{"id":"b7417c8e-0080-419a-9e04-7a8b201340ae","arxiv_id":"2507.00019","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The reported 40 to 60 percent encoding-time reduction from three quantum-inspired relabeling strategies is contradicted by the paper's own measurements for several embedding types.","lead":"This paper proposes three simple data-transformation schemes that relabel feature values as symbolic 'quantum' states and then feeds the relabeled data to ordinary classifiers. The claimed speedup in encoding time conflicts with the paper's own timing table, so the core result fails on internal evidence.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline time-saving claim is contradicted by the paper's own Table 1: several ILS/GDS/CC-ILS entries fall outside or reverse the stated 40–60% reduction range.","rationale":"The reader's strongest_claim identifies exactly the contradiction I consider most load-bearing: the paper's own Table 1 shows multiple cases where ILS, GDS, or CC-ILS increases encoding time relative to DE, and most cases where it does decrease time fall well outside the claimed 40–60% band. The central conclusion of the paper is that large time savings come with negligible accuracy cost; if the time-savings claim is false, the entire comparative evaluation loses its main justification. The undefined encoding function is also a serious reproducibility concern, but the headline time claim is already falsified internally, so it is the decisive issue. My concrete check is arithmetic on the published table, so it can be run immediately and settles the matter. I therefore agree with the reader's rejection and recommend no change to the verdict.","tokens_in":23036,"tokens_out":2556,"duration_ms":25702,"concrete_test":"Independently compute the percentage change (DE − strategy)/DE × 100 for every row of Table 1 and compare each entry against the claimed 40–60% band. The comparison already shows failures: QAOA ILS (+5.8%), Displacement GDS (+52.9%), Squeezing GDS (+27.9%), Angle ILS (−22.7%), and CC-ILS values of −26.3%, −6.0%, −7.2%, and −1.1%. If the authors instead intended a different claim, such as 'GDS is fast for discrete encodings' or 'some strategies sometimes reduce time', the Results section must be revised to state and bound that narrower claim before any accuracy trade-off can be evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Results and Discussion section is that 'ILS, GDS, and CCILS consistently reduce the encoding time by approximately 40 to 60% compared to Direct Encoding (DE) across all six quantum inspired embedding methods.' This is the main quantitative contribution of the paper. It fails against the paper's own measurements in Table 1. For example: QAOA ILS is 67171.5900 s versus DE 63505.5914 s, an 5.8% increase, not a 40–60% reduction. Displacement GDS is 63.3668 s versus DE 41.4430 s, a 52.9% increase. Squeezing GDS is 66.9308 s versus DE 52.3153 s, a 27.9% increase. Angle ILS is 6086.5100 s versus DE 7877.3700 s, a 22.7% reduction; IQP ILS gives a 16.0% reduction; CC-ILS gives reductions of 26.3% (Basis), 6.0% (Angle), 7.2% (IQP), and 1.1% (QAOA), all far below 40%. Only a minority of rows fall in the stated band, and the direction of the effect is wrong in several critical cases. Since the paper's own data directly contradict the headline claim, the core empirical result is not supported regardless of how the encoding function is later specified. This is an internal inconsistency, not a mere disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three quantum-inspired data encoding strategies — Instance Level Strategy (ILS), Global Discrete Strategy (GDS), and Class Conditional Value Strategy (CCVS, presented as CC-ILS and CC-GDS) — for transforming classical tabular data into symbolic quantum-state representations intended for use with purely classical machine learning classifiers. The authors evaluate the strategies on a telecom churn dataset with six embedding types (Basis, Angle, IQP, QAOA, Displacement, Squeezing) and report encoding times and classification metrics for several classical models. The central claim is that ILS, GDS, and CC-ILS consistently reduce encoding time by 40–60% relative to Direct Encoding (DE) while keeping accuracy within ±1–2%.","tokens_in":23270,"tokens_out":3896,"duration_ms":37400,"significance":"If the 40–60% time