{"id":"83f069d6-8a15-4b5d-8620-92228666fcc9","arxiv_id":"2508.14906","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors train a variational quantum Hopfield associative memory with an autoencoder and K-Means clustering to classify MovieLens 1M users, reporting ROC 0.9795 in ideal simulation and 0.9177 under noise.","lead":"A hybrid model combining a classical autoencoder, K-Means clustering, and a variational quantum Hopfield associative memory is applied to the MovieLens 1M dataset. The paper reports high classification metrics in simulation, but the method is described too vaguely to reproduce and lacks a same-setup classical baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unclear whether k-means archetypes are fit before the train/test split, so the reported test metrics may be in-sample.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the order of k-means clustering relative to the train/test split is not specified, so the reported test metrics may not be true generalization estimates. I agree with this reading. The missing specification of the variational QHAM circuit is a serious reproducibility defect, but it does not by itself invalidate the numerical results; the leakage concern would make the reported numbers meaningless as evidence for the central claim. Therefore, the REJECT verdict is not changed by this stress-test pass. If a re-run with training-only clustering reproduces the reported metrics, the concern would be resolved, but as written the paper does not support the claim.","tokens_in":8996,"tokens_out":6599,"duration_ms":74716,"concrete_test":"Re-run the full pipeline with k-means restricted to the training users only, then assign test users to the nearest training-derived archetype, and recompute ROC, accuracy, and F1 on the test set. If those metrics drop materially from the reported values (e.g., ROC below 0.92 or accuracy below 0.84), the original numbers were inflated by using test-user information in archetype construction. Ideally, the authors should also release the exact split indices and fitting code so the check is fully reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2.1 says 'We used all the users' data for the simulation' immediately before describing the train/test split and then the k-means clustering step, but it never states whether k-means is fit on the training split only or on the full user set. If the archetype centers are computed from all users, or if the autoencoder is trained on all users before the split, then held-out test users contributed to the very patterns the model is tested against. Since the model's output is a softmax over archetype categories derived from those patterns, the 'test' labels are then defined by a clustering that saw the test users, and the reported ROC, accuracy, and F1 become in-sample fit statistics rather than generalization estimates. This directly undermines the central quantitative claim and the Section 3.4 comparison to classical baselines, because an accuracy computed on leaked labels is not comparable to a properly evaluated baseline. The text is ambiguous, but the ambiguity is itself load-bearing: as written, the paper does not rule out the in-sample interpretation, and no code or data is provided to resolve it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid recommendation system for MovieLens 1M that combines a deep autoencoder, k-means clustering of users into archetypes, and a variational quantum Hopfield associative memory (QHAM). The encoder compresses user rating vectors; k-means produces archetype labels; the QHAM is trained to store and retrieve these archetypes; a softmax layer classifies the user. The authors report ideal simulation ROC 0.9795, accuracy 0.8841, and F1 0.8786, and noisy simulation ROC 0.9177, accuracy 0.8013, and F1 0.7866 after 35 epochs, and claim these results are comparable or superior to classical collaborative filtering baselines.","tokens_in":9107,"tokens_out":6208,"duration_ms":62795,"significance":"If the results are reproducible, this would be one of the first applications of quantum associative memory to collaborative filtering at the scale of MovieLens 1M and would provide evidence that variational quantum models can tolerate realistic noise. The claimed qubit-overhead reduction by updating one random qubit is also potentially interesting. However, the manuscript currently does not supply enough detail to verify these claims: the QHAM circuit is not specified, the preprocessing order is ambiguous, and no code or statistical uncertainty is given. The conceptual contribution is interesting but not yet substantiated.","major_comments":[{"comment":"The variational QHAM circuit is not described. The text only gives the amplitude encoding of Eq. (1) and refers to Mottonen state preparation; it does