{"id":"2c479a9d-2cbb-421b-a9df-c6230f0588bd","arxiv_id":"2607.23683","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Qutrit neural quantum kernels beat matched QNN baselines on four benchmarks, with gains that depend on feature budget, register size, and SU(3) parameterization, under noiseless simulation.","lead":"This paper extends neural quantum kernels from qubits to qutrits and tests how feature count, system size, kernel construction, and SU(3) parameterization affect binary and three-class accuracy. It is a careful design study showing qutrit kernels usually beat matched quantum neural nets and stay competitive with classical SVMs under ideal simulation.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The QNN-to-NQK lift may be entirely a readout-capacity effect: the baseline QNN is limited to a fixed threshold/argmax decision rule, while the NQK gets a tuned SVM over the kernel — no matched-capacity readout control is run.","rationale":"The paper is a careful, well-hedged simulation study: cross-validation without leakage, error bars, ablations over parameterization and cost function, heavy baseline optimization (7000 restarts), an honest classical RBF-SVM comparison, and explicit listing of the noiseless-simulation limitation. The reader correctly identified the idealization risk, and I partially agree — but that risk is openly declared as a limitation and the claims are explicitly scoped to \"settings considered\" under \"ideal simulation,\" so it does not undermine internal validity. The less-examined soft spot is internal: the headline comparative claim (NQK > QNN) is generated by an asymmetric comparison in which only the NQK side receives a capacity-tuned classifier head. The paper narrates this mechanism but never measures its magnitude. This does not make the paper wrong — the NQK principle from Ref. [54] is precisely this readout upgrade — but it bears directly on how much of the evidence supports the design conclusions about qutrit-specific ingredients (SU(3) parameterization, p- and n-scaling) versus generic classifier-head effects. A single cheap control experiment would settle it, and the result would either strengthen the paper (lift persists → kernel geometry matters) or require re-scoping the central claim. This keeps the reader's CONDITIONAL verdict appropriate; I would not move it.","tokens_in":28384,"tokens_out":2154,"duration_ms":59995,"concrete_test":"For each fold and dataset, freeze the trained QNN embedding and record the full single-qutrit output probability vector (and, for n>1, the marginals on all qutrits). Train (i) multinomial logistic regression and (ii) an RBF-SVM on these classical feature vectors with the same nested C/gamma grid used for the NQK and RBF-SVM baselines. Compare against Table II / Table VI entries. If the matched-capacity readout closes the QNN-to-NQK gap (e.g., reaches ~90.7% binary / ~87.2% three-class on Fashion-MNIST), the lift is readout capacity rather than kernel geometry and the central comparative claim needs re-scoping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that \"qutrit NQKs improve over the corresponding QNN baselines in nearly all settings considered.\" But the comparison is structurally asymmetric in classifier capacity. The QNN baseline makes predictions from a fixed, hand-chosen decision rule on a single qutrit: a threshold at zero on ⟨Sz⟩ (Eq. 20) for binary tasks, or argmax over three computational-basis probabilities (Eq. 21) for three-class. The NQK takes the same trained circuit and replaces this head with an SVM whose regularization C is grid-searched, operating on a kernel matrix over the full n-qutrit feature state. The paper itself frames the mechanism this way (Sec. II A: \"replacing the QNN's fixed measurement-based decision rule with an optimal linear readout in the induced feature space\"), so the lift is essentially guaranteed whenever the QNN was not trained to optimality — which, despite 7000 restarts, is selected on training loss, not decision quality.\n\nWhat is missing is the control that would make the claim informative: a matched-capacity classical readout on the frozen trained embedding. For example, train a linear or kernel SVM on the full measurement probability vector of the trained QNN (all three outcome probabilities of the readout qutrit, or marginals on all qutrits), or apply the same threshold-vs-SVM comparison. If such a readout matches the NQK accuracy, then \"NQKs beat QNNs\" reduces to \"SVMs beat fixed thresholds,\" which says nothing about the qutrit kernel construction, SU(3) parameterization, or scaling with n and p — the quantities the paper's design-study conclusions rest on. The parameterization ablation (Sec. III E 1) partly inherits this issue, since optimization diagnostics (App. D) are measured at the QNN stage while headline accuracies are post-SVM. The paper is honest about the mechanism in prose, but no experiment quantifies how much of the reported lift survives once readout capacity is equalized, and the central empirical claim is stated as a property","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript extends neural