{"id":"0e2f8784-9490-4659-8f28-95a27fcf5e53","arxiv_id":"2506.03339","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Permutation-invariant quantum circuits label cliques in small random graphs more accurately than cyclic-invariant or standard ansatze in simulation.","lead":"This paper tests whether quantum circuits built to respect permutation symmetry can learn to identify cliques in random graphs. In simulations on 6 and 8 qubits, the permutation-invariant circuit reaches high accuracy while standard ansatze lag.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparison is uninterpretable because the paper never defines the accuracy metric while labels are drawn randomly among possibly multiple cliques; the reported near-perfect accuracy is either impossible under an exact-label metric or measures something other than the claimed…","rationale":"I agree with the reader's weakest assumption. The paper provides no code, data, or hyperparameters, so a reproduction is necessary for any of the empirical claims. The conceptual symmetry discussion is standard group-equivariant QML; calling the circuit 'invariant' while using it equivariantly is a terminology issue rather than the deciding flaw. The deciding flaw is that the empirical headline cannot be interpreted without the missing metric and dataset statistics. If the dataset has unique cliques and the metric is exact label match, the claim may be recoverable; if not, the comparison is meaningless. Because the manuscript does not address this, the rejection verdict stands without adjustment.","tokens_in":6607,"tokens_out":7321,"duration_ms":92524,"concrete_test":"Reproduce the dataset construction of Section IV and count, for the 6- and 8-qubit cases, the fraction of graphs that contain more than one valid k-clique. For each such graph, compute the best possible exact-label accuracy of any deterministic predictor (max over cliques of the label probability; for exact node-wise matching this is below 1 whenever ties exist). If the reported Fig. 1 accuracy exceeds that bound, the metric cannot be exact label matching. Then either (a) re-run the benchmark with an explicitly stated metric, e.g., node-wise F1 or 'predicted set is any valid clique', reporting results on graphs with unique cliques separately, or (b) if the original metric was 'any valid clique', rerun all three ansatze under that metric and check whether the permutation-invariant advantage persists. This check settles whether the reported advantage is an evaluation artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV says graphs are separated into label groups, graphs without a clique of the desired size get a blanket -1 label, and 'For graphs containing larger cliques, the contained clique is chosen randomly.' No definition of 'accuracy' is given in Section V or anywhere else. If a graph contains two distinct 4-cliques and the label is uniformly one of them, no deterministic circuit can match the random label on both draws; expected exact-label accuracy is bounded by the maximum probability of a single clique. Yet Fig. 1 shows the permutation-invariant ansatz reaching near 100% accuracy. The only ways to reconcile this are that the dataset was filtered to graphs with a unique clique (not stated) or that the metric credits any valid clique rather than the randomly chosen label. In the latter case, the comparison is not 'labeling the randomly selected clique' but a weaker clique-existence or clique-prediction test, and the performance gap between ansatze could be an artifact of that metric. This is load-bearing because the paper's central claim is an empirical superiority claim resting entirely on Fig. 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports numerical experiments comparing three quantum circuit ansatze - a permutation-invariant circuit, a cyclic-invariant circuit, and a standard strongly-entangling ansatz - on a supervised clique-labeling task. Graphs are Erdős-Rényi random graphs encoded via a graph embedding, and the target label is a measurement pattern marking the vertices of a 4-clique (6 qubits) or 5-clique (8 qubits). The paper claims that the permutation-invariant ansatz reaches near-perfect validation accuracy within a few epochs, that the cyclic-invariant ansatz converges more slowly, and that the standard ansatz performs at chance level. The authors attribute this to an intrinsic symmetry of the problem: the feature and label transform together under simultaneous permutation of the vertices. The central empirical claim rests on Figure 1, with no formal derivation and only a qualitative symmetry argument in Section VI.","tokens_in":6733,"tokens_out":6905,"duration_ms":90606,"significance":"If the reported results are correct and the comparison is fair, the paper would provide a useful demonstration that symmetry-restricted quantum circuits can exploit the equivariance of a graph-labeling task to achieve faster and more accurate training than less symmetric ansatze. The hypothesis in Eq. (1) is clearly stated, and the use of quantum natural gradient with error bars over ten seeds is a sensible experimental design. However, the paper as written does not define the accuracy metric, does not establish that the baselines are comparably trained, and provides no code or data. These omissions leave the central performance claim unverified, so the significance cannot currently be assessed beyond the qualitative observation that permutation-invariant circuits converge quickly on this particular dataset.","major_comments":[{"comment":"The accuracy metric is never defined, and the label-generation rule makes the reported near-100% accuracy impossible to interpret. Section IV states that for graphs containing larger cliques, \"the contained clique is chosen randomly.