{"id":"2c95c21f-6d0f-46b7-8959-e990327283e6","arxiv_id":"2412.12484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"EvoQAS-ED uses an evolutionary algorithm to search for variational quantum circuits with high effective dimension, reporting high model capacity in simulations.","lead":"This paper evolves variational quantum circuit designs, scoring each candidate by a mathematical measure of model capacity called effective dimension. The evolved circuits score higher on this measure than classical neural networks of similar size, but they are not tested on real learning tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract claims EvoQAS-ED enhances QML performance, but the paper never trains any discovered circuit on a task; the effective-dimension fitness proxy is the sole evidence, making the central claim unsupported without task-level validation.","rationale":"The reader's weakest_assumption already identifies the ED-proxy issue; my stress-test concurs and sharpens it as an unsupported causal leap from the optimized fitness metric to real QML performance. There is genuine value in the paper: the differentiable circuit representation, mutation operator, and ED-based fitness objective are coherent, and the ED curves in Figs. 4-5 demonstrate that the evolutionary search optimizes its stated objective. No internal inconsistency in the optimization itself was found. However, the advertised payoff requires demonstrating that the optimized proxy transfers to an actual learning task, and that demonstration is absent. The concrete test above would settle this: if ED rankings do not predict task accuracy, the central claim collapses; if they do, the conditional verdict can be upgraded. Because the reader's CONDITIONAL already reflects this gap, the verdict should remain UNCHANGED rather than escalate.","tokens_in":8380,"tokens_out":4032,"duration_ms":42153,"concrete_test":"Train the three evolved circuit configurations from Fig. 6 and a control set of random circuits with the same qubit count, encoding map, and parameter count on a fixed binary classification task (e.g., two-moons or the Wine dataset under the paper's angle encoding), using identical optimizer, learning rate, and a held-out test set across at least 5 seeds. Compare test accuracy against ED rankings. If the evolved circuits do not outperform the random controls, or if test accuracy is not positively correlated with the circuits' effective dimension, the claim that ED-optimized architectures enhance QML performance is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the unstated equation 'higher effective dimension implies better QML performance.' Section V uses dγ,n as the fitness function, and Section VI validates the search only by plotting effective dimension curves (Fig. 7) and Fisher spectra (Fig. 8). No discovered circuit is ever trained on a classification or regression task, so the abstract's phrase 'thereby enhancing the overall performance of QML models' is not directly evidenced. This matters because dγ,n is a property of the model family defined by a parameter-space average in Eq. 3, not a measured learning outcome. A circuit family can have high effective dimension yet be hard to optimize, exhibit barren-plateau-like spectra on actual data, or overfit; conversely a lower-ED circuit can generalize better on a finite sample. The comparison in Fig. 7 is also selection-biased: the QNNs are the products of 1000 generations of ED maximization, whereas the classical NNs are fixed, hand-chosen baselines, one deliberately without a nonlinear activation. The reported ED gap may reflect optimization effort rather than a reliable capacity signal or a quantum advantage. Without task-level benchmark evidence, the central claim is an extrapolation from an unvalidated proxy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EvoQAS-ED, an evolutionary quantum architecture search method that encodes quantum circuit components (encoding layer and variational layers) as real-valued vectors, mutates them with Gaussian noise, and converts them into discrete circuits via softmax sampling and one-hot selection. The fitness function is the effective dimension of the resulting QNN model, computed from the empirical Fisher information matrix as defined in Eq. (3). Experiments evolve a population of 50 agents for 1000 generations at dataset sizes n=1000 and n=2000, showing convergence of the effective dimension, and then analyze three discovered circuits by plotting effective dimension versus dataset size and comparing Fisher information eigenvalue spectra against two classical feedforward networks. The paper concludes that the method discovers circuits with high model capacity and improved learning capabilities, thereby enhancing QML performance.","tokens_in":8676,"tokens_out":6177,"duration_ms":58811,"significance":"If validated against task-level benchmarks, EvoQAS-ED would be a useful, relatively simple method for replacing expert ansatz design with an automated search over circuit architectures, and it extends quantum architecture search to a capacity-based objective. The real-valued representation