{"id":"582db7c1-f7ad-40d5-be16-ce5f102ad210","arxiv_id":"2411.13520","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A quantum orthogonal neural network attention layer matches a classical ViT on quark-gluon jet classification (test AUC ~0.737 vs ~0.739), with no demonstrated efficiency gain.","lead":"This paper builds a hybrid quantum-classical vision transformer that uses quantum circuits inside the attention mechanism to classify quark and gluon jets from CMS Open Data. It reports that the quantum version matches a small classical transformer in accuracy, but does not show a real speed or scalability benefit.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. 2.2 unary loader is internally inconsistent with the stated RBS matrix, so the attention coefficients used in the central comparison may not be the claimed linear forms.","rationale":"The reader's weakest_assumption correctly targets the unary amplitude loader, but identifies nonnegativity as the main limitation. The more precise and more severe problem is a sign inconsistency in the loader as written: even for nonnegative inputs, the RBS matrix in Sec. 2.2 produces the wrong sign on the second amplitude. This is an internal inconsistency, not merely an underspecified edge case, and it directly undermines the central empirical claim because the attention coefficients of Fig. 5 depend on the loaded amplitudes. The concrete check is a two-line simulation that settles whether the printed RBS convention is actually used. If the check fails, the paper needs either a corrected RBS convention or a different loader description before the QViT results can be interpreted. Since the paper is otherwise a plausible small-scale study and the issue is fixable with a clarification, the reader's CONDITIONAL verdict remains appropriate; no verdict adjustment is needed.","tokens_in":7364,"tokens_out":12337,"duration_ms":130384,"concrete_test":"Implement the 2-qubit loader of Sec. 2.2 exactly as defined: start with |0>|0>, apply X to the first qubit, then RBS(θ) with the matrix printed in Sec. 2.2. For x=(1/√2, 1/√2), compute α0=arccos(x0)=π/4 and compare the resulting state to x. If the |01> amplitude is −1/√2, the loader fails; repeat for a vector with a negative last component to test the nonnegativity limitation. Passing both checks would clear the concern; failing means the paper's attention circuit cannot compute the claimed coefficients.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (quantum attention as effective as classical) requires the attention circuit of Fig. 5 to compute x_i^T W x_j from patch embeddings. That computation relies on the unary amplitude loader of Sec. 2.2. As written, the loader is inconsistent with the RBS gate definition. The RBS matrix in Sec. 2.2 maps |10> to cosθ|10> − sinθ|01>. Starting from X|0...>, after the first RBS the state is cos α0|10...> − sin α0|01...>. The recursive formula sets α0 = arccos(x0), implying the second amplitude should be +sin α0, but the circuit yields −sin α0. Thus even a nonnegative vector such as x = (1/√2, 1/√2) is loaded as (1/√2, −1/√2), not x. If this sign convention is actually used, every attention coefficient is computed from a corrupted embedding, so the AUC comparison in Table 1 does not test the proposed quantum attention. The paper gives no procedure for handling signs, negative components, or choosing an alternative RBS orientation, and no code is provided to disambiguate the convention.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid quantum-classical vision transformer (QViT) in which the query–key attention coefficients are computed by a quantum circuit built from reconfigurable beam-splitter (RBS) gates, forming a quantum orthogonal neural network (QONN). The model is trained and evaluated on 50,000 CMS Open Data jet images for quark–gluon classification, with the class token and two jet-level features passed to a classical multilayer perceptron head. The central empirical claim is that the QViT matches a classical ViT in performance, with test AUC 0.7369 vs 0.7385 and test accuracy 0.6755 vs 0.6788 (Table 1). The paper also claims that the quantum attention mechanism is as effective as the classical mechanism and that the approach offers promising scalability.","tokens_in":7544,"tokens_out":5007,"duration_ms":51470,"significance":"If the implementation is correct and reproducible, the paper would provide a useful benchmark for hybrid quantum attention in HEP jet classification, with the virtue of using a public dataset and an independently implemented classical baseline. The claimed AUC parity is a concrete, falsifiable result. The paper, however, does not ship code, does not report statistical uncertainties, and contains an internal inconsistency in the unary loader/RBS sign convention that prevents verification of the central claim. The strengths are the use of public CMS Open Data, the explicit classical baseline comparison, and the clear identification of the orthogonal-layer parameter count.","major_comments":[{"comment":"The stated RBS matrix in Section 2.2 maps the basis state |10> to cos(theta)|10> - sin(theta)|01>. Starting from the initialized state X|0...