{"id":"4d6536a0-83ac-4cc9-b806-d5df2b0b7e46","arxiv_id":"2509.07633","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Variational quantum circuits can guide particle swarm optimization to find better industrial control configurations in an offline contextual bandit setting, with performance comparable to classical neural networks.","lead":"This paper tests whether variational quantum circuits (VQCs) can serve as reward-prediction models in offline contextual bandit problems, using a simulated industrial benchmark. The authors report that quantum models, paired with particle swarm optimization, can find better control settings than the training data, though classical neural networks perform similarly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline ROG values (35.45 vs 38.37) are maxima selected over 1,152 hyperparameter configurations using the very ROG metric being reported; without a permutation null or confidence intervals, the positive gain and the quantum-vs-classical comparison are not established.","rationale":"The reader's weakest assumption is stationarity of the IB snapshot reward. I do not choose that as the single most load-bearing concern: the very low test MSE (~2e-4) reported in Table II suggests the grid labels are largely deterministic under the data-collection protocol, so the static-snapshot assumption is empirically plausible. The more fragile link is the inferential step from 1,152 searched configurations to the claim that VQCs 'successfully identified' superior configurations and are competitive with classical models. The ROG values quoted as headline results are selected maxima of the same metric used for selection. This is a textbook selection-on-the-test-statistic problem. It biases both the level and the comparison: the best of many noisy ROG estimates will be positive even under a null model, and the difference between two maxima (35.45 vs 38.37) has unknown variance. The paper does report 263 vs 133 positive ROG counts, but without a significance test or error bars this does not resolve the issue. A permutation test directly targets the selection mechanism and would settle whether the reported gains exceed what random label-noise plus hyperparameter search would produce. If the null max ROG is below 35.45, the central feasibility claim survives; if not, the empirical support collapses to 'some models found better points under an unevaluated selection procedure.' This does not require rejecting the paper; it requires the conditional acceptance to include the null analysis and release of code/data, which the reader already requested.","tokens_in":13087,"tokens_out":13244,"duration_ms":167503,"concrete_test":"Permutation-null test: randomly permute the target rewards y across the 14,641 dataset rows (breaking any x->y relationship, preserving the marginal distribution), retrain the same 1,152 quantum hyperparameter configurations, rerun PSO, and recompute ROG with the same 1,000-trajectory ground truth evaluation. Repeat for at least 100 permutation seeds and record the maximum ROG across configurations. If the 95th percentile of this null max ROG is comparable to or greater than 35.45, the reported quantum ROG is consistent with selection over hyperparameters plus evaluation noise, and the central claim is unsupported. Complementary check: compute ROG for PSO using a linear-regression surrogate or a constant predictor to gauge how much of the gain is just continuous search over a coarse grid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the ROG of 'best quantum models' (35.45) versus 'best classical models' (38.37), with the additional assertion that more quantum configurations (263/1152) than classical (133/1152) exceed ROG>0 (Section V.B, VI.A). These numbers are the result of a two-stage selection: (i) 1,152 hyperparameter configurations per model family are trained; (ii) the configurations with highest ROG are chosen and their ROG is then reported. Maximizing over 1,152 noisy evaluations inflates the expected best ROG even if no configuration has genuine optimization ability. No confidence intervals, significance tests, or multiple-comparison corrections are provided, so the 35.45/38.37 gap and the 263-vs-133 fraction could be selection noise. Because the paper explicitly switches from MSE to ROG for model selection (Section V.B) and then uses ROG as the outcome, the headline numbers cannot be taken at face value as evidence of successful optimization or of quantum-classical competitiveness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid classical-quantum pipeline for offline contextual bandits in the Industrial Benchmark (IB). The authors discretize the four-dimensional state/action space into a static regression dataset of 14,641 points, train VQC and feedforward-NN reward models over 1,152 hyperparameter configurations each, and then use particle swarm optimization (PSO) on the learned reward surrogate to propose continuous steering configurations, which are evaluated in the IB simulator. Performance