and mutation scheme are clearly described, and the paper explicitly reports convergence curves for the evolutionary optimization. However, the evaluation is entirely based on the same metric (effective dimension) that serves as the fitness function, so the optimization curves are partly a consequence of the selection rule. The Fisher-spectrum comparison is suggestive but does not establish improved learning capability. No code or exact numerical settings for the effective-dimension computation are provided, limiting reproducibility.","major_comments":[{"comment":"The central claim that the discovered circuits 'offer improved learning capabilities, thereby enhancing the overall performance of QML models' is unsupported because no discovered circuit is ever trained on an actual learning task. All reported evidence consists of effective dimension curves and Fisher eigenvalue spectra. High effective dimension does not by itself guarantee trainability, good optimization landscapes, or good generalization on a finite dataset; a circuit family with high ED can still suffer from barren-plateau-like spectra on specific data or overfit. To support the claimed performance enhancement, add task-level experiments (e.g., classification or regression benchmarks) comparing EvoQAS-ED circuits against fixed hand-designed ansatze and against randomly sampled circuits, and report training curves, test accuracy/loss, and variance over independent runs.","section":"Abstract and Section VI, Figure 7"},{"comment":"The evaluation is circular in an important sense: effective dimension is used both as the fitness function being optimized and as the success metric, so the rising top-agent curves in Figures 4 and 5 demonstrate only that the search maximizes dγ,n. The comparison in Figure 7 is also selection-biased: the three QNNs are products of 1000 generations of ED maximization, while the two classical NNs are fixed, hand-chosen baselines, one deliberately without a nonlinear activation. To establish that EvoQAS-ED discovers circuits with exceptional capacity, compare against quantum baselines under identical conditions: for instance, hardware-efficient ansatze, random circuits from the same search space, and random search with the same evolution budget. Report means and standard deviations over multiple independent evolutionary runs.","section":"Sections V and VI, Figures 4–7"},{"comment":"The manuscript does not specify the statistical model p(x,y;θ) used in the Fisher information matrix, the data distribution from which (x,y) pairs are drawn, or the measurement observables. Equation (3) defines the effective dimension with respect to E_{(x,y)∼p}, but p is never made concrete for the simulated QNNs. Without this information, the reported effective dimension values are not reproducible, and the dependence of the results on the choice of p (including whether y are labels of a classification task or continuous targets) is unknown. Please state the model family, the data distribution, the number and nature of observables, and how the expectation values in the Fisher information are computed.","section":"Section V, Eq. (3) and surrounding text"},{"comment":"The numerical procedure for evaluating the effective dimension for each agent is not described. It is unclear how the integral over the parameter space Θ in Eq. (3) is discretized, how many parameter samples are used, how the empirical Fisher information matrix is estimated, and how gradients of the QNN expectation values are obtained (e.g., via the parameter-shift rule). These details are essential for reproducing Figures 4 and 5 and for assessing the computational cost of running 1000 generations with a population of 50. Please provide the algorithm and, if possible, release code or a detailed pseudocode.","section":"Section V, fitness computation"}],"minor_comments":[{"comment":"The notation 'idd' in the determinant most likely denotes the d-dimensional identity matrix; please typeset it as \\(\\mathbb{I}_d\\) and define the symbol.","section":"Section V, Eq. (3)"},{"comment":"The dimensions of the representation vectors are introduced with symbols such as NUM_H_LAYERS and NUM_ROTATIONS but are not explicitly defined in the text; for clarity, state that NUM_H_LAYERS = 2, NUM_ROTATIONS = 3, and NUM_ENTANGLING = 2, matching Table I.","section":"Section IV"},{"comment":"The x-axis label 'Dataset Size n 1e6' is ambiguous; suggest 'Dataset size n (×10^6)' or a similar unambiguous label.","section":"Figure 7"},{"comment":"The axis label 'Eigenvalue size' is unclear; use 'Eigenvalue magnitude' or 'Eigenvalue value'. Also, the x-axis ranges of the subplots are not fully consistent, which makes visual comparison difficult.","section":"Figure 8"},{"comment":"The phrase 'After 1000 generations of training' is a misnomer since the process is evolutionary optimization, not training of model parameters; suggest 'After 1000 generations of evolution'.","section":"Section VI"},{"comment":"Several author names contain spacing artifacts (e.g., 'Y .