{"id":"3642d384-bf71-4e43-8c9e-43d210546921","arxiv_id":"2501.08193","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A Qiskit-based benchmark of four QML classifiers on genomic data shows near-chance test accuracy with strong feature-map dependence, but lacks baselines and statistical rigor.","lead":"This paper applies standard quantum machine learning classifiers to a genomic sequence benchmark and reports that feature map choice affects accuracy, recall, and overfitting. The results are near random on test data, and the paper lacks baselines, error bars, and a working code link.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pegasos-QSVC's near-perfect recall matches a majority-class classifier, so the headline 'excelling in recall' is an artifact rather than evidence of QML capability.","rationale":"I read the paper as claiming two things: that feature map choice changes QML performance, and that QML improves genomic classification, with Pegasos-QSVC notably having high recall. For the second claim to hold, the reported metrics must be informative. The pattern in Table I for Peg-QSVC, accuracy near 51, precision near 51, recall near 99 across three unrelated feature maps, is exactly the signature of a model emitting a constant positive label on a roughly balanced test set. The paper gives no confusion matrix, no class distribution for the subset used, and no baseline, so the reader cannot distinguish 'excelling in recall' from 'predicting everybody into the positive class'. This is more load-bearing than the Algorithm 2 representational concern: even if one accepts a dual or kernel implementation of Pegasos, the numbers in Table I still would not show QML improvement. The issue also explains why all test accuracies hover at chance: the models are not learning a separating rule that transfers. I therefore agree with the reader's REJECT verdict, but for a different primary reason than the reader's weakest-assumption check, hence partial agreement. I do not see the paper as fraudulent; the issue is an evidentiary and interpretive failure in the reported experiments.","tokens_in":15376,"tokens_out":5129,"duration_ms":53952,"concrete_test":"Re-run the claimed Pegasos-QSVC experiment, or inspect the authors' Qiskit code, and report the full test confusion matrix plus the class proportions of the democoding versus intergenic subset. Then compare against a dummy classifier that always predicts the majority class. Concretely: if the fraction of positive test predictions made by Peg-QSVM is at least 0.99, or if accuracy and precision match the majority-class base rate within one percentage point while recall is near 100%, the high recall is a trivial artifact and cannot support the paper's claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is not the Hilbert-space representation of w in Algorithm 2, although that is also unresolved; it is that the paper's flagship quantitative result is indistinguishable from a trivial constant classifier. In every feature-map row, Peg-QSVM has test accuracy near 51%, precision near 51%, and recall near 99% (Table I, rows 2, 6, 10). On a roughly balanced binary set, always predicting the positive class yields exactly recall near 100%, precision near 51%, accuracy near 51%, and F1 near 67.4%. That is what the table shows. The paper never reports the class distribution of the chosen subset, nor a majority-class baseline, so the 'high sensitivity' claim has no discriminating content. The discussion's explanation that Pegasos 'prioritizes recall' is also wrong: Pegasos minimizes regularized hinge loss (Eq. 11), which has no recall bias; the observed pattern is a degenerate boundary or prediction collapse. Since the conclusion rests on this result to claim that QML enhances genomic classification, the central claim collapses unless this artifact is ruled out.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Manuscript arXiv:2501.08193 evaluates four QML classifiers (QSVC, Pegasos-QSVC, VQC, and QNN) combined with three feature maps (ZFeatureMap, ZZFeatureMap, and PauliFeatureMap) for binary classification of genomic sequences from the democoding-vs-intergenomic benchmark. It presents pseudocode for the classifiers, a convergence-analysis section, and simulation results on a small PCA-reduced subset using Qiskit AerSimulator. The main conclusions are that feature-map choice materially affects performance, that Pegasos-QSVC achieves near-perfect recall, that QNN achieves the highest training accuracy, and that QML “enhances genomic classification performance.” The paper also claims an open-source implementation and superior results over prior work.","tokens_in":15610,"tokens_out":5994,"duration_ms":57419,"significance":"If