{"id":"69cc6655-781e-419f-9f65-b1540afa0f1e","arxiv_id":"1908.04480","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"QAML-Z, an iterative zooming and classifier-augmentation upgrade to QAML, improves Higgs classification AUROC over QAML and matches a deep neural network at small training set sizes, but a classical simulated-annealing variant with the same ingredients performs equally.","lead":"The authors extend an earlier quantum-annealing machine learning algorithm, QAML, by iteratively zooming in on real-valued classifier weights and adding threshold-shifted weak classifiers. They test the resulting QAML-Z method on Higgs boson classification with a D-Wave 2X annealer, where it beats QAML and approaches deep neural network accuracy on small datasets, though a classical simulated-annealing version matches it.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9) omits cross-offset couplings between augmented weak classifiers, so the optimized Hamiltonian does not match the claimed QAML-Z squared-error objective; this can be settled by code inspection.","rationale":"The reader's conditional verdict rests primarily on the 5% pruning of the D-Wave Hamiltonian. I agree that pruning is a real limitation, but my reading points to a more fundamental issue: even the unpruned Hamiltonian in Eq. (9) ignores cross-offset couplings among the augmented weak classifiers. The abstract's central claim depends on the iterative zooming objective being the squared-error loss of the augmented ensemble, and Eq. (9) does not establish that. This is an internal-consistency concern, not a disagreement with consensus, and it is directly testable because the paper provides a public repository. The empirical comparison is otherwise reasonable: baselines are optimized, error bars are reported, and Figure 4 honestly shows that the classical SAE-Z reproduces QAML-Z, which already weakens the quantum-specific interpretation. My concern sharpens the conditions under which the paper should be accepted, but it does not by itself overturn the reported empirical curves, so the appropriate verdict remains conditional.","tokens_in":776,"tokens_out":817,"duration_ms":147762,"concrete_test":"Clone the qaml-z repository and inspect the function that builds the QUBO/Ising Hamiltonian from the augmented classifiers. Check whether quadratic terms are generated for pairs with different offset indices l != l', or only within the same offset. Independently expand sum_tau (y_tau - sum_{i,l} mu_il c_il(x_tau))^2 and compare the resulting pair coefficients with Eq. (9). If the code includes cross-offset terms, rerun the Figure 2 experiment with the full Hamiltonian, e.g. using SAE-Z, and see whether the AUROC shifts by more than the error bars. If they are absent from both paper and code, the reported results are for a different algorithmic objective and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the Hamiltonian minimized on the annealer is the correct Ising encoding of the QAML-Z objective for the augmented classifier set. Eq. (9) does not appear to satisfy this. With the augmented strong classifier R = sum_{i,l} mu_il c_il, the squared-error expansion contains pairwise terms C_{(i,l),(j,l')} mu_il mu_jl' for l != l'. Eq. (9), with C_ijl defined in Eq. (10), only couples classifiers within the same offset l; all cross-offset couplings are absent. The augmented classifiers c_il = sgn(h_i + delta_l)/N are sign functions of the same h_i with shifted thresholds, so they are strongly correlated across offsets; these omitted terms are not negligible. The iterative minimization therefore solves a different problem from the one derived in Section II B. This is separate from, and more fundamental than, the 5% pruning of couplings in Section III A, which is an additional hardware approximation. If the supplied code follows Eq. (9), the reported AUROC curves are for an algorithm other than the one claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QAML-Z, an iterative extension of the QAML algorithm in which binary weak-classifier weights are replaced by continuous weights via a zooming search, and the weak-classifier set is augmented by threshold offsets. The algorithm is implemented on a D-Wave 2X quantum annealer and benchmarked on Higgs-boson classification, reporting AUROC values that match a deep neural network at small training-set sizes and reduce the QAML-to-DNN gap by about 47% at large training-set sizes. The authors also provide simulated-annealing and logistic-regression baselines, report error bars, and promise public source code and