{"id":"bd234696-e2c1-46c8-b1d9-acce0bb1b1b5","arxiv_id":"2505.01735","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid QSNN and QLSTM model, trained in three phases, is claimed to beat classical and quantum baselines on imbalanced credit card fraud data, but the benchmark lacks matched controls and statistical rigor.","lead":"This preprint combines quantum spiking neural networks with quantum long short-term memory into a two-stage 'brain-inspired' model and tests it on credit card fraud detection. The authors report it outperforms classical and quantum baselines with fewer parameters, data points, and training iterations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority of QSNN-QLSTM rests on an uncontrolled comparison: different data sizes, optimizers, batch sizes, and iteration counts per model, with no significance tests, so the reported edge may be an artifact of tuning.","rationale":"The reader's verdict is REJECT with high correctness risk, and this stress-test agrees. The weakest assumption is indeed the meaningfulness of the benchmark comparisons, and it is load-bearing: every component of the claimed advantage (superior F1/AUC/recall, fewer parameters, less training data, fewer iterations) is entangled with per-model choices made experimentally to achieve the best performance for each model. Even if the QSNN-QLSTM architecture were exactly as described, the empirical evidence does not establish that it outperforms, because the comparison is uncontrolled. A secondary internal inconsistency in the parameter count reinforces the need for code release: Section 4.2 includes a 10x2 linear layer in the combined model while the QLSTM input is 10-dimensional, which is difficult to reconcile with a split QSNN. This does not replace the protocol concern but adds a reproducibility issue that a controlled re-benchmark with released artifacts would settle. Therefore the verdict should remain REJECT (UNCHANGED), and the proposed controlled re-benchmark would determine whether the concern actually lands.","tokens_in":15640,"tokens_out":7100,"duration_ms":69854,"concrete_test":"Release code and data, then re-run a controlled benchmark: fix a single train/test split (or 5-fold cross-validation), use identical numbers of fraud and non-fraud training samples for every model, give each model a fixed hyperparameter budget tuned only on the validation fold, and report per-run predictions with a paired significance test (e.g., McNemar on predictions; bootstrap or Wilcoxon on F1 and AUC). If QSNN-QLSTM no longer significantly beats the best classical baseline under matched conditions, the headline superiority claim fails. Independently, verify that the QSNN-QLSTM computation graph matches the Section 4.2 parameter count, specifically whether the 10x2 QSNN output layer is present and, if the QLSTM input dimension is 10, how the QSNN output is converted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sections 4.3 and 5—that QSNN-QLSTM consistently surpasses all compared models in F1, AUC, and recall—is not supported by the experimental design. Section 4.2 states that hyperparameters, optimizers, and iteration counts were 'conducted experimentally through multiple trials to achieve the best performance for each model.' Classical models were trained on 5,390 samples (390 fraud + 5,000 non-fraud) while quantum models used 1,390 samples (390 fraud + 1,000 non-fraud), with different optimizers (SGD, Adam, RMSprop), batch sizes (64–256), and iteration counts (40–700). No statistical test, confidence interval, or paired comparison is reported; results are 10-run boxplots only. Because the test set contains only 101 fraud positives, the observed F1 and recall gaps between high-scoring models could be produced by a few changed predictions and may lie within sampling noise. Moreover, no code or data are released, so the reported 40-iteration QSNN-QLSTM run cannot be independently checked. A secondary internal inconsistency reinforces this: the Section 4.2 parameter list includes a 10x2 linear layer in the QSNN-QLSTM graph, yet the QLSTM input is described as 10-dimensional, which is hard to reconcile with a 'split' QSNN and raises questions about whether the evaluated graph matches the paper's description. The superiority claim therefore does not follow from the evidence as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid quantum spiking neural network (QSNN) combined with a quantum long short-term memory (QLSTM) model for credit card fraud detection, and compares it with classical