acceptodds
Under review as a conference paper at ICLR 2027

C-QOBRA: A Study of Quantum Conditioning in Specification-Driven Molecular Generation

Abstract

C-QOBRA is a conditional variational autoencoder (cVAE) for molecular design, whose encoder and decoder are parameterized quantum circuits of eight to ten qubits with at most trainable parameters. Read directly from the decoder circuit, with no classical network, it is already an effective unconditional generator: on QM9, it yields more unique-and-valid and more novel-and-valid molecules than %the fully quantum QVAE-Mole while using fewer quantum parameters, fully quantum existing VAEs while using fewer quantum parameters, and most of its novel molecules stay within QM9's size limit. Conditional control is weaker under the same direct readout: requested logP values come out in the right order but are compressed to about a third of their range, and conditioning on QED barely moves the output. Because the circuit generates well, this points to decoding rather than conditioning as the bottleneck. Adding a classical post-processing head restores controllability and reveals how little classical capacity is needed: a head roughly the size of the largest tested already recovers most of the attainable logP control, with nearly five times fewer classical parameters than the pre-existing equivalent, which generates only unconditionally. Overall, the quantum circuits store the latent representation, and a small classical head suffices to decode it into conditioned property values.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum machine learning, parameterized quantum circuits, conditional generative models, variational autoencoder, molecular generation, de novo design, QM9, conditioning

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_kVe4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces C-QOBRA, a conditional quantum variational autoencoder for molecular generation that injects property conditions into parameterized quantum circuits without adding extra qubits. The experiments are conducted on the QM9 dataset.

Critical strengths

The paper is generally easy to understand, and experimental settings are described relatively clearly.

Critical weaknesses
  1. The methodological novelty is quite limited. Most components are adopted from existing quantum autoencoders, data re-uploading, and standard Transformer decoding, while the main new element is a relatively small modification of how conditions are injected into the circuit.

  2. The main hybrid model may not be feasible on the hardware as intended. The Transformer requires the full decoder statevector, which is not directly available on real quantum hardware and would require costly state reconstruction or tomography.

  3. The core experimental design may have a serious confounding problem. The register-readout and Transformer models use different qubit counts and are trained separately, while conditioning also enters through a learned classical prior, so the experiments do not cleanly isolate either the decoder effect or the quantum conditioning effect.

  4. The evaluation fails to establish a convincing quantum advantage. There is no parameter-matched classical baseline, the classical parameter accounting is incomplete, and the analysis largely equates fewer trainable parameters with better quantum efficiency without measuring the actual quantum computational cost. The experiments are conducted only on the toy dataset QM9. The generation performance appears to be much lower than that reported by published classical methods.

  5. The conditional generation results are too weak to support the proposed method. The fully quantum readout provides little or no control for several properties, and the paper does not demonstrate simultaneous multi-property conditioning despite presenting fixed-width multi-condition control as a key advantage.

  6. There might be serious issues in the model formulation. The latent-space argument relies on problematic assumptions about qubit degrees of freedom and distributions, and directly encoding categorical token IDs as amplitudes introduces an arbitrary numerical ordering into the molecular representation.

Questions for authors
  1. What is the actual novelty of this paper?

  2. Is the proposed method actually correct and feasible to implement?

  3. What is the motivation for using quantum computing? Does the proposed method truly effective and provide any real advantage?

1 comment

Weakness 1 response: The reviewer is very keen on noticing the standard components of our construction, though the final product does present novelty compared to before. For instance, the fixed-width conditioning mechanism removes the need for extra qubits to embed chemical property conditions, as was done in QVAE-Mole.

Weakness 2 response: We thank the reviewer and separate the three settings the submission conflated. (i) Statevector simulation: the published slopes. (ii) Finite-shot simulation: the same checkpoints decoded from N simulated measurements per molecule; 1,000 shots is where reconstruction comes within 1% of exact (logP slope 0.736 ± 0.025), a criterion chosen on reconstruction because the slope is non-monotone in N. (iii) Noisy simulation at Quantinuum H1-1/H2-1/H2-2 published error rates: the Transformer arm retains 0.93–0.97 of its slope, and the register arm 0.39–0.63. State preparation is training-only, since generation samples from the prior and never runs the encoder; generation costs ~1,000 shots/molecule (4.4 QPU-min per 1,000) for the transformer head against a median of 33,618 (2.4 h) for the register readout.