quantum kernels from qubits to qutrits. It generalizes the data-reuploading embedding to local SU(3) blocks, introduces qutrit SUM and controlled-rotation entanglers, and studies both 1-to-n and n-to-n kernel constructions for binary and three-class classification. Noiseless simulations on Fashion-MNIST, HAR, MAGIC, and Covertype examine feature budget, system size, three SU(3) parameterizations, and cost-function sensitivity. The authors report that the resulting NQKs outperform matched QNN baselines in nearly all settings, that gains from additional features and qutrits are dataset-dependent and often saturate, and that the Geometric SU(3) parameterization usually performs best and shows more favorable training diagnostics. The NQKs are broadly competitive with, but do not clearly exceed, an RBF-SVM baseline.","tokens_in":28799,"tokens_out":6196,"duration_ms":407302,"significance":"If the reported comparisons hold up, this is a useful systematic design study for qudit quantum machine learning and a natural extension of NQKs beyond qubits. Its strengths are the breadth of the empirical protocol: two kernel constructions, binary and multiclass tasks, four datasets, three SU(3) parameterizations, matched circuit baselines, stratified five-fold evaluation with preprocessing fitted only on training folds, standard errors, a classical RBF-SVM comparison, and informative cost-function and training-diagnostics appendices. The claims are generally hedged and the paper does not overstate them as a quantum advantage. Its significance is primarily methodological guidance for qutrit model design under ideal simulation, rather than evidence of near-term practical superiority over classical methods.","major_comments":[{"comment":"The evaluation protocol does not state where the SVM regularization C—and the RBF baseline’s C and γ—are selected. If the grid search uses the held-out fold’s test accuracy, Tables II and VI and the QNN-versus-NQK comparisons are optimistically biased; if an inner training split or nested cross-validation was used, this needs to be stated explicitly. Because several reported differences are only about 1–3 percentage points, the selection protocol is load-bearing. Please use nested model selection or a fixed validation split, and preferably report paired fold-level differences in addition to fold means and standard errors.","section":"Sec. III B; Tables II and VI"},{"comment":"The central QNN-to-NQK comparison is asymmetric in readout capacity. The QNN predicts from a fixed threshold on one ⟨Sz⟩ value, or an argmax over three probabilities on one qutrit, whereas the NQK receives a tuned SVM over a kernel of the full n-qutrit feature states. The manuscript itself characterizes the mechanism as replacing the QNN’s fixed measurement rule with an optimized readout. Consequently, the lift may reflect readout capacity or access to richer state information rather than the qutrit kernel construction specifically. A frozen-embedding control is needed: for the same trained circuits, train a logistic/linear or kernel SVM on all single-qutrit outcome probabilities or a richer fixed set of low-order moments, using the same nested tuning; a validation-tuned binary threshold would also be useful. If this control matches the NQK, the conclusion should be narrowed to an SVM-re","section":"Sec. II A; Eqs. (20)–(21); Figs. 4–7; Table II"},{"comment":"The parameterization ablation compares Geometric, Euler, and Givens coordinates at the same learning rate and with the same sequential assignment of the first p coordinates, but these coordinates have different nonlinear roles, effective ranges, and Jacobian scales. The larger gradient norms reported for the Geometric form may therefore indicate a different parameter-space metric or effective optimization step size rather than intrinsically more favorable optimization. To support the comparatively strong conclusion that the unitary representation affects optimization behaviour, please add a learning-rate/optimization-budget sweep or a Jacobian-normalized encoding-scale comparison, or explicitly restrict the claim to the present fixed optimizer and coordinate-scaling protocol.","section":"Sec. III E 1, Eqs. (10)–(12); Appendix D, Fig. 20"}],"minor_comments":[{"comment":"Please state in the abstract that all experiments are ideal, noiseless state-vector simulations at fixed depth L=6. The limitation is appropriately acknowledged later, but including it in the abstract would prevent the performance claims from being read as hardware results.","section":"Abstract and Sec. IV"},{"comment":"The statement that features enter “symmetrically” in the Geometric parameterization is potentially misleading because the su(3) generators do not commute. “Jointly through a linear combination before exponentiation” would be more precise.","section":"Secs. II B 4 and III E 1"},{"comment":"Please specify how the fixed per-class subsamples are drawn, whether the same subsamples and fold assignments are used across all models, and how many independent optimization runs are used at each progressive-growth step for n>1. A code/data availability statement