is quantified by Relative Optimization Gain (ROG), a cumulative difference in ground-truth reward between PSO-discovered configurations and the best dataset configuration across 11 setpoints. The paper reports that the best quantum model reaches ROG 35.45 versus 38.37 for classical, and that a larger fraction of quantum configurations achieve ROG>0 (263/1152 vs 133/1152). It concludes that VQCs are a viable and competitive alternative for offline contextual bandit optimization in industrial settings.","tokens_in":13345,"tokens_out":4575,"duration_ms":53302,"significance":"If the findings are supported by appropriate statistical analysis, this is a useful proof-of-concept: it extends VQC benchmarking beyond regression accuracy to downstream optimization, uses a realistic simulator, and includes a large hyperparameter search with repeated retraining of selected configurations. The practice of grounding model selection in downstream optimization value (ROG) is a good idea; the issue is that the same metric is used as the reported outcome without selection correction. The paper is also honest about limitations such as simulation-only results and computational cost. It could be a contribution to quantum machine learning for industrial control, but the central comparative claim is currently not robustly established.","major_comments":[{"comment":"The headline ROG values (35.45 vs 38.37) are maxima over 1,152 hyperparameter configurations selected using the same ROG metric that is reported as the outcome. This two-stage selection inflates the expected best ROG even under a null hypothesis where no configuration has genuine optimization ability. No permutation null, bootstrap confidence intervals, or multiple-comparison correction is provided. The 263-vs-133 fraction, while less directly selected, still compares two samples without uncertainty. Please add (a) a null distribution for the best ROG obtained by permuting model-family labels or by drawing random hyperparameter subsets, (b) bootstrap CIs for the difference in mean/median ROG, and (c) the full distribution of ROG values rather than top-10 summaries. Without these, the central claim that quantum models are competitive with classical models is not established.","section":"Section V.B and VI.A"},{"comment":"The classical and quantum hyperparameter search spaces are not matched in several structural dimensions: layer counts (1-4 vs 20-60), presence of activation functions vs parallel encoding/output scaling, optimizer sets (Adam/SGD vs Adam), and resulting parameter counts (Table II: 4801-5359 vs 968-1448). The comparison therefore confounds model family with search-space richness and capacity. Please either match the spaces more closely (e.g., same parameter budget, same optimizer choices, classical analogues of data re-uploading) or explicitly restrict the conclusion to 'within our chosen search spaces.' The abstract and conclusion should be softened accordingly if the spaces remain unmatched.","section":"Section IV.B and Table I"},{"comment":"The regression target is a 100-step average after a 100-step swing-in, and this static snapshot is the basis for both training and PSO. If the IB dynamics have not reached stationarity within that window, the target is contaminated by transient effects and the 'ground truth' used in ROG is not a well-defined function of (p,v,g,h). The paper provides no diagnostic supporting the swing-in length. Please add a stationarity check (e.g., compare first-half vs second-half averages, or vary the swing-in and averaging windows for a subset of grid points) or otherwise justify that the condensed mapping is stable. If the target is non-stationary, the positive ROG could reflect artifacts of the snapshot rather than genuine optimization.","section":"Section III.A and III.D.2"}],"minor_comments":[{"comment":"The abstract says models 'generalize well in noisy and sparse datasets,' but no explicit noise injection is described; clarify that noise arises from the stochastic IB dynamics and sparsity from the 11-level grid sampling.","section":"Abstract"},{"comment":"The labels 'Classical (MSE)' and 'Classical (ROG)' are unclear; define in the caption that these are configurations selected by validation MSE and by ROG, respectively.","section":"Table II"},{"comment":"The caption states 'zoomed-in region' but the region is not clearly marked; specify how many points lie outside the displayed range and whether the fraction 263/1152 vs 133/1152 is computed on the full or zoomed set.","section":"Figure 5"},{"comment":"The text states that quantum models favor deeper circuits and parallel encoding, but no corresponding hyperparameter-frequency figure is shown for quantum models (Fig. 9 covers only classical models). Add a quantum version or explicitly state that the observation is qualitative.","section":"Appendix B"},{"comment":"Reference [15] is incomplete ('B. Metevier et al., in Advances in Neural Information Processing Systems' with no title or year); provide a full