Weakness 3 response: Both confounds are now measured. For the prior, freezing it showed little effect (logP 0.886 ± 0.025 vs. 0.879 ± 0.028), while the opposite is true for freezing the circuit, so the conditioning passes through the circuit. On the other hand, for the width and readout, a transformer arm at ten qubits gives a logP slope of 0.729 ± 0.026 against 0.742 ± 0.015 at eight, and a frozen register-trained circuit read by a newly trained head reaches 0.728 ± 0.011 against 0.374 ± 0.094 under the register readout. The register arm's weak control comes from the readout, not the circuit or the width.

Weakness 4 response: We thank the reviewer. The register arm’s “0 classical parameters” omitted the prior MLP (458) and the log-variance head (32): the correct count is 490 classical / 204 quantum, so the defensible claim is no trainable classical readout, not a quantum core alone. The same omission affects every arm, and the corrected table gives per-block counts. We withdraw the claims of superior generation and efficiency over QVAE-Mole, whose margins fall inside our seed spread and whose representation differs. What survives is circuit-level: 82 compiled CX against 366, and 14.9× less two-qubit depth, with both methods reading the same amplitude-magnitude vector, so readout is at parity. Published classical rows are not parameter-matched, and our absolute generation quality is lower.

Weakness 5 response: Thanks to the reviewer for noticing one request we did not address in our original response. A single model conditioned on logP and Fsp3 simultaneously (one 8-qubit circuit, no added width) retains most of its single-property control: sweeping logP at pinned Fsp3 gives slopes of 0.76-0.86 (vs. 0.879 single-property), and sweeping Fsp3 at pinned logP gives 0.85-0.91 (vs. 0.963). These two conditions are not fully independent: while pinned Fsp3 is essentially undisturbed by the logP target (cross-slope <= 0.01), pinned logP does drift with the Fsp3 target (cross-slope up to 0.32). Joint hit rates are 12-26% across 5 seeds. We will add this experiment and state the asymmetry as a limitation rather than claiming clean joint control.

Weakness 6 response: We thank the reviewer; both points are correct and neither changes a reported number. The purity penalty encourages purity in transformer arms (0.963–0.980) and the register (0.985–0.993), so we will state the one-coordinate-per-qubit description as approximate, not exact. For real quantum states, the density is arcsine, so we withdraw that derivation while keeping the prior itself. θ = arccos(tanh z) confines each latent qubit to a half great circle on which ⟨Z⟩ ↦ θ is a bijection. Relabelling token IDs 1–24 before amplitude encoding gives a logP slope of 0.726 ± 0.007 against 0.742 ± 0.015 over 5 training seeds—a mean shift of 0.016, with permutation spread 0.45× the seed spread; reconstruction and NUVR are unchanged.

Question 1 response: (1) Conditional embedding does not require auxiliary qubits to enlarge the circuit. A parameter-free readout keeps a condition's direction but not its scale, and reading the same circuit with a head restores the scale, so the information is in the circuit; (2) about 1,000 classical parameters recover most of the attainable control, as they do for an equally small classical bottleneck.

Question 2 response: Neither changes a reported number. On feasibility, we now measure what the paper lacked: 1,000 shots per molecule for the head and 93–97% of the slope retained under trapped-ion noise. All results will be reproducible from the released code.

Question 3 response: The compact circuit effectively allows condition information to pass through the circuit; it is as effective as an equally small classical bottleneck but not more, and its readout costs ~1,000 shots per molecule.

Reviewer_D5VsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C-QOBRA, a conditional variational autoencoder (cVAE) for molecular design. The decoder and encoders are variational quantum circuits with 8 - 10 qubits and at most 204 trainable quantum parameters. The paper showed that only a very small classical post-processing is need to recover the performance close to the best previous works.

