{"id":"0cf7c2bc-b7c9-436d-8d4a-232a6f35a93f","arxiv_id":"2412.02285","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A graph transformer that uses learnable quantum-walk encodings as attention biases reports modest, statistically unvalidated accuracy gains on five TUDatasets.","lead":"This paper introduces GQWformer, a graph transformer that adds a learned quantum-walk encoding to attention and a recurrent module for local structure. The authors report small accuracy gains on five benchmark datasets, but the gains are within one standard deviation, and no code is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported accuracy gains over RWC are within the reported standard deviations on all five datasets; the claim of significant improvement is unsupported.","rationale":"My stress-test focuses on the empirical claim rather than the engineering ambiguity highlighted by the reader. The under-specification of the quantum-walk encoding (how M_t is formed, how the shift operator handles variable degrees, what p_ij represents) is a serious reproducibility defect; without a precise formula for M_t, the experiments cannot be independently reimplemented. However, the paper's own reported statistics already undermine the central claim even if every pipeline detail were filled in. The authors claim 'significantly outperforms RWC,' but the reported margins are 0.5–2.3 percentage points with standard deviations around 2–4 points. With 10-fold CV, the standard error of each difference is roughly 1.2–2.0, so none of the differences is statistically distinguishable from zero. This is a direct, objective falsification of the paper's headline statement, and it is independent of any subjective judgment about the quantum-walk construction. It is also a concern the authors could address with a simple re-analysis or by releasing per-fold results. Therefore, the reader's REJECT verdict stands, but the primary load-bearing issue is statistical, not definitional.","tokens_in":12273,"tokens_out":6358,"duration_ms":65782,"concrete_test":"Request the per-fold accuracy vectors for GQWformer and RWC on all five datasets and run a paired Wilcoxon signed-rank test (or paired t-test) on the per-fold differences. If per-fold data are unavailable, compute an approximate two-sample t-test from the reported means and standard deviations with n=10 per group: calculate the standard error of the difference as sqrt(s_G^2/10 + s_R^2/10) and derive the t-statistic for each dataset. If any dataset yields p ≥ 0.05, the claim of significant superiority is not supported; the headline conclusion must be revised unless statistically meaningful margins are demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports GQWformer vs RWC as mean ± std across 10-fold CV: MUTAG 95.2±3.0 vs 94.7±3.7, PTC 76.7±4.2 vs 74.4±4.4, PROTEINS 80.7±4.3 vs 79.4±4.7, IMDB-B 79.3±1.7 vs 78.8±3.1, IMDB-M 55.3±2.1 vs 54.5±2.9. For each dataset, the standard error of the difference (approximately sqrt(s1^2/10 + s2^2/10), which ignores the positive correlation of paired folds) yields t-values of 0.33, 1.20, 0.64, 0.42, and 0.66 respectively. None of these differences reach statistical significance at conventional levels. The paper explicitly states that GQWformer 'consistently and significantly outperforms RWC on all five datasets by a large margin' (Experiments section). This is an empirical claim directly contradicted by the paper's own numbers. Since the central claim of the abstract is that GQWformer 'outperforms existing state-of-the-art graph classification algorithms,' a non-significant difference does not establish this. Even if the quantum-walk encoding were fully specified and reproducible, the reported evidence would not justify the headline conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GQWformer, a graph transformer that injects structural information into self-attention via attribute-aware discrete-time quantum walks (QWs). A learnable coin operator, constructed with Householder reflections from an attention function over node features, drives T steps of a multi-walker QW; the resulting sequence of matrices is used both as an additive attention bias (GQW-Attn, Eq. 10) and as input to a bidirectional GRU module (GQW-Recu). The model is evaluated on five TUDataset graph classification benchmarks, with a claim of consistent and significant improvement over the RWC baseline. The manuscript also includes an ablation study on PTC and a sensitivity analysis over walk length.","tokens_in":12538,"tokens_out":3408,"duration_ms":39289,"significance":"If the method were fully specified and the empirical gains were significant, the idea of using learnable quantum-walk encodings as structural biases for graph transformers would be a potentially interesting contribution to graph representation learning. The paper identifies a real limitation of standard graph transformers — the lack