\" If a graph contains multiple distinct cliques of the target size and the label is the randomly chosen one, no deterministic circuit can match that random label on expectation over draws, so exact-label accuracy would be bounded well below the values shown in Fig. 1. If the metric instead accepts any valid clique, then the task is not \"labeling the randomly selected clique\" and the relationship between the figure and the claim in the abstract must be restated. Please define the accuracy metric precisely, and either filter the dataset to graphs with a unique target clique or report both exact-label and any-valid-clique accuracy.","section":"Section IV and Section V, Fig. 1"},{"comment":"The comparison between ansatze is not shown to be fair. The text says the target is 120 parameters, giving \"40 repetitions in the case of the permutation-invariant quantum circuits, 30 in the case of the cyclic-invariant quantum circuits and three for the strongly-entangling layer in the standard ansatz.\" A single PennyLane strongly-entangling layer on 6 qubits typically contains 18 parameters (three rotation angles per qubit), and on 8 qubits it contains 24, so three layers would give 54 or 72 parameters, not 120. If the standard ansatz is trained with far fewer parameters, its chance-level performance in Fig. 1 may be an artifact of underparameterization. Please report the actual parameter count for each ansatz and rerun the standard ansatz with a comparable number of parameters.","section":"Section III, parameter matching"},{"comment":"The manuscript does not state the loss function, the measurement/readout scheme, the optimizer hyperparameters, or the exact training/validation split beyond \"100 graphs per epoch\" and a balanced dataset of 3000 graphs. Without these details, the convergence curves in Fig. 1 cannot be reproduced or independently assessed. The absence of code and data exacerbates this problem. Please provide the omitted training specifications or release the code and dataset so that the central empirical claim can be checked.","section":"Section IV and Section V, training details"},{"comment":"The explanation in Section VI restates the design principle rather than independently explaining the observed convergence. Eq. (1), F -> l iff pi_k F -> pi_k l, asserts the equivariance of the task under simultaneous permutation of feature and label; the permutation-invariant circuit is constructed to respect exactly this symmetry. The experiment therefore does not distinguish the hypothesis that symmetry alignment is causally responsible for the performance from other differences between the ansatze, such as parameter count, trainability, or landscape structure. Please add a control that breaks the symmetry (for example, randomly permuting the label order relative to the feature) or compare against a classical equivariant baseline to make the symmetry explanation testable.","section":"Section VI, Eq. (1)"}],"minor_comments":[{"comment":"The circuit diagrams in Table I are not legible in the provided manuscript, so the parameter-sharing structure of the permutation-invariant and cyclic-invariant layers cannot be verified; please provide a higher-resolution figure or a formal gate-by-gate specification.","section":"Table I"},{"comment":"The phrase \"labels indicate the clique position\" is inconsistent with the later statement that graphs without a clique are labeled \"-1 expected measurement on all qubits\"; clarify whether the -1 label means all qubits are in state |-1> or whether it denotes a separate class.","section":"Section IV"},{"comment":"The notation in Eq. (1) is undefined: the arrow '->' and the action of the permutation pi_k on both the graph feature F and the label l should be specified explicitly.","section":"Section VI, Eq. (1)"},{"comment":"The sentence \"Other clique sizes show similar results\" is unsupported because no results for other clique sizes are shown; either include them or remove the claim.","section":"Section IV"},{"comment":"The validation set is said to contain 2900 graphs per epoch, but the number of validation samples used to compute each accuracy point and the way the 95% confidence intervals are computed over the 10 runs should be stated.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"This is a very short application note whose central claim is currently unverifiable because the accuracy metric is undefined and the baseline training is not demonstrably matched. The issues are fixable in principle, but the authors should be told that the next version must include a precise metric definition, a matched-parameter control, full training details, and ideally code/data. I would also ask the editor to consider whether the incremental contribution, even after these fixes, rises to the level of a full paper for this journal, given that the circuits and the symmetry framework come from the authors' own prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the central comparison in Fig. 1 is not something you can trust as reported. The paper says graphs with multiple 4-cliques get a randomly selected clique as the label, and never defines accuracy. If the metric demands exact match to that random label, no deterministic circuit can get near 100% on such graphs; expected accuracy would be bounded by the max label probability. The near-perfect curves then imply either the dataset was filtered to unique-clique graphs (not stated) or the metric credits any valid clique (not stated). Either way, the empirical claim is uninterpretable. The stress-test note is correct and it lands.