acceptodds
Under review as a conference paper at ICLR 2027

Quantum Feature Accessibility as a Mechanism for Shortcut Learning under Distribution Shift.

Abstract

Which features do quantum machine learning (QML) models learn when stable and shortcut signals compete? We show that statistical predictiveness alone does not explain feature reliance: quantum processing makes some predictive signals more accessible than others. We develop a learning-theoretic account of quantum feature accessibility and prove, under explicit assumptions, that population-risk minimizers of local quantum models with affine heads rely only on a competing cue, even when joint features encode label information. Controlled interventions and encoding ablations reveal how the same processing bias can favor stable signals or shortcuts. Experiments on realistic shifts show task-dependent generalization, while MonarQ evaluations reveal hardware-dependent changes in stable-signal prediction. Together, these results link quantum processing to learned reliance and show why out-of-distribution (OOD) generalization depends on the stability of the signals used.

Primary area:
learning theory
Keywords:
Out of distribution generalization, quantum machine learning, quantum feature accessibility, shortcut learning, spurious correlations, inductive bias

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3jWmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper defines quantum feature accessibility as whether predictive information can be used under specified processing and learning resources. It distinguishes encoded information, readout visibility, and learned reliance. For product-encoded local circuits with local measurements and affine heads, it proves additive logits and, under conditional-marginal assumptions, cue-only population-risk minimizers even when relational coordinates jointly contain the label.

However, the core population theorem is an additive-model result rather than an intrinsically quantum phenomenon, and only the constructed Q-LINK setting closely instantiates its assumptions.

Critical strengths

The cue-only theorem links a circuit restriction to shortcut dependence under explicit assumptions. This is a cute and novel results at least to me.

The shared splits and representations in the realistic suite, absence of OOD-based selection, and extensive caveats improve credibility.

Critical weaknesses

The main theorem applies to any additive model satisfying the closure assumptions; several propositions are standard trace-distance, calculus, least-squares, or concentration results. The specifically quantum theoretical novelty is limited.

Some comparisons are confounded by preprocesing. In controlled spectral tasks, classical models receive full inputs while VQCs receive six-dimensional PCA inputs. In realistic scaling, increasing the dimension d changes both representation and circuit width, right?

Questions for authors

Can a pre-training accessibility metric predict reliance across architectures?

How does a classical additive model compare on Q-LINK under identical optimization and interventions?

Do controlled spectral conclusions persist with exactly matched input representations?

Can hardware findings be replicated across sessions or devices with confidence intervals so we have some uncertainty quantifications?

11 comments

Thank you for highlighting the theorem's link to shortcut reliance and the shared inputs, splits, and absence of OOD-based selection.

1. Response to the critical weaknesses

1. What is theoretically new. Theorem 4 uses a standard convexity argument that also applies to classical additive models (Appendix A.9). The new result is a circuit-level prediction of which competing feature a VQC uses at a population optimum and how that choice affects test risk. Product encoding with one coordinate per qubit, local input-independent channels and readouts, and an affine head force additive logits (Proposition 3). In Q-LINK, the stable relative phase predicts the label, while each phase coordinate has a fixed conditional marginal given the label and cue (Appendix C.9). Under the theorem's conditional-mean and closure assumptions, every attained finite population optimum is cue-only. Under Proposition 6's support condition, test risk then depends on the cue–label law, so reversing the accessible cue can cause failure with the stable predictor still encoded.

With basis encoding and target-ZZ measurement fixed, a CNOT changes the effective observable from Z2Z_2 to Z1Z2Z_1Z_2 and raises affine label recovery from 0 to 1 under the cue-balanced law (Appendix A.4). Swapping feature roles at fixed Local VQC architecture makes the accessible cue stable. OOD reversed accuracy rises from 9.8% on Q-LINK to 100% on Q-LINK-Reversed (Tables 69 and 123). Other tasks probe feature reliance empirically.

2. Preprocessing and scaling. We refit spectral baselines on the exact six-angle VQC inputs. The realistic suite shares frozen inputs and splits at each dd. Across d=6,12,24d=6,12,24, the sweep changes representation and circuit width together (§5, Appendix E.3).

2. Answers to the questions

(1) Can a pre-training accessibility metric predict reliance across architectures?

