{"id":"4ce997d6-6555-40f8-bc2c-a9d774d43207","arxiv_id":"2509.05924","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid CV-quantum neural network with a classical readout is shown in simulation to classify two- and three-mode states as separable or entangled with over 99% accuracy, and a universal-approximation theorem is stated for informationally complete readouts.","lead":"This paper trains a continuous-variable quantum neural network to classify quantum states as entangled or separable, reporting over 99% accuracy on simulated two- and three-mode states. It claims a performance gap over classical machine learning baselines that grows with system size, plus a theorem about the network's expressive power.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unreported matched-feature baseline and contradictory feature descriptions leave the claimed scalable performance gap unverifiable.","rationale":"I agree with the reader's weakest assumption. The claim that matters for the paper's thesis is not the universality theorem (which is a standard Stone–Weierstrass argument and is not the bottleneck), but the empirical 'scalable performance gap.' That gap is entirely dependent on how the classical baselines were configured. The manuscript explicitly promises a matched-feature comparison in §III.C.2 and never reports it; the Results and Conclusion give incompatible descriptions of what the baselines received. This is not a minor omission: if the baselines had the engineered negativities, the dataset labels (constructed via negativity checks in Algorithm 1) make the problem nearly linearly separable, so the reported ~75% accuracy is not credible; if they had the matched IC probabilities, then the text's 'comprehensive engineered features' description is wrong. Either way the reader cannot independently assess the central claim. The photon-loss comparison compounds this by omitting any noisy classical baseline. I see no issue with the theoretical approximation result itself; the problem is the missing evidence for the empirical headline. Therefore I recommend keeping the reader's REJECT verdict unchanged.","tokens_in":16563,"tokens_out":6543,"duration_ms":74914,"concrete_test":"Run the matched-feature baseline experiment specified in §III.C.2: for the identical 2- and 3-mode datasets, feed the classical SVM and MLP the exact probability vectors obtained from the fixed (untrained) CV-QNN circuit with IC readout, including the same finite-shot sampling (N_shots=1000) and photon-loss channel used for the quantum model, and report test accuracy with 95% CIs. If these matched-feature classical baselines match the CV-QNN's ~99% accuracy—or if the noiseless engineered-feature version already exceeds 99%—the observed gap is an artifact of unfair feature/input comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim—a >99% CV-QNN accuracy with a scalable gap over SVM/MLP that persists under photon loss—depends on a fair baseline comparison, and that comparison is not actually reported. Section III.C.2 promises a 'Direct Comparison (Matched Features)' where classical baselines receive exactly the IC-readout probability vectors fed to the CV-QNN's classical head. The Results (Table II, §IV.B) instead describe baselines given 'a comprehensive set of engineered features' (purity, entropies, all bipartite negativities, flattened density matrix), while the Conclusion asserts the baselines had 'the same post-measurement information as the quantum model's classical head.' These are different inputs. The discrepancy matters quantitatively: if engineered features include bipartite negativities, the dataset labels were constructed by requiring positive negativity for entangled states, so a threshold on max negativity should separate the classes almost perfectly; the reported collapse to ~75% would be implausible. If the baselines actually received matched IC features, the text misdescribes them. Either way, the central 'robust scalable performance gap' cannot be checked from the paper. The photon-loss experiment adds a second unfair asymmetry: the noisy quantum model is compared only to noiseless classical baselines (§IV.C), so even a large gap may reflect noise asymmetry rather than native advantage. These are not stylistic issues; they determine whether the headline claim is true.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid quantum-classical model for learning nonlinear entanglement witnesses from continuous-variable quantum data. The architecture combines a CV-QNN (interferometer, squeezers, displacements, Kerr gates) with an informationally complete POVM readout and a classical MLP head. The authors report numerical experiments on two- and three-mode state datasets (Gaussian, non-Gaussian, pure, mixed), claiming >99% accuracy with a 'robust scalable performance gap' over SVM and MLP baselines, robustness to photon loss under finite shots, and a universal approximation theorem for continuous witness functionals when the measurement is informationally complete.","tokens_in":16871,"tokens_out":4907,"duration_ms":53564,"significance":"If the empirical claims were well supported, the work would provide a useful benchmark for quantum machine learning in entanglement certification. The paper contains reproducible architectural