s.d. ≤ 2.7 × 10⁻⁵.

Table 1. Multi-seed comparisons on the independent Blobs test set.

Study (submitted table)MethodParamsSeed 05-seed meanWithin-backbone init. s.d.
Table 8 (joint sharing)Shared/private, q=641,2320.0099530.009960≤ 0.000027
Table 8 (joint sharing)Independent offsets6,2720.0148260.0143250.00067–0.00163
Table 2 (extension)Shared, 64 new params640.0111870.0111920.000005–0.000009
Table 2 (extension)Independent cold start7840.0116760.0118120.000115–0.000849

Expanding the fitted shared solution into its exact 6,272-parameter representation confirms that the independent model can represent it, but cold-start optimization in the unconstrained 6,272-parameter space plateaus at a higher training MMD² (0.014051 vs 0.010086 for shared), and continuing optimization of all 6,272 parameters from the shared solution does not improve validation fit. Thus, coordinate-wise sharing acts as a structural constraint that improves optimization stability on the fixed 784-support dictionary, rather than implying extra representational capacity.

AuthorsOpenReview ↗

rebuttal for Q2

The note in Appendix E.1 refers to the fixed 10,000-sample Blobs test split (Appendix B.1) used across the Blobs evaluations, which had been evaluated during earlier development iterations (while checkpoint selection within each run used the validation split; MNIST and Shapes20 are unaffected). To eliminate any potential test-reuse bias, we generated a completely new, independent 10,000-sample Blobs test set with a fresh seed and re-evaluated all locked Blobs checkpoints without retraining or reselection.

Every method ranking holds on the independent test set. Most unfiltered configurations move by about 1–2%: main ConQuER MMD² changes from 0.010494 to 0.010598, Table 8 shared vs independent changes from 0.009838/0.014669 to 0.009953/0.014826, and Table 2 extension obtains 0.011187 vs 0.011676 (winning 18/24 pairs vs 19/24). The frozen + postselection reference moves by about 13% (0.00420 → 0.004747) but remains the lowest MMD² among the compared methods. The revision replaces all Blobs distributional numbers with those on the independent test set.

AuthorsOpenReview ↗

rebuttal W2/Q3

When M⊆Gbase\mathcal{M}\subseteq\mathcal{G}_{\mathrm{base}}, our composition is indeed algebraically equivalent to conditional sparse phase adaptation; keeping the composed circuit strictly within the IQP family is what enables polynomial-time stochastic MMD estimation once n exceeds the limit of exact state-vector simulation. To demonstrate this beyond exact simulation, we trained and evaluated ConQuER on MNIST-3/5 at 16, 36, and 64 qubits using 1Q+2Q backbones and purely stochastic IQP MMD estimation (without state vectors, probability enumeration, or Hellinger projection). At 36 and 64 qubits (2^36 and 2^64 states), full state-vector simulation is infeasible; 16 qubits is included as a reference.

Table 2. Non-full-state resolution scaling on MNIST-3/5. MMD² in 10-3; mean ± s.d. across three pipeline seeds.

Resolution / qubitsKFrozen MMD²Factorized paramsFactorized MMD²Direct paramsDirect MMD²
4 × 4 / 1613636.808 ± 0.5251,3802.273 ± 0.1402722.059 ± 0.094
6 × 6 / 3639217.994 ± 0.1353,9401.496 ± 0.1237841.523 ± 0.107
8 × 8 / 6478415.078 ± 0.0637,8601.436 ± 0.0911,5681.439 ± 0.055

At 36 and 64 qubits, ConQuER reduces conditional MMD² by 91.7% and 90.5% on CPU, directly validating the scalability motivation of preserving IQP structure.

AuthorsOpenReview ↗

rebuttal for W1

(W1) Baselines and Pretrained Phases

Postselection vs. pre-measurement control. Table 3 compares unfiltered ConQuER against frozen-backbone postselection across all three tasks.

Table 3. Conditional quality and raw-sampling cost (N_raw is condition-averaged 1/acceptance probability; Blobs uses the independent test set).

TaskFrozen + postselection (MMD² / Raw draws)ConQuER, unfiltered (MMD² / Raw draws)
MNIST-3/50.01761 / 2.070.00227 ± 0.00020 / 1.00
Binary Blobs0.004747 / 8.020.010598 / 1.00
Shapes200.00729 / 4.010.02416 / 1.00
If postselection is additionally applied on top of ConQuER, raw draws drop from 8.02 to 1.28 on Blobs (MMD² 0.020672; 0.02079 under the sharing checkpoint triplet) and from 4.01 to 1.56 on Shapes20 (MMD² 0.01546).

On MNIST-3/5, where classifier boundaries are imperfect, unfiltered ConQuER achieves both an 87% lower MMD² than postselection (0.00227 vs 0.01761) and zero rejection overhead (1.00 vs 2.07 draws). On the synthetic tasks, postselection achieves lower MMD² at the cost of 8.02× and 4.01× raw circuit draws.