Answer. At initialization, the Taylor diagnostic uses the first labeled minibatch's Adam step to estimate the change in stable-minus-shortcut logit contrast on a balanced 64-example Random probe. It matches 24/27 first-update directions versus 18/27 for always predicting shortcut. Final contrast agreement is 16/27 versus 15/27 for always predicting stable. First-update/final matches are 8/9 and 9/9 for Local, 7/9 and 6/9 for Entangling, and 9/9 and 1/9 for Reuploading. The diagnostic predicts early shifts in reliance, with little evidence of predicting final reliance (§4.3, Figure 5b, Table 126).

(2) How does a classical additive model compare on Q-LINK under identical optimization and interventions?

Answer. The sine–cosine head matches the Local VQC's angles, source examples, initial logits, minibatches, logistic loss, Adam settings, and interventions. Adam updates coefficients or VQC parameters. Table 123 reports OOD reversed mean accuracies over three seeds.

ModelFull competing-feature taskStable-only
Local VQC9.8%51.6%
Classical additive9.8%49.9%
Entangling VQC98.6%99.9%
Classical interactions99.8%100.0%

Both additive models recover to 90.2% when only the shortcut is flipped. Adding 16 cross-block products at fixed inputs enables reliance on the stable relation, as in the Entangling VQC (§4.3, Figure 5a).

(3) Do controlled spectral conclusions persist with exactly matched input representations?

Answer. On Spurious-Fourier, matching inputs lowers logistic regression and MLP accuracy by 4.1 and 4.5 points to 95.9% and 95.0% OOD reversed, near the Local VQC's 95.6%. On Q-FOS, matched classical models reach 99.8–100% OOD reversed versus 45.2–59.5% for default-scale VQCs. Classical stable-only accuracy is 100%, showing the projection retains the stable signal. Halving the rotation scale on the same PCA inputs brings Local and Entangling VQCs to 100% OOD reversed, tying use of the signal to encoding (§4.3, Tables 52 and 124–125).

(4) Can hardware findings be replicated across sessions or devices with confidence intervals so we have some uncertainty quantifications?

Answer. Appendix F.3 pairs 21,600 rows from 18 clean-trained runs on Q-LINK and Spurious-Fourier with matched checkpoints, examples, and 1,000 shots. Entangling Q-LINK OOD reversed accuracy falls from 98.67% in simulation to 50.33% on MonarQ (−48.33 pp, 95% CI [−53.67, −42.67]). The drop is 44–55 pp across the three checkpoints and persists on stable-only inputs (−49.33 pp, 95% CI [−54.67, −44.00], Table 127). Exact and 1,000-shot simulation give the same full-input accuracy at each checkpoint. The wider 405-run archive spans June 30–August 22, 2026 (Appendices D and F.3). The clustered bootstrap quantifies example-sampling uncertainty for the archived runs. Session and device repeatability require matched reruns across identified sessions or devices.

If these clarifications address your concerns, we would appreciate any updates to your assessment and rating.

Thank you for the revision, I will keep the original score, the concern is partially addressed but insufficient.

Thank you for taking the time to review the revision. Could you clarify which concern remains unresolved and what you found insufficient in our response? This would help us address the specific issue during the discussion.

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

The paper studies how quantum models rely on stable features and shortcuts under distribution shift. This paper distinguishes information present in the encoded input from information accessible through the circuit, measurements, and prediction head. Its central theorem identifies conditions under which every attained population optimum of a local model with an affine head relies only on a competing cue, despite predictive information being available in joint features. Experiments cover five controlled tasks, ten realistic distribution-shift settings, encoding and noise ablations, and execution of simulation-trained checkpoints on MonarQ hardware.

Critical strengths
  1. The controlled feature interventions provide useful evidence beyond aggregate accuracy.
  2. The hardware evaluation adds practical value.
Critical weaknesses
  1. The quantum-specific contribution of the central explanation remains unclear. Theorem 4 follows from additivity, a conditional-mean assumption, and convexity of the loss; Appendix A.9 explicitly states that it also applies to classical additive models. Moreover, Appendix A.3 expresses the local model using classical sine and cosine features. These results explain why a restricted model cannot exploit relational information, but the paper does not clearly establish what additional insight is provided by calling this “quantum feature accessibility.”
  2. The experiments are extensive, but the connection to the proposed mechanism remains mostly qualitative.
Questions for authors
  1. Which findings remain distinctive when the local VQC is compared with a classical model using the corresponding trigonometric features, and the interacting VQC with a classical model containing explicit interaction features?
  2. Can you explain more how your paper is a quantum specific paper?
9 comments

Thank you for highlighting the feature interventions as evidence beyond aggregate accuracy and the practical value of the MonarQ evaluation.

