{"id":"cfd30265-ae0d-4ee5-b335-c4a975c4d942","arxiv_id":"2508.19437","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Quantum kernels can beat an untuned classical kernel on datasets whose labels the authors deliberately construct from the quantum kernel's own spectrum, showing data efficiency by construction.","lead":"The authors build a tool that rewrites the answers (labels) of ordinary tabular datasets so they line up with a quantum kernel's strongest directions, and show that on these rewired datasets the quantum kernel learns from fewer examples than an untuned classical kernel. They also bring a classical formula that predicts kernel learning curves into the quantum setting and check it against simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Data-efficiency gap is not controlled: labels are constructed from the QKM eigenbasis (Eq. 15) and compared against an untuned RBF, so the observed advantage may be an artifact of target alignment rather than a quantum-specific property.","rationale":"The reader identifies the same core weakness: the relabeling construction plus an untuned RBF baseline is not a fair test of data efficiency. I agree that this is the single most load-bearing concern because it directly undermines the abstract's 'clear empirical evidence' claim. The paper is transparent about the construction and its limitations, so the existence claim likely holds, but the strength of the evidence is overclaimed. The recommended test—using an RBF-aligned labeling or a tuned RBF—would settle whether the gap reflects a quantum-specific data-efficiency property or merely target-kernel alignment. I also note the metric-validation concern (Appendix A divergence) as a secondary issue, but the baseline fairness is more central to the paper's main message. Since the reader's verdict of CONDITIONAL already accounts for this, no verdict change is needed.","tokens_in":11434,"tokens_out":5478,"duration_ms":58368,"concrete_test":"Re-run the experiment with two additional baselines: (1) generate labels using the RBF kernel's eigenbasis (replace V_γ with V_RBF in Eq. 15, same step-function ĉ) and train both QKM and RBF on these labels; (2) on the original quantum-aligned labels, tune the RBF bandwidth (e.g., cross-validated over a logarithmic grid) and report the best-performing RBF. If the QKM data-efficiency gap persists against the RBF-aligned labels and the tuned RBF, the claim survives; if the gap shrinks or disappears, the observed advantage is an artifact of target alignment and the abstract must be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Fig. 3) compares QKM and RBF on datasets whose labels are generated by Eq. (15): ỹ = V_γ^{-T}√ĉ, where V_γ is the eigenvector matrix of the quantum kernel Gram matrix and ĉ is a step function concentrating target power in the top n eigenmodes. This construction directly forces the target to be highly aligned with the QKM's kernel eigenbasis, and by Eq. (10) guarantees fast learning of those modes as N grows. The classical RBF kernel is used with fixed (untuned) hyperparameters, and the authors admit (Section V) that the procedure 'did not explicitly account for a classical kernel (RBF).' Consequently, the comparison measures the effect of aligning the labels to one particular kernel, not an intrinsic data-efficiency advantage of quantum models. A classical kernel given the same alignment (labels built from its eigenbasis) or a bandwidth-tuned RBF could plausibly exhibit the same low-error-at-small-N behavior. Thus the abstract's 'clear empirical evidence' is not established by the current experimental protocol; the result is an existence proof of datasets favorable to QKMs, not evidence of a quantum-specific data-efficiency property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses whether quantum kernel methods (QKMs) can be more data-efficient than classical kernel methods on classical data. The authors adapt a spectral-bias generalization metric from Canatar et al. [8] to QKMs, and use it to construct 'semi-artificial' datasets: classical features are kept fixed, while new labels are generated as y~ = V_γ^{-T}√ĉ (Eq. 15), where V_γ is the eigenvector matrix of the quantum kernel Gram matrix and ĉ is a step function concentrating target power in the top n eigenmodes. On three clinical datasets, they report that QKMs achieve lower error with fewer training points than an RBF kernel, and that the predicted loss from Eq. (7) tracks empirical loss. The paper claims this is clear empirical evidence that data-efficient learning with quantum models is possible on classical data, and positions the label-generation procedure as a tool for systematic dataset design.","tokens_in":11649,"tokens_out":3057,"duration_ms":33617,"significance":"If the empirical claims held under an unbiased comparison, the paper would make a useful contribution: it would provide a concrete, tunable construction of datasets where quantum kernels appear to require fewer samples, and it would offer a validation of a practical generalization metric for QKMs. The paper also has strengths: it engages seriously with the spectral-bias literature, it clearly states its protocol, and it is unusually candid about limitations, including the admission in Section V that the procedure 'did not explicitly account for a classical kernel (RBF)' and the divergence between predicted and empirical loss in Appendix A. These candid statements, however, highlight that the central claim is currently not supported by the experimental