of attribute-aware structural inductive biases — and proposes a mechanism that is conceptually novel. However, the contribution as presented is not yet substantiated: the central empirical claim is contradicted by the paper's own reported error bars, the core quantum-walk encoding pipeline is underspecified at several load-bearing steps, and no code is provided to disambiguate the construction. The ablation and sensitivity studies are useful in intent but are reported without standard deviations or repeated runs, limiting their evidentiary value.","major_comments":[{"comment":"The claim that GQWformer 'consistently and significantly outperforms RWC on all five datasets by a large margin' is not supported by the numbers in Table 1. The accuracy differences over RWC are: MUTAG 95.2±3.0 vs 94.7±3.7 (diff 0.5), PTC 76.7±4.2 vs 74.4±4.4 (diff 2.3), PROTEINS 80.7±4.3 vs 79.4±4.7 (diff 1.3), IMDB-B 79.3±1.7 vs 78.8±3.1 (diff 0.5), IMDB-M 55.3±2.1 vs 54.5±2.9 (diff 0.8). Using the reported standard deviations in a two-sample comparison with n=10, the resulting t-values are approximately 0.33, 1.20, 0.64, 0.42, and 0.66 respectively; none approach conventional significance levels. No significance tests, confidence intervals, or seed-level variance information are reported. Furthermore, the baseline numbers appear to be copied from prior publications rather than obtained by rerunning the baselines under identical conditions; this is not stated in the paper. Since the abstract's headline claim is that GQWformer 'outperforms existing state-of-the-art graph classification algorithms,' the reported evidence does not establish this.","section":"Experiments, Table 1"},{"comment":"The construction of the encoding matrices {M0, ..., MT} is not specified. The paper states that 'by summing the squares of the spin states of each superposition state, we derive a sequence of matrices' and then treats MT as a structural encoding matrix whose (i,j)-th element p_ij 'encodes the distance between node vi and node vj.' No equation defines M_t from the superposition tensor |ϕt⟩ ∈ C^{n×n×d}, nor is it stated whether p_ij is real-valued, normalized, or bounded. Since p_ij is added directly to the attention logit q_i^T k_j in Eq. (10) without any scaling or calibration, the magnitude and sign of p_ij can dominate or be negligible relative to the semantic term depending on the unspecified construction. Without a precise definition of M_t and p_ij, the central mechanism of GQW-Attn is not well-defined and the method cannot be reproduced.","section":"Method, 'Attribute-aware Graph Quantum Walk'"},{"comment":"The quantum-walk update is underspecified for graphs whose nodes have degrees smaller than the maximum degree d. The coin space H_c is defined as the span of {|i⟩, i ∈ [1,...,d]}, and in Eq. (9) g(vi) is computed against Xi_d, a matrix with d rows equal to the feature of vi. For a node of degree deg(vi) < d, it is not explained how the attention function a handles nonexistent neighbors, whether features are padded, or how the Householder coin C_i = I - 2g(vi)g(vi)^T/(g(vi)^T g(vi)) is formed when the effective dimension of g(vi) is deg(vi) rather than d. Similarly, the shift operator S is said to 'swap the states of vertices connected by edges,' but for a node with degree less than d, the coin basis includes directions that do not correspond to any edge; the action of S on those components is not defined. These are not cosmetic details: they determine whether the walk is a well-defined unitary process on the given graph.","section":"Method, Eqs. (8)-(9), and shift operator"},{"comment":"The paper repeatedly describes the QW encoding as an 'inductive bias' that provides 'tailored inductive bias' and 'structural bias.' However, the coin operator C_i is generated from g(vi), which depends on the learnable weight matrix W (Eq. 9), and W is trained end-to-end on the classification objective. Thus the structural encoding is fitted to the labels rather than being a fixed, independent prior. This weakens the conceptual claim that QWs inject an a priori structural inductive bias; the encoding is better described as a learned structural feature. This issue should be explicitly acknowledged, and the ablation rows 'vanilla QW' and 'QW(inv)' only partially address it because they compare against other learned or fixed coin variants, not against the absence of the QW bias with the same number of parameters.","section":"Method, 'Attribute-aware Graph Quantum Walk'; Introduction; Ablation"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and grammatical errors, including 'chanllenging', 'Auther', 'intrinsic', 'similarly', 'infromation', 'precess', 'oppsite', 'datsets', 'captical', 'drived', and 'concatebate'. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The self-attention formulation in Eq. (1) is nonstandard: a_i^l is a weighted sum of value vectors v_j^(l-1), but then h_i^l = a_i^l + γ h_i^(l-1) omits the usual output projection and residual connection structure. The notation also overloads a_i^l as an attention vector rather than a scalar attention weight. Clarify the exact transformer variant being used.","section":"Preliminary, Eq. (1)"},{"comment":"The ablation results are reported as single accuracy numbers without standard deviations or repeated trials. Given that the main results carry ±1.7 to ±4.3 percentage point standard deviations, single-run ablations are not informative for comparing modules or encoding variants.","section":"Ablation Studies, Table 2"},{"comment":"The sensitivity analysis varies walk length from 3 to 8 on PTC only and reports one run per setting. The differences, e.g., 76.7 at length 4 vs 73.0 at length 3, are not accompanied by variance information, so it is unclear whether the pattern is stable or noise.","section":"Sensitivity Analysis, Table 3"},{"comment":"The virtual-node readout is mentioned briefly but not specified: how the virtual node is initialized, whether its features are learned, and how its hidden state is obtained after the GQWformer blocks are not described.","section":"Readout"}],"recommendation":"reject","confidential_remarks":"The paper's central empirical claim is not supported by its own reported statistics, and the core algorithmic construction is not fully specified. The baseline column in Table 1 appears to reproduce previously published results without stating that these numbers were taken from the literature, and no code is provided; a non-expert reader cannot implement the method from the text. If the authors can supply a complete formal specification of the QW encoding, rerun baselines under identical conditions, and report significance tests or seed-level variance, the work could be resubmitted as a new contribution. As it stands, I do not see how the current manuscript can be accepted or meaningfully revised within a normal revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick version. This is basically Random Walk Conformer (RWC) with the random-walk structural encoding swapped for a learnable discrete-time quantum walk, and the RW-Conv local module replaced by a bidirectional GRU over the walk sequence. That exact combination is new as far as I can tell, and the paper cites the relevant prior work (RWC, Dernbach et al., Rohde et al.) properly.\n\nWhat it does well: the idea of generating the coin operator from node features via a Householder reflection is a clean way to make the walk attribute-aware, and the ablation on PTC gives some evidence that both the attention bias and the recurrent module contribute. Running a sensitivity analysis over walk length is also the right instinct.\n\nThe soft spots are serious. The core encoding step is never defined. The paper says 'by summing the squares of the spin states of each superposition state, we derive a sequence of matrices {M0,...,MT}' and then uses the (i,j) entry of MT as a distance in an attention bias, but it never gives the formula for Mt, never explains how the shift operator handles nodes with degree less than the maximum degree d, never specifies how the Householder coin is built for variable-degree neighborhoods, and never justifies why pij is a calibrated additive bias. As written, the central mechanism is not reproducible.\n\nThe empirical claim is also unsupported. On all five datasets the accuracy difference over RWC is within one reported standard deviation. A quick t-test on the 10-fold means (sqrt(s1^2/10 + s2^2/10)) gives t-values of 0.33, 1.20, 0.64, 0.42, and 0.66. None are significant. The text nevertheless says GQWformer 'consistently and significantly outperforms RWC on all five datasets by a large margin.' That is contradicted by the paper's own table. There are no significance tests, no seed variance, and the baselines appear to be taken from prior papers rather than re-run under identical conditions.\n\nThere is also a conceptual wrinkle: the 'inductive bias' is trainable. W in g(vi) is learned on the classification objective, so the structural encoding is fitted to the labels. That does not automatically invalidate the approach, but it undercuts the narrative that the quantum walk independently captures structure.