{"id":"ad5eedc9-9e90-41b5-a7cf-b75679257e23","arxiv_id":"2412.00286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A genetic algorithm over feature-to-qubit permutations improves quantum embedding fitness scores by small margins over random selection in simulated QML classifiers, but several comparisons omit negative results and lack error bars.","lead":"This paper uses a genetic algorithm to search for the best way to assign classical data features to qubits in a quantum machine learning circuit, instead of training the embedding. On MNIST and Tiny ImageNet classification tasks, the search finds feature layouts that score higher and sometimes run faster than random layouts, but the reported gains are small and not consistent across all tested settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Optimal-embedding claim rests on single-seed fitness evaluations; without initialization-sensitivity tests, the 720-permutation sweep and GA-vs-random comparison do not establish a stable, intrinsic advantage.","rationale":"The paper's central claim has two parts: (1) different feature-to-qubit mappings cause significant performance variation, with an optimal embedding existing; (2) GA-based search finds better mappings than random selection. Both parts depend on the fitness evaluation being a reliable indicator of embedding quality. The manuscript uses a single fixed model initialization and only 5 epochs, and reports no repetition of the noiseless sweep. Under these conditions, the observed 20-point spread could arise from the interaction between the embedding and the specific random initialization, not from the embedding alone. This is precisely the reader's weakest_assumption, and it is load-bearing because if the ranking of embeddings is not stable across seeds, the notion of an 'optimal' embedding is ill-defined and the GA's fitness signal is noise. The comparison with random selection is further weakened by the many negative improvements in Table I, which the abstract omits. These issues are correctable by adding multi-seed repetitions and statistical reporting, matching the reader's CONDITIONAL verdict. Therefore, the stress-test does not change the verdict: it remains CONDITIONAL, requiring sensitivity analysis and transparent reporting of all results before the central claims can be accepted.","tokens_in":13776,"tokens_out":6245,"duration_ms":55427,"concrete_test":"Run the 720-permutation sweep for MNIST (2,6), 3 qubits, with 10 different random PQC initializations (same 5 epochs, lr=3e-3). Compute the Spearman rank correlation of combined scores between seeds. If the rank correlation is low (e.g., <0.5) or the top-10 permutations are disjoint, the existence of a stable optimal embedding is unsupported. Additionally, run GA and random selection 10 times each for (0,1,4) with different seeds and compare distributions of best fitness; if the GA improvement is not significant (e.g., overlapping interquartile ranges), the headline claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.A describes the full 720-permutation sweep with 'fixed model initialization' and '5 epochs, 3e-3 learning rate', with no repetition over PQC initializations for the noiseless case. The authors conclude from a ~20-point spread that an optimal embedding exists. But with only 5 epochs, final training accuracy is heavily influenced by the random initial parameters of the strongly entangling layers; a single seed cannot separate the effect of the embedding from the effect of initialization. If the ranking of permutations changes across seeds, the 'optimal' embedding is an artifact of one seed, and the GA (which uses the same single-run fitness) is selecting for that artifact rather than an intrinsically good mapping. The central comparative claim is also not robust: Table I lists many negative improvements (GA best minus random best), e.g., -3.8 for (0,3,5) noiseless and -3.1 for (2,6,3) noisy, yet the abstract quotes only the positive range 0.33-3.33, which corresponds to selected high-qubit averages. Without repeated runs or error bars, these negative values show the claimed advantage is not consistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes framing the selection of a quantum embedding (feature-to-qubit mapping) for QML as a discrete search problem, solved with a genetic algorithm (GA). Experiments on MNIST and Tiny ImageNet binary classification, using angle embedding and a strongly entangling PQC, compare GA-selected embeddings with randomly selected mappings, report runtime improvements, and compare against QEK, QAOA-embedding, and QRAC. The authors also present a full 720-permutation sweep on 3 qubits to argue that an optimal embedding exists.","tokens_in":14021,"tokens_out":7976,"duration_ms":68563,"significance":"The idea of using a genetic algorithm to search over embedding permutations is reasonable and addresses a real practical question in QML. The complete sweep of all 720 permutations on 3 qubits is a useful empirical observation, and the authors provide code on GitHub. However, the current evidence does not support the paper's main comparative claims: the GA-vs-random comparison is compromised by selection on the test metric, the noiseless sweep