with seeds would also improve reproducibility.","section":"Sec. III B"},{"comment":"Several qualitative phrases such as “saturates,” “non-monotonic,” and “marginal” would be easier to assess if the captions or tables reported the numerical paired differences and uncertainties at the relevant values of n and p.","section":"Figs. 4–7 and Appendix B"},{"comment":"Typographical issues include “97.69±00.33,” “qutrit NQK’s,” and the missing period in “Fig 15.” There is also a duplicated comma in the reference to “Sec. II A 1,,” in Sec. III B.","section":"Table VI and Appendix B"},{"comment":"Refs. [11] and [50] appear to cite the same Nature Physics paper by Liu, Arunachalam, and Temme. Please consolidate duplicate references and check the reference list for similar redundancies.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The study is broad and the results are plausible, but I would want the model-selection protocol resolved before acceptance. If SVM hyperparameters were selected using the outer test folds, the main quantitative tables would need to be regenerated. The readout-capacity control is also important for interpreting the central QNN-to-NQK lift, although it should be feasible within the manuscript’s existing simulation framework."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a careful empirical extension of the authors’ own neural quantum kernels to qutrits, with the real value in the design ablations (p, n, 1-to-n vs n-to-n, and especially SU(3) parameterization), not in a claim of quantum advantage.\n\nWhat is actually new is the qutrit instantiation plus the systematic sweeps. The NQK idea, progressive growth, and 1-to-n / n-to-n constructions are from their prior qubit papers; single-qudit re-uploading and qutrit classifiers already exist in the cited literature. Still, they do the work: four benchmarks, binary and three-class, matched QNN baselines, classical RBF-SVM, 5-fold CV with error bars, cost-function sensitivity, and a clean Geometric vs Euler vs Givens comparison that actually moves both optimization diagnostics and accuracy. Claims are hedged the way they should be—gains saturate, are dataset-dependent, and depend on parameterization. Tables III–V are useful.\n\nSoft spots, in proportion. Everything is noiseless simulation, L fixed at 6, n-to-n only to n=4, heavy restarts, PCA to [-1,1]. They say so in the Discussion; that caps how far you can push the results, it does not sink the design study. The stress-test point is fair: the QNN baseline uses a fixed threshold/argmax on one qutrit while the NQK gets a grid-searched SVM on the full embedding, so part of the “NQK beats QNN” lift is readout capacity. The paper is explicit about that mechanism in Sec. II A. What is missing is a matched classical readout on the frozen embedding (e.g. SVM on outcome probabilities). That would tighten the central claim; it does not erase the parameterization and scaling results inside the kernel models, where the comparison is cleaner. Classical RBF remains competitive throughout—again, they report it.\n\nMath and citations look fine for this genre; circularity is low. No code is a practical annoyance for a methods paper.\n\nWho it is for: people already building qudit or kernel QML who want concrete SU(3) and scaling choices under ideal simulation. Not a must-read outside that lane. I would send it to peer review; it is serious, proportionate work. Engage if you care about qudit embeddings; skim the ablations and Table V if you only need the takeaways.","headline":"Solid, honest design study extending NQKs to qutrits; useful ablations, incremental novelty, and a real but overstated QNN-vs-kernel comparison caveat.","tokens_in":30144,"tokens_out":624,"would_cite":true,"duration_ms":21798,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Qutrit neural quantum kernels beat matched quantum neural nets on binary and three-class tasks, and the SU(3) parameterization choice strongly shapes both training and accuracy.","keywords":["neural quantum kernels","qutrits","quantum neural networks","data re-uploading","SU(3) parameterization","quantum kernel methods","multiclass classification","qudit machine learning"],"falsifier":"Rerun the same Fashion-MNIST and Covertype binary and three-class protocols with realistic noise and finite-shot overlap estimates: if the 1-to-4 and 4-to-4 NQKs no longer beat the matched QNN baselines, or if the geometric SU(3) edge disappears, the central empirical claim fails to transfer.","tokens_in":29782,"feed_emoji":"⚛️","tokens_out":1116,"duration_ms":23362,"temperature":0.7,"pith_summary":"This paper extends neural quantum kernels from qubits to qutrits: a quantum neural network is pretrained on the classification task, then frozen and reused as a task-adapted embedding that defines a quantum kernel for a classical support-vector classifier. The authors argue that three-level systems give richer local embeddings (via SU(3) unitaries with up to eight parameters) and a natural three-class readout, and they systematically vary feature budget, number of qutrits, kernel construction (1-to-n versus n-to-n), and three different