ANN, LSTM, SNN and quantum QNN, QLSTM, QSNN baselines on the standard Kaggle dataset. The authors report 10-run F1, AUC, recall, and precision results, and claim that QSNN-QLSTM consistently outperforms the other models while using fewer parameters, less training data, and fewer iterations. The paper also describes a three-phase training pipeline (pre-training QSNN, single-pass QLSTM training, then joint co-training) and discusses the limitations of quantum simulators.","tokens_in":15954,"tokens_out":5915,"duration_ms":50617,"significance":"The conceptual combination of a spiking quantum front-end and a recurrent quantum memory is a plausible research direction, and the paper's emphasis on data efficiency for quantum models is timely. However, the central empirical claim is not established because the comparison is uncontrolled: training set sizes, optimizers, batch sizes, iteration counts, and validation splits differ across models, and hyperparameters were tuned per model without a stated protocol or any significance testing. If the claim were substantiated with a matched experimental design and statistical analysis, the architecture would be a useful contribution to quantum machine learning and anomaly detection; as presented, the evidence is insufficient. The paper also provides no code or data release, which limits reproducibility.","major_comments":[{"comment":"The comparison is uncontrolled. Quantum models were trained on 1,390 samples (390 fraud + 1,000 non-fraud) whereas classical models used 5,390 samples (390 fraud + 5,000 non-fraud), with different optimizers (SGD, Adam, RMSprop), batch sizes (64-256), iteration counts (40-700), and validation splits of different absolute sizes. The paper states that hyperparameters, optimizers, and iteration counts were 'conducted experimentally through multiple trials to achieve the best performance for each model,' but it never specifies the data used in those trials or a protocol that prevents the test set from influencing the choices. The reported advantage of QSNN-QLSTM may therefore be an artifact of per-model tuning rather than an architectural property.","section":"§4.2, Tables 1-2"},{"comment":"Results are reported only as 10-run boxplots, without statistical significance tests, confidence intervals, or paired comparisons. Because the test set contains only 101 fraud positives, the observed differences in F1, recall, and AUC among the top models could be produced by a handful of changed predictions and may lie within sampling noise. The claim that QSNN-QLSTM 'consistently surpasses other models' is also contradicted by Fig. 14, in which SNN attains the highest precision and QSNN-QLSTM is a close second.","section":"§4.3, Figs. 11-14"},{"comment":"The architecture description for QSNN-QLSTM is internally inconsistent: the parameter list includes a linear layer of size 10×2+2 after the QLIF1 cell, yet the following QLSTM preprocessing layer is given as 10×20+20, implying a 10-dimensional input to the QLSTM. It is unclear whether the evaluated model matches the described graph, and the manuscript provides no code to resolve the ambiguity.","section":"§4.2, QSNN-QLSTM parameter list"},{"comment":"The claim that quantum models 'require fewer iterations' is tautological given the experimental design, because the iteration count for each model was intentionally set to the point at which that model converged. With per-model stopping criteria, the number of iterations is not an independent advantage, and no runtimes or convergence thresholds are reported to make the comparison meaningful.","section":"§4.2–§4.3"}],"minor_comments":[{"comment":"The manuscript contains numerous typographical errors and misspellings, e.g., 'Annomaly Detection' in the keywords, 'Presicion' in the Fig. 14 caption, 'clasification', 'hiperbolic', and 'subsenquently'; these should be corrected.","section":"Keywords and Figure 14 caption"},{"comment":"Equation (14) presents the chain rule with brace notation that is garbled in the PDF; the intended expression should be reformatted so that the surrogate gradient substitution is readable.","section":"§2.1, Eq. (14)"},{"comment":"The dataset is described as having 30 features including 28 PCA-transformed features and the transaction amount, which sums to 29; the standard Kaggle credit card dataset also includes a 'Time' feature. Please clarify the exact feature set used.","section":"§4.1"},{"comment":"References [11] and [37] appear to duplicate the same source (the snnTorch / training SNNs paper) in different