acceptodds
Under review as a conference paper at ICLR 2027

How Much Does the Quantum Part Do? Auditing Quantum Implicit Neural Representations

Abstract

Quantum implicit neural representations place a parametrised circuit inside a coordinate network and report gains over classical baselines. We rebuild two of them under a common evaluation harness at matched parameter budgets, recover the reported performance level for both, and then ask what produces it. For Quantum Visual Fields the answer is almost entirely classical: changing one activation in the classical embedder is worth  dB, while training the -parameter ansatz, of the model, is worth  dB at width and  dB at width , the largest effect we observe and the only multi-seed configuration whose paired confidence interval excludes zero, at – the training cost. A proposition explains the ceiling: the circuit's output is a quadratic form in the embedder's own output; each channel's form lies in a -dimensional symmetric-matrix space, while the jointly reachable family has intrinsic dimension at most . A budget-matched classical quadratic readout with the same trainable parameters matches the circuit while training roughly three orders of magnitude faster; an unconstrained relaxation performs  dB better. The same proposition predicts why the architecture is flat in classical width and why its one effective scaling direction costs . At matched parameter budgets SIREN leads in fitting quality, while QIREN requires three to four orders of magnitude more compute. We also document that the released artifact does not run as shipped and omits the variant producing its published numbers.

Primary area:
datasets and benchmarks
Keywords:
quantum machine learning, implicit neural representations, reproducibility, ablation study, evaluation protocols

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

This work evaluates two so-called quantum implicit neural representations (QINRs). The central question that is being discussed here is that whether the performance reported by these hybrid models can be attributed to their trainable quantum parameters or whether the quantum component is working. After reading this work and the mentioned QINR works, I understand the argument of the authors. However, I did not find the contents interesting and I did not learn anything novel. I am also not fully convinced by the conclusion and therefore did not see it of high value to the community. I gave my reasons below in corresponding sections.

Critical strengths

I view this work as an evaluation of existing so-called quantum implicit neural representations. The experiments cover different areas that are of interest.

Critical weaknesses

Here I gave my reasoning and doubts on this work. They are mainly from the following perspective:

[i] the hyperparameter tuning for the quantum INR seems asymmetric. Some potentially important settings do not seem to be carefully controlled or varied. For example, I did not find a sensitivity study of the temperature factor used in the transformation for amplitude encoding. It seems to be an important factor as it directly concerns the prepared state. The positional encoding also seems to be fixed, despite the paper (from what I see) itself emphasizing the importance of the accessible frequency spectrum.

[ii] are the learning rates same and fixed for both the classical neural model and the quantum circuit in the experiments? Recent works such as [1,2] show that the lr. should be treated as separate hyperparameters; [2] further shows that their relative learning rates can affect which component dominates training. Since the paper shows that the optimization choices can change a quantum INR result by more than 10 dB, I believe a more systematic optimization study of the quantum component is needed to have stronger conclusions about how much the ansatz contributes.

  1. Benchmarking Quantum Reinforcement Learning, ICML 2025
  2. Parallel Hybrid Networks: An Interplay between Quantum and Classical Neural Networks,, Intelligent Computing 2023

[iii] I am not fully convinced that the claimed frozen identity-initialized model measures the contribution of the entire quantum circuit. Identity initialization followed by freezing makes the trainable unitary exact, i.e. U=IU=I. This is indeed a clean study of the trainable quantum transformation. However, the hybrid model still retains the learnable amplitude encoding and the measurement/readout structure. In particular, the amplitude encoder still learns the input-dependent state, while the ansatz can equivalently be viewed as transforming this state or changing the effective measurement observables. Freezing UU therefore removes only the latter transformation. Thus, the experiment more directly supports the conclusion that training the ansatz provides limited additional benefit in several settings, rather than that the broader quantum component contributes little.

[iv] The analysis is also not fully established for the second quantum INR evaluated. The frozen-ansatz test and the structural analysis are performed for the first quantum INR, whereas the second quantum INR experiment studies reproduction, learning-rate tuning, and simulation cost. I did not see analogous component-level ablation showing how much of QIREN's performance is attributable to its quantum circuit.

Most importantly, this work is an evaluation of existing works from which I did not learn anything.

Overall, I do not find the current manuscript of potential value to the community.