1. Response to the critical weaknesses

Quantum-specific contribution. We show when a circuit and its measurements prevent a VQC from using a stable signal already encoded in its input. Theorem 4 links this restriction to cue-only optima through a risk argument shared by classical additive models (Appendix A.9). In Q-LINK, the stable relative phase and the local shortcut both predict the training label. Product encoding with one coordinate per qubit, input-independent local processing and measurements, and an affine head make the logit additive (Proposition 3). Each phase coordinate has a fixed conditional marginal given the label and cue, so the sum of the phase terms has constant conditional mean. Replacing that sum by its mean preserves the conditional mean logit. When the class permits this replacement, strict convexity of logistic loss makes every attained finite population optimum cue-only. Under Proposition 6's support condition, test risk then follows the cue–label law. The circuit can thus lead to reversal failure even when the stable relation remains encoded. Appendix A.3 gives the sine–cosine form of the local logit, which lets us test the same restriction in a classical control.

Quantitative link to the mechanism. We vary joint access at fixed inputs and swap feature roles at fixed Local VQC architecture. The role swap shifts the mean first-update change in stable-minus-shortcut contrast from −5.00 to +4.96 (10⁻³ logits). The relational contrast is numerically zero at initialization and after training in both tasks. OOD reversed accuracy rises from 9.8% to 100% when the accessible cue becomes stable (Tables 69 and 123). The first update favors the accessible cue on average in both tasks. Using the first labeled minibatch's Adam step, the Taylor diagnostic predicts 24/27 first-update directions versus 18/27 for an always-shortcut baseline. Final contrast agreement is 16/27 versus 15/27 for an always-stable baseline. The diagnostic is informative about early reliance, with limited evidence for predicting final reliance (§4.3, Figure 5b, Appendix F.2).

2. Answers to the questions

Answer to Question 1. The classical controls reproduce both reliance patterns. One head uses the Local VQC's sine–cosine form. The other adds 16 cross-products between the phase blocks. Both match the Local VQC's six angles, source examples, minibatches, logistic loss, Adam settings, and initial logits. Adam updates each model's own parameters. Table 123 reports OOD reversed means ± sample SD over three seeds.

ModelFull competing-feature accuracyStable-only accuracyLog lossShortcut-flip sensitivity
Local VQC9.8 ± 0.5%51.6 ± 1.0%3.269 ± 0.1620.932 ± 0.014
Classical additive9.8 ± 0.5%49.9 ± 2.6%3.287 ± 0.1600.932 ± 0.015
Entangling VQC98.6 ± 1.9%99.9 ± 0.1%0.064 ± 0.0490.033 ± 0.042
Classical interactions99.8 ± 0.1%100.0 ± 0.0%0.080 ± 0.0090.080 ± 0.009

Flipping only the shortcut restores both additive models to 90.2%. Both interaction-capable models predict well on stable-only inputs and show much smaller probability changes under the flip. Adding cross-products at fixed inputs tests the effect of giving the classical head access to the stable relation. The quantum-specific finding is how circuit operations make the joint relation accessible through local measurements (§4.3, Figure 5a, Appendix F.1).

(2) Can you explain more how your paper is a quantum specific paper?

Answer. Our contribution to quantum OOD generalization is to explain when a quantum architecture makes a stable predictor usable and when the accessible feature is a shortcut that reverses. For one upload followed by input-independent processing, Proposition 1 identifies which encoded-state differences the measured expectations can distinguish through the span of effective observables.

With the same basis encoding and target-ZZ measurement, adding a CNOT changes the effective observable from Z2Z_2 to Z1Z2Z_1Z_2. The same local readout reveals the joint label signal even with product basis states, raising affine label recoverability from 0 to 1 under the cue-balanced diagnostic law (Appendix A.4). For OOD diagnosis, effective observables tell us whether an encoded contrast can reach the readout. A gate or measurement change can expose that contrast. Stable-only and shortcut-flip interventions then test whether the trained model uses it.