the empirical claims were properly supported, the paper would provide a useful benchmarking reference for QML in genomics, especially the systematic comparison across feature maps. The manuscript is clearly organized and the simulation workflow (PCA to four qubits, three feature maps, four algorithms, and four metrics) is easy to follow. However, no machine-checked proofs, reproducible code link, or baseline comparisons are actually provided; the convergence section is standard material; and the key quantitative claims are not supported by the reported experiments. As it stands, the contribution is a collection of tentative observations rather than a validated benchmarking study.","major_comments":[{"comment":"Rows 2, 6, and 10 of Table I report Pegasos-QSVC with test accuracy near 51%, precision near 51%, recall near 99%, and F1 near 67.4% for every feature map. This is exactly the signature of a degenerate classifier that always predicts the positive class on a nearly balanced binary set: accuracy ≈ P(positive), precision ≈ P(positive), recall = 1, and F1 = 2P(positive)/(1+P(positive)), which equals 0.674 for P(positive) ≈ 0.51. The paper never reports the class distribution of the chosen subset, the subset size, or a majority-class baseline, so the high recall has no discriminating content. The explanation in Section IV-A that Pegasos “prioritizes recall” is also inconsistent with Eq. (11), which minimizes regularized hinge loss and has no recall bias. This artifact, if not ruled out, invalidates the central conclusion that QML enhances genomic classification.","section":"§IV, Table I"},{"comment":"Algorithm 2 lines 13–19 and Eqs. (13)–(15) update a weight vector w directly in the quantum feature space: w ← (1−ηλ)w + η y_i φ(x_i), with φ(x_i) an element of an exponentially large Hilbert space. The text itself acknowledges in Section II-C that “direct manipulation [of w] is challenging” and says the kernel will be used implicitly, but no finite representation of w or kernelized dual update is provided. As written, the pseudocode is not implementable, because w cannot be stored or updated explicitly. Since the Pegasos-QSVC row of Table I is the basis for the paper’s most prominent empirical claim, this unresolved algorithmic gap makes those results unreliable.","section":"§II-C, Algorithm 2"},{"comment":"The experimental section reports no error bars, no repeated-seed statistics, no classical baselines (e.g., linear SVM or random forest on the same PCA features), and no comparison numbers from the references [3, 9, 13, 22] that the paper claims to outperform. All accuracies in Table I lie within roughly 2–5 points of 50%, and the training/test differences are of the same size as the differences between algorithms; without variance estimates or significance tests, the observed differences cannot be distinguished from noise. The statement in Section IV that the results are “a significant improvement … outperforming the results of [13], [3], [22], and [9]” is therefore unsubstantiated.","section":"§IV, Experiments"},{"comment":"This section restates standard convergence results for convex SVM/Pegasos (Eqs. (39)–(43)) and standard barren-plateau scaling (Eq. (45)) rather than analyzing the customized quantum algorithms actually proposed. The assertion near Eq. (48) that “a positive semi-definite Hessian implies a smooth optimization landscape with no sharp minima” is not generally valid: a pointwise PSD Hessian does not imply global convexity or absence of sharp minima in the nonconvex VQC/QNN landscapes. This does not support the claimed “theoretical evaluations and demonstrations” contribution.","section":"§III, Convergence Analysis"}],"minor_comments":[{"comment":"There are multiple typos: “PualiFeatureMap” should be “PauliFeatureMap”, “dinmentionality reduction” should be “dimensionality reduction”, “asubset of dataset is chosen and slip into training” should be “a subset of the dataset is chosen and split into training”, and “the the Karush-Kuhn-Tucker” in Section III should be “the Karush-Kuhn-Tucker”.","section":"§II-A, §I-B, §IV"},{"comment":"Equation (18) defines the cost as the expectation value ⟨O⟩, while Eq. (19) defines it as the mean squared error; these are inconsistent and should be reconciled.","section":"§II-D, Eqs. (18)–(19)"},{"comment":"The paper states that an open-source Qiskit implementation is presented, but no repository URL or code-availability statement appears in the manuscript; this prevents reproducibility.","section":"§IV, Abstract"},{"comment":"The labels “Extended QSVC”, “Improved Pegasos-QSVC”, “Enhanced VQC”, and “Advanced QNN” are not justified by any comparison with the base versions; the