design. The label-generation procedure forces the target to align with the QKM eigenbasis, so the observed advantage is expected from Eq. (10) and does not demonstrate a quantum-specific data-efficiency property.","major_comments":[{"comment":"The central empirical claim is largely forced by construction. Equations (14)-(15) generate labels ỹ = V_γ^{-T}√ĉ, where V_γ are the eigenvectors of the quantum kernel Gram matrix and ĉ places all target power on the top n eigenmodes. Equation (10) states that modes with large γ_i are learned rapidly as N grows, so the QKM's low-error-at-small-N behavior is a direct consequence of the construction, not an emergent property of quantum models. The classical RBF baseline is used with fixed hyperparameters and, as the authors admit in Section V, the procedure 'did not explicitly account for a classical kernel (RBF)'. To support the paper's claim, the authors must provide a control where the same relabeling construction is applied to a classical kernel (e.g., labels built from the RBF eigenbasis), or where the classical kernel is tuned or otherwise given comparable target alignment. Without s","section":"Section IV-A, Eq. (15), and Fig. 3"},{"comment":"The validation of Eq. (7) as a predictive tool is also not yet convincing. The predicted loss is computed from the same γ and α that define the labels, so agreement between Eq. (7) and the QKM empirical loss is partly tautological. More importantly, Appendix A shows that for the alternative geometric-difference relabeling of Huang et al., the predicted loss 'shows a notable divergence from the empirical loss' (Fig. 4). Thus the metric's agreement in Fig. 3 is not evidence of general predictive validity for QKMs. The authors should either quantify the agreement (e.g., error bars, correlation, or normalized error) in the main construction, or explain why the divergence in Appendix A does not undermine the claim that Eq. (7) 'remains consistent with empirical behavior.'","section":"Section V and Appendix A"},{"comment":"The experimental reporting is under-specified. Algorithm 1 uses 'while n < size(X)' but n is not initialized, and the same symbol n is used for the hyperparameter in ĉ_x^n, creating ambiguity. The main plots (Fig. 3) show no error bars or confidence intervals despite the text stating that the experiment is repeated 50 times. Since the data-efficiency gap is the central quantitative claim, the spread across repetitions must be shown; otherwise the reader cannot assess whether the gaps in Fig. 3(a)-(d) are statistically meaningful.","section":"Algorithm 1 and Fig. 3"}],"minor_comments":[{"comment":"The text says 'n and v are tunable hyperparameters', but the step function uses x, not v. The notation should be corrected (and the same for the description of Fig. 2).","section":"Eq. (14)"},{"comment":"Equation (11) defines C(i) as a cumulative distribution, but the text does not state whether the eigenvalues γ_i are sorted in descending order before this definition is applied. Since the construction of ĉ depends on this ordering, it should be made explicit.","section":"Section III-B"},{"comment":"The description of Fig. 5 says 'Relabeled - KM' and 'Relabeled - qKM' but the text states that the correlation values for ĉ_1^20 and ĉ_10^20 are identical. If the figure shows overlapping curves, this should be stated clearly, and the claim that x controls only scale is presented without quantitative support.","section":"Appendix B"},{"comment":"The phrase 'one of the first evidence' (Section I) is grammatically awkward; more importantly, the abstract's 'clear empirical evidence' overstates what the current protocol demonstrates (see major comment above).","section":"Abstract and Section I"},{"comment":"Reference [31] is formatted with the data-set contributor names as authors ('Kolby Nottingham Markelle Kelly Rachel Longjohn'). This should be corrected to the standard UCI repository citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main claim is an existence proof of classical datasets where QKMs are more data-efficient than an untuned RBF kernel. As it stands, the construction is essentially tailored to the quantum kernel's own eigenbasis, so the paper demonstrates a property of the label-generation method rather than of quantum models. The authors are candid about this limitation, and the manuscript could be revised to include the necessary controls (e.g., same relabeling applied to classical kernels, tuned RBF, or at least a matched target-alignment comparison). If such controls show that the gap persists, the paper would be a solid contribution. The Appendix A divergence is a second concern that should be addressed before the Eq. (7) validation is claimed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look for the construction alone: Eq. (15) inverts the target-alignment measure into a label-construction rule, giving the QML community a controllable way to generate datasets with tunable difficulty relative to a chosen kernel. That is genuinely new, and the authors are upfront that they are building favorable datasets, not discovering naturally occurring ones. The transfer test of Canatar's spectral-bias formula to quantum kernels is also a useful step, and the predicted-vs-empirical agreement in the main setup is a point in its favor.