> = |10...>, the cascade of RBS gates therefore produces a negative amplitude on the second unary basis state. With the recursive angle formula alpha_0 = arccos(x_0), a valid nonnegative input such as x = (1/sqrt(2), 1/sqrt(2)) is loaded as (1/sqrt(2), -1/sqrt(2)) rather than x itself. Unless the implemented circuit uses the opposite RBS sign convention, every attention coefficient computed by the circuit in Figure 5 is a function of a sign-corrupted embedding, so the AUC comparison in Table 1 does not test the proposed quantum attention mechanism. The manuscript gives no procedure for handling negative components or for choosing an alternative RBS orientation.","section":"Section 2.2, RBS gate matrix and vector loading circuit"},{"comment":"The circuit is claimed to compute |x_i^T W x_j|^2, but no derivation links the measurement probability of the first qubit in state |1> to this quantity, nor is it explained how loading |x_j>, applying W, and loading <x_i| implements the transpose operation. Additionally, the attention mechanism in Section 2.1 uses softmax(QK^T / sqrt(d_k)), a linear inner product, whereas the quantum circuit outputs a squared absolute value. Without an explicit mapping from the circuit output to the attention weights and a specification of how the softmax and scaling are applied, the forward pass of the model is undefined, and the comparison with the classical ViT is not a like-for-like test.","section":"Section 2.2, 'Attention Coefficient Circuit' and Figure 5"},{"comment":"The central claim that the QViT is 'as effective' as the classical ViT rests on test AUC 0.7369 vs 0.7385 and test accuracy 0.6755 vs 0.6788. No error bars, multiple seeds, or significance tests are reported, and no code is provided to reproduce the single run. Given that the differences are well within typical stochastic variation for a 15-epoch training run, the evidence as presented does not support the claim beyond an anecdotal single-run observation.","section":"Table 1 and Section 4"},{"comment":"The unary amplitude loader requires the input vector to be normalized and to have nonnegative components, as the recursive angles are defined through arccos of quotients of the vector entries. Patch embeddings are outputs of a learned linear projection E and are not constrained to be nonnegative or within [0,1]. The paper states that normalization does not degrade performance but does not address the sign restriction or provide a procedure for out-of-range values. Consequently, the loader as written cannot be applied to general patch embeddings, which is load-bearing because the attention coefficients in Figure 5 are computed from those embeddings.","section":"Section 2.2, vector loading circuit applicability"}],"minor_comments":[{"comment":"The heading 'Aknowledgments' should be spelled 'Acknowledgments'.","section":"Section 6"},{"comment":"The author name appears inconsistently as 'Alessandro Tesi' in the author list and 'Alessandro Tessi' in the acknowledgment footnote.","section":"Author line and footnote"},{"comment":"The figure shows a decomposition of the RBS gate using Hadamard, controlled-Z, and Ry rotations, but the text does not verify that this decomposition realizes the matrix given in Section 2.2; a short algebraic check would help.","section":"Figure 2"},{"comment":"The classical attention formula uses a scaling by sqrt(d_k), but the quantum attention coefficient circuit does not mention any such scaling; the relationship between the two should be clarified.","section":"Section 2.1 and Section 3.2"},{"comment":"The statement that 'the simulation requires 26 x 26 attention circuits per self-attention block' is not explicit about whether this is per image and per epoch; please clarify the counting.","section":"Section 3.2"},{"comment":"Reference [1] is a webpage without a stable identifier; please provide a persistent DOI, CERN document number, or accession date and URL.