\n\nWhat's genuinely new: applying the authors' permutation-invariant circuits to a clique-labeling task, extending their prior graph-problem work. That's fine as a small extension. The circuits themselves are from their earlier papers and the related work is cited properly. The observation that the joint feature-label permutation symmetry is what matters (Eq. 1) is a fair way to frame the task, though it's standard equivariance dressed up as 'invariance'—the paper's terminology is sloppy.\n\nSoft spots beyond the metric. The baseline comparison is unfair: the standard ansatz is only trained for the 50 epochs shown, while the cyclic ansatz additionally gets 200 epochs in a separate experiment. So 'standard ansatz shows no convergence' is not a settled result. No code, data, or hyperparameters are given, so nothing is reproducible. And the 'explanation' of the performance gap is essentially the design principle: the circuit was built to be permutation-equivariant and the task was chosen to be permutation-equivariant. That isn't circular in a harmful way—it's an expected result—but it's not independent evidence.\n\nWhat holds up: the claim that permutation-invariant circuits converge faster on this equivariant task is plausible, and the cyclic-vs-permutation comparison, if the metric were fixed, would be a reasonable sanity check. The paper is short, clearly written, and doesn't overclaim beyond the benchmark.\n\nWho this is for: QML practitioners interested in symmetry-informed ansatze. It's a conference-short-paper-level result, not a transformative one.\n\nRecommendation: send it to peer review, but with a referee who will demand the metric be defined, the dataset handling of multiple cliques be clarified, and the baselines trained to equal budget. The flaws are fixable; the core idea isn't wrong, it's just under-supported. I'd want to see a revised version before believing the figure.","headline":"The reported benchmark is not interpretable as written because the accuracy metric is never defined while labels are drawn randomly among multiple valid cliques; the symmetry story is plausible but the empirical claim needs a rework.","tokens_in":7328,"tokens_out":1901,"would_cite":false,"duration_ms":19699,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","05C69"],"pacs":[],"model":"deepseek-v4-flash","headline":"A permutation-invariant quantum circuit learns to label cliques in random graphs, while unsymmetric circuits do not converge.","keywords":["quantum machine learning","permutation-invariant quantum circuits","equivariant quantum circuits","clique problem","graph labeling","symmetry-restricted quantum circuits","variational quantum circuits","Erdős–Rényi graphs"],"falsifier":"Retrain the three circuits on graphs that contain exactly one clique of the target size, with that clique as the fixed label, and compare validation accuracy; the paper's symmetry explanation predicts the permutation-invariant circuit still wins, whereas the stochastic-label artifact predicts the gap shrinks or disappears.","tokens_in":6349,"feed_emoji":"⚛️","tokens_out":8638,"duration_ms":100715,"temperature":0.7,"pith_summary":"Quantum machine learning usually matches a circuit's built-in symmetry to a symmetry of the data, but this paper studies a labeling problem that looks asymmetric: given a random graph, label all nodes of a randomly chosen clique. The paper reports that a permutation-invariant circuit learns this task quickly, reaching high validation accuracy within a few epochs, while a weaker cyclic-invariant circuit converges slowly and a standard strongly entangling ansatz stays near random. The explanation offered is that the problem is nevertheless highly symmetric, because permuting the graph together with the label leaves the input–output relation unchanged, $F \\to \\ell \\iff \\pi_k F \\to \\pi_k \\ell$. If this explanation is right, symmetry-aware quantum circuits can be applied to equivariant graph-labeling problems, not just to invariant graph-property questions, and the symmetry to exploit is the joint symmetry of features and labels.","feed_headline":"Symmetry-matched quantum circuit learns clique labels","feed_subtitle":"A permutation-invariant circuit labels cliques in a few epochs while unsymmetric circuits stay near random.","key_machinery":"The load-bearing object is the permutation-invariant parameterized quantum circuit, constructed by restricting the generator algebra of a variational layer to the invariant subspace of the permutation group and realizing the resulting generators through SWAP-interconnected two-qubit gates; layers are repeated to match parameter counts. It is paired with a graph embedding that maps each node to a qubit and each edge to a CZ gate. The identity that carries the argument is the joint equivariance condition $F \\to \\ell \\iff \\pi_k F \\to \\pi_k \\ell$, which says that a graph and its clique label may be permuted together without changing the learning target; this