-C. Chen') likely from the source format; please proofread the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of quantum machine learning and quantum architecture search. The main concern is that the paper's headline claim about improved QML performance is not directly tested; the effective-dimension proxy is used as both fitness and evaluation metric. The absence of reproducibility details for the effective-dimension computation is also concerning. I would encourage the editor to request a revision with task-level benchmarking and appropriate baselines, as these additions are feasible within the manuscript's scope. I saw no indication of redundant or dual publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: EvoQAS-ED does what it says on the tin—it evolves variational circuits toward higher effective dimension—but the paper's bigger claim, that this enhances QML performance, is not backed by any experiment. The stress-test note is right: not one discovered circuit is trained on a classification or regression task. The abstract's 'thereby enhancing the overall performance' is an extrapolation from a proxy.\n\nWhat's actually new: representing a circuit architecture as a dictionary of continuous vectors and mutating by adding Gaussian noise is a clean handle for evolutionary search. Using effective dimension as the fitness function instead of task accuracy is a sensible choice if you believe the proxy, and the optimization curves in Figures 4 and 5 show the expected convergence. The Fisher spectrum comparison is suggestive and consistent with earlier work.\n\nWhere it goes soft: the evaluation is circular in a mild way—effective dimension is both the objective and the success metric, so seeing it rise is partly by construction. The comparison against classical NNs in Figure 7 is selection-biased: the QNNs are the products of 1000 generations of ED maximization, while the NNs are fixed, hand-picked baselines, one without an activation function. That gap doesn't tell you much about quantum advantage or even about general capacity. Also, the reproduction details are thin: how exactly the empirical Fisher is computed, what data distribution is used for the expectation, and how the gradients are obtained are not fully specified.\n\nNone of this kills the core optimization result. The method does find circuits with high effective dimension, and the continuous representation is a useful practical contribution. But the jump from that to 'improved learning capabilities' is unsubstantiated. A careful reviewer should ask for a task-level benchmark (e.g., classification on a standard dataset) where the evolved circuits are actually trained and compared against fixed baseline circuits and classical NNs with matched parameter counts.\n\nMy take: worth a serious referee, but the author needs to either add that experiment or rewrite the claims to stay within what the data shows.","headline":"EvoQAS-ED does evolve circuits toward higher effective dimension, but the paper never trains a discovered circuit on a task, so its central claim about enhanced QML performance is unsupported.","tokens_in":9109,"tokens_out":1642,"would_cite":false,"duration_ms":15309,"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":"EvoQAS-ED uses evolutionary search with effective dimension as fitness to discover variational quantum circuits with high model capacity.","keywords":["quantum machine learning","quantum architecture search","evolutionary algorithm","effective dimension","variational quantum circuits","model capacity","Fisher information"],"falsifier":"Train the evolved high-effective-dimension circuits and the classical baselines on the same supervised learning task, such as classifying a standard dataset like MNIST or a synthetic nonlinear function, and compare test accuracy or generalization gap. If the high-effective-dimension circuits do not outperform the baselines, the central claim that effective dimension indicates performance fails.","tokens_in":8182,"feed_emoji":"🧬","tokens_out":5551,"duration_ms":45622,"temperature":0.7,"pith_summary":"The paper tries to establish that an evolutionary algorithm can automatically design variational quantum circuits whose effective dimension, a measure of model capacity, exceeds that of hand-designed or classical networks. If true, this would remove a key expert bottleneck in quantum machine learning, letting a search process guided by the Fisher information metric find circuits that can approximate complex functions. Numerical simulations on small circuits show the evolved circuits' effective dimension grows and then saturates with dataset size, and their Fisher spectra remain spread out rather than collapsing toward zero as qubit count grows.","feed_headline":"Evolution finds high-capacity quantum circuits","feed_subtitle":"A fitness function based on effective dimension yields QNNs that outpace classical nets on capacity metrics.","key_machinery":"The central object is the effective dimension of a statistical model, defined through the normalized Fisher information matrix and a dataset-size parameter $n$; it quantifies how many independent parameter directions a model effectively explores at a given data budget. The paper