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closely related to the authors' prior QViT papers (refs [5,6]), and the incremental novelty here is the QONN-based attention layer applied to CMS Open Data. The main issue is that the circuit description and the empirical claim cannot be reconciled without code: the sign inconsistency in the unary loader, the undefined mapping from the circuit output to attention weights, and the absence of uncertainty estimates collectively prevent verification of the central result. If the authors can provide code and correct the circuit description, the paper could be salvageable as a benchmark study; otherwise the central claim is unsupported. I would also flag that the tiny performance differences in Table 1 should be accompanied by error bars or repeated-seed statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, the paper is a straightforward benchmark: a quantum vision transformer with QONN pyramid attention versus a classical ViT on CMS quark/gluon jet images. Test AUC is 0.7369 versus 0.7385, essentially a tie. Second, the quantum attention circuit as written does not compute the stated dot product. The RBS matrix in Sec. 2.2 puts -sin on the |10> to |01> crossing, but the recursive loader assumes a positive sine. For x = (1/sqrt2, 1/sqrt2), the circuit loads (1/sqrt2, -1/sqrt2), not x. So the attention coefficients in the central comparison are likely computed from corrupted embeddings. The stress-test note holds up. Also, the paper says the measured quantity is |x_i^T W x_j|^2, a squared coefficient, while the classical baseline uses softmax attention; that is a different model, and the text never reconciles the two.\n\nWhat the paper does well: the authors run a genuine head-to-head training comparison, use public CMS Open Data, report both validation curves and test metrics, and keep the setup simple. The self-citations to earlier QViT papers are natural; this is a continuation, not a disguised repeat. The prose is readable and the computational-cost caveats are at least stated, even though the abstract's \"promising scalability\" is not measured anywhere.\n\nSoft spots, in proportion: (1) no error bars or multiple seeds; a 0.0016 AUC gap is noise without a variance estimate. (2) no code and no verified data loader; the sign issue cannot be resolved from the manuscript. (3) the HL-LHC scalability claims are speculative, and the paper's own Sec. 3.2 says the simulation is computationally prohibitive. (4) the classical baseline is a weak single-block ViT with 25 patches, so \"matching the classical ViT\" is a low bar.\n\nWho this is for: the QML-for-HEP subgroup, and maybe people tracking hybrid vision transformers. It is not a result that changes practice. My recommendation: conditional acceptance at best, and only if the authors release code, fix or clarify the loader convention, add error bars, and drop or substantiate the scalability claims. As it stands, I would not cite it, but it is worth a serious referee's time because the flaw is concrete and fixable.","headline":"A clean benchmark undermined by a sign inconsistency in the unary loader that likely invalidates the claimed quantum attention computation; worth a real referee only if code and error bars are supplied.","tokens_in":8169,"tokens_out":2829,"would_cite":false,"duration_ms":30162,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a","07.05.Mh"],"model":"deepseek-v4-flash","headline":"A quantum-enhanced vision transformer with orthogonal attention layers matches the classical model on quark-gluon jet classification (test AUC 0.7369 vs 0.7385).","keywords":["quantum vision transformer","quantum orthogonal neural network","quark-gluon jet classification","CMS Open Data","RBS gate","high-energy physics","quantum machine learning","attention mechanism"],"falsifier":"Run a statevector simulation of the trained QViT, record the attention coefficients actually produced from a batch of test jet images, and compare them with the classical formula $x_i^T W x_j$ evaluated on the same embeddings; any disagreement—expected if embeddings contain negative components—would show the quantum and classical attention mechanisms are not computing the same quantity.","tokens_in":7141,"feed_emoji":"⚛️","tokens_out":9871,"duration_ms":103937,"temperature":0.7,"pith_summary":"The paper aims to show that a vision transformer whose attention mechanism is built from quantum orthogonal neural networks (QONNs) performs as well as an ordinary