\n\nBottom line: this is a paper for graph-transformer researchers who want to see one more way to inject structural bias into attention, but only after the encoding is fully specified and the experiments are redone with proper significance testing. In its current state the headline claim does not hold up. I would not send it to peer review; I'd return it to the authors with a request to define Mt, handle the degree mismatch, report multiple seeds with significance tests, and compare against RWC under a shared protocol.","headline":"A quantum-walk variant of Random Walk Conformer whose reported gains over RWC are within noise on every dataset, and whose core encoding step is undefined.","tokens_in":13063,"tokens_out":3799,"would_cite":false,"duration_ms":39068,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GQWformer injects attribute-aware quantum-walk distances into transformer attention and reports state-of-the-art graph classification accuracy on five benchmarks.","keywords":["Graph Transformer","Quantum Walk","Structural Encoding","Inductive Bias","Graph Classification","Self-Attention","Recurrent Module","Attribute-Aware Coins"],"falsifier":"Compute the exact map from $|\\phi_t\\rangle$ to $M_T$ as the authors intend and re-run the five classification benchmarks under the stated ten-fold protocol; the central claim is false if the reported accuracies cannot be reproduced, or if zeroing out the quantum-walk bias $p_{ij}$ changes accuracy by less than one standard deviation.","tokens_in":12077,"feed_emoji":"⚛️","tokens_out":8041,"duration_ms":75676,"temperature":0.7,"pith_summary":"GQWformer is a graph transformer for graph classification whose central claim is that discrete-time quantum walks, made sensitive to node attributes through learnable feature-dependent coin operators, can supply the structural inductive bias that plain self-attention lacks. The paper argues that this attribute-aware quantum-walk encoding captures both topology and node features, and that adding the resulting pairwise distances as an additive bias inside the attention softmax, together with a bidirectional recurrent module over the walk sequence, improves classification. On five public benchmarks the authors report accuracies of 95.2% on MUTAG, 76.7% on PTC, 80.7% on PROTEINS, 79.3% on IMDB-B, and 55.3% on IMDB-M, each above the random-walk transformer baseline they compare against. If correct, this establishes quantum-walk encodings as a practical structural prior for graph transformers rather than a purely theoretical device.","feed_headline":"Quantum-walk graph transformer beats prior graph classifiers","feed_subtitle":"Attribute-sensitive quantum-walk distances added to attention improve accuracy on five benchmarks.","key_machinery":"The attribute-aware graph quantum walk is the load-bearing object. It is a discrete-time quantum walk on an arbitrary graph with a position Hilbert space over nodes and a coin Hilbert space of dimension equal to the maximum degree; the coin at each node is a Householder reflection built from a trainable attention vector $g(v_i)$ over that node's neighbors, making the walk sensitive to node features as well as topology. The paper's key identity is the evolution $U = S(I \\otimes C)$, where $C$ is the collection of feature-dependent coin operators and $S$ is the shift operator; iterating $U$ for $T$ steps yields the superposition tensor $|\\phi_t\\rangle$, and the extracted pairwise bias matrix $M_T$ is what carries structural information into the self-attention bias of Eq. (10).","core_discovery":"The central claim is that a graph transformer can be made structure-aware by running multiple non-interacting discrete-time quantum walks on the node-attributed graph, one starting at each node, and using the resulting superposition sequence as an inductive bias. A learnable, feature-dependent coin operator $C_i = I - 2g(v_i)g(v_i)^\\top/(g(v_i)^\\top g(v_i))$ is generated from neighbor attention scores, the walk is evolved for $T$ steps, and the squared spin amplitudes are aggregated into matrices $\\{M_0, \\dots, M_T\\}$. The matrix $M_T$ is treated as a pairwise distance between nodes and added as a bias $p_{ij}$ inside the attention softmax, so attention weights reflect both semantic similarity $q_i^\\top k_j$ and the quantum-walk distance. A bidirectional GRU then consumes the sequence $\\{M_0, \\dots, M_T\\}$ to strengthen local, order-sensitive information, and its pooled output is combined with the attention output. The paper reports that this design outperforms state-of-the-art graph classifiers on MUTAG, PTC, PROTEINS, IMDB-B, and IMDB-M.","pith_inferences":["If the encoding is as transferable as it appears, the same $M_T$ bias could be dropped into other attention-based graph models, extending the result beyond the exact GQWformer stack.","A classical diffusion encoding with a learnable pairwise bias might reproduce most of the gain; comparing the two would isolate what the quantum machinery adds.","Because the paper fixes $T$ and the block count for all datasets, per-dataset tuning of walk length and number of blocks could push the reported accuracies higher.","The reported gap over the random-walk transformer baseline is sharpest on PTC, so a replication study focused on PTC would be the cheapest way to test the mechanism's contribution."],"forward_implications":["Graph transformers gain a principled structural bias by feeding quantum-walk distances into the attention bias, so node pairs that are close in the quantum-walk metric receive