Classical AR and backbone structure. While the from-scratch 290-parameter classical AR model beats the submitted ConQuER configuration on Shapes20 (0.01233 vs 0.02416), pairing ConQuER with a 1Q+2Q frozen IQP backbone (changing both the backbone and support dictionary; exploratory) lowers MMD² to 0.003821 on Shapes20 (outperforming AR) and 0.001222 on Blobs, while solving the distinct problem of adapting a frozen quantum generator rather than training a classical model from scratch.

Zero-phase ablation. In Table 6, zeroing the backbone phases raises mode hit (81.70% vs 79.16%) because hit only counts total probability mass inside a mode's Voronoi region, ignoring within-mode distribution shape. On MMD²—which measures full within-mode distributional fit—retaining the pretrained phases reduces error by 25% (0.009882 vs 0.013223 on the independent test set), even though the zero-phase run still benefits from reusing the support dictionary discovered on the pretrained backbone.

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

QP circuits are a family of quantum circuits that can sample from learned probability distributions using commuting phase gates. ConQuER enables conditional generation by keeping a pretrained IQP generator frozen and training a controller that supplies condition-dependent phase adjustments. The controller changes the sampling distribution before measurement, and some configurations share parameters across conditions. The method is compared with filtering samples from an unconditional generator, training separate conditional generators, and classical generation on three small, classically simulated benchmarks.

Critical strengths

The proposed approach improves on existing methods as described below.

Compared with postselection, ConQuER avoids discarding generated samples, while compared with training separate generators for each condition, it requires fewer newly trained parameters by reusing a frozen backbone and enabling parameter sharing across conditions.

Critical weaknesses

The experiments are too small for it to be fully convincing.

Questions for authors

The overall paper is interesting but needs additional work:

  1. I would like to see a baseline MNIST, comparison of the 1,380-parameter controller with a directly trained 272-parameter phase table, using the same supports and comparable hyperparameter tuning

  2. Exclude a held-out condition from both backbone pretraining and support discovery, then adapt using its labeled examples so that it genuinely generalizes to new settings.

5 comments
AuthorsOpenReview ↗

rebuttal for Q1

We thank the reviewer for the insightful suggestions. We have completed the 272-parameter direct-table comparison, the strict unseen-condition generalization experiment, and 64-qubit scaling evaluations.

We compared the 1,380-parameter factorized map (d=10) and the 272-parameter direct phase table (2136) on the exact same frozen MNIST backbone, 136 supports, three-stage schedule, and early-stopping rule. Using a matched hyperparameter search (three learning-rate multipliers and two tuning seeds per method, selected by validation score), we evaluated both across five training seeds.

Table 1. Matched comparison of ConQuER phase mappings on MNIST-3/5.

Phase mappingParamsMMD² (U) ↓PMF MMD² (V) ↓TV ↓SVM consistency (%) ↑
ConQuER (Direct table, retuned)2720.00222 ± 0.000160.00415 ± 0.000150.2010 ± 0.004270.06 ± 0.32
ConQuER (Factorized, retuned)1,3800.00230 ± 0.000360.00428 ± 0.000210.2127 ± 0.005470.80 ± 0.60

The 272-parameter direct table matches the 1,380-parameter factorized map on the primary MMD² metric (paired t-test p=0.585), achieves slightly better exact-PMF MMD² and TV, and reduces newly trained parameters by 80.3% (improving over the frozen backbone by 94.2%). The two mappings target different regimes. The direct table requires CK total parameters (K=136 per new condition), whereas the factorized map uses d(C+K) parameters—requiring only d=10 new embedding parameters per new condition with the learned dK decoder held fixed. At C=2, the direct table is much more compact (272 vs 1,380), whereas factorization reduces total parameters once C11. Confirming that 272 direct parameters suffice on MNIST-3/5 makes our C=2 result 5× more parameter-efficient; we will report both and clarify this total-versus-marginal parameter distinction in Section 3.2 and Section 4.2.

AuthorsOpenReview ↗

rebuttal for Q2

Following the reviewer's suggestion, we rebuilt the frozen backbone, 784-support dictionary, and shared controller from scratch using only seven Blobs conditions, introducing labeled examples of the eighth condition only during final adaptation.

Table 2. Adapting to a held-out condition excluded from both backbone pretraining and support discovery (mean ± s.d. across three pipeline seeds on an independent 10,000-sample Blobs test set).