data.","tokens_in":11940,"tokens_out":10746,"duration_ms":109560,"significance":"If the reported AUROC results are trustworthy, the paper makes a useful contribution by showing that iterative zooming plus classifier augmentation materially improves QAML and brings quantum-annealing-based classification closer to classical benchmarks in a small-data regime. The manuscript has real strengths: it uses an independent simulated benchmark, compares against strong classical baselines whose hyperparameters were optimized, reports error bars, and promises reproducible code and data. It is also honest in reporting that simulated annealing with excited states (SAE-Z) reproduces QAML-Z performance to one standard deviation, which limits any claim of a quantum-specific advantage. The main caveats are that the implemented Hamiltonian deviates from the derived objective, the hardware results use a heavily pruned coupling graph, and the excited-state selection procedure uses validation-set labels in a way that may bias the comparison.","major_comments":[{"comment":"Equation (9) does not correspond to the squared-error objective for the augmented classifier set. Expanding the objective with the augmented strong classifier R = sum_{i,l} mu_il c_il yields pairwise couplings C_{(i,l),(j,l')} = sum_tau c_il(x_tau)c_jl'(x_tau) for l != l', as well as linear cross terms involving mu_jl' C_{(i,l),(j,l')}. Equation (9) contains only same-offset couplings C_ijl and only for j > i, so the minimization solves a different optimization problem. Because c_il = sgn(h_i + delta_l)/N are highly correlated across offsets, the omitted cross-offset terms are not negligible. The authors should either derive the full Ising mapping for the augmented set or explicitly state and justify the approximation, and the supplied code should be checked against the derivation.","section":"Section II B, Eq. (9)"},{"comment":"The D-Wave implementation retains only the largest 5% of Ising couplings, and every reported QAML-Z result on the quantum annealer is therefore obtained from a heavily truncated Hamiltonian. This means the comparison with DNN and XGBoost in Figure 2 is not a clean test of the QAML-Z objective defined in Section II B. The authors should quantify the effect of the pruning, for example by running SA-Z on both the full and the pruned Hamiltonian, or by reporting classical full-Hamiltonian QAML-Z results for a smaller problem size.","section":"Section III A"},{"comment":"The final classifier is not a single fixed ensemble but a per-threshold supremum over excited states selected on a validation set. This validation-set selection procedure is substantially more flexible than the validation-based hyperparameter tuning used for the DNN and XGBoost baselines, so the AUROC comparison in Figure 2 may be biased in favor of QAML-Z. The authors should report the number of excited states used, the selection criteria in detail, and compare with a fixed ensemble rule such as averaging over the selected excited states, to separate the benefit of the zooming algorithm from the benefit of validation-set model selection.","section":"Section II B and Section III A"}],"minor_comments":[{"comment":"Please state the number N of weak classifiers and the total size of the augmented classifier set; with A = 3, the augmented set contains 7N classifiers, and it is unclear how the problem is embedded into the 33 fully connected logical qubits available on the D-Wave 2X.","section":"Section III A"},{"comment":"The regularization schedules pf(t), qf(t) and the excited-state thresholds d and ne are described only as being specified in the supplementary code; these schedules should be given in the text or an appendix because they affect the reported results.","section":"Section II B"},{"comment":"Because SAE-Z achieves statistically equivalent AUROC to QAML-Z, the abstract and conclusion should qualify the role of quantum annealing more carefully; the improvement over QAML appears to be driven substantially by zooming and by excited-state selection rather than by the quantum device.","section":"Section III C, Figure 4"},{"comment":"The derivation from Eq. (7a) to Eq. (7b) should explicitly show how self-interaction terms and constant terms are handled, and the spin encoding should be stated unambiguously for the zooming iterations.","section":"Section II B, Eqs. (7a)-(7b)"},{"comment":"The promised source code and data should include version or commit identifiers and persistent DOIs so that the reported experiments can be reproduced exactly.","section":"References [61,62]"},{"comment":"The numerical AUROC values underlying Figure 2 should be given in