pseudocode content is essentially the standard algorithm.","section":"§II, Algorithms 1–4"},{"comment":"The caption of Fig. 2 says the ZFeatureMap circuit includes CNOT gates, but the text describes only independent Rz rotations for the ZFeatureMap; the caption and circuit should be checked for consistency.","section":"§II-A, Fig. 2"}],"recommendation":"reject","confidential_remarks":"I could not verify the open-source claim because no code link is provided. The most serious problem is the degenerate-classifier signature in the Pegasos-QSVC rows of Table I, which, together with the absence of any baseline, makes the central conclusion unsupportable. Even if the authors added a majority-class baseline, the reported accuracies near 50% would not support the claim that QML enhances genomic classification; this would require a fundamentally different experimental conclusion. For these reasons I recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's only distinctive numeric finding—Pegasos-QSVC's near-perfect recall—is almost certainly an artifact of predicting the majority class, and the algorithm pseudocode itself is not implementable as written. Nothing else in the paper rises above a routine benchmark scan.\n\nWhat it does well: the authors give clean circuit diagrams and a consistent experimental pipeline (PCA to four qubits, three feature maps, four QML algorithms) on a public genomic benchmark. If the numbers were trustworthy, the comparison would be a useful data point for practitioners. The convergence section is a readable restatement of standard results.\n\nSoft spots, in order of severity. First, the table: every Peg-QSVM row has test accuracy around 51%, precision around 51%, recall near 99%, F1 near 67%, AUROC around 50. On an even roughly balanced two-class set, that is exactly what a constant positive classifier produces. The paper never reports the class split or a majority-class baseline, so the \"excelling in recall\" claim has no discriminating content. The discussion's rationalization that \"Pegasos prioritizes recall\" is simply wrong—the algorithm minimizes regularized hinge loss, which has no recall bias. Second, Algorithm 2 updates a weight vector directly in exponentially large Hilbert space; the text admits this is \"challenging\" but never gives a finite representation, so the method is not reproducible from the paper. Third, no error bars, no classical baselines, no subset size, no code despite the \"open-source\" claim.\n\nFor a reader: this is a cautionary example of what happens when you benchmark without baselines or class priors. It is not a research contribution. My recommendation: desk reject. If the authors want to fix it, they need to report class distribution, add majority-class and classical SVM comparisons, provide the code, and rewrite the Pegasos-QSVC description to match a valid kernelized implementation.","headline":"A routine QML benchmark whose headline result—Pegasos-QSVC's high recall—is indistinguishable from a majority-class predictor, and whose algorithm pseudocode is not implementable as written.","tokens_in":16126,"tokens_out":3270,"would_cite":false,"duration_ms":31495,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the choice of quantum feature map — the way genomic data is encoded into quantum states — interacts with the choice of classifier to determine performance, with Pegasos-QSVC reaching near-perfect recall and QNN…","keywords":["Feature Map","Genomic Sequence Classification","Pegasos-QSVC","Quantum Machine Learning","Quantum Neural Networks (QNN)","Quantum Support Vector Classifier (QSVC)","Variational Quantum Circuits (VQC)"],"falsifier":"Re-run the Pegasos training with the weight vector stored implicitly as a kernel expansion, the standard way to avoid manipulating a vector in an exponentially large space, and compare the resulting metrics both to the paper's numbers and to a classical linear SVM on the same four PCA features; if a classical model already reproduces roughly 99 percent recall at 51 percent accuracy, the quantum encoding is not the source of the reported sensitivity, and if the kernel-expansion version cannot reproduce the paper's numbers, the reported results are not reproducible from the pseudocode.","tokens_in":15146,"feed_emoji":"🧬","tokens_out":14659,"duration_ms":123922,"temperature":0.7,"pith_summary":"This paper argues that quantum machine learning can be applied to binary genome-sequence classification and that the way data is encoded into quantum states — the feature map — interacts with the choice of classifier to determine performance. Simulating quantum circuits on the democoding-versus-intergenomic benchmark, the authors find that Pegasos-QSVC