\n\nThe soft spots are real but concentrated in the interpretation. The headline claim of data efficiency is largely forced: labels are built from the quantum kernel's eigenbasis, so the QKM is given the answer key, while the RBF baseline is used with fixed hyperparameters and no equivalent alignment. The authors admit they didn't account for the classical kernel, but the abstract still says 'clear empirical evidence' of data-efficient learning. That overshoots. A bandwidth-tuned RBF, or an RBF whose eigenbasis generated the labels, would likely show the same small-N behavior. So the paper is an existence proof of datasets favorable to QKMs, not evidence of a quantum-specific property.\n\nThere are also concrete technical issues. Algorithm 1 doesn't match ridge regression—the prediction step y_hat = K_n y is not the estimator that Eq. (7) describes—and the empirical error is defined ambiguously. No error bars, no code. Appendix A reports a divergence between predicted and empirical loss for the geometric-difference relabeling, so the metric's validity is not general. Appendix B finds no pattern in correlations, and Appendix C's bandwidth result contradicts the authors' earlier study; both are explicitly preliminary.\n\nNone of this kills the contribution. The construction is disclosed, the paper is clearly scoped as a proof-of-principle, and the spectral-bias transfer is meaningful even if the metric and target share spectral objects. The fix is straightforward: soften the abstract, add a properly tuned classical baseline, clarify the algorithm, and include error bars. If a referee pushes on those, the paper becomes a solid contribution to the QML toolkit.\n\nI'd send it to peer review, not desk reject. The tool deserves to be in the literature, and the flaws are fixable rather than fatal.","headline":"Useful tool paper with an honest existence proof, but the headline data-efficiency claim is weaker than the abstract implies because the quantum kernel is handed the labels and the classical baseline is untuned.","tokens_in":12282,"tokens_out":2088,"would_cite":true,"duration_ms":24738,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q12","81P68","68T05"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"Quantum models can learn from less data — when labels match the kernel","keywords":["quantum kernel methods","data-efficient learning","target alignment","spectral bias","generalization prediction","semi-artificial datasets","kernel ridge regression","quantum machine learning"],"falsifier":"Take the same three datasets and build labels from the classical RBF kernel's eigenbasis instead of the quantum kernel's, keeping everything else identical; if the quantum kernel still needs fewer points than RBF, the claim is not a construction artifact, but if RBF wins on its own engineered labels, the reported gap is an artifact of label construction. A second check: tune the RBF kernel's bandwidth per dataset and see whether the quantum advantage vanishes.","tokens_in":11191,"feed_emoji":"⚛️","tokens_out":6614,"duration_ms":69456,"temperature":0.7,"pith_summary":"The paper asks whether quantum models can learn from less data than classical models on classical datasets. It argues yes: by constructing semi-artificial labels that concentrate the target function in the top eigenmodes of a quantum kernel, quantum kernel methods reach low error with fewer training points than an untuned classical RBF kernel on three clinical datasets. The paper also transfers a spectral-bias generalization formula from classical kernel regression to quantum kernels and reports that its predicted loss tracks the measured learning curves. If correct, the contribution is an existence proof that data-efficient quantum learning on classical data is possible, plus a practical tool for designing datasets that favor quantum models and for screening real datasets for such opportunities.","feed_headline":"Data-efficient quantum learning exists on classical data","feed_subtitle":"A label-engineering tool aligns target functions with quantum kernel eigenmodes; predicted and measured losses match.","key_machinery":"The engine is target-alignment saturation: a label-construction rule (Eqs. 14–15) that uses the quantum kernel's eigen-decomposition to place all target power in the first n eigenmodes, making the cumulative alignment measure C(i) rise as steeply as possible. This forces the task onto the modes the quantum kernel learns fastest. The paper couples this with the spectral-bias generalization formula (Eq. 7), which predicts per-mode and total error from kernel eigenvalues γ_i, ridge parameter λ, dataset size N, and target coefficients α̂_i.","core_discovery":"The central claim is that the data-efficiency of a quantum kernel is governed by how well the target labels align with the kernel's eigenbasis, and that this alignment can be engineered. Concretely, the paper keeps the features of three clinical datasets fixed and replaces the labels with y~ = V_γ^{-T}√ĉ, where V_γ is the eigenvector matrix of the quantum kernel and ĉ is a step function placing nonzero target power only in the top n eigenmodes. On these relabeled datasets, the quantum kernel reaches low mean squared error with fewer points than the classical RBF kernel, and the paper calls this the first empirical evidence that data-efficient quantum kernel learning on classical data exists.","pith_inferences":["The gap is manufactured in the paper's favor: labels are built from the quantum kernel's eigenbasis while the classical RBF baseline is left untuned, so the experiment demonstrates feasibility, not a generic advantage.","A natural next test would repeat the construction blind to both kernels, or with a tuned classical kernel; if the quantum