{"id":"27376a2c-7d54-4069-aa00-b4390aee3f55","arxiv_id":"2505.10167","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Q-MEDLEY estimates global feature importance in hybrid quantum-classical models by averaging drop-column and permutation importance, re-evaluating the quantum feature map after each perturbation.","lead":"This paper introduces Q-MEDLEY, an explainer that measures how much each original data feature matters to a hybrid quantum-classical machine learning model by re-running the quantum feature encoding after perturbing the feature. It reports experiments showing the explainer can pick out informative features from synthetic noise and that it competes with classical explainers on benchmark datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative validation (§4.2–§4.3) explains only classical DT/RF models with no quantum feature map, so Q-MEDLEY's distinctive ΦQ re-encoding step is never measured; the only quantum results (Figs. 3–4) are visual, with no metric, ground truth, or uncertainty.","rationale":"The reader's CONDITIONAL verdict with medium correctness risk matches my reading, and my stress-test identifies the same load-bearing gap, sharpened: Figures 5–6 and Table 1 quantify Q-MEDLEY only on classical models with no quantum feature map, so the mechanism that distinguishes Q-MEDLEY from the classical MEDLEY of KAXAI [26] — re-running ΦQ on each perturbed input — is never quantitatively exercised. The only evidence involving real feature maps (Figures 3–4) is qualitative. The central claim therefore stands or falls on the correctness of an unmeasured code path. I agree with the reader that the results establish only that Q-MEDLEY explains classical models, and I concur that this is fixable: the provably-irrelevant-feature experiment proposed above would settle it in one controlled run, so REJECT is not warranted. Credit where due: the base formulas (Equations 1–3) are standard and implementable from the paper, the repository is open-sourced, and Section 5.3 honestly flags proxy limitations. Secondary concerns reinforce CONDITIONAL without changing it: (1) the Section 4.3 text overstates Table 1 — for Covtype-RF and Diabetes-RF the full configuration (0.83, 0.83) is worse than the baseline (0.85, 1.00), so the 'virtues of the composite structure' narrative is not monotonic in the data; (2) AdaptiveWeighting and InteractionPI are never defined, making the ablation irreproducible; and (3) no uncertainty quantification is given anywhere, so single-run Recall@3 differences are uninterpretable. None of these individually defeats the framework, which is why the appropriate verdict remains CONDITIONAL (UNCHANGED).","tokens_in":15663,"tokens_out":16065,"duration_ms":150667,"concrete_test":"A single controlled experiment with the quantum map present and a known ordering. In simulation, build amplitude-encoded HQML models on Noisy Iris and Noisy Wine with m extra input features that ΦQ never references (no gate acts on them), so MHQCL is mathematically independent of those features and their true importance is exactly zero. Run Q-MEDLEY (Algorithm 1, amplitude path) with K=5 repeats over at least 5 seeded train/test splits, and check: (i) the ignored features receive importance scores statistically indistinguishable from zero and are ranked below all encoded features in every seed; (ii) the Recall@3 and Spearman values reported for these quantum-containing models, with standard errors, are comparable to the classical-model values in Section 4.2. This directly executes the re-encoding step (Algorithm 1, lines 9–16) that the current validation never runs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Q-MEDLEY 'delineates influential classical aspects in HQML models, as well as separates their noise' (Abstract). The quantitative support for this claim is Section 4.2 (Recall@3 and Spearman, Figures 5–6) and the Section 4.3 ablation (Table 1). In every one of those experiments, the explained objects are classical Decision Tree and Random Forest models trained directly on raw features; the quantum feature map ΦQ of Equations (4)–(5) is not present. Q-MEDLEY's only distinctive mechanism — re-evaluating ΦQ(x') for each perturbed input (Algorithm 1, lines 9–16) — therefore reduces to the identity and is never exercised by any reported number. The experiments that do involve an actual feature map (Figures 3–4) are assessed only by visual inspection, with no metric, no known ordering, and no repeated-run variance. Section 5.3 concedes that 'our practical reliance on classical interpretable models as proxies for validation may not a perfect reflection of all the intricacies of feature influence in the quantum