Additional revisions. Appendix F also reports the narrower Spurious-Fourier gap on matched angles, matched Q-FOS controls, and paired MonarQ intervals and the June 30–August 22, 2026 acquisition window.

If these clarifications address your concerns, we would appreciate any updates to your assessment and rating.

Thank you again for your feedback. Our reply above addresses the points you raised. Are there any points in our response you would like us to clarify? Please let us know if you have any further comments or questions. We’d be happy to clarify anything that would be helpful.

Thanks for the detailed rebuttal. However, the quantum part still didn't convince me well enough to raise my score. I feel confident keeping my current score.

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

This paper studies how quantum machine learning (QML) extracts features for supervised learning in the presence of two types of features -- one that is useful for label prediction; the other that is not so useful. The authors assume some statistical model for the features as well as some function class for QML to analyze how QML would extract features during learning. However, I couldn't find any meaningful theoretical results related with quantum; the provided results seem to be trivial.

Critical strengths

None

Critical weaknesses

First, the paper is very hard to follow mainly due to the mere listing of propositions and theorems without any intuitive descriptions of them. As far as I understood, the core of this paper is Sec. 3.2 where the authors tried to show that: (i) consider quantum machine learning model that can express arbitrary functions of the form g(x)+b where x is input while the expressivity of g is dictated by the quantum circuits; (ii) under such function class, if there is some irrelevant feature for label prediction, not using it achieves the optimal risk; (iii) now suppose that there exists some model that can ignore such irrelevant features, then it would also perform well under some distributional shifts.

If my understanding is correct, Sec. 3.2 seems to be a trivial result without any meaningful relevance with quantum. The main reason I think in this way is due to the absence of comparison between classical learning and quantum; what quantum brings new to the story; what is the benefit of quantum compared to classical case.

That said, I think in Sec. 3, there should be some results like: "(i) given some feature-label modeling (e.g., define important/not important features); (ii) how classical learning would extract features and what would be their risk; (iii) how quantum learning would extract features and what would be their risk; (iv) when and why quantum learning has advantage than classical learning; (v) how does this result extends to the distribution shift"

However, I couldn't find any of these, which makes me think that this work is lack of novelty (at least based on the way it is currently written).

Questions for authors

If the authors can clearly state (I'm repeating my comments above):

(i) clearly define the considered feature-label modeling (e.g., define important/not important features) and why this is practically relevant

(ii) how classical learning would extract features and what would be their risk

(iii) how quantum learning would extract features and what would be their risk

(iv) when and why quantum learning has advantage than classical learning

(v) how does this result extends to the distribution shift

then I would be happy to change my assessment. It could be that, the authors may already have all the required contents somewhere in the draft, but the current way of writing was very hard to follow unfortunately.

13 comments

Thank you for asking for a worked example. We show when a circuit's readout forces cue reliance at a population optimum. In Q-LINK, we vary joint access at fixed inputs and swap feature roles at fixed architecture (§4.3, Appendix C.8).

1. Response to the critical weaknesses

Intuition. Think of two clock hands. In Q-LINK, aligned and opposite hands favor different labels. The common rotation makes either hand alone look uniform, even given the label and local cue. A separate local cue also predicts the training label.

Why this changes what the model learns. With one input coordinate per qubit, product encoding, local input-independent processing and readouts, and an affine head, the logit is a sum of single-coordinate functions (Proposition 3). The sum of the phase terms has a constant mean conditional on the label and cue. Replacing the varying sum by its mean leaves the conditional mean logit fixed and lowers expected logistic loss by strict convexity. If the model can remove the phase terms and absorb their mean into the bias, every attained finite population optimum is cue-only (Theorem 4). For bounded logits, Proposition 5 bounds mean squared deviation from the cue-only logit by excess population risk.

What the quantum analysis adds. The risk argument applies to classical additive models under the same assumptions (Appendix A.9). The quantum analysis derives the restriction from circuit and measurement choices and shows how a gate removes it. At fixed basis encoding and target-ZZ measurement, a CNOT changes the effective observable from Z2Z_2 to Z1Z2Z_1Z_2. The same local measurement exposes parity, raising affine label recoverability from 0 to 1 with the cue balanced independently of the label (Appendix A.4).