citation.","section":"References"},{"comment":"No code or data availability statement is provided. Given the large hyperparameter search and environment-specific pipeline, a repository with the IB wrapper, circuit definitions, and evaluation scripts would strengthen reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core idea is plausible, but the comparative claims are currently over-stated because of selection-on-outcome and unmatched search spaces. I recommend requiring the statistical analyses described in the major comments before acceptance. If the authors cannot provide selection-corrected evidence or choose to soften the claims to a purely descriptive study, the contribution may still be acceptable, but not with the current abstract and conclusion. The missing code/data statement is also likely to be a practical barrier for readers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the combination: VQCs as reward regressors in an offline contextual bandit, evaluated end-to-end with PSO on the Industrial Benchmark. That specific application hasn't been done before, and the paper is honest that it is a proof of concept. It also does a plausible hyperparameter search, retrains top configurations with different seeds, and evaluates candidates in the actual IB simulator rather than only on the surrogate. Credit where due: that is more careful than a lot of QML application papers.\n\nThe central feasibility claim — that some quantum models can guide PSO to configurations better than the dataset baseline — survives reading. The full-hypercube count (263/1152 quantum vs 133/1152 classical configurations with ROG>0) is a less selection-biased result than the headline numbers, and it points in the direction the authors claim.\n\nBut the headline numbers themselves are weaker than presented. The 35.45 vs 38.37 ROG values are maxima over 1152 hyperparameter configurations, selected using the ROG metric that is then reported as the outcome. Maximizing over 1152 noisy evaluations inflates the expected best even if no model has real optimization ability. There are no confidence intervals, no permutation null, and the quantum vs classical comparison rests on unmatched hyperparameter grids (classical has 1–4 layers, 40–128 neurons; quantum has 20–60 layers with different encoding and output scaling options). That asymmetrical grid makes the “quantum models generalize better” claim underdetermined.\n\nTwo smaller things. The static-snapshot reduction — swing-in for 100 steps then averaging 100 steps — is a real assumption about the IB dynamics being sufficiently stationary; the paper treats it as obvious and never checks whether transient effects have decayed or whether the reward is actually a function of the setpoint/steering variables alone. And there is no code or data, which matters for a benchmark-style empirical paper.\n\nThis is not a bad paper. It is a useful benchmark result for QML in offline decision-making, and the authors are appropriately cautious about not claiming quantum advantage. But the quantitative comparison should be treated as illustrative until the selection issue is addressed. For peer review, I would send it, but require a null model or permutation test for ROG significance, full distribution reporting rather than maxima, better matched hyperparameters, and a release of code/data.\n\nFor a reading group: maybe. For me personally, I would not cite the headline numbers but might cite it as an example of VQC offline bandit regression.","headline":"A solid proof of concept that VQCs can serve as offline reward models in a contextual bandit setting, but the headline ROG numbers are selected maxima and need a null model before they can support the quantum-vs-classical comparison.","tokens_in":13818,"tokens_out":1247,"would_cite":false,"duration_ms":16724,"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":"Variational quantum circuits, paired with particle swarm optimization, can discover industrial control settings better than any in their offline training data, rivaling classical neural networks.","keywords":["variational quantum circuits","offline contextual bandits","quantum regression","Industrial Benchmark","particle swarm optimization","reward modeling","quantum machine learning","generalization"],"falsifier":"Re-collect the dataset with a markedly longer swing-in (for example, 1000 steps) and re-run the same training and PSO pipeline; if the best configurations' ground-truth rankings change or the quantum models' positive ROG disappears, the static-snapshot collapse is the load-bearing assumption and the paper's result is an artifact of the 100-step window.","tokens_in":12995,"feed_emoji":"⚙️","tokens_out":6241,"duration_ms":62482,"temperature":0.7,"pith_summary":"This paper asks whether variational quantum circuits can serve as reward models in offline contextual bandit problems, where decisions must be made from a fixed dataset without interactive exploration. Using the Industrial Benchmark environment, the authors collapse its dynamics into a static