classical vision transformer on a high-energy physics classification task. The test case is distinguishing quark-initiated from gluon-initiated jets using multi-detector images from CMS Open Data. On a 50,000-image subset, the quantum model reaches a validation AUC of about 0.675, effectively the same as the classical model, and a test AUC of 0.7369 versus 0.7385 for the classical transformer. The authors read this near-parity as evidence that quantum attention can be substituted for classical attention without loss of accuracy, while inheriting the stability and parameter efficiency of orthogonal transformations. If true, this opens a concrete route for quantum-enhanced machine learning at the High Luminosity LHC, where data volumes outpace classical resources.","feed_headline":"Quantum attention matches classical on quark-gluon jets","feed_subtitle":"A quantum-enhanced vision transformer on CMS Open Data achieves nearly identical jet-classification AUC.","key_machinery":"The object that carries the argument is the QONN layer: a pyramid-shaped circuit of RBS gates that implements an $n \\times n$ orthogonal matrix using $n(n-1)/2$ tunable angles. A unary amplitude loader encodes each normalized input vector into a superposition with only $n-1$ RBS gates, and each attention coefficient is produced by a circuit that loads the key vector, applies the orthogonal weight matrix $W$, and overlaps the result with the query vector, with the measured first-qubit probability supplying the attention score. The orthogonality of $W$ is what the paper credits for stable gradients and efficient parameterization in high-dimensional spaces.","core_discovery":"The central claim of the paper is that replacing the linear projection layers that generate queries, keys, and values in a vision transformer with quantum orthogonal layers yields a model that is as effective as the classical ViT. The QViT uses pyramidal circuits of reconfigurable beam splitter gates to implement orthogonal weight matrices, and a quantum circuit to compute each attention coefficient from loaded query and key states. On quark-gluon jet classification, the authors report test accuracy 0.6755, test AUC 0.7369, and validation AUC converging to roughly 0.675, against 0.6788, 0.7385, and roughly 0.675 for the classical model. They conclude that the quantum attention mechanism matches the classical one and that orthogonal quantum layers offer a scalable alternative for the data demands of the HL-LHC.","pith_inferences":["The attention circuit's output is naturally $|x_i^T W x_j|^2$, a squared, sign-free kernel, while classical softmax attention uses signed dot products; the parity result therefore shows that a network with this squared kernel can match the classical AUC, not that the two models compute the same attention function.","A clean ablation would replace the quantum orthogonal layer with a classical orthogonal layer (for example, a weight matrix constrained by QR decomposition) while keeping the same nonnegative unary-loading constraint; if that model also matches the classical ViT, the quantum hardware is not essential to the observed parity.","Repeating the 15-epoch training with several random seeds would show whether the 0.0016 test-AUC gap is within run-to-run variance; the paper reports a single run for each model.","Inspecting the distribution of learned patch embeddings would settle whether the loader's nonnegativity requirement clips or distorts the features; if it does, the reported comparison is between classical attention and a quantum circuit acting on a constrained, altered representation."],"forward_implications":["If quantum attention is as effective as classical attention, QONN-based vision transformers become a candidate architecture for HL-LHC-era jet classification, where the orthogonal layers' parameter efficiency could reduce training cost.","The reported numbers—test AUC 0.7369 versus 0.7385, accuracy 0.6755 versus 0.6788—provide a concrete benchmark that future quantum transformer proposals can be compared against on the same CMS Open Data task.","The architecture's linear-depth data loading and $n(n-1)/2$-parameter orthogonal layers imply that the quantum resources per image grow with the square of the number of patches (25 patches require 26×26 attention circuits), so the scalability argument is architectural rather than a demonstrated speedup on current simulators.","Replacing generic variational circuits with orthogonal quantum layers should improve gradient behaviour during training, which