converts a superficially asymmetric labeling task into one whose symmetry matches the circuit.","core_discovery":"On its own terms, the paper's central discovery is that the clique-labeling task is equivariant under simultaneous permutations of the input graph and the output clique label, and that a circuit hard-wired to be permutation-invariant exploits this equivariance effectively. In the experiments, the permutation-invariant ansatz converges to high validation accuracy within a few epochs for both six and eight qubits; the cyclic-invariant ansatz converges more slowly, reaching roughly 65% accuracy after 200 epochs; and the unconstrained strongly entangling ansatz shows no convergence, staying near chance. The paper reads this ordering as evidence that matching the circuit symmetry to the joint permutation symmetry of the problem, rather than raw circuit capacity, is what makes the task learnable.","pith_inferences":["Editorial extension: the same simultaneous-permutation argument should transfer to other subset-labeling problems, such as independent set or vertex cover, as long as a random optimal subset is drawn as the label; this predicts a similar accuracy ordering and can be tested without new machinery.","Editorial extension: because the accuracy metric is not defined, the ranking could depend on how alternative valid cliques are scored; a benchmark using graphs with a unique clique of target size would separate the symmetry benefit from any benefit of label stochasticity.","Editorial extension: the gradual degradation from full permutation symmetry to cyclic symmetry suggests a spectrum of optimization difficulty indexed by subgroup size, which could be probed with intermediate symmetry groups between $C_n$ and $S_n$."],"forward_implications":["Symmetry-aware quantum circuits can be applied to equivariant graph-labeling tasks, not only to invariant graph-property decisions such as connectivity or bipartiteness.","A practical design rule follows: identify the largest symmetry under which both features and labels transform together, then restrict the ansatz to that symmetry.","The slow partial success of the cyclic-invariant circuit indicates that weaker symmetries give partial learnability, so symmetry content, not parameter count, is the primary control on convergence.","The joint equivariance condition can be checked before training, giving a criterion for when a permutation-invariant ansatz is worth trying."],"supporting_citations":[{"why":"Provides the standard strongly entangling layer used as the unsymmetric baseline and the simulation environment for all experiments.","marker":"[1]"},{"why":"Defines the Erdős–Rényi random graph model from which the dataset is sampled.","marker":"[7]"},{"why":"Supplies the group-invariant quantum machine learning formalism used to restrict circuit generators to the symmetry.","marker":"[12]"},{"why":"Gives the construction of permutation-invariant circuits through symmetrized generators and SWAP operations.","marker":"[14]"},{"why":"Provides the graph-to-qubit embedding with a CZ gate per edge and the prior permutation-invariant graph-problem benchmark this work extends.","marker":"[16]"},{"why":"Establishes the framework for exploiting symmetry in variational quantum machine learning that motivates matching circuit symmetry to problem symmetry.","marker":"[17]"},{"why":"Supplies the quantum natural gradient optimizer used for training all ansatze.","marker":"[25]"}],"fun_headline_variants":["Permutation-symmetric circuit learns clique labels in few epochs","Symmetry-matched quantum circuit beats cyclic ansatz on cliques","Hardwired permutation symmetry aids clique classification","Quantum circuit with built-in symmetry solves clique task"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison rests on the assumption that the undisclosed accuracy metric faithfully rewards correct clique labeling, even though graphs can contain several valid cliques and the target label is chosen randomly.","fun_headline_variants_meta":{"raw":{"variants":["Permutation-symmetric circuit learns clique labels in few epochs","Symmetry-matched quantum circuit beats cyclic ansatz on cliques","Hardwired permutation symmetry aids clique classification","Quantum circuit with built-in symmetry solves clique task"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1479,"prompt_tokens":739,"completion_tokens":740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":355,"completion_tokens_details":{"reasoning_tokens":677}},"tokens_in":355,"tokens_out":740,"duration_ms":9337,"temperature":1.0,"reasoning_tokens":677,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:05:41.093269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the three circuits on graphs that contain exactly one clique of the target size, with that clique as the fixed label, and compare validation accuracy; the paper's symmetry explanation predicts the permutation-invariant circuit still wins, whereas the stochastic-label artifact predicts the gap shrinks or disappears.","supporting_citations":[{"cited_title":"Group-Invariant Quantum Machine Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the group-invariant quantum machine learning formalism used to restrict circuit generators to the symmetry."},{"cited_title":"Quan- tum Natural Gradient,","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum natural gradient optimizer used for training all ansatze."}],"review_version":1}