uses a single model initialization, and the abstract omits numerous negative results. With additional experiments and more careful statistical reporting, the approach could become a solid contribution.","major_comments":[{"comment":"The noiseless 720-permutation sweep uses a single fixed model initialization and only 5 training epochs. With such short training, final accuracies are strongly influenced by the random PQC initialization. The claim that the ~20-point spread \"proves that there exists an optimal embedding configuration\" is not supported without demonstrating that the ranking of permutations is stable across multiple initializations; otherwise the best permutation may be an artifact of one seed. Please repeat the sweep over several initializations and report rank correlations or the distribution of best/worst permutations.","section":"III.A-B"},{"comment":"The fitness function is the combined score, defined as the mean of training and inferencing accuracies. Using test-set accuracy in the fitness function means the GA is directly selecting on the test metric, while the random baseline performs no such selection. The comparison is therefore biased in favor of GA, and the reported fitness improvements are inflated. Use a separate validation set for selection and report test accuracy only after the final embedding is chosen.","section":"IV.B-C, Table I"},{"comment":"The abstract states that GA outperforms random selection by 0.33-3.33 for MNIST, but Table I lists many negative improvements (GA best minus random best), e.g., -3.8 for (0,3,5) noiseless and -3.1 for (2,6,3) noisy. The quoted range is from the high-qubit MNIST subset only; the low-qubit noiseless results are mostly negative. Please report all results and make the abstract consistent with the full table.","section":"Abstract, Table I"},{"comment":"The comparison with QEK, QAOA-embedding, and QRAC is not controlled: QEK and QRAC results are taken from the original papers without re-running under identical conditions, and no error bars are given for any method. The improvements (1.003X-1.06X) are small and may be within noise. Please either run all methods under the same setup or clearly label the comparison as indicative.","section":"IV.D, Table III"},{"comment":"The GA hyperparameters (population size 20, generations 5, crossover 0.8, retention 0.1, mutation 0.001) are fixed without sensitivity analysis, and no repeated GA runs are reported, so the reported fitness scores have no error bars. Additionally, the runtime advantage of GA over random selection is not explained given that both evaluate 100 permutations; please clarify how GA runtime can be lower and report variance across repetitions.","section":"IV.A, IV.C"}],"minor_comments":[{"comment":"There are several typos: \"existance\" should be \"existence\", \"indiviuduals\" should be \"individuals\", \"Addtionally\" should be \"Additionally\", and \"varation\" should be \"variation\".","section":"III.B, IV.A, IV.C"},{"comment":"The header \"mean best runtime best runtime\" is ambiguous; please label columns clearly as RS and GA, and specify whether runtime refers to the best individual or the mean.","section":"Table I"},{"comment":"The description of reference [3] as \"quantum metric learning\" does not match the cited paper (Lloyd et al., \"Quantum embeddings for machine learning\"); please correct the description.","section":"II.D"},{"comment":"Variable names such as \"f s\" and \"os\" are not descriptive, and the \"random(parents, 2)\" notation is unclear; please use more explicit naming and explain the selection step.","section":"Algorithm 1"},{"comment":"In the sentence \"where perform training for proposed GA and QAOA-embedding-based methods\" there is a missing pronoun; should be \"we perform training\".","section":"IV.D"},{"comment":"The \"1X\" in the improvement column should be written as \"1.0X\" for consistency with the other ratios, and the decimal formatting should be uniform.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper's abstract and conclusion overstate the results; the negative improvements in Table I should have been reported prominently. The comparison with existing works is not rigorous enough for the claims made. The authors should be encouraged to add multi-seed experiments and a proper validation/test split, which would make the paper much stronger. The idea is worth pursuing, but the current version needs substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper establishes one thing convincingly and one thing weakly. The strong part is the 720-permutation sweep on 3 qubits: embedding order moves the combined train/inference score by roughly 20 points under fixed hyperparameters. That is a concrete, reproducible observation and it justifies caring about the mapping problem. The weak part is the central claim that the GA beats random selection. The evidence for that is not there yet: each embedding is evaluated once, with one model initialization and 5 epochs; the GA is stochastic and no repeated runs are reported; and Table I shows several negative improvements (-1.7 to -3.8) that disappear from the abstract's \"0.33-3.33\" range. The stress-test note is on target: a single seed cannot separate the embedding effect from initialization noise, so the \"optimal\" permutation might be an artifact of that seed. The authors' own Section IV.E admits the search is backend-specific, which is honest but also means the claimed advantage is conditional on a particular noise model.