a table with confidence intervals, since the small-training-set region appears to have large error bars and the claim of matching DNN performance should be supported by explicit numbers.","section":"Section III B"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the central performance claim is plausible and the experimental setup is potentially reproducible, but the derivation mismatch in Eq. (9) is the main technical risk. I recommend asking the authors to reconcile the code and the derivation, quantify the effect of the 5% coupling pruning, and address the validation-set selection bias in the excited-state ensemble. The paper is within scope for a quantum machine learning venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a real algorithmic contribution with honest, reproducible benchmarks, but the derivation of the augmented Hamiltonian does not match the stated squared-error objective, and the authors' own classical baseline undercuts the quantum-specific claim. It deserves a serious referee, not a desk reject.\n\nWhat's new: the zooming iteration—converting QAML's binary on/off weights into continuous weights via a binary search over the energy landscape—and the threshold-augmented weak classifiers are both genuine additions over the original QAML. The code and data links are provided, and the classical baselines (optimized DNN, XGBoost, logistic regression, and simulated annealing variants) are appropriate. The main empirical result, roughly a 47% reduction in the AUROC gap to a DNN at large training sizes and parity at small sizes, is internally consistent with the figures and given with error bars.\n\nNow the soft spots, in increasing order of severity. Minor: the derivation of Eq. (7b) is compact and drops constants, but that part checks out. More serious: the D-Wave implementation prunes 95% of the Ising couplings, so the hardware runs on an approximation of the model; and excited states are selected by maximizing validation AUROC, which is a selection bias the paper acknowledges but does not quantify. Most serious: Eq. (9) only couples augmented classifiers within the same threshold offset. The squared-error objective for R = sum mu_il c_il includes cross-offset pair terms (i,l)-(j,l') for l != l', and those are absent from Eq. (9). So the Hamiltonian actually minimized on the annealer is not the Ising encoding of the QAML-Z objective derived in Sec. II B. This is a genuine mathematical gap, separate from the pruning. The public code can settle it quickly: if the code follows Eq. (9), the reported AUROC curves are for a different algorithm than the one claimed. The paper should either include the missing terms or explicitly state that cross-offset interactions are dropped as an approximation. Finally, the abstract and conclusion imply quantum annealing drives the improvement, but their own Figure 4 shows SAE-Z matches QAML-Z within one sigma. That claim should be reframed.\n\nWho this is for: researchers in quantum annealing for machine learning, especially near-term quantum classifiers. They will want the zooming idea and the benchmark methodology, but they need to fix or acknowledge the Hamiltonian discrepancy and tone down the quantum claim. I'd engage with it, ask for a code check of Eq. (9) and a revised discussion of the SAE-Z comparison, and then likely accept it as a solid algorithmic contribution.","headline":"QAML-Z is a genuine algorithmic extension with reproducible benchmarks, but the augmented Hamiltonian omits cross-offset couplings and the quantum advantage disappears against SAE-Z.","tokens_in":12578,"tokens_out":2652,"would_cite":false,"duration_ms":27773,"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":"QAML-Z, a zooming variant of quantum-annealing machine learning, matches optimized classical deep neural networks on small Higgs training sets and almost halves the large-set AUROC gap to the DNN.","keywords":["quantum annealing","machine learning","QAML-Z","Ising model","Higgs boson classification","weak classifiers","continuous optimization","ROC curve"],"falsifier":"Run QAML-Z on the exact, unpruned Ising model for the same Higgs data (in simulation, with all cross-terms and hundreds of logical qubits) and compare AUROC against the 5%-pruned version used on the annealer. If the full-model AUROC no longer matches the deep network at small training-set sizes, or no longer cuts the large-size gap by roughly half, then the claimed improvement is an artifact of the pruning rather than of the zooming algorithm.","tokens_in":11487,"feed_emoji":"⚛️","tokens_out":11068,"duration_ms":98388,"temperature":0.7,"pith_summary":"QAML-Z