grid of setpoint-and-steering combinations, train quantum and classical regression models on the resulting rewards, and then let particle swarm optimization search the continuous steering space using each model as a surrogate. They report that the best quantum models found steering configurations whose true measured reward exceeded the best in the original dataset by a cumulative margin of 35.45, close to the classical models' 38.37, and that a larger fraction of quantum models than classical models beat the dataset baseline (263 of 1152 versus 133 of 1152). The point is a proof of concept: quantum regression, despite higher training cost and slightly worse pointwise error, can generalize beyond a sparse, noisy grid well enough to guide optimization.","feed_headline":"Quantum models find better settings than their training data contains","feed_subtitle":"Offline bandit test shows variational quantum circuits rival classical networks at tuning industrial controls.","key_machinery":"The central object is the variational quantum circuit as a regressor: an angle-encoding feature map (Pauli-X rotations with trainable scaling weights), a parameterized ansatz drawn from a hardware-efficient design or from the paired Circuit 11 and Circuit 9 designs, and a Pauli-Z expectation measurement. The circuit is trained by mean squared error on the reward data, and its output is then handed to particle swarm optimization, which searches the continuous steering space separately for each setpoint. Success is quantified by Relative Optimization Gain (ROG), the cumulative ground-truth reward improvement of the PSO-discovered configurations over the best configuration in the original grid","core_discovery":"The central claim is that variational quantum circuits, when combined with particle swarm optimization, are viable reward models for offline contextual bandits in an industrial setting, and that they exhibit a generalization advantage over classical networks trained on the same data. The quantum models do not achieve lower regression error—their best test MSE is 2.066e-4 versus 1.877e-4 for the best classical model—but a much larger share of quantum configurations improves on the best grid-sampled configuration under ground-truth evaluation. The paper attributes this to better extrapolation beyond the training distribution, enabled by design choices such as parallel encoding, deeper circuit","pith_inferences":["The reported gains suggest the true reward surface's optima lie between the discretized grid points; quantum models' periodic, frequency-rich encoding may be especially suited to interpolating such surfaces, a hypothesis the paper does not directly isolate.","A natural next experiment is to vary the swing-in duration and grid resolution to measure how much of the quantum models' ROG advantage depends on the static-snapshot construction; the paper does not report this sensitivity.","The classical baselines include sine activations, which are periodic like quantum rotations, but the paper does not compare quantum models directly against periodic-activation classical networks; such a comparison would clarify whether the advantage is quantum-specific.","If quantum hardware matures, the same surrogate-plus-swarm loop could be augmented with a quantum search routine, as the paper mentions, though the continuous three-dimensional action space would need careful discretization."],"forward_implications":["If the result holds, offline contextual bandits in industrial control do not require online quantum exploration; a static dataset plus a quantum surrogate can already find actions better than any recorded action.","Quantum reward models can be competitive with feed-forward neural networks in noisy, sparse regression-to-optimization pipelines even when their pointwise prediction error is slightly worse.","Hyperparameter design matters more than architecture class: parallel encoding, deeper circuits, and trainable output scaling determine whether a VQC generalizes past its training grid.","The ROG metric provides a practical model-selection criterion that can disagree with MSE-based selection, favoring models that optimize well rather than merely fit well.","Because offline contextual bandits reduce to supervised regression, any improvement in quantum regression transfers directly to offline bandit decision-making."],"supporting_citations":[{"why":"Defines the Industrial Benchmark environment whose dynamics and reward (consumption and fatigue weighted 1:3) supply the dataset and the ground-truth evaluator.","marker":"[25]"},{"why":"Establishes the reduction of offline contextual bandits to supervised regression oracles, the premise for framing reward prediction as regression.","marker":"[14]"},{"why":"Provides the variational quantum circuit learning framework used for the quantum regression models.","marker":"[9]"},{"why":"Shows that data encoding controls the accessible frequency spectrum, grounding the angle-encoding and re-uploading design choices.","marker":"[12]"},{"why":"Introduces