not specify the ansatz, the number of qubits, the number of layers, the parameterized gates, the update rule, or the objective function used to train the QHAM. The abstract's claim of 'updating only one random targeted qubit' appears nowhere in the methods. Without this information, the reported performance in Table 1 cannot be reproduced or evaluated.","section":"§2.1.2"},{"comment":"The data-preprocessing order is ambiguous and potentially circular. The sentence 'We used all the users' data for the simulation on the ideal and noisy environments' appears immediately before the train/validation/test split description and the k-means clustering description. If k-means archetypes or autoencoder training use all users before the split, then the test labels are derived from test inputs and the metrics in Table 1 are in-sample. The authors must state explicitly that all clustering and encoder fitting are performed on the training split only, and if that is not the case, all reported results must be recomputed on a properly split pipeline.","section":"§2.2.1"},{"comment":"The comparison with Bobadilla et al. [26] misstates the reported accuracy. The text says 'our model demonstrated the same accuracy' while the cited paper reports about 90% and Table 1 reports 88.41% in the ideal case and 80.13% in the noisy case. The claim should be corrected, and any performance comparison should include confidence intervals derived from multiple random seeds.","section":"§3.4"},{"comment":"The noise model is underspecified. The text references [20] and mentions bit-flip and readout errors 'with the same probabilities as in real quantum hardware', but no error probabilities, hardware target, or circuit-level noise model are given. The noisy results cannot be reproduced without these parameters.","section":"§2.2.4"},{"comment":"The number of archetype clusters K is never stated, even though K defines the classification task and the size of the softmax output. Without K, the reported accuracy and F1 are not meaningful, and the reader cannot tell what fraction of the reported error comes from the quantum memory versus the clustering choice.","section":"§2.2.1"}],"minor_comments":[{"comment":"The text says the noise model 'identifies three main sources of errors' but never lists them; add the enumeration.","section":"§2.2.4"},{"comment":"Equation (1) should clarify the allowable range of x_i after min-max normalization and Tanh, since the encoding assumes x_i ∈ [-1,1].","section":"§2.1.2"},{"comment":"The description 'split the test-validation subset into validation and test subsets using the same ratio' is ambiguous; specify the exact split fractions.","section":"§2.2.1"},{"comment":"Informal phrases such as 'way below' and 'can be considered random' should be replaced with exact numeric comparisons and statistical tests.","section":"§3.4"},{"comment":"The paper would benefit from a data and code availability statement; none is provided.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central quantitative claims rest on an ambiguous preprocessing pipeline and an underspecified quantum circuit. If the authors cannot provide code or clarify that clustering and encoder training are restricted to the training split, the paper should be rejected rather than revised. The current major_revision recommendation assumes these issues are addressable with additional details and experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper couples a classical autoencoder with a variational quantum Hopfield associative memory (QHAM) for collaborative filtering on MovieLens 1M, and it claims to be the first to use quantum associative memory for recommendation. The combination is genuinely new: Miller and Mukhopadhyay's QHAM, a standard autoencoder, k-means archetypes, and a softmax classifier. If the numbers held up, it would be a useful data point that a variational QHAM can behave reasonably on a benchmark dataset under noise.\n\nWhere it falls apart is in the missing details. The QHAM circuit is not specified—no ansatz, qubit count, layers, or update rule—so the central object of the paper is undescribed. The \"qubit-overhead optimization\" is one sentence. There are no error bars or multiple seeds. The comparison to Bobadilla et al. says 88.41% accuracy is \"the same\" as ~90%, which is not the same.\n\nThe most serious problem is the potential data leak. Section 2.2.1 says \"We used all the users' data for the simulation\" right before describing the split, and the k-means step is described after the split without stating whether clusters are computed on the training split only or on the full user set. If k-means sees all users, the archetype labels for test users are derived from test inputs, and the reported ROC/accuracy/F1 become in-sample fit statistics, not generalization estimates. The text does not rule this out, and no code or data is provided to check. That ambiguity is load-bearing, not a minor omission.