{"id":"98b30897-5edf-4de9-b297-dfe95b2992a5","arxiv_id":"2505.16484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"L-QMVKL trains view-specific quantum kernels and blends them with a hybrid global-local alignment objective, reporting modest accuracy gains on the Mfeat dataset over single-view and untuned classical baselines.","lead":"This paper combines multiple quantum kernels, one per data view, with a training step that uses local neighbor information, and tests the result on a six-view handwritten digit dataset. Its reported accuracy is a few points above single-view, global-only, and an untuned classical baseline, but the gains are small and the classical comparison is not fully optimized.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 1.44% accuracy advantage over the classical baseline is not established: the classical multi-view model is untrained while L-QMVKL optimizes both circuit parameters and kernel weights, and the difference is within the reported standard deviations (Table V).","rationale":"The central claim of the paper is that L-QMVKL achieves higher classification accuracy than its classical counterpart. The most load-bearing assumption is that the fixed-bandwidth, untrained Gaussian multi-kernel in Section IV.E constitutes a fair classical baseline. Because L-QMVKL optimizes both the quantum circuit parameters (Stage 1) and the combination weights (Stage 2) while the classical baseline is not trained or tuned at all, any accuracy advantage could simply reflect unequal optimization effort. This concern is reinforced by the paper's own observation that untrained quantum base kernels do not consistently outperform the classical kernels, meaning the training procedure, not quantum encoding, drives the gains. Additionally, the reported 1.44% multi-view advantage is smaller than the standard error of the difference computed from the published standard deviations, so the headline result is not statistically significant even taken at face value. The selection of λ, k, and P based on test accuracy further undermines the generalizability of the claimed improvements. These issues do not invalidate the proposed method as a plausible heuristic, but they prevent the central claim from being fully supported. The reader's verdict of CONDITIONAL is therefore appropriate; the paper needs a trained classical baseline, a proper validation split, and significance testing to substantiate the claimed quantum advantage.","tokens_in":15007,"tokens_out":8624,"duration_ms":71060,"concrete_test":"Re-run the comparison of Section IV.E with a trained classical multi-view baseline: optimize each view's Gaussian bandwidth (and, optionally, the combination weights η) by maximizing the same hybrid global-local kernel-target alignment (Eq. 12) on the training set, with identical λ, k, train/test splits, and 20 repeats. Also report a paired significance test (e.g., Wilcoxon or t-test) on the 20 accuracy differences. If the trained classical model reaches accuracy within one standard error of L-QMVKL's 91.25%, or if the p-value for the quantum-vs-classical difference exceeds 0.05, the central claim of quantum superiority is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.E compares L-QMVKL to a classical counterpart whose Gaussian base kernels use a fixed bandwidth equal to the mean Euclidean distance, with no training or tuning of bandwidths or combination weights, whereas L-QMVKL optimizes PQC parameters via hybrid global-local kernel-target alignment (Eq. 12) and weights η via Eq. (22). The paper itself notes untrained quantum base kernels do not consistently beat classical kernels (Section IV.E), suggesting the observed improvement after training reflects optimization effort rather than a quantum-specific benefit. Table V shows the multi-view classical accuracy is 89.81±4.63% vs trained quantum 91.25±4.56%; with 20 repeats the standard error of the difference is ≈1.45%, so the 1.44% gap is not significant (p≈0.3, no test reported). Likewise, hyperparameters λ, k, P are selected on the test set (Sections IV.C–D), further inflating the reported gains. The load-bearing assumption for the 'superior accuracy' claim is that an untrained, fixed-bandwidth Gaussian multi-kernel is the fair classical counterpart of a fully trained quantum multi-kernel; that assumption is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes L-QMVKL, a quantum multi-view kernel learning method that combines view-specific trainable quantum base kernels into a single multi-kernel and classifies with an SVM. The circuit parameters of each base kernel are trained by maximizing a hybrid global-local kernel-target alignment, and the combination weights are then learned by an alternating quadratic-programming procedure. Numerical simulations on the Mfeat dataset, with features reduced by PCA, are used to support claims that multi-view fusion, local information, and the quantum model itself improve accuracy over single-view, global-only, and classical Gaussian multi-kernel baselines.","tokens_in":15216,"tokens_out":7133,"duration_ms":63106,"significance":"The algorithm is coherently constructed and the two-stage training scheme is a reasonable extension of multiple kernel learning and local kernel alignment to quantum kernels. The paper provides analytic gradients, the parameter-shift rule for quantum gradient estimation, and a quadratic-programming formulation for the combination weights, which are useful methodological details. If the numerical claims were established with fair baselines and proper