achieves recall above 99 percent under every feature map, QNN attains the highest training accuracy across all feature maps, and VQC performs consistently but weakly. The paper's stated conclusion is that QML enhances genomic classification, with the caveat that feature-map-dependent swings in test accuracy signal a risk of overfitting. All reported test accuracies fall between about 49.6 and 51.7 percent, near the chance level of this two-class benchmark, a fact the paper reports alongside its positive conclusion.","feed_headline":"Pegasos quantum classifier hits 99% recall on genome data","feed_subtitle":"Feature-map choice swings training scores a few points while test accuracy stays near chance.","key_machinery":"The carrying object is the quantum feature map and the kernel it induces. Each genome sequence is reduced by PCA to four coordinates, which are encoded into four qubits by single-qubit rotations (ZFeatureMap), pairwise interaction terms (ZZFeatureMap), or rotations about all axes with entangling gates (PauliFeatureMap); the quantum kernel $K(\\mathbf{x}_i,\\mathbf{x}_j) = \\langle \\phi(\\mathbf{x}_i) | \\phi(\\mathbf{x}_j)\\rangle$, estimated by measuring state overlap, is the similarity measure that QSVC and Pegasos-QSVC use to separate the two classes. The variational methods (VQC, QNN) replace the kernel with parameterized circuits trained by the parameter-shift rule, while the Pegasos variant carries its classifier as a weight vector updated by stochastic sub-gradient steps inside the quantum feature space, with a projection step that enforces regularization.","core_discovery":"On the paper's own terms, the discovery is that the interplay between feature mapping and QML algorithm is the main factor in genomic classification performance: Pegasos-QSVC is highly sensitive, producing recall of 99.12 percent (ZFeatureMap), 99.76 percent (ZZFeatureMap), and 99.41 percent (PauliFeatureMap) while holding test accuracy near 51 percent; QNN reaches the highest training accuracy of every feature map tested (up to 55.02 percent with ZZFeatureMap); and ZZFeatureMap generally improves training accuracy over the other two maps without improving test accuracy, which the authors attribute to overfitting. The paper further contends that QML enhances genomic classification performance and offers convergence arguments — convexity of the QSVC dual problem, an $O(1/\\lambda\\epsilon)$ iteration bound for the Pegasos solver, and barren-plateau scaling for the variational methods — as theoretical support for these empirical rankings.","pith_inferences":["Reading the table, I infer that the near-chance test accuracies qualify the conclusion that QML enhances genomic classification: on this two-class benchmark, a constant positive-class predictor would already achieve roughly 50 percent accuracy and 100 percent recall, so the Pegasos recall claim should be checked against such a trivial baseline on the same four PCA features (my inference, not the p","The $O(1/\\lambda\\epsilon)$ convergence rate quoted for Pegasos is the classical result from the original solver; carrying it over to the quantum-kernel setting requires a finite representation of the weight vector, which the paper does not supply, so a natural extension is to restate the algorithm with the weight vector kept implicitly as a kernel expansion and verify that the reported metrics are","A systematic sweep over circuit depth, entanglement structure, and qubit count would reveal whether the observed ordering — ZZ best for training fit, Pegasos recall edge, VQC consistently weak — holds up or is an artifact of the four-qubit PCA compression (my inference, beyond the paper's experiments).","The comparison that would sharpen the paper's claim is a classical kernel SVM or random forest on the same PCA-reduced features; without such a baseline, 'QML enhances genomic classification' stands as an assertion rather than a demonstrated advantage (my inference, beyond the paper's text)."],"forward_implications":["If the central claim holds, feature-map selection becomes the first performance lever for QML in genomics: switching from the Pauli to the ZZ map shifts training accuracy by roughly two to four points, and no single map wins on both training and test metrics.","Pegasos-QSVC's consistently near-perfect recall makes it the natural candidate for genomic screening tasks in which missing a positive sequence costs more than a false alarm, at the price of precision around 51 percent.","QNN's top training accuracy across all feature maps suggests variational circuits can fit genomic data well but need explicit regularization or early stopping to carry that fit to unseen sequences.","The authors' conclusion