details, finite-shot simulation, and statistical bootstrap CIs. However, the central scalability claim rests on a baseline comparison that is internally inconsistent and partially missing; the photon-loss comparison is asymmetric; and the theoretical theorem is a corollary of the classical universal approximation theorem. As presented, the significance of the claimed advantage cannot be assessed.","major_comments":[{"comment":"The paper promises a 'Direct Comparison (Matched Features)' in §III.C.2 but never reports it. The Results (Table II, §IV.B) say classical baselines were given 'a comprehensive set of engineered features' (purity, entropies, all bipartite negativities, flattened density matrix), while the Conclusion says they had 'the same post-measurement information as the quantum model's classical head.' These are different inputs. The matched-feature experiment is essential to isolate the effect of training the quantum circuit; its absence makes the reported performance gap impossible to verify.","section":"III.C.2, IV.B, Conclusion"},{"comment":"The entangled labels in Algorithm 1 are defined by requiring all bipartite negativities to be positive (step 9). The engineered feature vector in §III.C.2 includes those same negativities. A simple threshold on the maximum negativity would then classify the training/test sets almost perfectly, contradicting the reported 76% (SVM) and 74.5% (MLP) three-mode accuracies. Either the engineered features were not actually supplied to the baselines, or the dataset labels are not consistent with the feature description. Neither case supports the claimed comparison.","section":"Algorithm 1, §III.C.2"},{"comment":"The photon-loss experiment compares the noisy CV-QNN (with per-layer loss and Nshots=1000 shot noise) against noiseless classical baselines (SVM/MLP). Classical baselines are not subjected to the same photon-loss channel or shot noise. The observed gap under noise may therefore be an artifact of asymmetric noise treatment rather than a native advantage. A matched-noise comparison is required.","section":"IV.C, Fig. 6"},{"comment":"The theorem is essentially the classical universal approximation theorem applied to the image of an IC-POVM. The proof fixes a single Θ* with injective gΘ*, then invokes UAT for MLPs; the variational circuit U(Θ) and its trainability play no role. This does not establish any approximation advantage specific to CV-QNNs. The statement should be substantially qualified or replaced with a result that uses the variational degrees of freedom.","section":"Theorem II.1, Appendix A"}],"minor_comments":[{"comment":"The statement that non-overlapping 95% CIs imply p<0.001 is not justified; bootstrap CI non-overlap does not directly give a p-value. Please either report a proper significance test or remove the p-value.","section":"IV.B"},{"comment":"The reference to 'Algorithm 7' should be 'Algorithm 1'.","section":"Appendix C"},{"comment":"The symbol K is used both for the number of IC measurement settings and for the Kerr gate; consider different notation.","section":"II.C, Eq. (7)"},{"comment":"The paper would benefit from stating explicitly whether the classical baselines were given access to the full density matrix in the three-mode experiments; the text is inconsistent on this point.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"This manuscript's central numerical claim is not verifiable from the provided text: the matched-feature baseline experiment is promised but absent, and the engineered-feature description is inconsistent with the label-generation rule. The photon-loss experiment is also asymmetric. These are not stylistic but load-bearing. I recommend rejection, though the authors could resubmit after a major revision with fair comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the three-mode performance gap is the whole point of this paper, and I can't verify it from the text. The two-mode benchmark is clean and the universality theorem is correct, but it's just IC-POVM injectivity plus the classical UAT — the variational circuit does no work in the proof. The numerical claim is what would make this a useful result, and that's where the paper falls apart.\n\nOn the positive side, the data generation pipeline is described carefully (Algorithm 1), the training details are concrete, and the two-mode results look sensible: all methods near 99%, with a slight edge for the CV-QNN. The authors also made an explicit effort to include heterogeneous state families and to verify entanglement by negativity. That is real work.\n\nWhere it gets soft: Section III.C.2 promises a 'Direct Comparison (Matched Features)' where classical baselines get exactly the IC probability vectors that feed the CV-QNN's classical head. That experiment is never reported. The Results (Section IV.B) say the baselines were given 'a comprehensive set of engineered features' including the full density matrix and all bipartite negativities. The Conclusion, however, claims the baselines received 'the same post-measurement information as the quantum model's classical head.' These are different inputs, and the distinction is load-bearing. Worse, if the engineered features really included all bipartite negativities, a simple threshold on max negativity should separate the classes almost perfectly, because entangled states are admitted only if they have positive negativity and separable states are product states. The reported collapse to 74-76% is implausible under that description. The matched-feature experiment might fix this, but it's absent. The photon-loss comparison has a separate asymmetry: the noisy quantum model is compared only to noiseless classical baselines, so the gap there could just reflect noise asymmetry. Data and code are not released, so nothing can be checked independently.