validation, this would be a valuable benchmark contribution to quantum kernel methods. However, the current empirical support is not yet convincing: the classical baseline is untrained, hyperparameters are selected on test accuracy, and key accuracy gaps are within the reported statistical uncertainty. The significance of the proposed contribution therefore rests on a reworked experimental evaluation.","major_comments":[{"comment":"The claim that L-QMVKL outperforms its classical counterpart by 1.44% is not supported by the reported data. The classical multi-view baseline uses Gaussian base kernels whose bandwidth is fixed to the mean Euclidean distance, with no training or tuning of the bandwidth or of the combination weights, whereas L-QMVKL optimizes the circuit parameters through hybrid global-local alignment and learns the weights via Eq. (22). The paper itself notes that untrained quantum base kernels do not consistently beat classical kernels, so the observed gain may reflect unequal optimization effort rather than a quantum-specific benefit. Moreover, with 20 repeats the unpaired standard error of the difference between 89.81±4.63% and 91.25±4.56% is about 1.45%, so the 1.44% gap is not statistically significant; no paired test or confidence interval is reported. The comparison should be made fair by giving the classical model the same optimization budget, and the difference should be reported with paired statistics.","section":"§IV.E, Table V"},{"comment":"The headline hyperparameter settings, including λ=0.125, k=8 and P=6, are selected using test-set accuracy. The text in §IV.C states that 'the optimal hyperparameters corresponding to the maximum accuracy vary across different feature views,' which shows that model selection is performed on the test set rather than on a separate validation split. This selection bias inflates the reported gains, including the 10.06% improvement quoted in the introduction and the 1.44% classical comparison in §IV.E. The authors should either fix the hyperparameters a priori, use a validation split, or report nested cross-validation results so that the reported test accuracy corresponds to a single pre-specified configuration.","section":"§IV.C–D"},{"comment":"The experimental conclusions are drawn from 20 repeats with standard deviations of roughly 4–7%, but no statistical significance tests are reported. For example, the claimed local-information gains of 0.19% and 0.56% at λ=0.25 in Table II are far smaller than the standard errors of the means, and the 3.81% MOR-feature improvement is also within the range that could arise from sampling noise. The paper consistently uses words such as 'significant' (abstract and Section I) without supporting tests. For the central comparisons, the authors should report paired tests or bootstrap confidence intervals rather than only mean ± standard deviation.","section":"§IV, Tables II–V"}],"minor_comments":[{"comment":"In the Multi-view row at P=6, the entry '80.56.94 ± 3.59' appears to be a typo and should likely read '90.56 ± 3.59'; this should be corrected because the tables are the main evidence for the accuracy claims.","section":"Table IV"},{"comment":"It is not specified whether the 20 repeats use independent random training/test splits or a single fixed split; this detail is needed to interpret the reported standard deviations.","section":"§IV.A"},{"comment":"The classical multi-view baseline is described only as using Gaussian base kernels; the paper does not state how the view-specific classical kernels are combined (for example, equal weights or learned weights), which is necessary for a fair and reproducible comparison.","section":"§IV.E"},{"comment":"The text says 'σx q and σx q represent the PauliX operator and the Pauli Z operator', but the second symbol should be σ_z; the Z-operator symbol is missing a subscript.","section":"§II.B, Eq. (24)–(25)"},{"comment":"The captions of Fig. 4(a) and 4(b) do not explain the 'untrain' horizontal line or the meaning of the separate single-view curves; please clarify the legend and the untrained baseline.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's algorithmic contribution is reasonable, but its main selling point is the reported accuracy improvement over classical and global-information baselines. That empirical case currently rests on an untrained classical baseline, test-set hyperparameter selection, and comparisons without significance tests. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also encourage the editor to ask for a clear statement of the classical multi-view combination method and for paired statistical comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible incremental algorithm paper. The new bit is the two-stage training of view-specific quantum kernels by hybrid global-local kernel alignment, then learning combination weights with a dynamically updated neighbor set. That combination is new relative to the cited quantum multi-kernel works. The math is clearly laid out, the parameter-shift gradients and quadratic programming step check out, and the paper is honest that untrained quantum kernels do not consistently beat classical ones. That honesty is a credit.\n\nThe soft spots are mostly in the experiments. The classical baseline in Section IV.E is not trained: Gaussian base kernels use a fixed bandwidth equal to the mean Euclidean distance, and there is no mention of learning the combination weights. The quantum base kernels are trained via Eq. (12), so the 1.44% average accuracy advantage is an unfair comparison and likely reflects optimization effort, not any quantum-specific benefit. The stress-test note is right: with 20 repeats and the reported standard deviations, that gap is within noise. The authors also select lambda, k, and P using test accuracy (Sections IV.C-D), which further inflates the reported gains. There are no significance tests anywhere.