Questions for authors

See weakness. I do not have further questions.

Reviewer_pd7VOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors audit the quantum expressivity (“attribution”) of QINRs (i.e. QVFs and QIRENs). They do this by constructing a fair benchmarking framework with a frozen-circuit control that is exactly classical, so the paired difference isolates the contribution of quantum parameters. From this framework, they conclude that QVF’s reported gains over classical baselines come mainly from its classical components, specifically the activation function in the classical embedder. They show that QVF’s circuit reduces to a fixed quadratic form of the embedder’s output, read out through Pauli-Z expectation values, and that a classical quadratic readout with the same number of parameters can beat it. They do observe a small but measurable quantum contribution.

Critical strengths
  1. Attempts to answer an important question in QML.
  2. Generates a theory that produces benchmarking framework which is testable.
  3. Open and honest about Limits in Experimentation.
Critical weaknesses
  1. The Text is somehow disorganized, with too little background and analysis; needs to establish context in setup/methods before going to results, deviations from the original are not shown with respect to the results.
  2. Doesn’t describe the far-reaching implications of QINR in representation learning (i.e., helps mitigate spectral bias)
  3. No QINR-specific failure points motivate the audit (i.e., missing parameter splits, quantum-part ablations, or untuned baselines)
  4. Evaluation framework ignores task complexity (i.e image in-painting vs. random dropout). Target signal shapes the training signal, which directly affects the way any component contributes.
  5. Not generalizable to all QINRs. Quantum contribution (by nature of QVF and QIREN) depends on how the circuit is coupled to the classical part, so the results for QVF may not necessarily apply to QIREN.
Questions for authors
  1. Prop. 2 shows the trained model’s output is D·(Uψ)², where (Uψ)² is itself a probability vector, and the frozen model outputs D·p, where the softmax can produce any distribution. This seems to imply the frozen model can reach every output the trained one can at each pixel. Do you agree? If so, is any gain from training the circuit purely an inductive bias, i.e., a change in how learning proceeds rather than in what the model can represent? If so, make sure to state this explicitly.
  2. Your evaluation framework separates the quantum and classical parts of QVF through identity block initialization. The title and recommendations address QINRs in general, but your conclusion seems to hold only for QVF. From my understanding and reading the original paper (Zhao et. al., 2024), the coordinate x_i uses a hybrid layer, and this hybrid layer directly determines the encoding gates. Since it doesn't seem possible that the implementation has an identity-block control (like QVF), how do you separate the classical component from the quantum one?
  3. Why is task complexity not considered at all in the study?
Reviewer_t7EqOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper audits image fitting with Quantum Visual Fields and QIREN. The authors repair the released code, compare classical INRs at similar parameter budgets, and freeze QVF's ansatz to separate its contribution from the classical embedder's. They also express the QVF readout as a quadratic form and test classical replacements, including one with the same parameter budget. The results suggest that much of the fitting performance comes from the classical architecture, at far lower cost than the simulated quantum implementations.