venues; these should be unified.","section":"References"},{"comment":"Figure 15 uses different y-axis scales across panels, so the visual claim of 'consistent progression' and 'variance decreases' is not supported; provide comparable axes or quantitative convergence statistics.","section":"Fig. 15"}],"recommendation":"reject","confidential_remarks":"The central empirical claim is not supported by the experimental design. The issues are not local presentation problems but concern the validity of the benchmark itself: unequal training data sizes, per-model hyperparameter tuning without a stated protocol, and absence of significance tests. The internal inconsistency in the QSNN-QLSTM architecture description further undermines confidence in what was actually evaluated. I would be open to a resubmission that addresses these issues with a controlled comparison, pre-registered hyperparameters, and statistical analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper puts a new three-phase QSNN-QLSTM pipeline on a fraud-detection task and claims it consistently beats classical and other quantum baselines with fewer parameters, fewer data, and fewer iterations. The architecture combination is genuinely not in the cited literature, and the authors are transparent about parameter counts and training settings. But the central comparison is not credible as reported. Section 4.2 shows that each model was tuned separately with different training set sizes (1,000 vs. 5,000 non-fraud samples), different optimizers, batch sizes, and iteration counts, all selected experimentally to maximize performance for that model. Section 4.3 reports only boxplots over ten runs, with no significance tests, and the test set contains just 101 fraud positives, so a few prediction flips could move the F1 and recall numbers. No code or data are released, so the 40-iteration QSNN-QLSTM result cannot be independently checked. There is also a small internal inconsistency: the parameter list for QSNN-QLSTM includes a 10x2 linear layer that is hard to reconcile with the described 10-dimensional QLSTM input. That is minor compared to the experimental protocol, but it adds to the difficulty of taking the numbers at face value.\n\nWhat the paper does well: the LIF derivation and the LSTM background are competent, the parameter-counting is explicit, and the idea that a hybrid quantum spiking plus quantum recurrent model might be data-hungry in the low-data regime is worth a look. The three-phase training (pre-train QSNN, single-pass QLSTM, co-train) is clearly described, even if the neuroscience framing is decorative rather than explanatory.\n\nThe soft spots are load-bearing. The claimed superiority is fitted, not predicted: hyperparameters were chosen after multiple trials for each model, so the comparison is closer to a model-selection exercise than a controlled benchmark. That alone is enough to reject the paper in its current form. The architecture might still be useful, but the evidence would need a proper re-benchmark with matched data and iteration budgets, paired significance tests, and ideally released code.\n\nWho is this for? Quantum machine learning folks who want a concrete example of a hybrid quantum spike-and-recurrence design on tabular data. It belongs in a reading group as a case study in benchmarking pitfalls, not as an accepted result. A serious editor could send it to peer review because the architectural idea deserves scrutiny, but acceptance would require a substantially revised experimental section and code release. As it stands, this is a reject with a specific path to resubmission.","headline":"Plausible hybrid architecture, but the headline superiority claim is not supported by an uncontrolled benchmark with no significance tests and no code.","tokens_in":16502,"tokens_out":2009,"would_cite":false,"duration_ms":21572,"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":"A two-stage hybrid of quantum spiking and quantum recurrent networks outperforms six classical and quantum baselines on credit-card fraud detection.","keywords":["Anomaly Detection","Quantum Neural Networks","Quantum Spiking Neural Networks","Quantum Long Short-Term Memory","Brain-inspired models","Credit card fraud","Variational quantum circuits"],"falsifier":"Run the same seven models on the same credit-card fraud data under identical data budgets, optimizer families where possible, and a fixed early-stopping rule, evaluating by nested cross-validation on the training folds; if QSNN-QLSTM does not rank first in