2. Answers to questions (i)–(v)

(i) Feature–label model and practical relevance. A visual analogue is classifying objects by relative orientation under an unknown scene rotation, with background as a source cue. In Q-LINK, the phase–label relationship stays fixed while cue–label agreement reverses on the OOD reversed split (Figure 1, Appendix B.6).

(ii) Classical feature use and risk. In the exact example, let YY be balanced and UU a fair sign independent of (Y,C)(Y,C). Then X1=UX_1=U and X2=YUX_2=YU jointly encode the label as X1X2=YX_1X_2=Y. For a sign cue CC with Pr⁡(C=Y∣Y)=p∈(1/2,1)\Pr(C=Y\mid Y)=p\in(1/2,1), the cue-only optimum, if representable, is Clog⁡[p/(1−p)]C\log[p/(1-p)] with risk H(p)=−plog⁡p−(1−p)log⁡(1−p)H(p)=-p\log p-(1-p)\log(1-p). A classical model with the product feature can use MX1X2MX_1X_2 with M>0M>0 for zero classification error and risk log⁡(1+e−M)\log(1+e^{-M}), approaching zero as MM grows (Appendices A.4–A.5).

(iii) Quantum feature use and risk. The basis-encoded Local VQC can read CC with a cue-qubit ZZ measurement and attain H(p)H(p). With CNOT, the target readout exposes X1X2X_1X_2, so the affine head can achieve the same risk as the classical product-feature model (Appendix A.4).

On Q-LINK, the classical heads match the Local VQC's six angles, source examples, minibatches, logistic loss, Adam settings, and initial logits. They differ by 16 cross-products between phase blocks. Table 123 gives means ± sample SD over three seeds. Pooled ID overlaps training. Both accuracy columns use OOD reversed examples.

ModelPooled ID log lossOOD reversed log lossOOD reversed accuracyStable-only accuracy
Local VQC0.346 ± 0.0183.269 ± 0.1629.8 ± 0.5%51.6 ± 1.0%
Classical additive0.348 ± 0.0193.287 ± 0.1609.8 ± 0.5%49.9 ± 2.6%
Entangling VQC0.034 ± 0.0140.064 ± 0.04998.6 ± 1.9%99.9 ± 0.1%
Classical interactions0.009 ± 0.0010.080 ± 0.00999.8 ± 0.1%100.0 ± 0.0%

(iv) When and why quantum helps. The circuit helps by exposing the stable relation. Flipping only the shortcut raises both additive models from 9.8% to 90.2%. Mean absolute probability change is 0.932 for both, 0.033 for the Entangling VQC, and 0.080 for the classical interaction head (Table 123). High stable-only accuracy and low flip sensitivity support stable reliance in both models with joint access. Quantum advantage over interaction-capable classical models remains open.

(v) Distribution shift. Under Proposition 6's absolute-continuity condition, a cue-only predictor remains cue-only at test time. Its risk depends on the joint cue–label law. For the exact example, changing cue agreement from 90% to 10% moves the same predictor's accuracy from 90% to 10% and logistic risk from 0.325 to 2.083 (Appendix A.5). The product-feature predictor retains its risk as long as parity still determines the label. We also swap the roles of the two signals at fixed Local VQC architecture. With the accessible local cue made stable, OOD reversed accuracy rises from 9.8% on Q-LINK to 100% on Q-LINK-Reversed (Appendix C.8, Tables 69 and 123).

If these clarifications address your concerns, we would appreciate any updates to your assessment and rating.

Thank you, but I'm afraid that without clear comparison with classical approaches, it is hard to appreciate the results shown in this work. For instance, what if the results that the authors are claiming in this work can be trivially achieved by classical approaches? Of course, even in such a case, quantum machine learning itself can have some advantages so being not worse than classical case can be meaningful. But at the moment, such kind of comparison is missing which makes it hard to grasp the main point of this work.

So, as I clarified in my previous comments, there should be (i) define the problem setting mathematically; (ii) when solving such a problem, how classical approaches would perform and how this is compared with the authors' result. Is it better? Is it worse? Of course, depending on the class of the classical learners, the comparison would differ. So for what classes of classical learners are comparable with the authors' quantum; what are worse; what are better? And given such a clear comparison, what is the reason for considering quantum learners?

Without such kind of clear comparisons, I'm afraid that it is hard to see the clear innovation of this work, unfortunately. Thank you.