\n\nI want to be fair: the noise study is a real attempt to evaluate robustness, and the authors are transparent that this is a simulation with no speedup claim. But as submitted, the central quantitative claim is not verifiable. I would not accept this for peer review yet; I'd encourage the authors to release code, specify the circuit in full, clarify the split/clustering order, and add a same-pipeline classical baseline. After that, it could be a serious candidate. For now, the right call is a desk reject with a clear route to revision.","headline":"A plausible hybrid quantum-classical recommendation pipeline whose headline numbers are unverifiable: the QHAM circuit is never specified, and the text does not rule out a train/test leakage through k-means clustering.","tokens_in":9742,"tokens_out":2633,"would_cite":false,"duration_ms":26295,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid quantum Hopfield associative memory pipeline performs collaborative filtering on MovieLens 1M with ROC 0.98 and F1 0.88 in ideal simulation, and degrades gracefully under hardware-like noise.","keywords":["Quantum Machine Learning","Quantum Hopfield Associative Memory","Collaborative Filtering","Recommendation System","Hybrid Quantum-Classical Model","Noisy Quantum Computing","MovieLens 1M"],"falsifier":"Re-run the pipeline with k-means clustering applied only to the training subset before any test user is observed, then recompute ROC, accuracy, and F1 on the untouched test split; if those metrics fall substantially below the reported 0.9795, 0.8841, and 0.8786, the central generalization claim is falsified by the data-split order alone.","tokens_in":8724,"feed_emoji":"⚛️","tokens_out":9700,"duration_ms":88107,"temperature":0.7,"pith_summary":"The paper claims that a recommender system can be built on a quantum associative memory instead of on matrix factorization, and that the resulting hybrid classical-quantum pipeline is competitive with classical deep-learning recommenders on the MovieLens 1M benchmark. The pipeline compresses user rating vectors with an autoencoder, clusters users into archetypes with k-means, stores those archetypes as amplitude-encoded states in a variational quantum Hopfield associative memory, and retrieves the best-matching archetype for a new user. In ideal simulation the authors report ROC 0.9795, accuracy 0.8841, and F1-score 0.8786 on the test set; under a custom noise model with bit-flip and readout errors, the numbers become ROC 0.9177, accuracy 0.8013, and F1-score 0.7866. They also claim that updating a single randomly targeted qubit reduces qubit overhead relative to earlier QHAM implementations. If these results hold, quantum associative memories could serve as the retrieval core of practical recommendation systems on near-term noisy hardware.","feed_headline":"Quantum-assisted recommender scores 0.98 ROC on MovieLens 1M","feed_subtitle":"A quantum Hopfield associative memory matches classical deep learning accuracy even under simulated hardware noise.","key_machinery":"The load-bearing component is the variational Quantum Hopfield Associative Memory (QHAM), a trainable quantum circuit that stores user-archetype patterns as quantum states and retrieves the stored state closest to an input. Each archetype's real-valued vector is encoded into qubit amplitudes by a cosine/sine mapping, the full state is prepared with a standard uniformly-controlled-rotation procedure, and the circuit parameters are trained end-to-end with mean squared error loss. A classical autoencoder reduces each user's sparse rating vector to a compact latent code, k-means supplies the archetypes, and a final softmax layer converts the retrieved quantum state into a predicted user category. The paper's claimed novelty is updating only one randomly targeted qubit during training, which is said to cut qubit overhead compared with earlier QHAM implementations.","core_discovery":"The central claim is that a hybrid architecture consisting of a deep autoencoder, k-means clustering, and a variational Quantum Hopfield Associative Memory (QHAM) can perform collaborative filtering at a level comparable to state-of-the-art classical neural recommenders, while training in only 35 epochs. User rating vectors are compressed by the encoder, archetype patterns are extracted by k-means and converted to polar form, and the QHAM stores them as amplitude-encoded quantum states. Retrieval of the archetype most similar to an encoded user is trained end-to-end under mean squared error loss. The authors report that the model achieves ROC 0.9795, accuracy 0.8841, and F1-score 0.8786 on the MovieLens 1M test set in an ideal simulator; under