Critical strengths
  1. The reason question is reasonably well motivated. In addition to comparing accuracies, the authors also study the effect of classical decoder capacity.
  2. The ablation studies on post-processor is systematic. Comparing 82 - 198784 classical parameters and multiple random seeds.
  3. The limitations are acknowledged.
Critical weaknesses
  1. In the paper the authors allow the transformer to read the full quantum statevector. In general, the full statevector is not possible to get unless expensive quantum state tomography is implemented.
  2. Lack parameter-matched classical comparison. Can a small classical VAE do similar thing?
  3. The cost of amplitude encoding should be discussed. It is very simple on a simulator, however, very difficult on a real device.
Questions for authors
  1. Does the reported parameter efficiency remain meaningful after accounting for state-preparation gates, measurement shots, and quantum-to-classical readout costs?
  2. How can the proposed architecture be implemented efficiently on real quantum devices, and does its reported parameter efficiency remain meaningful after accounting for this overhead?
  3. How to efficiently implement this on real quantum computer when full statevector is required?
1 comment

Weakness 1: In the paper the authors allow the transformer to read the full quantum statevector. In general, the full statevector is not possible to get unless expensive quantum state tomography is implemented.

Response. We thank the reviewer and withdraw the slight advantage claim. We used a 166-parameter classical cVAE as the control, with post-decoder width, prior, training protocol, and evaluation kept identical. Across four properties and three post-processor widths, logP for d8 has a t-value of 1.76, while fScafHetero d8 t = −1.94. Elsewhere, |t| < 1 and its sign varies. We report the two as performing comparably at matched parameter count.

Weakness 2: Lack parameter-matched classical comparison. Can a small classical VAE do a similar thing?

Response. We thank the reviewer and withdraw “slight advantage.” The control is a classical twin: the 160-parameter PQC is replaced by a 166-parameter classical cVAE, everything else identical—post-decoder width, prior, training protocol, evaluation—with totals matched to within 6 parameters. Across four properties and three widths, the difference is within seed noise and changes sign (logP d8 Δslope = +0.012, t = 1.76; fScafHetero d8 Δ = −0.124, t = −1.94; |t| < 1 elsewhere). We report that the two perform comparably at matched parameter count.

Weakness 3: The cost of amplitude encoding should be discussed. It is very simple on a simulator; however, it is very difficult on a real device.

Response. The reviewer raises a valid point. In terms of hardware application, the generation of de novo molecules relies on angular encoding in the latent space, which is then processed by the fully parameterized decoder. Amplitude encoding matters most in training, which can be done on classical simulation. Once trained, the parameters can be saved and loaded onto a real device, where there will be no need for complex quantum amplitude setting.

Question 1: Does the reported parameter efficiency remain meaningful after accounting for state-preparation gates, measurement shots, and quantum-to-classical readout costs?

Response. We thank the reviewer and separate the three settings that the submission conflated. (i) Statevector simulation: the published slopes. (ii) Finite-shot simulation: the same checkpoints decoded from N simulated measurements per molecule; 1,000 shots is where reconstruction comes within 1% of exact (logP slope 0.736 ± 0.025), a criterion chosen on reconstruction because the slope is non-monotone in N. (iii) Noisy simulation at Quantinuum H1-1/H2-1/H2-2 published error rates: the Transformer arm retains 0.93–0.97 of its slope, and the register arm 0.39–0.63. No device run is claimed. State preparation is training-only, since generation samples from the prior and never runs the encoder; generation costs ~1,000 shots/molecule (4.4 QPU-min per 1,000) for the Transformer head against a median of 33,618 (2.4 h) for the register readout.

Question 2: How can the proposed architecture be implemented efficiently on real quantum devices, and does its reported parameter efficiency remain meaningful after accounting for this overhead?

Response. We thank the reviewer and separate the three settings the submission conflated. (i) Statevector simulation: the published slopes. (ii) Finite-shot simulation: the same checkpoints decoded from N simulated measurements per molecule; 1,000 shots is where reconstruction comes within 1% of exact (logP slope 0.736 ± 0.025), a criterion chosen on reconstruction because the slope is non-monotone in N. (iii) Noisy simulation at Quantinuum H1-1/H2-1/H2-2 published error rates: the Transformer arm retains 0.93–0.97 of its slope, and the register arm 0.39–0.63. No device run is claimed. State preparation is training-only, since generation samples from the prior and never runs the encoder; generation costs ~1,000 shots/molecule (4.4 QPU-min per 1,000) for the Transformer head against a median of 33,618 (2.4 h) for the register readout.

Question 3: How to efficiently implement this method on a real quantum computer when a full state vector is required?