that QML enhances genomic classification carries the corollary that larger datasets and multiclass tasks would preserve the observed ordering among feature maps and algorithms, which the paper lists as future work.","Because the experiments run on a noiseless simulator, the authors expect device noise and crosstalk to alter the rankings and defer that evaluation to future work."],"supporting_citations":[{"why":"Provides the quantum support vector machine formulation and complexity context from which the paper's QSVC and Pegasos-QSVC variants are drawn.","marker":"[3]"},{"why":"Supplies the quantum neural network model whose training accuracy the paper reports as the highest across feature maps.","marker":"[13]"},{"why":"Defines the ZZFeatureMap and PauliFeatureMap plus the quantum kernel approach that all four classifiers consume.","marker":"[14]"},{"why":"Defines the ZFeatureMap whose single-qubit Z-rotations are the baseline encoding in Table I.","marker":"[15]"},{"why":"Supplies the seminal QSVC algorithm and kernel-based decision rule the paper extends.","marker":"[17]"},{"why":"Gives the variational quantum classifier framework and its expressiveness analysis that underlies the VQC implementation.","marker":"[20]"},{"why":"Provides the earlier QML classifier performance analysis that the paper claims to outperform on genomic data.","marker":"[22]"},{"why":"Supplies the Pegasos stochastic sub-gradient solver and its convergence rate, which the Pegasos-QSVC variant embeds.","marker":"[23]"},{"why":"Establishes the parameter-shift rule used to compute the analytic gradients for VQC and QNN training.","marker":"[28]"},{"why":"Supplies the benchmark genomic dataset (democoding vs. intergenomic) on which every reported metric is measured.","marker":"[34]"}],"fun_headline_variants":["Quantum classifier hits 99% recall, but test accuracy stays at chance","Feature maps sway quantum training scores while test accuracy stays flat","QML genome test: high recall, low test accuracy, overfitting risk","Pegasos QSVC: 99% recall, yet test accuracy near 50%","Quantum vs genome: training soars, testing stalls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's strongest numerical result depends on the assumption that Pegasos-QSVC's internal model can be stored and updated directly inside the quantum feature space, a step the paper itself admits is challenging and never shows how to carry out.","fun_headline_variants_meta":{"raw":{"variants":["Quantum classifier hits 99% recall, but test accuracy stays at chance","Feature maps sway quantum training scores while test accuracy stays flat","QML genome test: high recall, low test accuracy, overfitting risk","Pegasos QSVC: 99% recall, yet test accuracy near 50%","Quantum vs genome: training soars, testing stalls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3087,"prompt_tokens":893,"completion_tokens":2194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":509,"tokens_out":2194,"duration_ms":14487,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:28:54.664840+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Pegasos training with the weight vector stored implicitly as a kernel expansion, the standard way to avoid manipulating a vector in an exponentially large space, and compare the resulting metrics both to the paper's numbers and to a classical linear SVM on the same four PCA features; if a classical model already reproduces roughly 99 percent recall at 51 percent accuracy, the quantum encoding is not the source of the reported sensitivity, and if the kernel-expansion version cannot reproduce the paper's numbers, the reported results are not reproducible from the pseudocode.","supporting_citations":[{"cited_title":"Universal expressiveness of variational quantum classifiers and quantum kernels for support vector machines,","cited_arxiv_id":null,"evidence_quote":"Gives the variational quantum classifier framework and its expressiveness analysis that underlies the VQC implementation."},{"cited_title":"Performance analysis of quantum machine learning classifiers,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier QML classifier performance analysis that the paper claims to outperform on genomic data."},{"cited_title":"Pegasos: Primal estimated sub-gradient solver for svm,","cited_arxiv_id":null,"evidence_quote":"Supplies the Pegasos stochastic sub-gradient solver and its convergence rate, which the Pegasos-QSVC variant embeds."},{"cited_title":"Genomic benchmarks: a collection of datasets for genomic sequence classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark genomic dataset (democoding vs. intergenomic) on which every reported metric is measured."}],"review_version":1}