acceptodds
Withdrawn from ICLR 2027

Resource-Constrained Quantum Modulation of Liquid Temporal Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

What functional role can a small parameterised quantum circuit play when embedded inside a classical dynamical model? We study this question through resourceconstrained quantum modulation of irregular temporal dynamics, where a lowqubit circuit modulates the effective time scale and nonlinear drift of a liquid recurrent update rather than acting as a stand-alone predictor. We characterise the induced function class and its resource sensitivity. Without entanglement, local measurements factorise into bounded nonlinear ridge functions of learned input projections; repeated data re-uploading can expand the potentially accessible local Fourier support under the stated angle-encoding assumptions, while entangling gates permit mixed-coordinate spectral interactions. We further derive a finite-shot trajectory perturbation bound showing that O(S−1/2) measurement error is filtered through temporal spacing and recurrent state sensitivity. Experiments on a controlled irregularised Room Occupancy benchmark support this mechanistic distinction. Removing entanglement while retaining re-uploading improves the classical liquid backbone from 0.3065 to 0.4018 accuracy and also improves macro-F1, negative log-likelihood, Brier score, and expected calibration error relative to the full circuit where available, whereas removing re-uploading worsens most reported metrics. Strong MLP and GRU references remain substantially better overall predictors, so the results do not support a claim of predictive quantum advantage. Finite-shot evaluation produces non-monotonic operatingpoint changes, motivating measurement budget as an explicit property of recurrent quantum models. Overall, the results support viewing low-qubit quantum circuits as structured dynamical modulators whose usefulness depends on the function class required by the downstream temporal system.

Primary area:
learning on time series and dynamical systems
Keywords:
Quantum Machine Learning, Irregular Time Series, Continuous-Time Models, Liquid Neural Networks, Variational Quantum Circuits

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_b8FBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the role of a small parameterised quantum circuit play when embedded inside a classical dynamical model.

Critical strengths

1.The research problem is clear.

Critical weaknesses

1.The manuscript fails to deliver new knowledge or meaningful insights.

2.The experiments lack clear motivation and the rationale and purpose for designing these experiments are not elaborated.

Questions for authors

Please see the weaknesses.

Reviewer_g3HfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper discusses embedding small parametrised quantum circuits into a liquid recurrent cell. The authors argue that this could, potentially, improve model performance and provide a near-term way to incorporate quantum machine learning into otherwise classical methods. They find that the introduction of these cells does not improve performance. Further, they show in ablation studies in which they argue that removing entanglement appears to improve model performance.

Critical strengths
  • The framing of the paper is honest. The authors present a genuine attempt at introducing a quantum component to classical algorithms to identify performance improvement. They don't attempt to oversell the benefits or the results. The idea of introducing quantum algorithms into classical ones is interesting and a potentially near-term way to finding quantum advantage.
  • The paper is well written and clearly explained.
Critical weaknesses
  • Removing entanglement also changes the optimisation landscape, and may simply widen the space of reachable solutions. This isn't really discussed or pulled apart in the paper, just left as a result.
  • The liquid models are confidently wrong (NLL ≈ 10 vs. ln 4 ≈ 1.39 for uniform prediction) and likely below a majority-class predictor. Ablations inside such a model are hard to interpret.
  • The results appear to have been taken from a single run. That isn't sufficient to quote a reliable outcome.
  • The appendix theory is a bit overdone for a conference paper and for the topic.
Questions for authors
  • I wonder if the lack of entanglement is the direct cause of the improved performance. By removing entanglement, or a demand for it, you allow for more flexibility in the representation space, which could account for improved performance. The question is, can you pinpoint the cause of the improvements?
  • Did the authors run the experiments multiple times?

Minor: broken "Appendix ??" reference in G.3.

Reviewer_Bk4COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors explored the question what functional role can a small parameterised quantum circuit play when embed- ded inside a classical dynamical model. Specifically, they study this question through resource constrained quantum modulation of irregular temporal dynamics, where a low qubit circuit modulates the effective time scale and nonlinear drift of a liquid recurrent update rather than acting as a stand-alone predictor. Then, the authors particularly explored what would happen when removing the entanglement. They also conducted a theoretical analysis of the functions represented by different quantum circuit structures (Proposition 1 to 3). The experiments are conducted on the Room Occupancy Estimation dataset from the UCI Machine Learning Repository, by constructing an irregularised benchmark through controlled observation thinning, as the source observations are regularly timestamped. In addition, classical baselines are used, and the authors claimed that: Strong MLP and GRU references remain substan- tially better overall predictors, so the results do not support a claim of predictive quantum advantage.

Critical strengths

The paper is addressing important and interesting problem. The finding of the paper is potential useful for the dynamical model. The paper is easy to follow. The paper has some useful theoretical analysis.

Critical weaknesses

It is nice to see the theoretical analysis in the paper, however, while I can see that these analysis (proposition 1 to 3) provide some useful interpretation of the proposed quantum modulator. I do have some concern how and whether they represent substantial new theoretical contributions. Let’s take Proposition 2 as an example, it is about Re-uploading expands potentially accessible spectral support, I feel this result has certain connection (at conceptual level) with what found in Schuld et al. (2021). For this work about classical dynamic model, I thin what particularly interesting should be something more clear more clear about what new theoretical understanding is obtained specifically from the interaction between quantum modulation and the recurrent temporal system.

Another question/concern is about the big gap between the quantum models and the classical liquid backbone. While the author did noticed and claimed that this paper is not about a claim of predictive quantum advantage. I still felt this is a big concern, as this differences is too large to ignore what is the causes. I would like the authors to explore/explain why the liquid backbone performs so poorly compared with the GRU.

It is also noticed that: (Table 1) the classical liquid backbone has 3,396 trainable parameters, while the quantum variants have 4,652 parameters. Therefore, is it possible that the improvement of the no-entanglement quantum model over the classical liquid model come (or partially) from the additional nonlinear transformations and trainable parameters? If not, can the authors please show this with further exploration?

The authors concluded that data re-uploading and entanglement contribute different functional structures. However, the experiment results mainly compare the final classification performance of different circuit variants. I do not think these results are sufficient to show that the proposed theoretical mechanisms are actually responsible for the observed performance differences. While the authors did provide some theoretical analysis (as discussed before), can any corresponding experiments be designed and concluded to support this particularly?

Finally, the experiments are conducted on only one dataset, which is originally regularly sampled (as shown in the summary of the paper at beginning) and then modified to create irregular temporal observations. Although this provides a controlled setting, I think it is not sufficient to demonstrate the general usefulness of the proposed approach for irregular time-series learning. Why not use real irregular datasets? Or at least on both modified and real datasets and at least two each category.

Questions for authors

Please see details in the above weakness section.

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

Related papers

Discussion (0)

Sign in to comment.