reduction at near-constant accuracy held, the paper would offer a practically useful low-cost preprocessing recipe for quantum-inspired encoding workflows. The paper also provides a structured taxonomy of three encoding strategies with illustrative examples, which may be pedagogically useful. However, the study does not provide code, machine-checked proofs, or parameter-free derivations; the complexity expressions O(n·d), O(m), and O(c·m) merely restate the definitions of the strategies. The main empirical claims are unsupported by the paper's own tables, and the encoding function is never specified, so the results are not reproducible. The significance of the contribution as presented is therefore low.","major_comments":[{"comment":"The headline claim that 'ILS, GDS, and CCILS consistently reduce the encoding time by approximately 40 to 60% compared to Direct Encoding (DE) across all six quantum inspired embedding methods' is directly contradicted by Table 1. For QAOA, ILS takes 67171.5900 s versus DE's 63505.5914 s, a 5.8% increase. For Displacement, GDS takes 63.3668 s versus DE's 41.4430 s, a 52.9% increase; for Squeezing, GDS takes 66.9308 s versus DE's 52.3153 s, a 27.9% increase. Several CC-ILS reductions are far below 40%: Basis 26.3%, Angle 6.0%, IQP 7.2%, QAOA 1.1%. Only a minority of rows fall in the stated 40–60% band, and the direction of the effect is reversed in critical cases. The stated conclusion is not supported by the paper's own measurements.","section":"Results and Discussion; Table 1"},{"comment":"The encoding function E is never defined. Equations (1)–(4) introduce the notation E(x_i) → |ψ_i⟩ and E(u_k) → |ψ_k⟩, and Section 4 says each value is 'replaced with its quantum representation,' but the paper does not specify how a classical value is mapped to a concrete quantum state, how that state is represented numerically, or how the symbolic states are converted into feature vectors for the purely classical classifiers. Without this specification, the classification pipeline and every accuracy number in Table 2 are unreproducible, and the results could change arbitrarily under a different implementation of E.","section":"Sections 3.2 and 4"},{"comment":"Table 2 contains multiple internal inconsistencies that undermine the reliability of the accuracy results. For example, several IQP DE rows report accuracy values greater than 1 (49.3316 for Logistic Regression and SVM Linear, 62.0321 for KNN, 67.2460 for AdaBoost), and the SVM Linear Angle DE row shows '66.8449' in the accuracy column, which appears to be a percentage-like value rather than a proportion. Additionally, the DE rows for Basis and Angle are identical across every classifier (accuracy 0.6684, precision 0.6513, etc.), which is implausible unless the two embeddings produce the same features or the table has been mis-copied. These data-integrity issues mean Table 2 cannot be used to support the claimed ±1–2% accuracy invariance.","section":"Table 2"},{"comment":"The claim that accuracy 'remains within a small margin of variation, typically ±1 to 2%,' is also contradicted by the paper's own numbers. For instance, Logistic Regression with Basis encoding shows DE accuracy 0.6684 versus ILS accuracy 0.7433, a 7.5-percentage-point change; KNN with Basis encoding shows DE 0.5067 versus ILS 0.6941, an 18.7-percentage-point change. These differences are an order of magnitude larger than the stated margin, further invalidating the paper's central accuracy-efficiency trade-off claim.","section":"Table 2; Results and Discussion"}],"minor_comments":[{"comment":"There are pervasive typos and nomenclature inconsistencies: 'Sqeezing' in Table 2, 'keggal' in the Data Availability Statement, and the abstract uses 'CCVS' while the main text uses 'CC-ILS' and 'CC-GDS'; please align the abbreviations.","section":"Throughout"},{"comment":"The text says the classifiers used were Logistic Regression, KNN, SVM, and ensemble methods such as Random Forest, LightGBM, AdaBoost, and CatBoost, but Table 2 also reports Decision Tree, Extra Trees, Gradient Boosting, and XGBoost; clarify which models were actually run and why the list differs.","section":"Section 4"},{"comment":"The GDS Basis entry reports 2 unique values and an encoding time of 0.0000 seconds; since the PCA-transformed features are continuous, it is unclear how only two unique values