is a proposed upgrade to quantum-annealing machine learning that replaces the binary on/off weights of weak classifiers with continuous weights found by repeatedly annealing inside a shrinking region of the energy landscape. On a Higgs-boson classification benchmark, the authors report that QAML-Z matches an optimized deep neural network at the smallest training-set sizes and reduces the area-under-the-ROC-curve (AUROC) gap between the earlier QAML algorithm and the DNN by 47% at the largest training-set size. The paper also shows that the zooming procedure drives the Ising energy down monotonically on the test set, and that a classical simulated-annealing version with excited states reproduces the quantum results to one standard deviation. The authors present these results as evidence that near-term quantum annealers can be competitive with classical machine learning in small-data regimes and that discrete annealers can be applied to continuous optimization.","feed_headline":"Zooming quantum annealer nearly halves gap to deep nets on Higgs data","feed_subtitle":"Continuous-weight annealing matches deep nets on small training sets and cuts the large-set AUROC gap by 47%.","key_machinery":"The load-bearing mechanism is the zooming Hamiltonian. Starting from the QAML objective, each weak classifier's contribution is shifted and scaled by the current weight $\\mu_i(t)$ and the search breadth $\\sigma(t)$, giving $H(t)=\\sum_i(-C_i+\\sum_j\\mu_j(t)C_{ij})\\sigma(t)s_i+\\sum_{i<j}C_{ij}\\sigma^2(t)s_i s_j$, up to constants. This turns one discrete optimization into a sequence of binary searches over the real-valued weight vector, with spin-up/spin-down outcomes moving each weight by an exponentially decaying step. The threshold-augmented classifier set provides the extra weak classifiers that make the continuous ensemble expressive.","core_discovery":"QAML-Z extends the QAML algorithm in two steps: it augments the weak-classifier set by threshold shifts $c_{il}(x_\\tau) = \\operatorname{sgn}(h_i(x_\\tau)+\\delta_l)/N$, and it turns each binary weight into a continuous weight by iterating the substitution $s_i c_i(x_\\tau) \\to \\sigma(t)s_i c_i(x_\\tau)+\\mu_i(t)c_i(x_\\tau)$ with shrinking breadth $\\sigma(t)=b^t$. Each anneal updates $\\mu_i(t+1)=\\mu_i(t)+s_i\\sigma(t+1)$, so the final classifier is a real-weighted ensemble. On the annealer, the authors report AUROC gains over QAML at every training-set size, parity with the DNN at small sizes, and a 47% reduction of the QAML-DNN gap at the largest size. Because only 33 fully connected logical qubits were available, the implemented Ising Hamiltonian kept only the largest 5% of cross-terms, and the authors also show that simulated annealing with zooming and excited states matches QAML-Z, which locates the algorithm's performance gain in the zooming procedure rather than to quantum dynamics specifically.","pith_inferences":["My reading: the 5% cross-term pruning means the experimental numbers characterize a truncated Hamiltonian, so a natural check is to simulate the full QAML-Z objective on the same data and compare AUROC curves before attributing the gain to the algorithm.","My reading: zooming is a generic way to convert binary Ising optimization into continuous optimization, so the same shrinking-sigma schedule should transfer to other annealer tasks such as feature selection or regularized regression, not just classification.","My reading: since logistic regression on the augmented classifiers already matches the DNN, much of the power may come from the physics-based kinematic classifiers plus threshold augmentation; an extension would test whether zooming still helps when weak classifiers are generic stumps or random features.","My reading: the equivalence with SAE-Z implies that any future quantum-speedup or quantum-advantage claim for QAML-Z must be benchmarked against a classical counterpart with excited states, not against plain simulated annealing."],"forward_implications":["QAML-Z improves AUROC over the original QAML at every tested training-set size on the Higgs benchmark, so the zooming-plus-augmentation recipe is a working upgrade for this problem.","At training sets of order 1,000 examples, QAML-Z reaches deep-network-level AUROC, indicating that annealer-based classifiers can be competitive in small-data, online-learning settings.","At the largest training set, the AUROC gap between QAML and the deep network shrinks by 47%, meaning the quantum-classical performance margin is roughly halved.","The Ising energy on the test set decreases monotonically across zooming