domain,' but the gap is larger than proxy fidelity: the quantitative loop contains no quantum component at all, so the results cannot support the HQML-specific claim regardless of how accurate the classical ground truth is. The claim that Q-MEDLEY explains hybrid quantum-classical models therefore rests on an untested code path; at best the numbers validate the generic DCI-plus-PI aggregation (inherited from the classical MEDLEY, KAXAI [26]) on classical models. The ablation claim is additionally irreproducible because AdaptiveWeighting and InteractionPI are never defined in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces QuXAI, a framework for training and explaining hybrid quantum-classical machine learning (HQML) models of two types: amplitude-encoded models, where classical inputs are mapped by a quantum feature map Phi_Q into state-vector amplitudes, and quantum-kernel models, where a fidelity kernel is fed to a classical learner. The central contribution is Q-MEDLEY, an explainer that computes global feature importance by averaging Drop-Column Importance and Permutation Importance, re-evaluating the quantum feature map for each perturbed input. The paper reports qualitative feature-importance bar charts for amplitude-encoded HQML models on noisy Iris and Wine datasets, quantitative comparisons against DCI, PI, TreeSHAP, and LogRegL1 on classical Decision Tree and Random Forest models, and an ablation study of Q-MEDLEY's components. The authors claim that Q-MEDLEY delineates influential classical features in HQML models, separates them from noise, and competes well with established XAI techniques.","tokens_in":15952,"tokens_out":4012,"duration_ms":39041,"significance":"If validated, Q-MEDLEY would address a real and currently underserved need: model-agnostic global feature attribution for HQML architectures that use quantum feature encoding followed by a classical learner. The paper has genuine strengths: the algorithmic formulation is explicit and reproducible (Eqs. (1)-(6), Algorithms 1-2), the idea of re-evaluating the quantum feature map or kernel after each perturbation is well motivated and correctly formalized, and the code is open-sourced. However, the quantitative evidence as presented does not exercise the quantum-specific path at all: the benchmark experiments in Section 4.2 and the ablations in Section 4.3 explain only classical Decision Tree and Random Forest models trained directly on raw features, with no quantum feature map present. The only experiments that involve a quantum feature map are evaluated by visual inspection. Because the central claimed capability is specifically about HQML models, the current significance is not established by the reported measurements.","major_comments":[{"comment":"The quantitative evaluation explains only classical Decision Tree and Random Forest models trained directly on raw features, so the quantum feature map Phi_Q of Eqs. (4)–(5) never appears in any measured result. In these experiments, Algorithm 1's PredictAdapted branch (lines 2–3) reduces to identity, and the distinctive re-evaluation of Phi_Q(x') on perturbed inputs (lines 9–16) is never exercised by a reported number. Consequently, the Recall@3 and Spearman results can support at most the generic DCI+PI aggregation, not the paper's central claim that Q-MEDLEY delineates influential classical aspects in HQML models. The authors' own limitation in Section 5.3, that reliance on classical interpretable models as proxies 'may not a perfect reflection of all the intricacies of feature influence in the quantum domain,' understates the gap: the issue is not proxy fidelity but the complete absence of a quantum component from the measured loop.","section":"§4.2–4.3, Figs. 5–6, Table 1"},{"comment":"Table 1 shows that the full configuration Q-MEDLEY + AdaptiveWeighting + InteractionPI underperforms the baseline Q-MEDLEY (DCI+PI Avg) on Random Forest for Covtype (0.83 vs. 0.85) and for Diabetes (0.83 vs. 1.00); similarly, Q-MEDLEY with InteractionPI alone underperforms the baseline on RF-Covtype (0.84 vs. 0.85) and RF-Diabetes (0.89 vs. 1.00). This contradicts the Section 4.3 statements that adding components yields 'increasing performance' and that the full configuration 'could obtain the maximum or among the maximum Recall@3 scores.' These differences may be within noise, but no error bars or repeated runs are reported, so the claimed monotonic benefit of the composite structure is not supported by the table.","section":"§4.3, Table 1"},{"comment":"The only experiments that actually involve a quantum feature map are Figures 3 and 4, but these are assessed exclusively by visual inspection. No quantitative