\n\nWhat's genuinely new: framing embedding selection as a discrete search and applying a GA to feature-to-qubit permutations is a natural extension of existing meta-heuristic kernel search, and the paper does the first full sweep for angle embedding that I know of. The SC trend in the histograms is a nice sanity check. The code is supposedly on GitHub, though I didn't verify it.\n\nThe comparison with QEK, QAOA, and QRAC is not controlled: different encoding methods, different feature counts, and QRAC numbers are lifted from the original paper. The 1.003X/1.03X/1.06X ratios are essentially meaningless as stated.\n\nBottom line: this is a workable engineering idea with an interesting observation, but the headline claim needs repeated runs over seeds, error bars, and a fairer baseline before I'd trust the GA advantage. It deserves a serious referee because the core observation is real and the flaws are fixable, not because the current evidence is convincing.","headline":"Solid empirical observation that embedding permutation matters, but the GA-vs-random claim needs repeated runs and error bars before it holds.","tokens_in":14580,"tokens_out":1914,"would_cite":false,"duration_ms":18119,"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":"Genetic algorithm finds better quantum embeddings than random choice","keywords":["quantum embeddings","genetic algorithm","quantum machine learning","feature-to-qubit mapping","angle embedding","variational quantum classifier","NISQ devices","permutation search"],"falsifier":"For a fixed binary-class pair such as MNIST classes 0 and 1 on 4 qubits, rerun the GA search multiple times with different random seeds and model initializations, or train the resulting embeddings for many more epochs. If the embeddings found by GA under the original fixed setup do not consistently outperform randomly sampled embeddings when the training configuration changes, the claim that GA finds a task-optimal embedding would be falsified.","tokens_in":13552,"feed_emoji":"🧬","tokens_out":6188,"duration_ms":47609,"temperature":0.7,"pith_summary":"Quantum machine learning performance depends on which classical features are mapped to which qubits in the embedding circuit, not just on which embedding type is used. The authors frame the selection of a feature-to-qubit mapping as a discrete search problem and propose a genetic algorithm to solve it, instead of the usual continuous optimization of trainable embeddings. On binary classification tasks from MNIST and Tiny ImageNet, the GA finds embeddings with higher combined training and inference accuracy than randomly chosen mappings, by margins of 0.33–3.33 and 0.5–3.36 fitness points, while reducing runtime by up to 15% and 8.8%, respectively. The paper also reports that the GA scales with qubit count and outperforms three existing embedding methods, quantum embedding kernels, QAOA-based embedding, and trainable QRACs, by 1.003X, 1.03X, and 1.06X.","feed_headline":"Genetic algorithm finds better quantum embeddings than random choice","feed_subtitle":"On MNIST and Tiny ImageNet, the GA lifts fitness by up to 3.36 points and cuts runtime by up to 15 percent","key_machinery":"The machinery is the angle-embedding circuit that pairs two features per qubit, one encoded in an RX gate and one in an RY gate, making a feature-to-qubit mapping a permutation of feature indices. The fitness score is the mean of training and inference accuracy after training the QNN for a fixed number of epochs. The genetic algorithm explores the permutation space with a population of 20, 5 generations, tournament selection of size 2, a crossover that takes a prefix from one parent and fills the remaining positions in the order they appear in the other parent, and a low-probability swap mutation (rate 0.001). The search is run under a fixed training configuration, and the best-scoring permutation is the proposed optimal embedding.","core_discovery":"The central discovery is that feature-to-qubit permutations are a meaningful axis of performance in quantum neural networks, and that a genetic algorithm can search this discrete space effectively. Evidence for the existence of an optimal mapping comes from a full sweep of all 720 permutations on 3 qubits for six MNIST binary-class pairs: under fixed data, model initialization, and hyperparameters, the combined accuracy score spans about 20 points, with the best embedding scoring 93.18 and the worst 74.75 on classes 2 and 6. The GA, using tournament selection, order-based crossover, and swap mutation over a population of 20 permutations for 5 generations, matches or exceeds random selection in both noiseless and noisy (FakeBrisbane) simulations; in the noisy case it often beats even the best randomly chosen embedding. On higher qubit counts (6–8 qubits) and on Tiny ImageNet, the advantage persists, with fitness-score gains up to 3.33–3.36 and runtime savings up to 8.8–15%.","pith_inferences":["One