\n\nThe theory is fine but nearly tautological: an injective feature map plus the classical universal approximation theorem gives density in C(X). The authors themselves admit the variational circuit plays no role in the proof. That's a statement about expressivity, not a new framework.\n\nOverall, the paper has a solid two-mode proof-of-concept and a standard theory section, but the headline three-mode gap and the robustness claim are not supported by what's actually reported. A referee could fix this by requiring the matched-feature results, applying identical loss and shot noise to the classical baselines, and releasing data and code. As it stands, the central claim is unverifiable. I'd send it to review rather than desk reject, because the question is worth asking and the authors did real work — but I'd expect major revision, and I wouldn't cite it until the numbers are reproducible.","headline":"The two-mode demo and the theory are fine; the three-mode gap is the whole story and it rests on a baseline comparison the paper contradicts itself about and never actually reports.","tokens_in":17397,"tokens_out":4587,"would_cite":false,"duration_ms":48177,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P40","81P68"],"pacs":["03.65.Ud","03.67.Mn"],"model":"deepseek-v4-flash","headline":"A hybrid continuous-variable quantum neural network can be trained to act as a nonlinear entanglement witness, classifying separable versus entangled states with above 99% accuracy and keeping that accuracy in three-mode problems where clas","keywords":["entanglement witness","continuous-variable quantum neural networks","quantum machine learning","hybrid quantum-classical","multipartite entanglement","photon loss","universal approximation","state classification"],"falsifier":"Train the SVM and MLP on the same shot-noise-limited features from the untrained IC measurement circuit used for the quantum head; if either reaches about 99% accuracy on three-mode states, the claimed scalable performance gap disappears.","tokens_in":16434,"feed_emoji":"⚛️","tokens_out":6083,"duration_ms":62542,"temperature":0.7,"pith_summary":"This paper tries to establish that a continuous-variable quantum neural network (CV-QNN) can be trained as a nonlinear entanglement witness that classifies quantum states as separable or entangled directly from measurement data. The authors report above 99% classification accuracy on two- and three-mode datasets covering Gaussian and non-Gaussian, pure and mixed states, with a performance gap that widens as the system grows: classical SVM and MLP baselines drop to roughly 75% in the three-mode case. The paper also proves that, when the measurement readout is informationally complete, the hybrid architecture can approximate any continuous witness-like functional on compact sets of states. A sympathetic reader would care because separability is NP-hard in general, so a trainable, hardware-native witness that scales better than classical models could offer a practical route to entanglement certification on near-term photonic devices.","feed_headline":"Quantum neural net holds 99% accuracy where classical models collapse","feed_subtitle":"Hybrid quantum-classical witness holds 99% accuracy on three-mode entanglement, while SVM and MLP baselines drop to 75%.","key_machinery":"The core object is the learned witness functional W(ρ)=f_Φ(g_Θ(ρ)). The quantum stage is a two-layer CV-QNN built from four gate blocks per layer: an interferometer that mixes modes, single-mode squeezing and displacement that act as affine scaling and bias, and a Kerr gate that supplies non-Gaussian nonlinearity. The readout is an informationally complete POVM, implemented in simulation by measuring Fock-basis probabilities after several fixed pre-measurement unitaries; injectivity of this feature map is what lets the classical head distinguish any two distinct states. The classical head is a small MLP that maps the concatenated probability vector to a scalar witness value. The universal-ap","core_discovery":"The central claim is that a variational continuous-variable circuit composed of interferometers, squeezers, displacements, and Kerr gates, followed by an informationally complete measurement and a small classical neural head, learns a nonlinear witness functional W(ρ)=f_Φ(g_Θ(ρ)) that separates entangled from separable states with high accuracy. In numerical experiments, the model achieves 99.02% accuracy on two-mode states and 99.00% on three-mode states, while the best classical baseline drops from about 98% to 76%. The paper further claims that when the measurement map is injective, the witness functional class is dense in the continuous functions on any compact set of states, so the arch","pith_inferences":["If the classical baselines were instead