\n\nNone of this kills the method. The multi-view improvement over single-view models is more robust (around 10%), and the idea of using local structure in quantum kernel training is reasonable. But the central claim as stated—superiority over the classical counterpart—is not supported by the evidence. The paper needs a properly tuned classical baseline (e.g., learning Gaussian bandwidths and kernel weights on validation data), a separate validation split for hyperparameters, significance tests, and ideally another dataset. Code and data are also missing, which weakens reproducibility.\n\nWho is this for? Someone working on quantum multiple kernel learning or on multi-view quantum classifiers would get a useful algorithmic reference. It is not a broad conceptual advance. I would not cite it in my own work in the next year, but I would send it to a referee: the method is concrete, the presentation is competent, and a serious revision could make it a solid contribution. My recommendation is to accept for peer review.","headline":"A plausible incremental algorithm with an unfair classical baseline and test-set hyperparameter selection; the quantum-vs-classical claim is not supported, but the multi-view local training idea is worth a serious revision.","tokens_in":15749,"tokens_out":2994,"would_cite":false,"duration_ms":24402,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"L-QMVKL trains one quantum kernel per data view on a hybrid global-local alignment objective and beats single-view and classical multi-kernel baselines on Mfeat.","keywords":["quantum kernel","multiple kernel learning","multi-view learning","kernel-target alignment","local information","parameterized quantum circuits","support vector machine","Mfeat dataset"],"falsifier":"Re-run the Mfeat comparison with the classical multi-kernel given the same training budget: optimize its bandwidth by the same hybrid global-local kernel-target alignment, select $\\lambda$, $k$, and $P$ on a validation split, and compare test accuracy. If the tuned classical model reaches or exceeds 91.25%, the claimed quantum advantage is an artifact of unequal optimization effort.","tokens_in":14776,"feed_emoji":"⚛️","tokens_out":13035,"duration_ms":95782,"temperature":0.7,"pith_summary":"This paper proposes L-QMVKL, a quantum kernel method that fuses view-specific quantum kernels instead of relying on a single feature representation. It argues that training each quantum base kernel with a hybrid global-local kernel-target alignment objective, and then learning the fusion weights, lets the model capture local structure that global-only quantum kernels miss. On the Mfeat handwritten-digit benchmark the method reports large accuracy gains over single-view quantum models and a 1.44% average advantage over an untrained classical Gaussian multi-view kernel at $\\lambda=0.125$, $k=8$. The claimed significance is that quantum kernel methods traditionally plateau on heterogeneous data, and this is a concrete recipe for combining multiple views and local structure inside a trainable quantum kernel.","feed_headline":"Multi-view quantum kernels beat single-view and classical baselines","feed_subtitle":"A hybrid global-local training objective lifts the multi-view model to 91.25 percent accuracy on a six-view digits benchmark.","key_machinery":"The load-bearing object is the hybrid global-local kernel-target alignment $HTA(K,K^*)=(1-\\lambda)LTA(K,K^*)+\\lambda TA(K,K^*)$, where $TA$ is the usual kernel-target alignment against the ideal label kernel $K^*$ and $LTA$ averages alignment over local kernel matrices built from each point's $k$ nearest neighbors. Training is sequential: first, each view's parameterized-circuit parameters $\\theta^m$ are updated by gradient ascent on $HTA$ using analytic gradients from the parameter-shift rule; second, the combination weights $\\eta$ are optimized by alternating a quadratic program with a dynamically updated nearest-neighbor set defined through the combined quantum kernel. This two-stage scheme is what allows the model to fuse complementary views while still fitting local structure.","core_discovery":"The central claim is that the bottleneck of quantum kernels on heterogeneous data is single-view representation plus global-only training, and that both can be fixed in one construction. L-QMVKL assigns each of the M views its own trainable quantum kernel $\\kappa_{\\theta_m}^m$, trains the circuit parameters $\\theta_m$ by maximizing $HTA(K,K^*)=(1-\\lambda)LTA(K,K^*)+\\lambda TA(K,K^*)$, and then learns combination weights $\\eta$ by the same hybrid objective with nearest-neighbor sets recomputed in the fused-kernel space. On the binary Mfeat task this reaches 91.25% mean test accuracy, against 89.81% for the classical Gaussian multi-kernel and 88.13% for the same quantum multi-kernel left untrained, and the paper reports at least 10.06% higher accuracy than the best single-view global-only quantum model at $\\lambda=0.25$, $k=8$, $P=6$. The paper states this as evidence that multi-view fusion and local information together improve quantum kernel learning on real data.","pith_inferences":["Inference: the Section IV.E comparison pairs trained, hyperparameter-selected