arise, and the zero time should be explained or rounded transparently.","section":"Table 1"},{"comment":"The caption for Figure 2 reads 'Flow Chart for Instance level strategy' but the figure illustrates the Global Discrete Strategy; correct the caption.","section":"Figure 2 caption"},{"comment":"The complexity expressions O(n·d·C_embed), O(m), and O(c·m) are presented as the paper's own results, but they are restatements of the definitions and omit the per-value embedding cost for GDS and CC-GDS; please specify these costs explicitly.","section":"Sections 3.1–3.5"}],"recommendation":"reject","confidential_remarks":"The paper's central empirical claim fails against its own Table 1, and the method description lacks the specification needed for reproducibility. In addition, the novelty is limited because the strategies reduce to value deduplication and class-wise grouping, and the 'quantum' aspect is purely symbolic with no actual quantum computation. The data-integrity issues in Table 2 (out-of-range accuracy values, identical DE rows) suggest the experimental reporting needs substantial verification before any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the one thing you should know: the paper's main claim—that ILS, GDS, and CC-ILS consistently cut encoding time 40–60% across all six embeddings—is contradicted by Table 1. This is not a subtle issue. QAOA ILS takes 67,172 s against 63,506 s for DE (5.8% slower), Displacement GDS is 63.4 s vs 41.4 s (52.9% slower), and Squeezing GDS is 66.9 s vs 52.3 s (27.9% slower). Several CC-ILS entries are 1–26% reductions. Only a few rows actually sit in the 40–60% band. So the headline result fails on the paper's own numbers.\n\nWhat's genuinely decent: the three strategies are defined concretely with worked examples. They are simple dictionary/relabeling operations—dedupe identical rows, dedupe unique values, dedupe values within a class—and the exposition is clear enough that a reader can see what was done. The literature survey of basis, angle, IQP, QAOA, displacement, and squeezing embeddings is serviceable.\n\nThe soft spots are large. The encoding function E is never specified. The paper says each value is 'replaced with its quantum representation' but never says how |ψ_k⟩ becomes a numeric feature vector for scikit-learn classifiers. That omission alone makes every accuracy number in Table 2 unreproducible. It probably also explains the tell: Direct Encoding gives identical metrics for Basis and Angle across every classifier, which suggests the embedding choice is not actually reaching the classifiers. There is no code, no processed data, no error bars, one dataset, and no statistical tests. The complexity expressions O(n·d), O(m), and O(c·m) are restatements of the definitions, not independent results.\n\nOn novelty, the three strategies are familiar preprocessing tricks in classical ML (label encoding, hashing, deduplication). The quantum notation is decorative. That is not a crime, but it means the contribution is a timing benchmark, not a new encoding idea.\n\nWho this is for: someone wanting a pedagogical walk-through of PennyLane embedding timings might skim it, but as a research paper it does not support its conclusions. My recommendation: desk reject. Even with heavy revision and a released implementation, the headline claim would need to be rewritten completely. I would not spend referee time on this version.","headline":"The 40–60% encoding-time claim is contradicted by the paper's own Table 1, and the encoding pipeline is too underspecified to support any accuracy comparison.","tokens_in":23845,"tokens_out":3983,"would_cite":false,"duration_ms":42828,"reading_group":"no","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 proposes three quantum-inspired encoding strategies—row-wise, global unique-value, and class-conditional—and claims they cut encoding time by 40-60% while keeping classifier accuracy within about 1-2 percentage points.","keywords":["quantum-inspired encoding","instance-level encoding","global discrete encoding","class-conditional encoding","encoding time reduction","classification accuracy","quantum data representation","customer churn classification"],"falsifier":"Rerun the six encodings with a written-down $E$ and numeric feature extraction: if QAOA ILS still exceeds QAOA DE in time, or if Displacement or Squeezing