F1, AUC, or recall across the ten seeds, the paper's central superiority claim collapses.","tokens_in":15420,"feed_emoji":"🧠","tokens_out":9866,"duration_ms":86786,"temperature":0.7,"pith_summary":"The paper tries to establish that a two-stage quantum architecture—a Quantum Spiking Neural Network front end followed by a Quantum Long Short-Term Memory module—detects credit-card fraud more accurately than six classical and quantum baselines. The design is explicitly brain-inspired: the spiking stage is compared with low-level sensory processing, and the recurrent stage with memory formation and recall. The reported evidence is a seven-model comparison on 284,807 credit-card transactions, in which the hybrid model reaches the highest F1, AUC, and recall while using fewer parameters, fewer training samples, and fewer iterations than the classical models. A sympathetic reader would take the paper's central claim to be that quantum brain-inspired hybrids are the best of the compared approaches for anomaly detection on small, imbalanced datasets.","feed_headline":"Quantum spiking-plus-LSTM beats six rivals on fraud data","feed_subtitle":"Hybrid quantum memory model tops classical nets on fraud data with fewer parameters and samples.","key_machinery":"The central object is the two-stage QSNN-QLSTM pipeline. A Quantum Spiking Neural Network—a spiking network whose leaky integrate-and-fire neurons have their membrane potential initialized by a variational quantum circuit, trained with surrogate gradients that replace the non-differentiable spike decision with a smooth approximation in the backward pass—forms the first stage and is pre-trained alone. A Quantum Long Short-Term Memory module, in which the forget, input, update, and output gates are each variational quantum circuits, forms the second stage. The training protocol has three phases: pre-train the QSNN; propagate data through the frozen QSNN into a fresh QLSTM for a single pass; then co-train both with two optimizers at different learning rates. This staged protocol is what the paper argues lets the model learn new patterns without erasing prior knowledge, and the variational circuits are what make the model quantum.","core_discovery":"On the paper's own terms, the central discovery is that the novel QSNN-QLSTM model consistently surpasses the other compared models in F1 score, AUC, and recall on the credit-card fraud dataset, while using 774 trainable parameters (against 43,457 for the classical LSTM), training on 1,390 samples instead of 5,390, and converging in 40 iterations instead of 700 for the ANN. The authors attribute this to a division of labor: the QSNN stage filters noisy and infrequent events and strengthens spatiotemporally correlated signals, while the QLSTM stage captures and memorizes correlated patterns, transforming short-term into long-term storage. The paper also reports that the hybrid model is second only to the SNN in precision. In the authors' reading, this supports the broader conclusion that quantum brain-inspired models can detect patterns that classical and simpler quantum models miss.","pith_inferences":["An ablation study that removes either the QSNN or the QLSTM stage would isolate whether the gains come from the spiking front end, the recurrent memory, or simply the extra trainable parameters.","A paired statistical test over the paper's ten random seeds would show whether the reported F1 and AUC gaps exceed run-to-run variability; the paper does not report such a test.","The same two-stage design could be tried on other imbalanced tabular benchmarks with different fraud ratios and feature counts, which would reveal whether the advantage is specific to this dataset.","Because all quantum models here run on simulators, the reported margins are not yet evidence about performance on noisy quantum hardware; that is an inference, not a claim the paper makes."],"forward_implications":["If the superiority claim is right, hybrid quantum spiking-plus-recurrent models are a viable alternative to classical networks for imbalanced tabular anomaly detection.","The three-phase training recipe provides a concrete template for adding a recurrent memory stage to a pre-trained spiking network without catastrophic forgetting.","The reported data efficiency suggests quantum models could be used where labeled examples are scarce, reducing reliance on synthetic data augmentation.","The iteration count of 40 versus 700 implies that, absent simulation overhead, the hybrid model would train far faster in wall-clock time on