Critical strengths
  1. The frozen control gets at the right question. Proposition 1 shows that this circuit is exactly the identity when frozen at initialization, giving a clean classical comparator. The authors report the positive ansatz effect at width 256 and explain why the Gaussian-initialization controls are harder to interpret. These details make the attribution more credible than a comparison of final PSNR values alone.
  2. The quadratic-form analysis leads to a strong comparison. Matching the readout's 80 parameters addresses the capacity objection to the unrestricted classical replacement. Table 3 also adds an important qualification: QVF generalizes better to held-out coordinates than SIREN, despite SIREN's better fitting results, but the trainable ansatz does not explain that advantage at width 128. Excluding the shape results that could not be reproduced is sensible.
Critical weaknesses
  1. The width-256 comparison needs to be reconciled. Table 5 gives 36.79 dB for the budget-matched readout and 34.63 dB for the circuit, a difference of 2.16 dB; Section 4, the abstract and the conclusion quote a paired gain of 2.85 dB. The text and caption state three seeds for the surrogate comparison, while Table 2 uses five for the circuit. The supplied circuit logs for seeds 0-2 average 33.94 dB, which could explain the paired gain, but that subset is not identified beside the five-seed mean. I also could not find the width-256 surrogate logs. Two protocol details need attention: bench.py uses 0.001 for SIREN and the configured 0.01 for QVF, whereas Appendix F.4 describes 0.01 except for the settings in Section 7; and the packaged harness has a missing os import and an unexpanded '${QINR_ROOT}/work34/config_img.yaml' path that does not match the README's harness directory. Different learning rates may be appropriate. Please state what was used and supply the missing run records. These appear to be reporting and packaging issues, rather than evidence that the measurements are wrong.
  2. The claim that width does not help is too strong. Appendix B bases it on the frozen model, then says that qubit count is the only intervention that moves performance. Yet the trained sin model improves from 33.06 to 34.63 dB when width doubles (Table 2), and the ansatz's measured contribution changes with width too. This is weaker scaling than SIREN over the tested range, but it does not establish that width is ineffective or that the architecture has reached a ceiling. Proposition 2 explicitly does not prove such a ceiling. Section 6 already treats the trend cautiously; the abstract, conclusion and Appendix B should do the same.
Questions for authors
  1. Which seed pairs give the width-256 gain of 2.85 dB and its confidence interval? Please show the circuit mean for those pairs separately from the five-seed mean and identify or supply the surrogate logs.
  2. What learning rates were used for each arm in Tables 1-5? Were SIREN's rates fixed or tuned? Please reconcile Appendix F.4 with bench.py and provide an entry point that runs from the packaged directory structure.
  3. Is the flat-width claim intended only for the frozen model? How does it account for the trained sin model's increase from 33.06 to 34.63 dB? Please separate the observed trend from a claim about representational capacity.

Comments on the paper

5 comments

I thank the authors for their responses but unfortunately I am still not convinced.

1. Comparison with INR baselines. It seems that these QINR architectures can accept different classical architectures in their framework instead of single MLPs. If the authors can present such results, I think it can be interesting. The reason is that: if we intepret quantum circuit from its spectral expressive power, its co-behavior with different classical neural models will be nice to see. A theoretical analysis will be beneficial. Also it seems that QVF discussed such options? They also tested integrating SIREN in the QVF flow and compared with classical SIREN. Such results are missing.

2. Related work and experimental fairness. Several relevant works on quantum implicit neural representations appear to be missing from the discussion, including:

  • [1] Bernstein–Vazirani Networks: Quantum Machine Learning by Interference
  • [2] Spectral Quantum Memory for Implicit Neural Representations (SQMM)
  • [3] Q-RUN: Quantum-Inspired Data Re-uploading Networks
  • [4] Quantum Implicit Neural Representations for 3D Scene Reconstruction and Novel View Synthesis (Q-NeRF)

Given these existing studies, I remain concerned that the experimental comparisons are insufficiently comprehensive and that the paper's positioning of its contributions is potentially misleading.

Unfortunately, the authors' responses have not resolved my main concerns or clarified what fundamentally new insights this work provides beyond the existing literature. I therefore maintain my original assessment.

Second response to Reviewer KWwR

Part 1 of 2

Thank you for the follow-up. We address the two points in turn. The revision adds a spectral corollary, a related-work paragraph, and a new experiment with a SIREN embedder inside QVF.