GDS still exceeds its corresponding direct baseline, the paper's blanket 40-60% time-reduction claim is false. Likewise, any concrete implementation that changes the accuracies in Table 2 by more than the claimed ±1-2% would falsify the accuracy-preservation claim.","tokens_in":22752,"feed_emoji":"⚛️","tokens_out":8699,"duration_ms":91363,"temperature":0.7,"pith_summary":"This paper tries to establish that the expensive step of turning classical data into quantum-style representations can be made far cheaper without sacrificing classification quality. It proposes three strategies—encode every row separately, encode each unique value once across the whole dataset, or encode unique values separately for each class—and applies them with six quantum-inspired embedding methods to a customer-churn classification task. The paper reports encoding-time reductions of roughly 40-60% compared with direct row-by-row encoding, with classifier accuracy staying within about ±1-2%. A reader should care because encoding time is often the practical bottleneck in quantum-inspired machine learning, and a strategy that preserves accuracy while cutting that cost would make such pipelines usable on larger data.","feed_headline":"Quantum-inspired encodings claim 40-60% faster data prep","feed_subtitle":"Value-reuse strategies keep classifier accuracy within 1-2 points while cutting encoding cost, the authors report","key_machinery":"The central object is the encoding map $E$ from a classical feature value to a quantum state $|\\psi\\rangle$. The three strategies differ only in the argument set fed to $E$: one row at a time (ILS), the set of unique values over the whole dataset (GDS), or unique values restricted by class (CC-GDS) and rows within a class (CC-ILS). The load-bearing move is deduplication: the complexity stated by the paper falls from $O(n\\cdot d\\cdot C_{\\mathrm{embed}})$ to roughly $O(m)$ or $O(c\\cdot m)$, where $m$ counts unique values rather than rows. Within the experiments, six embedding implementations supply concrete instances of $E$, and identical rows are reused to reduce ILS cost.","core_discovery":"The paper's central claim is that value-deduplication can replace brute-force direct encoding. Formally, it proposes replacing per-row maps $E(x_i)\\to|\\psi_i\\rangle$ with per-value maps $E(u_k)\\to|\\psi_k\\rangle$ (GDS) or per-class-per-value maps $E_y(u_k^y)\\to|\\psi_k^y\\rangle$ (CC-GDS), plus a row-level class-conditioned variant; because repeated values are encoded once, the embedding workload drops from $O(n\\cdot d)$ embedding calls to roughly $O(m)$ or $O(c\\cdot m)$ calls. The authors assert that across Basis, Angle, IQP, QAOA, Displacement, and Squeezing embeddings, this drop translates into 40-60% less encoding time, and that the accuracy of classical classifiers trained on the reconstructed data varies by only ±1-2%. They also single out CC-ILS with Squeezing embedding as the best accuracy configuration and GDS as the fastest.","pith_inferences":["Editorial: the paper's own Table 1 contains outliers that contradict the blanket 40-60% reduction claim—QAOA ILS takes 67,171.59 seconds versus 63,505.59 seconds for QAOA direct encoding, and Displacement and Squeezing GDS both take longer than their direct baselines.","Editorial: with $E$ and the state-to-feature conversion unspecified, the accuracy results in Table 2 are not reproducible; a concrete choice of implementation could shift the numbers well beyond the claimed ±1-2%.","Editorial: CC-GDS is defined but never tested, so the experiments cannot cleanly separate the benefit of class-aware encoding from the benefit of per-class deduplication; testing it would close that gap.","Editorial: a natural scaling law to check is whether time savings grow with value redundancy, i.e., with the ratio $m/(n\\cdot d)$; if so, GDS-style strategies would be most attractive for categorical or low-cardinality datasets."],"forward_implications":["If the reported trade-off holds, quantum-inspired embedding ceases to be the dominant cost in classical machine learning pipelines, and larger tabular datasets become feasible to encode.","Practitioners can choose GDS when encoding time is the binding constraint and accept that value-sharing may cost fidelity on complex data.","Class-conditional instance encoding, especially with squeezing-style embeddings, offers