AuthorsOpenReview ↗

# Part 1 of 2

Thank you for your follow-up and for raising these questions about the classical baselines and how the comparisons were designed. The initial submission compared logistic regression, an MLP, and three VQCs (Figures 2–3 and Tables 4 and 13). We chose logistic regression and the MLP as linear and nonlinear reference points for feature use under source training, since benchmarking larger networks or specialized OOD methods would also vary model capacity or training strategy. The revision adds matched classical heads with and without cross-block interactions (§4.3, Figure 5a, Appendix F.1, Table 123).

1. What is the mathematical problem? In the exact example (Appendices A.4–A.5), draw a balanced label Y∈{−1,+1}Y\in\{-1,+1\} and a fair sign UU independent of the pair (Y,C)(Y,C), where C∈{−1,+1}C\in\{-1,+1\} is a local cue. Set (X1,X2)=(U,YU)(X_1,X_2)=(U,YU). Each coordinate is uniform given (Y,C)(Y,C). The product X1X2=YX_1X_2=Y determines the label in both environments. For either class, let Pr⁡(C=Y∣Y)=0.9\Pr(C=Y\mid Y)=0.9 in training and 0.10.1 at test. A practical analogue is distinguishing objects by the relative orientation of two parts under an unknown common rotation, with a background cue whose association with the label changes.

2. How do the classical and quantum classes use these features? Consider unregularized population logistic risk for an additive logit a1(X1)+a2(X2)+g(C)a_1(X_1)+a_2(X_2)+g(C). The sum of the two relational terms has the same conditional mean for every (Y,C)(Y,C). Convexity ensures that replacing the sum by this mean can only lower or preserve risk. If the class permits the replacement, as when the relational weights can be zeroed and a free bias absorbs the mean, strict convexity implies that every attained finite optimum has a cue-only logit (Theorem 4). Proposition 3 gives this additive form for product encoding with one coordinate per qubit, local input-independent processing and measurements, and an affine head. The same risk argument applies to classical additive models.

When the cue logit is representable, the source optimum is Clog⁡9C\log 9. A basis-encoded Local VQC can attain it by measuring the cue qubit in ZZ. At cue agreement qq, its risk is qlog⁡(1+1/9)+(1−q)log⁡(1+9)q\log(1+1/9)+(1-q)\log(1+9). With the predictor fixed, accuracy falls from 90% to 10% and logistic risk rises from 0.325 to 2.083 nats. A classical head with X1X2X_1X_2 and a circuit whose readout exposes that product can both represent MX1X2MX_1X_2 for M>0M>0. In both environments, accuracy is 100% and logistic risk log⁡(1+e−M)\log(1+e^{-M}) tends to zero as MM grows. The binary example has common input support. Proposition 6 transfers cue-only behavior when the test input law is absolutely continuous with respect to the training input law. Preserving the joint cue–label law then preserves the fixed predictor’s risk.

3. Where are the comparisons, and which empirical models do better? Figure 1 and Appendix B.6 specify continuous Q-LINK. The noisy phase difference has class centers at 00 and π\pi. The stable signal is cos⁡(φA−φB)\cos(\varphi_A-\varphi_B). A common uniform rotation makes either phase marginal uninformative. Local cue-sign agreement drops from 90% to 10%.

The original §4.1 (Figure 2, Appendix C.3, Table 13) reports mean OOD reversed accuracies of 9.8% for logistic regression, 87.7% for the MLP, 9.8% for Local, 98.6% for Entangling, and 84.6% for Reuploading. Figure 3 and Appendix C.4, Table 19, already reported fixed-checkpoint interventions. With the cue masked on OOD reversed examples, the MLP scored 100%, logistic regression 53.5%, and Local 51.6%. The MLP therefore shows that the stable relation is usable by a classical learner.

In the revision, both classical heads use the same six angles, source examples, initial logits, minibatches, logistic objective, and Adam settings as the Local VQC. The additive head uses sine and cosine of each angle (Appendix A.3, Eq. 10). The interaction head adds 16 cross-block products. Table 123 reports three-seed means. Accuracy is on OOD reversed examples. Pooled ID overlaps training.