given the same shot-noise-limited features from an untrained circuit (the matched-feature scenario the paper describes but does not report), the performance gap would likely shrink; the claimed quantum-native advantage may reduce to an information-asymmetry effect.","The universal-approximation theorem covers expressivity, not trainability; nothing in the paper rules out barren-plateau-like behavior or poor sample complexity as the number of modes grows.","The same pipeline could be pointed at harder certification tasks, such as bound entanglement or non-Gaussian state verification, where no simple analytical witness exists; the paper does not test these.","A direct hardware test on a photonic chip with loss could settle whether the loss robustness persists when the noise is not simulated by the same model used for training."],"forward_implications":["A learned nonlinear witness can classify both Gaussian and non-Gaussian, pure and mixed two- and three-mode states with over 99% accuracy using a finite 1000-shot readout.","Scaling from two modes to three modes leaves the CV-QNN accuracy essentially unchanged (99.0%) while the best classical baseline drops to roughly 76%, giving an empirical scalable gap.","At 10% per-layer photon loss, the three-mode quantum model still classifies above 97%, outperforming noiseless classical baselines across the tested loss range.","With an informationally complete readout, the hybrid architecture is dense in the continuous functions on any compact set of states, so it can in principle represent any continuous witness functional."],"supporting_citations":[{"why":"Defines the layered CV-QNN ansatz with interferometer, squeezing, displacement, and Kerr gates that the paper trains as the witness.","marker":"[18]"},{"why":"Guarantees that an informationally complete POVM gives an injective measurement map, which the approximation theorem relies on.","marker":"[21]"},{"why":"Supplies the classical universal approximation theorem that completes the density argument in the proof.","marker":"[27]"},{"why":"Provides the Fock-basis simulator used to produce all numerical results in the paper.","marker":"[28]"},{"why":"Provides the parameter-shift rule used to train the quantum circuit parameters.","marker":"[23-25]"},{"why":"Establishes that the general separability problem is NP-hard, motivating the learned-witness approach.","marker":"[3]"},{"why":"Gives the beamsplitter-based decomposition for the programmable interferometer used in the circuit.","marker":"[20]"}],"fun_headline_variants":["CV-QNN hits 99% on entanglement, classical drops to 76%","Hybrid quantum net learns nonlinear witness: 99% vs classical 76% on 3 modes","CV-QNN keeps 99% on 3-mode entanglement; classical falls to 76%","Scaling to three modes: quantum net holds 99%, classical drops to 76%","Quantum neural net learns entanglement witness with 99% accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The classical baselines must receive genuinely equivalent information and noise conditions; the paper reports the full-density-matrix engineered-feature results but not the promised matched-feature baseline, so the fairness of the comparison is the load-bearing assumption.","fun_headline_variants_meta":{"raw":{"variants":["CV-QNN hits 99% on entanglement, classical drops to 76%","Hybrid quantum net learns nonlinear witness: 99% vs classical 76% on 3 modes","CV-QNN keeps 99% on 3-mode entanglement; classical falls to 76%","Scaling to three modes: quantum net holds 99%, classical drops to 76%","Quantum neural net learns entanglement witness with 99% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001217,"raw_usage":{"total_tokens":4837,"prompt_tokens":730,"completion_tokens":4107,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":3997}},"tokens_in":474,"tokens_out":4107,"duration_ms":32874,"temperature":1.0,"reasoning_tokens":3997,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:48:36.933493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the SVM and MLP on the same shot-noise-limited features from the untrained IC measurement circuit used for the quantum head; if either reaches about 99% accuracy on three-mode states, the claimed scalable performance gap disappears.","supporting_citations":[{"cited_title":"Continuous- variable quantum neural networks.Physical Review Research, 1(3):033063, 2019","cited_arxiv_id":null,"evidence_quote":"Defines the layered CV-QNN ansatz with interferometer, squeezing, displacement, and Kerr gates that the paper trains as the witness."},{"cited_title":"Strawberry fields: A software platform for photonic quantum computing.Quantum, 3:129, 2019","cited_arxiv_id":null,"evidence_quote":"Provides the Fock-basis simulator used to produce all numerical results in the paper."},{"cited_title":"Classical complexity and quantum entanglement.Journal of Computer and System Sciences, 69(3):448–484, 2004","cited_arxiv_id":null,"evidence_quote":"Establishes that the general separability problem is NP-hard, motivating the learned-witness approach."},{"cited_title":"Optimal design for universal multiport interferometers.Optica, 3(12):1460–1465, 2016","cited_arxiv_id":null,"evidence_quote":"Gives the beamsplitter-based decomposition for the programmable interferometer used in the circuit."}],"review_version":1}