is the main practical motivation the authors offer for preferring QONNs over other quantum attention designs."],"supporting_citations":[{"why":"supplies the quantum orthogonal neural network framework and the RBS-based orthogonal layers the QViT embeds in its attention mechanism.","marker":"[11]"},{"why":"demonstrates quantum RBS circuits for neural networks and image classification, the methodological basis for the circuit design.","marker":"[12]"},{"why":"prior hybrid quantum vision transformer for high-energy physics event classification that this architecture builds on and compares against.","marker":"[5]"},{"why":"prior quantum vision transformer applied to quark-gluon classification, establishing the task and the performance context.","marker":"[6]"},{"why":"defines the quantum vision transformer architecture that the paper extends with quantum orthogonal neural networks.","marker":"[7]"},{"why":"defines the classical vision transformer that serves as the performance baseline.","marker":"[8]"},{"why":"supplies the self-attention formula whose projection layers the quantum orthogonal layers replace.","marker":"[9]"},{"why":"provides the CMS Open Data jet-image dataset used for training and evaluation.","marker":"[13]"}],"fun_headline_variants":["Quantum vision transformer ties classical on jet tagging","Quantum orthogonal attention matches classical on CMS jets","QONN-based attention rivals classical in quark-gluon classification","Hybrid quantum-classical transformer matches ViT on jets","Quantum attention for jet classification: on par with classical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the circuit that turns data into quantum states can exactly represent any patch embedding produced by the network; the circuit's angle formula requires every entry to be nonnegative and no intermediate factor to vanish, but the embeddings are unconstrained real numbers, so this premise is not guaranteed by the architecture.","fun_headline_variants_meta":{"raw":{"variants":["Quantum vision transformer ties classical on jet tagging","Quantum orthogonal attention matches classical on CMS jets","QONN-based attention rivals classical in quark-gluon classification","Hybrid quantum-classical transformer matches ViT on jets","Quantum attention for jet classification: on par with classical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1498,"prompt_tokens":833,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":589}},"tokens_in":449,"tokens_out":665,"duration_ms":6456,"temperature":1.0,"reasoning_tokens":589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:18:53.569046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a statevector simulation of the trained QViT, record the attention coefficients actually produced from a batch of test jet images, and compare them with the classical formula $x_i^T W x_j$ evaluated on the same embeddings; any disagreement—expected if embeddings contain negative components—would show the quantum and classical attention mechanisms are not computing the same quantity.","supporting_citations":[{"cited_title":"Classical and quantum algorithms for orthogonal neural networks, 2022","cited_arxiv_id":null,"evidence_quote":"supplies the quantum orthogonal neural network framework and the RBS-based orthogonal layers the QViT embeds in its attention mechanism."},{"cited_title":"Quantum methods for neural networks and application to medical image classification","cited_arxiv_id":null,"evidence_quote":"demonstrates quantum RBS circuits for neural networks and image classification, the methodological basis for the circuit design."},{"cited_title":"Unlu, Marçal Comajoan Cara, Gopal Ramesh Dahale, Zhongtian Dong, Roy T","cited_arxiv_id":null,"evidence_quote":"prior hybrid quantum vision transformer for high-energy physics event classification that this architecture builds on and compares against."},{"cited_title":"Forestano, Sergei Gleyzer, Daniel Justice, Kyoungchul Kong, Tom Magorsch, Konstantin T","cited_arxiv_id":null,"evidence_quote":"prior quantum vision transformer applied to quark-gluon classification, establishing the task and the performance context."},{"cited_title":"Quantum vision transformers","cited_arxiv_id":null,"evidence_quote":"defines the quantum vision transformer architecture that the paper extends with quantum orthogonal neural networks."},{"cited_title":"About the CMS Experiment, 2024","cited_arxiv_id":null,"evidence_quote":"provides the CMS Open Data jet-image dataset used for training and evaluation."}],"review_version":1}