quantum hardware.","The parameter-count ordering reported (QNN below QSNN below ANN below QSNN-QLSTM below SNN below QLSTM below LSTM) indicates that hybrid quantum models can stay lightweight despite combining two architectures."],"supporting_citations":[{"why":"Supplies the leaky integrate-and-fire neuron equations, surrogate-gradient training, and the spiking-network background used to build the QSNN.","marker":"[11]"},{"why":"Provides the spiking-neural-network software the authors extend with quantum leaky integrate-and-fire neurons.","marker":"[37]"},{"why":"Defines the LSTM gate structure that the QLSTM replaces with variational quantum circuits.","marker":"[23]"},{"why":"Outlines the variational quantum circuit workflow—encoding, ansatz, measurement, classical optimization—that every quantum model here follows.","marker":"[30]"},{"why":"Catalogs data-encoding strategies; the paper selects amplitude encoding for QSNN and QNN and angle encoding for QLSTM.","marker":"[31]"},{"why":"Establishes parameterized quantum circuits as trainable machine-learning models, the basis of the quantum gates.","marker":"[25]"},{"why":"Provides the circuit-centric quantum classifier paradigm used as a comparison model (QNN).","marker":"[2]"}],"fun_headline_variants":["Quantum spiking-LSTM hybrid beats rivals on fraud data","Hybrid quantum model tops fraud detection with 1/56 parameters","QSNN-QLSTM wins fraud test, uses fewer samples and params","Quantum brain-inspired net surpasses classical on fraud"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the experimental settings—quantum models trained on 1,390 samples and classical on 5,390, with different optimizers, batch sizes, and iteration counts—were chosen fairly and not tuned to the test set, so the reported wins reflect the architecture rather than the configuration.","fun_headline_variants_meta":{"raw":{"variants":["Quantum spiking-LSTM hybrid beats rivals on fraud data","Hybrid quantum model tops fraud detection with 1/56 parameters","QSNN-QLSTM wins fraud test, uses fewer samples and params","Quantum brain-inspired net surpasses classical on fraud"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1511,"prompt_tokens":892,"completion_tokens":619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":548}},"tokens_in":508,"tokens_out":619,"duration_ms":6674,"temperature":1.0,"reasoning_tokens":548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:11:56.461849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same seven models on the same credit-card fraud data under identical data budgets, optimizer families where possible, and a fixed early-stopping rule, evaluating by nested cross-validation on the training folds; if QSNN-QLSTM does not rank first in F1, AUC, or recall across the ten seeds, the paper's central superiority claim collapses.","supporting_citations":[{"cited_title":"Eshraghian, Max Ward, Emre Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D","cited_arxiv_id":null,"evidence_quote":"Supplies the leaky integrate-and-fire neuron equations, surrogate-gradient training, and the spiking-network background used to build the QSNN."},{"cited_title":"Training spiking neural networks using lessons from deep learning.Proceedings of the IEEE, 111(9):1016–1054, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the spiking-neural-network software the authors extend with quantum leaky integrate-and-fire neurons."},{"cited_title":"Lstm can solve hard long time lag problems.Advances in Neural Information Processing Systems, 9:473–479, 1996","cited_arxiv_id":null,"evidence_quote":"Defines the LSTM gate structure that the QLSTM replaces with variational quantum circuits."},{"cited_title":"Elsevier, 2014","cited_arxiv_id":null,"evidence_quote":"Outlines the variational quantum circuit workflow—encoding, ansatz, measurement, classical optimization—that every quantum model here follows."},{"cited_title":"Expanding data encoding patterns for quantum algorithms","cited_arxiv_id":null,"evidence_quote":"Catalogs data-encoding strategies; the paper selects amplitude encoding for QSNN and QNN and angle encoding for QLSTM."},{"cited_title":"A variational algorithm for quantum neural networks","cited_arxiv_id":null,"evidence_quote":"Establishes parameterized quantum circuits as trainable machine-learning models, the basis of the quantum gates."},{"cited_title":"Svore, and Nathan Wiebe","cited_arxiv_id":null,"evidence_quote":"Provides the circuit-centric quantum classifier paradigm used as a comparison model (QNN)."}],"review_version":1}