1. QVF with different classical modules, and SIREN inside QVF.

  • The comparison you describe is in the paper.
    • QVF's "QVF with Siren" (its Tables 2 and 3) is its "+Sin" variant: the classical energy embedder uses sinusoidal activations. Its classical comparator, "MLP+Sin [46]", is SIREN.
    • This +Sin variant is the one we reconstruct (QVF+Sin; the release contains only the ReLU embedder, Section 3) and compare against a classical SIREN throughout (Tables 1, 2 and 4).
    • Under identity-block initialisation the circuit is the identity, so our frozen QVF+Sin arm is functionally QVF's own MLP+Sin ablation. It reaches 32.67 dB where QVF reports 29.2.
    • We have also run two classical modules (ReLU and sin) under the same protocol. The circuit's contribution is null with the ReLU embedder at both widths (Table 2).
  • A third classical module: a full SIREN embedder. The embedder now uses SIREN's sin⁡(ω0 ⋅)\sin(\omega_0\,\cdot) activations and SIREN's initialisation, with the same layer sizes, followed by QVF's amplitude encoding and circuit. We train it with the circuit trained, frozen, and replaced by the two classical readouts (new Table 14).
    • Setup. Width 256, three seeds. The setting was selected on seed 0 by the budget-matched readout, from a sweep over ω0∈{1,3,10,30}\omega_0\in\{1,3,10,30\}, the learning rate and TT.
    • ω0=30\omega_0=30 does not work inside QVF. SIREN's prescribed value trains unstably there (at best 32.7 dB), plausibly because the embedder's input is already a positional encoding.
SIREN embedderω0=3\omega_0=3, T=16T=16ω0=1\omega_0=1, T=64T=64
Circuit trained40.37 ± 0.2041.66 ± 0.35
Circuit frozen37.95 ± 1.1438.14 ± 0.96
Budget-matched MM (80 params)38.31 ± 1.5940.96 ± 0.43
Unconstrained MM41.56 ± 0.6240.28 ± 0.50
Trained − frozen+2.42 [−0.76, +5.60]+3.53 [+0.73, +6.33]
Budget-matched − trained−2.06 [−6.27, +2.15]−0.70 [−2.23, +0.83]
Unconstrained − trained+1.19 [−0.60, +2.98]−1.38 [−3.29, +0.52]
  • Training the ansatz helps with this module too, as it does with the sin module at a tuned temperature.
  • The unconstrained quadratic readout of the same features is level with the circuit at both settings. By Proposition 2 its function class contains the circuit's, whatever the embedder.
  • The budget-matched readout is behind in point estimate, by 0.7 to 2.1 dB, with both intervals containing zero. So with a SIREN module we do not claim that 80 classical parameters suffice; the revision says so.
  • Platforms. QVF reaches 41.66 dB with a SIREN module, against 42.19 dB with its own sin module at the tuned temperature. These two numbers come from different platforms. On its own platform, a tuned classical SIREN of the same size reaches 42.92 dB.

Part 2 of 2

  • Theory: what the circuit can add to the spectrum (new paragraph in Section 4). QVF loads the data once, as the amplitude vector ψ(x)\psi(x), and its circuit UU does not depend on xx.
    • By Proposition 2, every output channel lies in the fixed span of the 136 functions ψi(x)ψj(x)\psi_i(x)\psi_j(x), which the classical module alone determines.
    • Training the circuit changes the coefficients, never the functions. For every circuit setting, the Fourier support of the output therefore lies within that of {ψiψj}\{\psi_i\psi_j\}: the circuit cannot reach a frequency the classical module does not already supply, whichever classical module is used.
    • This is the opposite of data re-uploading (Schuld et al., 2021), where the encoding gates sit inside the circuit and the circuit generates its own spectrum, as in QIREN.
  • What the corollary predicts, and what we observe.
QVF (data loaded once)QIREN (data re-uploaded)
PredictionThe circuit is replaceable by a classical readout of the embedder's features, for any embedderThe circuit is not replaceable by a classical map without the corresponding frequencies
ObservationA classical quadratic readout of the same features is level with or ahead of the circuit with all three classical modules (ReLU, sin, SIREN; Tables 3, 6, 14); with the sin module the 80-parameter version suffices at every temperature (Table 3)Freezing costs 6.93 dB; every same-budget classical replacement is worse (Table 13)

The spectral view you suggest is therefore what separates the two published QINRs, and it explains why the same ablation gives opposite answers for them.

2. Related work. The Background now discusses the four works.

  • Data re-uploading, like QIREN.
    • Spectral quantum memory (SQMM) couples re-uploading layers through a shared quantum register. Its arXiv version (2611.04574) appeared on 3 November 2026, after the submission deadline.
    • Q-NeRF inserts QIREN modules into a Nerfacto radiance-field backbone.
    • Q-RUN carries the re-uploading construction over to a purely classical layer. Its results support our reading: the inductive bias of re-uploading can be realised classically.
  • Bernstein–Vazirani networks. These are non-variational and trained without gradients through interference. They belong to neither design, and we state that our analysis does not cover them.

On comprehensiveness:

  • What we audit. We audit one released model of each of the two variational designs, data loaded once (QVF) and data re-uploaded (QIREN). For each we report the parameter split, a component ablation, and a budget-matched classical replacement.
  • Why per model. The corollary above shows why these checks must be run per design: they give opposite component-level answers for the two designs. Both published models are nonetheless dominated at the model level by a tuned classical INR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.