a simulator-based noise model incorporating bit-flip and readout errors at hardware-like probabilities, it achieves ROC 0.9177, accuracy 0.8013, and F1-score 0.7866. They further report that their one-random-qubit updating scheme reduces qubit overhead relative to prior QHAM designs.","pith_inferences":["If the one-qubit-update strategy generalizes, it may reduce the gate depth of other variational quantum memory schemes, making them cheaper to run on real hardware.","Because the paper does not specify whether k-means clustering is applied before or after the train/test split, a re-run with clustering restricted to the training split would clarify whether the reported test metrics are truly out-of-sample.","A natural extension is to ablate the quantum memory by replacing it with a classical nearest-neighbor index over the same archetypes; the difference would isolate the empirical contribution of the quantum retrieval step."],"forward_implications":["Collaborative filtering can be implemented with a quantum associative memory as the retrieval core, without explicit matrix factorization.","The reported noise robustness (ROC falling from 0.9795 to 0.9177 under the custom error model) indicates the architecture may tolerate realistic hardware errors without catastrophic failure.","Training for 35 epochs to reach roughly 88% accuracy and 0.88 F1 suggests faster convergence than the 80-epoch classical baseline the paper compares against.","The model's performance on a standard benchmark positions it as a candidate for deployment on near-term quantum devices, though only as a simulated demonstration so far."],"supporting_citations":[{"why":"Establishes the quantum associative memory concept that the QHAM retrieval step builds on.","marker":"[8]"},{"why":"Provides the QHAM implementation and qubit-state encoding scheme that this paper adapts to its recommender pipeline.","marker":"[16]"},{"why":"Supplies the uniformly controlled rotations used to prepare the archetype quantum states.","marker":"[17]"},{"why":"Provides the software framework used to construct and train the hybrid quantum-classical network.","marker":"[18]"},{"why":"Supplies the error-model characterization that fixes bit-flip and readout error probabilities in the noisy simulation.","marker":"[20]"},{"why":"Provides the classical deep-learning collaborative filtering baseline that the paper compares against on accuracy and F1.","marker":"[26]"}],"fun_headline_variants":["Quantum recommender hits 0.98 ROC on MovieLens in 35 epochs","Hybrid QHAM matches classical recommenders on MovieLens","Variational QHAM powers collaborative filtering on MovieLens","Quantum recommender robust to noise, matches classical accuracy","Qubit-efficient QHAM hits 0.88 F1 under simulated noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported test scores assume the user archetypes stored in the quantum memory were formed from training data only, because the paper does not state whether k-means clustering is applied before or after the train/test split, and applying it to the full dataset would make the test metrics in-sample.","fun_headline_variants_meta":{"raw":{"variants":["Quantum recommender hits 0.98 ROC on MovieLens in 35 epochs","Hybrid QHAM matches classical recommenders on MovieLens","Variational QHAM powers collaborative filtering on MovieLens","Quantum recommender robust to noise, matches classical accuracy","Qubit-efficient QHAM hits 0.88 F1 under simulated noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000506,"raw_usage":{"total_tokens":2545,"prompt_tokens":1097,"completion_tokens":1448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":1356}},"tokens_in":713,"tokens_out":1448,"duration_ms":10903,"temperature":1.0,"reasoning_tokens":1356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:33:18.177626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with k-means clustering applied only to the training subset before any test user is observed, then recompute ROC, accuracy, and F1 on the untouched test split; if those metrics fall substantially below the reported 0.9795, 0.8841, and 0.8786, the central generalization claim is falsified by the data-split order alone.","supporting_citations":[{"cited_title":"Scientific Reports 11(1) (2021) https: //doi.org/10.1038/s41598-021-02866-z","cited_arxiv_id":null,"evidence_quote":"Provides the QHAM implementation and qubit-state encoding scheme that this paper adapts to its recommender pipeline."},{"cited_title":"International Journal of Interactive Multimedia and Artificial Intelligence 6(1), 68–77 (2020) https://doi.org/10.9781/ijimai.2020.02.006","cited_arxiv_id":null,"evidence_quote":"Provides the classical deep-learning collaborative filtering baseline that the paper compares against on accuracy and F1."}],"review_version":2}