gap survives, the conclusion strengthens, and if not, the observed data efficiency is partly an artifact of the construction.","The metric's predictive power is not yet general: Appendix A shows predicted loss diverging from empirical loss under the geometric-difference relabeling, so Eq. (7) needs validation on naturally labeled datasets before being used to screen for quantum advantage.","Inverting the tool suggests a practical screening procedure: compute the alignment measure C(i) on a real dataset and identify cases where a large fraction of target power sits in the top quantum-kernel eigenmodes; those are the datasets worth testing for quantum data efficiency."],"forward_implications":["If the claim holds, quantum kernel methods can beat classical kernels on data efficiency only when the target labels align with the quantum kernel's high-eigenvalue modes; alignment, not raw expressivity, is the operative resource.","The data-generation tool yields tunable semi-artificial datasets (via cutoff n and scale x), enabling controlled studies of which dataset characteristics favor quantum models.","The validated Eq. (7) predictor lets practitioners estimate quantum kernel learning curves from a kernel spectrum and target coefficients before committing to large training runs.","Because supervised quantum models can be viewed as kernel methods, the same spectral-bias analysis may extend to quantum neural networks via neural tangent kernels, as the paper notes.","The existence proof redirects the search for quantum advantage from ad hoc benchmarks toward principled dataset design and label-kernel alignment screening."],"supporting_citations":[{"why":"Supplies the spectral-bias generalization formula (Eq. 7), the task-model alignment measure C(i), and the per-mode error decomposition the paper transfers to quantum kernels.","marker":"[8]"},{"why":"Provides the geometric-difference relabeling used as the comparison method in Appendix A and the data-power framing that motivates data-efficiency questions.","marker":"[5]"},{"why":"Establishes the inductive-bias view of quantum kernels by linking kernel eigenvalues to learnable subspaces, motivating the target-alignment construction.","marker":"[15]"},{"why":"Establishes that supervised quantum models can be viewed as kernel methods, justifying the transfer of classical kernel regression theory.","marker":"[7]"},{"why":"Documents the Hamiltonian-evolution feature map used for the quantum kernel and its effectiveness in quantum kernel settings.","marker":"[28]"},{"why":"Supplies the three clinical datasets whose features are reused in the semi-artificial relabeling experiments.","marker":"[30]"},{"why":"Applies the alignment measure to quantum kernel bandwidth, providing background for using alignment to shape quantum kernel training.","marker":"[26]"}],"fun_headline_variants":["Quantum kernels beat classical at low data with tuned labels","Data-efficient quantum learning shown possible on classical data","Label engineering enables data-efficient quantum kernels","Eigen-aligned labels cut training data for quantum kernels"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that constructing labels directly from the quantum kernel's eigenbasis and comparing against an untuned classical kernel is a fair test of data efficiency; if labels were chosen without knowing the quantum kernel, or the classical kernel were tuned, the data-efficiency gap could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Quantum kernels beat classical at low data with tuned labels","Data-efficient quantum learning shown possible on classical data","Label engineering enables data-efficient quantum kernels","Eigen-aligned labels cut training data for quantum kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1422,"prompt_tokens":748,"completion_tokens":674,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":492,"tokens_out":674,"duration_ms":7112,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:49:38.733104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same three datasets and build labels from the classical RBF kernel's eigenbasis instead of the quantum kernel's, keeping everything else identical; if the quantum kernel still needs fewer points than RBF, the claim is not a construction artifact, but if RBF wins on its own engineered labels, the reported gap is an artifact of label construction. A second check: tune the RBF kernel's bandwidth per dataset and see whether the quantum advantage vanishes.","supporting_citations":[{"cited_title":"The inductive bias of quantum kernels","cited_arxiv_id":null,"evidence_quote":"Establishes the inductive-bias view of quantum kernels by linking kernel eigenvalues to learnable subspaces, motivating the target-alignment construction."},{"cited_title":"Importance of kernel bandwidth in quantum machine learning","cited_arxiv_id":null,"evidence_quote":"Documents the Hamiltonian-evolution feature map used for the quantum kernel and its effectiveness in quantum kernel settings."},{"cited_title":"Clinical data classification with noisy intermediate scale quantum computers","cited_arxiv_id":null,"evidence_quote":"Supplies the three clinical datasets whose features are reused in the semi-artificial relabeling experiments."},{"cited_title":"Bandwidth Enables General- ization in Quantum Kernel Models","cited_arxiv_id":null,"evidence_quote":"Applies the alignment measure to quantum kernel bandwidth, providing background for using alignment to shape quantum kernel training."}],"review_version":1}