LearnerFull accuracyStable-only accuracyOOD full log lossPooled ID full log loss
Local VQC9.8%51.6%3.2690.346
Classical additive9.8%49.9%3.2870.348
Entangling VQC98.6%99.9%0.0640.034
Classical interactions99.8%100.0%0.0800.009

Classical additive matches the Local VQC on reversal accuracy and closely matches its log loss. Entangling improves on both. The classical interaction head has the highest mean reversal accuracy at this setting, while Entangling has the lowest mean reversal log loss. Flipping only the shortcut raises both additive models to 90.2% accuracy. The mean absolute probability change is 0.932 for both additive models, compared with 0.033 for Entangling and 0.080 for classical interactions. High stable-only accuracy and lower flip sensitivity show that both joint-access models learned to use the stable relation.

Comments on the paper

1 comment

We thank all reviewers for their comments and feedback. Compared with the original submission, the revised manuscript adds Section 4.3 and Figure 5, and the appendix adds Section F.

  • Theory and intuition. We added a short explanation after Theorem 4 that connects the result to Q-LINK. The relative phase predicts the label, and each individual phase is uniform. A new worked example links the existing binary parity construction to the continuous Q-LINK task, showing how the classical interaction head and the fixed-measurement CNOT example expose joint information. (§3.2 and Appendix F.1, with the construction in Appendix A.4.)

  • Matched Q-LINK classical baselines. We added classical additive and interaction heads on the VQCs’ six input angles, matching the Local VQC’s initial logits and training protocol. The comparison reports full, stable-only, shortcut-only, and shortcut-flip results. The Local VQC and classical additive head both score 9.8% OOD reversed, and the classical interaction head reaches 99.8%. This gives a direct control for the effect of access to the stable relation. (§4.3, Figure 5a, and Appendix F.1, Table 123.)

  • Matched-input spectral comparisons. We refit logistic regression and the MLP on the exact six clipped PCA angles used by the VQCs. Matching inputs largely closes the Spurious-Fourier gap to the Local VQC, and the matched classical heads reach 99.8–100% on Q-FOS. Section 4.3 connects this result to the Q-FOS rotation-scale experiment, which changes encoding while holding the PCA coordinates fixed. (§4.3 and Appendix F.1, Figure 36 and Tables 124–125. See also Table 52.)

  • First-update diagnostic. We added a test of whether the first Adam step predicts changes in feature reliance. The diagnostic agrees with 24 of 27 observed first-update directions, compared with 18 of 27 for always predicting shortcut reliance. Final agreement is 16 of 27 versus 15 of 27 for always predicting stable reliance. We also updated the theory-to-evidence map to reflect the scope of the matched controls and the first-update test. (§4.3, Figure 5b, Appendices C.1 and C.9, Table 7, and Appendix F.2, Figure 37 and Table 126.)

  • Optimizer sensitivity and full-batch controls. We added learning-rate sweeps for Adam, momentum SGD, RMSProp, and plain SGD, tests of training budget and other hyperparameters, checkpoint comparisons at nearby source losses, and full-batch classical fits. With source-selected rates, both additive models score 9.80% OOD reversed across all four optimizers, and Entangling reaches 97.93–99.90%. The full-batch fits confirm that the classical additive failure persists when changing the optimization protocol. (§4.3 and Appendix F.4, Figure 39 and Tables 130–132.)

  • Feature-role reversal and training trajectories. We extended the role swap to the classical additive head alongside the Local VQC. When the accessible cue is stable, both models improve from 9.80% to 100% OOD reversed in all four matched optimizer settings. We also analyze 75 matched trajectory pairs to compare feature reliance under distinct training paths. (§4.3 and Appendix F.4, Figure 40 and Table 133.)

  • Paired analysis of archived MonarQ results. We added a reanalysis of 21,600 paired evaluations from 18 clean-trained runs on Q-LINK and Spurious-Fourier. It pairs exact and 1,000-shot simulation with MonarQ on the same checkpoints and examples, documents the acquisition scope, and reports paired 95% intervals for example-sampling uncertainty. (§4.2 and Appendix F.3, Figure 38 and Tables 127–128.)

  • New MonarQ reruns. We repeated Q-LINK on the same device for Entangling and Reuploading with both full and stable-only inputs. At matched checkpoints, inputs, and 1,000 shots, the stable-only gaps relative to simulation are −47.33 and −44.33 percentage points, respectively. The reruns provide point estimates showing that the hardware loss persists with the shortcut masked for both architectures. (§4.2 and Appendix F.3, Table 129.)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.