data re-uploading, used to expand the quantum models' expressive frequency spectrum.","marker":"[28]"},{"why":"Motivates trainable scaling weights in encoding rotations and output scaling, design elements the paper identifies as beneficial.","marker":"[30]"},{"why":"Introduces the hardware-efficient ansatz used as one of the two circuit architectures.","marker":"[31]"},{"why":"Supplies the Circuit 11 and Circuit 9 ansatz designs and the expressibility/entangling-capability comparison underlying the second architecture.","marker":"[32]"},{"why":"Demonstrates that jointly applying the two ansatz designs yields good expressibility and entangling capability, motivating the chosen combination.","marker":"[33]"},{"why":"Supplies the particle swarm optimizer used to search the continuous steering space over each surrogate reward model.","marker":"[39]"}],"fun_headline_variants":["Quantum circuits extrapolate past training data in bandits","Despite higher error, quantum models pick better settings","VQCs out-generalize classical nets on offline bandits","Quantum bandit models find better configs than data shows","Variational quantum circuits beat classical at offline tuning"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The Industrial Benchmark's average reward after a fixed 100-step swing-in is a well-defined, memory-free function of the setpoint and steering variables, so the discretized grid can be regressed and the surrogate's optima correspond to realizable improvements.","fun_headline_variants_meta":{"raw":{"variants":["Quantum circuits extrapolate past training data in bandits","Despite higher error, quantum models pick better settings","VQCs out-generalize classical nets on offline bandits","Quantum bandit models find better configs than data shows","Variational quantum circuits beat classical at offline tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000598,"raw_usage":{"total_tokens":2554,"prompt_tokens":586,"completion_tokens":1968,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":330,"completion_tokens_details":{"reasoning_tokens":1891}},"tokens_in":330,"tokens_out":1968,"duration_ms":14754,"temperature":1.0,"reasoning_tokens":1891,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:55:37.084437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-collect the dataset with a markedly longer swing-in (for example, 1000 steps) and re-run the same training and PSO pipeline; if the best configurations' ground-truth rankings change or the quantum models' positive ROG disappears, the static-snapshot collapse is the load-bearing assumption and the paper's result is an artifact of the 100-step window.","supporting_citations":[{"cited_title":"A benchmark environment motivated by industrial control problems,","cited_arxiv_id":null,"evidence_quote":"Defines the Industrial Benchmark environment whose dynamics and reward (consumption and fatigue weighted 1:3) supply the dataset and the ground-truth evaluator."},{"cited_title":"Practical contextual bandits with regression oracles,","cited_arxiv_id":null,"evidence_quote":"Establishes the reduction of offline contextual bandits to supervised regression oracles, the premise for framing reward prediction as regression."},{"cited_title":"Variational quantum circuits for deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Provides the variational quantum circuit learning framework used for the quantum regression models."},{"cited_title":"Effect of data encoding on the expressive power of variational quantum-machine-learning models,","cited_arxiv_id":null,"evidence_quote":"Shows that data encoding controls the accessible frequency spectrum, grounding the angle-encoding and re-uploading design choices."},{"cited_title":"Parametrized quantum policies for rein- forcement learning,","cited_arxiv_id":null,"evidence_quote":"Motivates trainable scaling weights in encoding rotations and output scaling, design elements the paper identifies as beneficial."},{"cited_title":"Hardware-efficient variational quan- tum eigensolver for small molecules and quantum mag- nets,","cited_arxiv_id":null,"evidence_quote":"Introduces the hardware-efficient ansatz used as one of the two circuit architectures."},{"cited_title":"Express- ibility and entangling capability of parameterized quan- tum circuits for hybrid quantum-classical algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the Circuit 11 and Circuit 9 ansatz designs and the expressibility/entangling-capability comparison underlying the second architecture."},{"cited_title":"Benchmarking quantum surrogate mod- els on scarce and noisy data,","cited_arxiv_id":null,"evidence_quote":"Demonstrates that jointly applying the two ansatz designs yields good expressibility and entangling capability, motivating the chosen combination."},{"cited_title":"Nevergrad: Black-box optimization platform,","cited_arxiv_id":null,"evidence_quote":"Supplies the particle swarm optimizer used to search the continuous steering space over each surrogate reward model."}],"review_version":1}