iterations, which the paper reads as evidence against overfitting during the repeated anneals.","Because simulated annealing with zooming and excited states matches QAML-Z within one standard deviation, the reported gain does not by itself demonstrate a quantum advantage over classical annealing."],"supporting_citations":[{"why":"Supplies the original QAML algorithm and the Higgs optimization benchmark that QAML-Z extends and compares against.","marker":"[33]"},{"why":"Provides the quantum adiabatic machine-learning formulation and the excited-state/regularization ideas QAML-Z inherits.","marker":"[22]"},{"why":"Defines quantum annealing in the transverse Ising model, the theoretical basis for the anneal step and for the simulated-annealing analogue.","marker":"[17]"},{"why":"Describes the 1098-qubit annealer architecture whose 33 fully connected logical qubits motivate the 5% cross-term pruning.","marker":"[28]"},{"why":"Supplies the gradient-boosted tree baseline included in the classical benchmark comparison.","marker":"[65]"},{"why":"Provides the Higgs kinematic data set used for training and testing all classifiers.","marker":"[62]"}],"fun_headline_variants":["Zooming quantum annealer narrows deep-net gap on Higgs data by 47%","QAML-Z: zooming quantum annealer cuts Higgs AUROC gap to DNN by half","Quantum zooming matches deep nets on small Higgs sets, halves gap","Annealing with zooming: continuous weights cut deep-net gap by 47%","Zooming QAML-Z: quantum annealer beats QAML, nears DNN on Higgs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experimental result assumes that pruning the Ising Hamiltonian to the largest 5% of its cross-terms leaves the energy landscape close enough to the full QAML-Z objective that the zooming anneals still find essentially the same good classifiers; if pruning distorts which low-energy configurations are found, the reported AUROC numbers are for a different, truncated problem rather than for the algorithm being proposed.","fun_headline_variants_meta":{"raw":{"variants":["Zooming quantum annealer narrows deep-net gap on Higgs data by 47%","QAML-Z: zooming quantum annealer cuts Higgs AUROC gap to DNN by half","Quantum zooming matches deep nets on small Higgs sets, halves gap","Annealing with zooming: continuous weights cut deep-net gap by 47%","Zooming QAML-Z: quantum annealer beats QAML, nears DNN on Higgs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3346,"prompt_tokens":953,"completion_tokens":2393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2283}},"tokens_in":569,"tokens_out":2393,"duration_ms":14855,"temperature":1.0,"reasoning_tokens":2283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:41:34.604112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run QAML-Z on the exact, unpruned Ising model for the same Higgs data (in simulation, with all cross-terms and hundreds of logical qubits) and compare AUROC against the 5%-pruned version used on the annealer. If the full-model AUROC no longer matches the deep network at small training-set sizes, or no longer cuts the large-size gap by roughly half, then the claimed improvement is an artifact of the pruning rather than of the zooming algorithm.","supporting_citations":[{"cited_title":"Solving a higgs optimization problem with quantum annealing for machine learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the original QAML algorithm and the Higgs optimization benchmark that QAML-Z extends and compares against."},{"cited_title":"Quantum adia- batic machine learning,","cited_arxiv_id":null,"evidence_quote":"Provides the quantum adiabatic machine-learning formulation and the excited-state/regularization ideas QAML-Z inherits."},{"cited_title":"Quantum annealing in the transverse ising model,","cited_arxiv_id":null,"evidence_quote":"Defines quantum annealing in the transverse Ising model, the theoretical basis for the anneal step and for the simulated-annealing analogue."},{"cited_title":"Archi- tectural considerations in the design of a superconduct- ing quantum annealing processor,","cited_arxiv_id":null,"evidence_quote":"Describes the 1098-qubit annealer architecture whose 33 fully connected logical qubits motivate the 5% cross-term pruning."},{"cited_title":"Xgboost: A scal- able tree boosting system,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-boosted tree baseline included in the classical benchmark comparison."},{"cited_title":"Kinematics variables of higgs and background data,","cited_arxiv_id":null,"evidence_quote":"Provides the Higgs kinematic data set used for training and testing all classifiers."}],"review_version":1}