extension the paper leaves implicit is using the GA to search not only the feature order but also which embedding circuit template to use, since the same permutation representation extends to any circuit with a discrete parameter-assignment step.","The optimality claim rests on a single training configuration, so testing whether the GA-selected embedding remains optimal under different hyperparameters or seeds would tell whether the search is finding a property of the data or an artifact of the fixed training setup.","The reported gains over QEK, QAOA embedding, and QRAC are small in multiplicative terms, so the practical significance may lie less in raw accuracy and more in the ability to find a good embedding quickly without gradient-based training of the embedding itself."],"forward_implications":["The feature-to-qubit mapping is a first-class design choice in quantum embedding, not a detail to be left to chance.","Because the GA only requires a fitness score from training, it can be applied to other embedding circuit families and to other classical datasets without changing the search procedure.","Under noisy simulation, the GA found embeddings that beat the best of 100 randomly chosen permutations, suggesting that evolutionary search is particularly useful when noise makes performance fluctuate.","The runtime of GA search is comparable to or lower than random selection at the same evaluation budget, so the search can be run as a standard part of QML model development on NISQ hardware."],"supporting_citations":[{"why":"Supplies the genetic algorithm methodology with selection, crossover, and mutation operators that the paper adapts for permutation search.","marker":"[6]"},{"why":"Provides the Quantum Embedding Kernel baseline that the proposed GA is compared against.","marker":"[2]"},{"why":"Provides the QAOA-based embedding approach used as a comparison baseline.","marker":"[3]"},{"why":"Provides the trainable QRAC embedding approach used as a comparison baseline.","marker":"[4]"},{"why":"Defines the Silhouette Coefficient used to select binary class pairs with varying separability for the experiments.","marker":"[12]"},{"why":"Provides the Strongly Entangling Layer used as the parameterized quantum circuit in the QNN.","marker":"[11]"}],"fun_headline_variants":["GA beats random quantum embedding mappings by up to 3.36 points","Genetic algorithm finds quantum embeddings superior to random","GA boosts quantum embedding fitness by 3.36 over random choice","Quantum embedding search: GA beats random mapping on MNIST and ImageNet","GA finds better quantum embeddings, 15% faster than random"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All fitness evaluations use one fixed training setup, 5 epochs, a learning rate of 3e-3, a fixed model initialization, and for noisy runs the FakeBrisbane noise model, so the best embedding found is assumed to be the best for the task under that single configuration.","fun_headline_variants_meta":{"raw":{"variants":["GA beats random quantum embedding mappings by up to 3.36 points","Genetic algorithm finds quantum embeddings superior to random","GA boosts quantum embedding fitness by 3.36 over random choice","Quantum embedding search: GA beats random mapping on MNIST and ImageNet","GA finds better quantum embeddings, 15% faster than random"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001604,"raw_usage":{"total_tokens":6397,"prompt_tokens":962,"completion_tokens":5435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":5348}},"tokens_in":578,"tokens_out":5435,"duration_ms":34953,"temperature":1.0,"reasoning_tokens":5348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:32:13.297298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed binary-class pair such as MNIST classes 0 and 1 on 4 qubits, rerun the GA search multiple times with different random seeds and model initializations, or train the resulting embeddings for many more epochs. If the embeddings found by GA under the original fixed setup do not consistently outperform randomly sampled embeddings when the training configuration changes, the claim that GA finds a task-optimal embedding would be falsified.","supporting_citations":[{"cited_title":"Quantum embeddings for machine learning, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the QAOA-based embedding approach used as a comparison baseline."},{"cited_title":"Invited: Trainable discrete feature embeddings for quantum machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the trainable QRAC embedding approach used as a comparison baseline."},{"cited_title":"Genetic algorithms","cited_arxiv_id":null,"evidence_quote":"Supplies the genetic algorithm methodology with selection, crossover, and mutation operators that the paper adapts for permutation search."},{"cited_title":"Finding groups in data: an introduction to cluster analysis","cited_arxiv_id":null,"evidence_quote":"Defines the Silhouette Coefficient used to select binary class pairs with varying separability for the experiments."},{"cited_title":"Circuit-centric quantum classifiers","cited_arxiv_id":null,"evidence_quote":"Provides the Strongly Entangling Layer used as the parameterized quantum circuit in the QNN."}],"review_version":1}