Adaptation on frozen 7-condition backboneNewly trained paramsProjected-backbone recipe1Q+2Q backbone + new growth (exploratory)
Unadapted 7-condition shared mean00.481519 ± 0.0075500.504799 ± 0.001593
Full dictionary adaptation (cold start)7840.023417 ± 0.0020470.001296 ± 0.000068
Full dictionary adaptation (shared-mean start)7840.023834 ± 0.0015200.001288 ± 0.000068
Existing private coordinates only (q=64)640.207754 ± 0.0155450.193152 ± 0.000127

Even when the target mode is completely unseen during backbone pretraining and support discovery, adapting the 784 controller offsets on the frozen 7-condition backbone reduces held-out MMD² by 95% on the projected backbone (0.4815 → 0.0234) and by 99.7% in the exploratory 1Q+2Q branch (0.5048 → 0.001296), while leaving all seven existing condition distributions unchanged. As expected, updating only the q=64 coordinates pre-selected for the seven seen conditions does not suffice for an unseen mode (0.2078), whereas in our original pooled-backbone interface-extension setting (where the backbone already covers all modes), 64 parameters match 784 (0.01119 vs 0.01168 on the independent test set). Thus, the 91.8% marginal-parameter saving is supported in the pooled interface-extension setting, while full 784-offset adaptation enables generalization to unseen modes. We will report the two protocols separately.

AuthorsOpenReview ↗

rebuttal for W

We extended MNIST-3/5 to 16, 36, and 64 qubits using frozen 1Q+2Q backbones and stochastic IQP MMD estimation without full-state simulation.

Table 3. Resolution scaling on MNIST-3/5. MMD² in 10-3; mean ± s.d. across three pipeline seeds.

Resolution / qubitsKFrozen MMD²Factorized paramsFactorized MMD²Direct paramsDirect MMD²
4 × 4 / 1613636.808 ± 0.5251,3802.273 ± 0.1402722.059 ± 0.094
6 × 6 / 3639217.994 ± 0.1353,9401.496 ± 0.1237841.523 ± 0.107
8 × 8 / 6478415.078 ± 0.0637,8601.436 ± 0.0911,5681.439 ± 0.055

Across 16, 36, and 64 qubits, both parameterizations reduce conditional MMD² by 90.5% to 94.4% without constructing state vectors. At 64 qubits across 8 MNIST classes (K=784), independent and shared (q=64) controllers likewise reduce MMD² by 96.7% and 90.3% (65.44 → 2.16 and 6.36 × 10⁻³).

Reviewer_U3AgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a modular framework that adds conditional generation to pre trained and frozen IQP generators without retraining the base model. The control mechanism learns phase shifts at selected points in the circuit. Experiments on Binary Blobs, binarized mnist, and shapes20 show an improvement in conditional MMD, and good parameter efficiency. It is interesting that adapting to new conditions may require far fewer additional parameters than a complete independent adaptation. Overall, this is a post-training solution suggestion for reusing quantum generative models.

Critical strengths

a. The method proposes a new paradigm of post-training conditional generation for IQP generators without changing the pre-trained backbone. b. The significance lies in enabling the reuse of a single model across multiple classes or modes which can be essential for generative model applications. c. Leave one condition out results show that new conditions can be added with up to 91% fewer parameters. d. Inclusion of multiple datasets, comparisons, ablation studies, and trade-off analyses between quality and parameter budget.

Critical weaknesses

a. The experiments are based on small or highly simplified datasets, such as mnist 4^2 thus scalability has not been demonstrated. b. A clearer comparison is needed with robust classical conditional generative methods and with other quantum generative models of comparable capacity, even at model parameter matching. c. The success of the controller likely depends on the quality and expressiveness of the initially pre trained iqp model, a point that is not fully analyzed. d. The results are mixed, perhaps if the comparison were better standardised.

Questions for authors

a. How do performance and the required number of supports/parameters change when the number of qubits, the data resolution, or the number of conditions increases?

b. To what extent can ConQuER correct an inadequately trained or limited-expressiveness IQP backbone without requiring retraining?

3 comments
AuthorsOpenReview ↗

rebuttal for Q1/W1

We thank the reviewer for the constructive comments. We have added 64-qubit scaling experiments, a backbone-trajectory analysis, and parameter-matched baselines.

Resolution. Table 1 extends MNIST-3/5 to 16, 36, and 64 qubits using frozen one- and two-qubit backbones. All stages—backbone pretraining, support scoring, controller adaptation, and evaluation—use stochastic IQP MMD estimation without full state vectors or probability enumeration.

Table 1. Resolution scaling on MNIST-3/5. MMD² in 10-3; mean ± s.d. across three pipeline seeds.

Resolution / qubitsKFrozen MMD²Factorized paramsFactorized MMD²Direct paramsDirect MMD²
4 × 4 / 1613636.808 ± 0.5251,3802.273 ± 0.1402722.059 ± 0.094
6 × 6 / 3639217.994 ± 0.1353,9401.496 ± 0.1237841.523 ± 0.107
8 × 8 / 6478415.078 ± 0.0637,8601.436 ± 0.0911,5681.439 ± 0.055

Factorized control reduces MMD² by 93.8%, 91.7%, and 90.5% across the three scales; direct control achieves 94.4%, 91.5%, and 90.5% reductions.