quantum kernels against a classical Gaussian multi-kernel whose bandwidth is not optimized, so that asymmetry, rather than a quantum-specific effect, may explain part of the 1.44% gap.","Inference: Sections IV.C and IV.D select $\\lambda$, $k$, and $P$ by the test accuracies reported in the tables, with no separate validation split, so the headline gains are likely optimistic for unseen data.","Inference: the same hybrid global-local alignment objective could be tested on other trainable kernels and on kernel-based clustering, where local structure is known to matter; those settings would show whether the mechanism, not the particular SVM-plus-Mfeat combination, is responsible.","Inference: the divergence between rising $HTA$ and falling accuracy at $P=8$ suggests that maximizing alignment alone may overfit the kernel to the training labels, a possibility the paper does not develop."],"forward_implications":["Multi-view quantum kernels trained this way beat every single-view quantum kernel on Mfeat; at circuit depth $P=6$ the independent multi-view advantage is at least 9.5%.","Local information helps beyond global-only training: on three views (FOU, ZER, MOR) the locally trained kernels outperform global-only kernels, with a peak gain of 3.81 percentage points at $k=8$, $P=6$.","Trained quantum base kernels outperform the untrained classical Gaussian baseline, and the assembled L-QMVKL averages 1.44% higher accuracy than the classical multi-view model at $\\lambda=0.125$, $k=8$; accuracy generally rises with $k$, reaching 90.56% at $k=12$.","Increasing the QAOA-style circuit depth $P$ raises the kernel-target alignment value but not accuracy monotonically: mean accuracy peaks at 90.75% for $P=7$ and falls at $P=8$.","The construction accepts any parameterized data-encoding operation $W(x,\\theta)$, so the same recipe could be applied with other ansatz families and other kernel-based classifiers."],"supporting_citations":[{"why":"defines kernel alignment and kernel-target alignment, the base of the training objective.","marker":"[49]"},{"why":"introduces local kernel-target alignment with nearest-neighbor local kernel matrices, from which the hybrid objective is assembled.","marker":"[42]"},{"why":"supplies the multiple kernel learning framework used to combine view-specific kernels into a multi-kernel.","marker":"[4]"},{"why":"prior trainable quantum multiple kernel learning that the paper extends to multi-view data and local information.","marker":"[35]"},{"why":"provides the QAOA-style parameterized circuit structure behind the paper's quantum base kernels.","marker":"[46]"},{"why":"the Mfeat dataset is the experimental benchmark for every accuracy comparison in the paper.","marker":"[53]"},{"why":"the parameter-shift rule supplies the analytic gradients used to train the circuit parameters.","marker":"[51]"},{"why":"defines the SVM classifier that consumes the trained combined kernel in the classification stage.","marker":"[48]"}],"fun_headline_variants":["Quantum multi-view kernels with local info hit 91.25%","Local info boosts quantum multi-view kernels past classical","Hybrid global-local training lifts quantum kernels to 91%","Multi-view and local structure improve quantum kernel accuracy","Quantum kernel fusion with local alignment beats classical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline quantum advantage assumes that a classical Gaussian multi-kernel with bandwidth fixed to the mean pairwise distance, and never trained or tuned, is a fair baseline for quantum kernels that receive full optimization and hyperparameter selection.","fun_headline_variants_meta":{"raw":{"variants":["Quantum multi-view kernels with local info hit 91.25%","Local info boosts quantum multi-view kernels past classical","Hybrid global-local training lifts quantum kernels to 91%","Multi-view and local structure improve quantum kernel accuracy","Quantum kernel fusion with local alignment beats classical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1282,"prompt_tokens":965,"completion_tokens":317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":240}},"tokens_in":581,"tokens_out":317,"duration_ms":2744,"temperature":1.0,"reasoning_tokens":240,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:59:21.464337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Mfeat comparison with the classical multi-kernel given the same training budget: optimize its bandwidth by the same hybrid global-local kernel-target alignment, select $\\lambda$, $k$, and $P$ on a validation split, and compare test accuracy. If the tuned classical model reaches or exceeds 91.25%, the claimed quantum advantage is an artifact of unequal optimization effort.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines kernel alignment and kernel-target alignment, the base of the training objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces local kernel-target alignment with nearest-neighbor local kernel matrices, from which the hybrid objective is assembled."},{"cited_title":"G¨ onen and E","cited_arxiv_id":null,"evidence_quote":"supplies the multiple kernel learning framework used to combine view-specific kernels into a multi-kernel."},{"cited_title":"Duin, Multiple Features, UCI Machine Learning Repository (1998)","cited_arxiv_id":null,"evidence_quote":"the Mfeat dataset is the experimental benchmark for every accuracy comparison in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the SVM classifier that consumes the trained combined kernel in the classification stage."}],"review_version":1}