couples this metric with a dictionary-based circuit representation (Equation 2) whose floating-point vectors are softmax-sampled into one-hot choices for gate types and entangling layers, and with a mutation operator that adds small Gaussian noise to the representation. The evolutionary loop selects the top agents by effective dimension each generation, so the search optimizes model capacity directly.","core_discovery":"EvoQAS-ED represents a quantum circuit as a dictionary of continuous vectors, mutates the representation by adding Gaussian noise, and selects successive generations by effective dimension. Applying this to a search space of encoding and variational layers yields circuits whose effective dimension, evaluated on dataset sizes of 1000 and 2000, is significantly higher than that of classical neural networks with comparable parameter counts. The discovered circuits also show Fisher information eigenvalue spectra that do not concentrate near zero, in contrast to classical networks, suggesting they are less prone to barren-plateau-like trainability issues.","pith_inferences":["The paper's evidence is entirely capacity-based; a direct comparison of training accuracy and generalization error on a concrete task would be needed to confirm that high effective dimension translates to real learning performance.","The continuous representation could be extended to evolve data-encoding angles, measurement observables, or even the number of layers, turning the search into a joint architecture-and-hyperparameter optimizer.","Combining this capacity-driven search with a trainability-aware penalty (such as a variance-based cost) might yield circuits that are both high-capacity and easy to optimize.","A natural stress test is to run the same evolutionary search with a different Fisher approximation (for example, an empirical Fisher computed from a single batch) and see whether the selected architectures remain stable."],"forward_implications":["The same evolutionary setup can be reused with any differentiable model metric by swapping the fitness function, as the paper states in its conclusion.","If effective dimension tracks learning capability, the discovered architectures should outperform typical hand-designed variational circuits on complex quantum machine learning tasks.","The observed Fisher spectra suggest the evolved circuits are less likely to suffer from vanishing gradients as the qubit count scales, which is a key obstacle in variational quantum algorithms.","The method requires no task-specific labels during the search phase, so it can pre-screen architectures before they are trained on a concrete dataset."],"supporting_citations":[{"why":"Defines the effective dimension metric used as the fitness function and establishes its behavior for quantum neural networks.","marker":"[39]"},{"why":"Provides the universal statistics of Fisher information in deep neural networks, which serves as the classical baseline spectrum the paper contrasts with.","marker":"[40]"},{"why":"Cited to support the claim that highly degenerate Fisher spectra impede training and lead to suboptimal optimization outcomes.","marker":"[44]"}],"fun_headline_variants":["Evolved quantum circuits beat classical capacity","Evolution crafts high-capacity quantum circuits","Effective dimension steers quantum circuit design","Quantum evolution optimizes circuit capacity","Fitness-based evolution for quantum circuits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that a circuit's effective dimension, computed from the empirical Fisher information at a fixed dataset size, faithfully predicts how well that circuit will learn and generalize on real machine-learning tasks.","fun_headline_variants_meta":{"raw":{"variants":["Evolved quantum circuits beat classical capacity","Evolution crafts high-capacity quantum circuits","Effective dimension steers quantum circuit design","Quantum evolution optimizes circuit capacity","Fitness-based evolution for quantum circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000459,"raw_usage":{"total_tokens":2228,"prompt_tokens":801,"completion_tokens":1427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":1367}},"tokens_in":417,"tokens_out":1427,"duration_ms":10340,"temperature":1.0,"reasoning_tokens":1367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:02:07.441188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the evolved high-effective-dimension circuits and the classical baselines on the same supervised learning task, such as classifying a standard dataset like MNIST or a synthetic nonlinear function, and compare test accuracy or generalization gap. If the high-effective-dimension circuits do not outperform the baselines, the central claim that effective dimension indicates performance fails.","supporting_citations":[{"cited_title":"Universal statistics of fisher information in deep neural networks: Mean field approach,","cited_arxiv_id":null,"evidence_quote":"Provides the universal statistics of Fisher information in deep neural networks, which serves as the classical baseline spectrum the paper contrasts with."}],"review_version":1}