SU(3) parameterizations. Across Fashion-MNIST, HAR, MAGIC, and Covertype, the kernel models improve on the corresponding QNN baselines in nearly every setting, can gain from more features and more qutrits until gains saturate, and remain competitive with a classical RBF-SVM. An ablation shows the geometric (Lie-algebra exponential) parameterization trains more favorably and usually outperforms Euler-angle and Givens forms. A sympathetic reader cares because the work treats qudit models as designable systems whose building blocks—not only scale—decide whether the quantum embedding helps.","feed_headline":"Qutrit kernels beat matched quantum nets on four datasets","feed_subtitle":"How you write SU(3) gates and how many features you encode both change accuracy and training","key_machinery":"Neural quantum kernels (NQKs): pretrain a data-reuploading qutrit QNN (optionally grown progressively), freeze its circuit as the embedding, and form an embedding quantum kernel from state overlaps; two lifts are used—1-to-n (replicate a trained single-qutrit map with fixed entanglement) and n-to-n (train the full multi-qutrit embedding).","core_discovery":"On four standard benchmarks, for both binary and three-class problems, qutrit neural quantum kernels improve over the matched pretrained QNN classifiers in nearly all configurations tested, and performance can rise with encoded feature count and with system size, though gains often saturate and depend on dataset and on how SU(3) unitaries are parameterized; the unitary parameterization itself materially changes optimization behavior and final accuracy.","pith_inferences":["If noise and shot noise erase the kernel lift first on n-to-n (deeper joint training) rather than 1-to-n, hardware-era designs may prefer replicated single-qutrit maps with fixed entanglers.","The geometric parameterization’s more symmetric generator coupling may explain its smoother feature scaling; testing whether random or learned feature-to-coordinate assignments close the gap for Euler and Givens forms would isolate inductive bias from expressivity.","Extending the same NQK pipeline to d>3 would test whether extra local generators buy multiclass headroom or mainly increase barren-plateau and calibration cost.","Because kernel construction reduced sensitivity to the pretraining loss in the three-class ablations, hybrid “train embedding, convex readout” pipelines may be a general stabilizer for qudit classifiers beyond this architecture."],"forward_implications":["Task-adapted qutrit embeddings used as fixed kernels are a practical alternative to end-to-end variational readout at matched resources.","Feature budget and register size are useful but saturating knobs; more is not automatically better once p and n are moderate.","SU(3) parameterization must be treated as a first-class design choice, not an implementation detail, because it changes both trainability and accuracy.","1-to-n and n-to-n constructions can reach similar accuracy at n=4 despite different trainable freedom, so the simpler replicated map remains competitive at moderate size.","Qutrit NQKs can stay competitive with classical RBF-SVMs on the same splits while offering a direct three-class computational-basis interface."],"fun_headline_variants":["Qutrit neural kernels top matched QNNs on four benchmarks","SU(3) parameterization shifts qutrit NQK accuracy and training","Feature count and qutrit count can lift NQKs, gains often saturate","Qutrit NQKs beat pretrained QNNs in nearly all tested settings","Qutrit kernels improve on QNNs for binary and three-class tasks"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"All reported gains come from ideal noiseless simulations with fixed circuit depth, nearest-neighbour entanglement, heavy random restarts, and PCA-compressed features, so the claimed QNN-to-kernel lift may not survive noise, finite shots, or different preprocessing.","fun_headline_variants_meta":{"raw":{"variants":["Qutrit neural kernels top matched QNNs on four benchmarks","SU(3) parameterization shifts qutrit NQK accuracy and training","Feature count and qutrit count can lift NQKs, gains often saturate","Qutrit NQKs beat pretrained QNNs in nearly all tested settings","Qutrit kernels improve on QNNs for binary and three-class tasks"]},"model":"grok-4.5","effort":"low","cost_usd":0.003135,"raw_usage":{"total_tokens":1151,"prompt_tokens":834,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":31348000,"prompt_tokens_details":{"text_tokens":834,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":230,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":834,"tokens_out":87,"duration_ms":5822,"temperature":1.0,"reasoning_tokens":230,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T15:58:25.224747+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Rerun the same Fashion-MNIST and Covertype binary and three-class protocols with realistic noise and finite-shot overlap estimates: if the 1-to-4 and 4-to-4 NQKs no longer beat the matched QNN baselines, or if the geometric SU(3) edge disappears, the central empirical claim fails to transfer.","supporting_citations":[],"review_version":1}