Response. In terms of hardware application, the generation of de novo molecules relies on angular encoding in the latent space, which is then processed by the fully parameterized quantum decoder. Furthermore, the Transformer reads amplitude magnitudes √p(k), not signed amplitudes, in every reported run. This process requires repeated measurements in one basis, not full tomography, though 256 probabilities must still be estimated from shots. Decoding the published checkpoints from finite-shot estimates, 1,000 shots per molecule keeps reconstruction within 1% of the exact readout (logP slope 0.736 ± 0.025 vs. 0.742 ± 0.015), about 4.4 QPU-minutes per 1,000 molecules. The revision will say “amplitude magnitudes” and report statevector and finite-shot results separately.

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

This paper proposes C-QOBRA, a conditional quantum variational autoencoder for molecular generation. The method introduces property conditioning through parameterized controlled rotations without requiring additional qubits. The authors investigate two decoding strategies: direct quantum register readout and a hybrid Transformer-based decoder.

Critical strengths
  1. The paper explores the division of responsibilities between quantum circuits and classical decoders in conditional molecular generation, addressing an interesting problem.
  2. The paper introduces a quantum conditioning mechanism based on controlled rotations and data re-uploading, enabling property conditioning without increasing the number of qubits.
Critical weaknesses
  1. Unclear contribution of the quantum circuit. The paper lacks parameter-matched classical baselines and ablation studies to demonstrate whether the quantum circuit provides meaningful benefits beyond the classical components. 2.Although the paper claims to avoid quantum state tomography, its Transformer decoder requires access to the full quantum statevector, potentially incurring substantial measurement overhead on real quantum hardware. This undermines the claimed efficiency advantage over QVAE-Mole.
Questions for authors
  1. The register readout and Transformer decoder use different numbers of qubits and are trained separately. Could the authors compare these decoding strategies while keeping the quantum circuit fixed to better isolate the effect of the decoder?
  2. Since QVAE-Mole generates 3D molecular structures while C-QOBRA generates SELFIES, how can the authors justify the claimed improvements over QVAE-Mole given their different molecular representations and computational costs?
1 comment

Weakness 1: Unclear contribution of the quantum circuit. The paper lacks parameter-matched classical baselines and ablation studies to demonstrate whether the quantum circuit provides meaningful benefits beyond the classical components. Response. We thank the reviewer and withdraw the slight advantage claim. We used a 166-parameter classical cVAE as the control, with post-decoder width, prior, training protocol, and evaluation kept identical. Across four properties and three post-processor widths, logP for d8 has a t-value of 1.76, while fScafHetero d8 t = −1.94. Elsewhere, |t| < 1 and its sign varies. We report the two as performing comparably at matched parameter count.

Weakness 2: Although the paper claims to avoid quantum state tomography, its Transformer decoder requires access to the full quantum statevector, potentially incurring substantial measurement overhead on real quantum hardware. This undermines the claimed efficiency advantage over QVAE-Mole. Response. We thank the reviewer for noticing this discrepancy. To clarify, the transformer reads amplitude magnitudes sqrt(p(k)), not signed amplitudes, in every reported run. This process does not need tomography, but only repeated measurements for all 256 probabilities to be estimated from shots. Decoding the published checkpoints from finite-shot estimates, 1,000 shots per molecule keeps reconstruction within 1% of the exact readout (logP slope 0.736 ± 0.025 vs. 0.742 ± 0.015), about 4.4 QPU-minutes per 1,000 molecules. The revision will say “amplitude magnitudes” and report statevector and finite-shot results separately.

Question 1: The register readout and Transformer decoder use different numbers of qubits and are trained separately. Could the authors compare these decoding strategies while keeping the quantum circuit fixed to better isolate the effect of the decoder? Response. We froze each register-trained circuit and its prior, then trained only a head on it. The slope rose from 0.374 ± 0.094 under the register readout to 0.728 ± 0.011 with a d8 head and 0.868 ± 0.017 with d128, matching heads trained end-to-end.

Question 2: Since QVAE-Mole generates 3D molecular structures while C-QOBRA generates SELFIES, how can the authors justify the claimed improvements over QVAE-Mole given their different molecular representations and computational costs? Response. In Table 1, we highlighted that SELFIES improves validity by construction. The idea was to follow Wu et al.’s (2024) logic of evaluating model performance using validity, validityuniqueness, and validitynovelty across various models, which they did across multiple hybrid quantum-classical ones like ours. By their rules of comparison, we stand out.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.