metric, no known ground-truth importance ordering, and no repeated-run variance are reported for these HQML models. Since the noisy and redundant features were synthetically added, a rank-correlation or top-k recall against the known importance structure could be computed for the amplitude-encoded HQML models, as the authors themselves do for the classical models in Section 4.2. Without such a measure, the claim that Q-MEDLEY 'separates their noise' in HQML settings is not quantitatively supported.","section":"§4.1, Figs. 3–4"}],"minor_comments":[{"comment":"The text uses 'Remember@3' where the metric is elsewhere called 'Recall@3'; the terminology should be made consistent.","section":"§4.2"},{"comment":"The contribution list describes Q-MEDLEY as a 'global and local feature importance explainer,' but Section 3 defines only the global importance score I_j and no local, instance-level explanation method is described or evaluated anywhere in the manuscript. The claim should be amended to 'global' unless a local method is actually provided.","section":"§1, Contributions"},{"comment":"The limitation paragraph notes that kernel-based HQML results 'were not vast,' but in fact no quantitative kernel-based experiments are reported at all. This should be stated plainly, given that the kernel branch of Algorithm 1 is part of the claimed contribution.","section":"§5.3"},{"comment":"The manuscript contains numerous grammatical and typographical errors ('an framework,' 'purposedly,' 'remonstrating,' 'the abundance of medley scores,' and several incomplete or run-on sentences). A careful copy-edit is needed.","section":"General"},{"comment":"Details needed for reproducibility are missing for the HQML experiments, including the number of qubits, the specific form of U_j(x_j) in Eq. (4), the simulation backend, and dataset sizes after the noise injection. Since the code is open-sourced, adding these parameters in the text or an appendix would make the experiments independently checkable.","section":"§3.2–3.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: Q-MEDLEY is a sensible and clearly specified adaptation of classic DCI+PI to hybrid quantum-classical models, with code and a clean pipeline. What is actually new is that the explainer re-evaluates the quantum feature map after each perturbation, and that is the right instinct for this architecture. But the paper never measures that mechanism. Every quantitative result in §4.2 and §4.3 comes from classical Decision Trees and Random Forests trained on raw features, with no ΦQ in the loop. The only quantum results are the bar charts in Figures 3–4, which are read by eye. So the paper's central claim — that Q-MEDLEY explains HQML models and separates their noise — is not supported by the data as presented.\n\nCredit where it is due: the method is simple enough to be implementable from Algorithm 1, the abstract and equations are internally consistent, the open-source repository is a plus, and the authors are honest in §5.3 that classical proxies may not reflect quantum feature influence. That caveat is more damaging than they let on, but it shows they see the issue.\n\nThe soft spots are mostly evaluation, not conception. The ablation table contradicts the text in several places (e.g., Diabetes RF goes 1.00 → 0.83 when moving from baseline to the full configuration, and Wine RF goes 0.82 → 1.00, yet the text says the full config is 'the only one' to reach perfection). The two headline components, AdaptiveWeighting and InteractionPI, are never defined, so the ablation is not reproducible. There are no error bars or repeated runs. These are fixable with effort, but they are not minor decorations; they are the evidence for the main claim.\n\nBottom line: this is a reasonable idea that needs a proper experiment. The fix is straightforward: construct a synthetic problem with known feature importance, run the explainer on an amplitude-encoded or kernel HQML model where ΦQ is actually exercised, and report the metric plus variance. If that works, the paper is worth publishing. Right now it is an extended proposal with placeholder validation. I would send it to review, because the idea deserves referee time, but I would expect heavy revision and likely another round.