Budgets. In a direct-map budget sweep at the same three resolutions, retaining just one quarter of the candidate pool (34/136, 166/666, and 520/2,080 supports; 68, 332, and 1,040 parameters) still reduces MMD² by 93.7%, 90.9%, and 90.0%.

Conditions. Table 2 scales the number of MNIST classes (C{2,4,8}) at 64 qubits with K=784.

Table 2. Condition-count scaling at 64 qubits. MMD² in 10-3.

Digits (C)Frozen MMD²Independent paramsIndependent MMD²Shared params (q=64)Shared MMD²
214.785 ± 0.0471,5681.490 ± 0.0848482.283 ± 0.114
456.278 ± 0.6843,1362.409 ± 0.0939765.793 ± 0.415
865.442 ± 0.2356,2722.155 ± 0.0391,2326.361 ± 0.285

At C=8, independent and shared control reduce MMD² by 96.7% and 90.3% relative to the frozen backbone, while sharing cuts trainable parameters by 80.4% (1,232 vs 6,272) with a corresponding trade-off in conditional fit.

AuthorsOpenReview ↗

rebuttal for Q2/WC

Pretraining quality. Table 3 evaluates adaptation from frozen checkpoints along the Blobs pretraining trajectory using a fixed support-growth schedule.

Table 3. Blobs conditional MMD² across pretraining checkpoints (mean ± s.d. across three backbones, three controller seeds each, evaluated on an independent 10,000-sample test set).

Backbone checkpointBefore adaptationAdapted, K=196Adapted, K=784
Pooled-MMD source0.3607450.045144 ± 0.0096040.030757 ± 0.002144
10% of projection0.3592730.021677 ± 0.0025250.015648 ± 0.002830
50% of projection0.3592080.018307 ± 0.0010610.012254 ± 0.001393
Selected final0.3592010.017782 ± 0.0008400.010824 ± 0.000403

ConQuER reduces MMD² by over 91% even on the inadequately trained source checkpoint (0.3607 → 0.0308 at K=784), though controller budget does not fully substitute for backbone quality (the final checkpoint already reaches 0.0178 at K=196).

Backbone structure. We also evaluated a one- and two-qubit frozen backbone (136 phases on Blobs, 210 on Shapes20) with a newly grown K=784 controller dictionary, reaching 0.001222 ± 0.000025 on Blobs and 0.003821 ± 0.000354 on Shapes20 in joint multi-condition adaptation (Table 4). Because both the backbone and the support dictionary change in this comparison, we treat it as exploratory: it shows the headroom of the method rather than isolating a backbone effect.

AuthorsOpenReview ↗

rebuttal for Wb/Wd

Table 4. Parameter-matched comparisons for joint multi-condition adaptation against from-scratch conditional MADE and per-condition QCBMs (MNIST uses 512-sample U-statistic MMD²; Blobs and Shapes20 use exact V-form MMD² on the independent Blobs test set and the submitted Shapes20 test set).

MethodMNIST-3/5 (params / MMD²)Blobs, joint 8-cond. (params / MMD²)Shapes20, joint 4-cond. (params / MMD²)
ConQuER (submitted backbone)1,380 / 0.00227 ± 0.000206,272 / 0.010598 ± 0.0007083,136 / 0.02416
ConQuER (1Q+2Q backbone, exploratory)272 / 0.00206 ± 0.00009*6,272 / 0.001222 ± 0.0000253,136 / 0.003821 ± 0.000354
Conditional MADE (from scratch)1,378 / 0.00218 ± 0.000426,269 / 0.001349 ± 0.0000273,125 / 0.001066 ± 0.000026
Per-condition QCBMs (from scratch)1,376 / 0.00186 ± 0.000246,272 / 0.001150 ± 0.0000763,120 / 0.001910 ± 0.000278
*272-parameter direct phase table from Table 1 (retuned on the submitted MNIST backbone gives 0.00222 ± 0.00016).

On MNIST-3/5, ConQuER is competitive with from-scratch MADE and QCBM at the same parameter budget (and achieves the same fit with just 272 parameters using a direct table). On Blobs and Shapes20, from-scratch models outperform the submitted Hellinger-backbone configuration; the exploratory 1Q+2Q configuration reaches 0.001222 on Blobs (between QCBM and MADE) and reduces the Shapes20 error from 0.02416 to 0.003821, still above MADE and QCBM. Crucially, MADE and QCBMs train new conditional models from scratch, whereas ConQuER equips an already deployed, frozen IQP generator with conditional control. We will explicitly standardize and label backbone structure, estimator type, and frozen-versus-trainable parameters across all tables.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.