the paper's best accuracy-runtime balance.","The proposed strategies are preprocessing schemes, so they compose with any downstream classical classifier a workflow already uses."],"supporting_citations":[{"why":"The authors' earlier comparative analysis of classical-to-quantum mapping techniques, which this study extends with the three strategies.","marker":"[6]"},{"why":"Supplies the parameterized-quantum-circuit framing that motivates quantum-inspired encodings for classical models.","marker":"[5]"},{"why":"Defines continuous-variable quantum neural networks and displacement-style embeddings used for the CV experiments.","marker":"[23]"},{"why":"Defines the QAOA embedding, one of the six compared encoding methods.","marker":"[27]"},{"why":"Provides the continuous-variable quadrature/displacement background for the Displacement and Squeezing encodings.","marker":"[19]"},{"why":"Documents the high computational cost of data encoding that the proposed strategies target.","marker":"[8]"},{"why":"Supports the premise that encoding choice materially affects efficiency, expressiveness, and model performance.","marker":"[26]"}],"fun_headline_variants":["Quantum-style encodings cut data prep time by up to 60%","Fast quantum-inspired encodings keep ML accuracy within 1-2 points","Value-reuse encodings slash encoding time, preserve classifier accuracy","Three quantum-inspired encodings speed up data prep with minimal accuracy loss","Encoding reuse cuts costs 40-60%, says quantum-inspired study"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the unshown encoding function $E$ and the unshown conversion from $|\\psi_k\\rangle$ to a numeric feature vector work as intended; the paper gives no definition of either, and every reported time and accuracy depends on those absent details.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-style encodings cut data prep time by up to 60%","Fast quantum-inspired encodings keep ML accuracy within 1-2 points","Value-reuse encodings slash encoding time, preserve classifier accuracy","Three quantum-inspired encodings speed up data prep with minimal accuracy loss","Encoding reuse cuts costs 40-60%, says quantum-inspired study"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000936,"raw_usage":{"total_tokens":3995,"prompt_tokens":928,"completion_tokens":3067,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2971}},"tokens_in":544,"tokens_out":3067,"duration_ms":22776,"temperature":1.0,"reasoning_tokens":2971,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:36:09.616756+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the six encodings with a written-down $E$ and numeric feature extraction: if QAOA ILS still exceeds QAOA DE in time, or if Displacement or Squeezing GDS still exceeds its corresponding direct baseline, the paper's blanket 40-60% time-reduction claim is false. Likewise, any concrete implementation that changes the accuracies in Table 2 by more than the claimed ±1-2% would falsify the accuracy-preservation claim.","supporting_citations":[{"cited_title":"Quantum data encoding: A comparative analysis of classical-to- quantum mapping techniques and their impact on machine learning accuracy.EPJ Quantum Technology, 11(1):72, 2024","cited_arxiv_id":null,"evidence_quote":"The authors' earlier comparative analysis of classical-to-quantum mapping techniques, which this study extends with the three strategies."},{"cited_title":"Continuous-variable quantum neural networks.Physical Review Research, 1(3):033063, 2019","cited_arxiv_id":null,"evidence_quote":"Defines continuous-variable quantum neural networks and displacement-style embeddings used for the CV experiments."},{"cited_title":"Braunstein and Peter Van Loock","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-variable quadrature/displacement background for the Displacement and Squeezing encodings."},{"cited_title":"Exponential data encoding for quan- tum supervised learning.Physical Review A, 107(1):012422, 2023","cited_arxiv_id":null,"evidence_quote":"Documents the high computational cost of data encoding that the proposed strategies target."},{"cited_title":"On fundamental aspects of quantum extreme learning machines","cited_arxiv_id":null,"evidence_quote":"Supports the premise that encoding choice materially affects efficiency, expressiveness, and model performance."}],"review_version":1}