\n\nRegards,","headline":"A sensible DCI+PI adaptation to HQML is undone by an evaluation that never exercises the quantum feature map; the idea deserves a careful rewrite, not a desk reject.","tokens_in":16544,"tokens_out":2181,"would_cite":false,"duration_ms":21321,"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":"Q-MEDLEY, an explainer for hybrid quantum-classical models, obtains global feature importance by perturbing classical inputs and re-evaluating the quantum feature map on each perturbation.","keywords":["quantum machine learning","explainable AI","hybrid quantum-classical algorithms","feature importance","drop-column importance","permutation importance","quantum feature encoding","trustworthy quantum AI"],"falsifier":"Train a kernel-based HQML model on a synthetic dataset whose label is a known function of a single feature, run Q-MEDLEY on it, and check whether that feature receives the top importance score; the paper's own evaluation is almost entirely amplitude-encoded, so a failure on the kernel path would show the claimed generality does not hold.","tokens_in":15451,"feed_emoji":"⚛️","tokens_out":4721,"duration_ms":40776,"temperature":0.7,"pith_summary":"This paper proposes QuXAI, a framework for explaining hybrid quantum-classical machine learning (HQML) models in which classical features are first mapped into quantum states and then fed to a classical learner. Its central contribution is Q-MEDLEY, an explainer that measures global feature importance by perturbing classical inputs, re-running the quantum feature map on every perturbed input, and averaging drop-column and permutation importance scores. The authors argue that this re-evaluation step is what makes the attributions faithful to the hybrid data flow, rather than treating the whole pipeline as a black box. On amplitude-encoded HQML models over noisy versions of Iris, Wine, and other benchmark datasets, Q-MEDLEY assigns higher importance to original informative features than to synthetic noise and redundancies, and in classical validation settings its rankings agree with Decision Tree and Random Forest ground-truth importances about as well as established explainers. The claim matters because, if correct, it gives practitioners a way to see which original classical inputs drive a quantum-enhanced model's decisions.","feed_headline":"New explainer traces feature importance through quantum encodings","feed_subtitle":"Q-MEDLEY re-evaluates the quantum feature map on every perturbation, separating real signals from noise in quantum-classical models.","key_machinery":"The central object is the Q-MEDLEY explainer, defined for HQML models $M_{\\mathrm{HQCL}}(x)=M_{\\mathrm{CL}}(f_Q(\\Phi_Q(x)))$. For feature $j$, it computes $I_j = \\frac{1}{2}(I^{\\mathrm{DCI}}_j + I^{\\mathrm{PI}}_j)$, where $I^{\\mathrm{DCI}}_j$ is the drop in accuracy when the column is neutralized ($x_j \\leftarrow 0$) and $I^{\\mathrm{PI}}_j$ is the average accuracy drop over $K$ random permutations of that column. The load-bearing adaptation is that evaluation of the perturbed dataset always recomputes the quantum feature map $\\Phi_Q(x')$ (amplitudes or kernel matrix) before the classical learner predicts, so perturbations propagate through the quantum stage exactly as real inputs would. The full version also adds adaptive weighting and an interaction-aware permutation importance, which the ablations credit for improved top-3 feature recovery.","core_discovery":"The paper's central claim is that a composite perturbation explainer that explicitly re-evaluates the quantum feature map when inputs are altered can produce reliable global feature-importance rankings for HQML models of the form $M_{\\mathrm{HQCL}}(x)=M_{\\mathrm{CL}}(f_Q(\\Phi_Q(x)))$. For each feature $j$, Q-MEDLEY computes a drop-column importance by neutralizing that feature and a permutation importance by shuffling it $K$ times, then averages the two. The critical design choice is that every perturbed input is passed through the quantum encoding $\\Phi_Q$ (or through a recomputed kernel matrix) before the classical learner is asked to score it, so the importance score reflects the feature's influence through the whole hybrid pipeline. The paper reports that on amplitude-encoded models, Q-MEDLEY separates informative from noisy features, and in controlled classical benchmarks it reaches Recall@3 and Spearman correlations comparable to or better than single-method baselines and a tree-ensemble Shapley explainer. Ablations show that adding adaptive weighting of the two components and an interaction-aware permutation step improves top-3 recovery, with the full configuration reaching perfect Recall@3 on some dataset-model pairs.","pith_inferences":["If the classical-proxy validation is not faithful, the reported benchmarks mostly establish Q-MEDLEY as a strong classical explainer; its quantum-specific value remains to be shown on kernel-based HQML and on real quantum hardware with noise.","The same perturbation-through-encoding mechanism could be adapted to local, instance-level explanations by restricting perturbations to a single input.","Because