different attention than pairs with similar features but different topology.","The walk length $T$ controls the model's receptive field; the sensitivity analysis shows performance peaks at $T=4$ on PTC, so choosing $T$ is a real hyperparameter affecting the global-local balance.","The ablation results imply both modules are needed: GQW-Attn captures global pairwise structure and GQW-Recu captures local, order-sensitive structure, and removing either one lowers PTC accuracy.","Attribute-aware coins matter: the proposed coin function outperforms both feature-blind and invariant quantum-walk encodings on PTC.","Consequently, quantum-walk encodings are a viable alternative to classical random-walk or shortest-path encodings in graph transformers, not just a theoretical curiosity."],"supporting_citations":[{"why":"Serves as the main baseline and source of the experimental protocol (ten-fold cross-validation, virtual-node readout); the random-walk structural encoding this method extends and outperforms.","marker":"(Yeh, Chen, and Chen 2023)"},{"why":"Supplies the transformer self-attention mechanism whose lack of graph inductive bias motivates the quantum-walk encoding.","marker":"(Vaswani et al. 2017)"},{"why":"Provides the multi-walker discrete-time quantum walk on arbitrary graphs, used to run one walker per node.","marker":"(Rohde et al. 2011)"},{"why":"Introduces feature-dependent coin quantum walk neural networks, the earlier attribute-aware coin idea this paper adapts and diverges from.","marker":"(Dernbach et al. 2019)"},{"why":"Supplies the discrete-time quantum walk formulation on general graphs, including the position and coin Hilbert spaces used in the encoding.","marker":"(Kendon 2006)"},{"why":"Provides the benchmark datasets used for the five graph classification experiments.","marker":"(Morris et al. 2020)"},{"why":"Introduces graph recurrent networks, the basis for the GQW-Recu module that processes the quantum-walk sequence.","marker":"(Huang et al. 2019)"}],"fun_headline_variants":["Quantum walks give graph transformers structural awareness","Graph transformer with quantum-walk bias beats state-of-the-art on five sets","Quantum-walk distances improve graph transformer accuracy","Structure-aware transformer via quantum walks outperforms graph classifiers","Quantum walk attention boosts graph classification benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction assumes that a well-defined pairwise distance matrix can be extracted from the quantum-walk superposition and used as a calibrated attention bias, but the paper does not specify how $M_t$ is computed from $|\\phi_t\\rangle$, how the shift operator handles nodes with degree below the maximum, or how the feature-dependent coin is built for variable-degree neighborhoods.","fun_headline_variants_meta":{"raw":{"variants":["Quantum walks give graph transformers structural awareness","Graph transformer with quantum-walk bias beats state-of-the-art on five sets","Quantum-walk distances improve graph transformer accuracy","Structure-aware transformer via quantum walks outperforms graph classifiers","Quantum walk attention boosts graph classification benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000897,"raw_usage":{"total_tokens":3888,"prompt_tokens":989,"completion_tokens":2899,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":2826}},"tokens_in":605,"tokens_out":2899,"duration_ms":22580,"temperature":1.0,"reasoning_tokens":2826,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:38:39.444590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact map from $|\\phi_t\\rangle$ to $M_T$ as the authors intend and re-run the five classification benchmarks under the stated ten-fold protocol; the central claim is false if the reported accuracies cannot be reproduced, or if zeroing out the quantum-walk bias $p_{ij}$ changes accuracy by less than one standard deviation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the main baseline and source of the experimental protocol (ten-fold cross-validation, virtual-node readout); the random-walk structural encoding this method extends and outperforms."},{"cited_title":"P.; Schreiber, A.; S tefa n \\'a k, M.; Jex, I.; and Silberhorn, C","cited_arxiv_id":null,"evidence_quote":"Provides the multi-walker discrete-time quantum walk on arbitrary graphs, used to run one walker per node."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces feature-dependent coin quantum walk neural networks, the earlier attribute-aware coin idea this paper adapts and diverges from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the discrete-time quantum walk formulation on general graphs, including the position and coin Hilbert spaces used in the encoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces graph recurrent networks, the basis for the GQW-Recu module that processes the quantum-walk sequence."}],"review_version":1}