each permutation requires re-evaluating the feature map, runtime grows linearly with the number of features and repeats; approximating the map's sensitivity analytically could make the explainer scalable to higher-dimensional inputs.","Q-MEDLEY scores could serve as a diagnostic for comparing quantum feature maps: maps that preserve the importance structure of the original features would score differently from maps that scramble it."],"forward_implications":["Q-MEDLEY gives global feature importance for amplitude-encoded HQML models without treating the quantum stage as a black box.","It can separate original informative features from synthetic noise and redundancies across multiple datasets and ten classical learners.","In classical validation, Q-MEDLEY's Recall@3 and Spearman correlations are comparable to or better than DCI, PI, LogRegL1, and a tree-ensemble Shapley explainer.","Ablations show the full Q-MEDLEY (adaptive weighting plus interaction-aware PI) achieves the best top-3 feature recovery on several dataset-model pairs, including perfect scores.","The framework provides a route for HQML debugging: detecting when noisy or irrelevant features exert disproportionate influence on predictions."],"supporting_citations":[{"why":"Supplies the MEDLEY explainer concept of combining multiple attribution techniques, which Q-MEDLEY adapts to the hybrid quantum-classical pipeline.","marker":"[26]"},{"why":"Extends Shapley values to quantum circuits, giving the quantum-XAI context that Q-MEDLEY builds on and contrasts with.","marker":"[1]"},{"why":"Provides Q-LIME, a local instance-based quantum explainer, against which the hybrid-specific design is positioned.","marker":"[9]"},{"why":"Frames the challenges of translating classical XAI to quantum systems, motivating the hybrid-specific adaptation.","marker":"[7]"},{"why":"Establishes the broader program of explainable quantum machine learning that this work contributes to.","marker":"[17]"},{"why":"Defines patterns for hybrid quantum algorithms, grounding the HQML architecture that Q-MEDLEY explains.","marker":"[37]"},{"why":"Describes quantum computing models for artificial neural networks, supporting the motivation for hybrid quantum-classical learning.","marker":"[32]"}],"fun_headline_variants":["Q-MEDLEY explains feature importance in quantum-classical models","Quantum-aware explainer separates signal from noise in hybrid ML","New explainer re-runs quantum encodings to rank features honestly","Q-MEDLEY: reliable feature rankings for quantum-enhanced AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation treats feature importances derived from classical Decision Tree and Random Forest models trained on the same data as ground truth for what an HQML model should consider important; the authors note this proxy may not capture how feature influence is reshaped by the quantum encoding.","fun_headline_variants_meta":{"raw":{"variants":["Q-MEDLEY explains feature importance in quantum-classical models","Quantum-aware explainer separates signal from noise in hybrid ML","New explainer re-runs quantum encodings to rank features honestly","Q-MEDLEY: reliable feature rankings for quantum-enhanced AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3318,"prompt_tokens":1052,"completion_tokens":2266,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":2194}},"tokens_in":668,"tokens_out":2266,"duration_ms":17265,"temperature":1.0,"reasoning_tokens":2194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:14:27.921026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a kernel-based HQML model on a synthetic dataset whose label is a known function of a single feature, run Q-MEDLEY on it, and check whether that feature receives the top importance score; the paper's own evaluation is almost entirely amplitude-encoded, so a failure on the kernel path would show the claimed generality does not hold.","supporting_citations":[{"cited_title":"Heese, T","cited_arxiv_id":null,"evidence_quote":"Extends Shapley values to quantum circuits, giving the quantum-XAI context that Q-MEDLEY builds on and contrasts with."},{"cited_title":"Pira and C","cited_arxiv_id":null,"evidence_quote":"Provides Q-LIME, a local instance-based quantum explainer, against which the hybrid-specific design is positioned."},{"cited_title":"Weigold, J","cited_arxiv_id":null,"evidence_quote":"Defines patterns for hybrid quantum algorithms, grounding the HQML architecture that Q-MEDLEY explains."},{"cited_title":"Mangini, F","cited_arxiv_id":null,"evidence_quote":"Describes quantum computing models for artificial neural networks, supporting the motivation for hybrid quantum-classical learning."}],"review_version":1}