{"id":"4eb6e8bc-efde-4e58-a54b-edc2acc71849","arxiv_id":"2505.05249","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A quantum classifier trained with a classical surrogate network and mid-circuit ancilla resets is claimed to avoid barren plateaus and beat baselines, but the supporting proofs and table values are inconsistent.","lead":"This paper proposes training a quantum image classifier through a classical surrogate network that predicts the quantum circuit's outputs, while mid-circuit measurements on extra qubits add non-unitary steps. The authors claim this avoids barren plateaus and beats classical baselines, but the proofs and the reported benchmark numbers contain contradictions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 assumes the surrogate gradient can be made arbitrarily accurate, which is precisely what barren plateaus prevent; no sample-complexity analysis shows this is achievable with polynomial resources.","rationale":"The reader's weakest_assumption correctly identifies the surrogate-gradient accuracy assumption in Theorem 2. I independently trace the chain: Theorem 2 assumption (2) (Section II-D) requires ε_t arbitrarily small; Lemma 2 requires ε ≪ ||∇L||. The paper never shows that a polynomial-sample surrogate can achieve this in the 15-qubit, 6-layer setting. In a barren plateau, ||∇L|| is exponentially small and m(θ) is concentrated, so the surrogate gradient from sparse samples is dominated by sampling noise; the condition ε_t = o(||∇L||) is exactly the absence of a plateau. The discussion after Lemma 2 and after Theorem 2 even concede a fixed surrogate 'may introduce some bias' and that idealized simulations were used, but no bound links sample count to ε_t. This is a genuine gap, not a disagreement with consensus: the proof is internally conditional on an unproven oracle. I also note the empirical self-contradiction (Table I vs text for QNN Direct Grad on CIFAR-10) and CIFAR-100 where the classical CNN beats the surrogate QNN, which weakens the 'superior accuracy' claim. Credit is due where the paper has independent support: Lemma 1's Kraus/CPTP derivation is standard and correct, and gradient-free surrogate strategies are legitimate in general; the issue is specifically the barren-plateau claim. A numerical sample-complexity test would settle whether the assumption is practical. Given that the load-bearing assumption is unquantified and the reported results are internally inconsistent, the REJECT verdict stands.","tokens_in":9635,"tokens_out":4056,"duration_ms":42507,"concrete_test":"Run a sample-complexity check for the exact 15-qubit, 6-layer, 4-ancilla circuit used in the paper. Fix a random parameter point θ and a perturbation δ with ||δ|| ≈ η = 7×10^-4. Estimate m(θ+δ) − m(θ) using N = 10^3, 10^4, and 10^5 circuit evaluations, and compare the empirical surrogate gradient error ||∇S − ∇L|| against an exact finite-difference gradient. If the error stays above ||∇L||, or the measured difference is below the sampling noise floor, for polynomial N, Theorem 2's assumption fails. A complementary analytical check is to derive Var(∇m(θ)) for this architecture and show whether it decays exponentially in n·L; if it does, the sample count needed to reach ε_t = o(||∇L||) is exponential.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that bypassing quantum gradients 'mitigates plateau issues' rests on Lemma 2 and Theorem 2. Theorem 2 assumption (2) requires that at every iteration the surrogate gradient error ||∇S(t)(θ(t)) − ∇L(θ(t))|| ≤ ε_t with ε_t arbitrarily small, and Lemma 2 requires ε ≪ ||∇L||. This is the load-bearing gap. In a barren-plateau landscape the true gradient is exponentially small in n and L, and the measurement outcomes m(θ) are exponentially concentrated. A surrogate trained on a polynomial number of circuit evaluations can only resolve gradients above the sampling noise floor; certifying ε_t = o(||∇L||) would require detecting exponentially small differences in L, hence exponentially many samples. The text after Lemma 2 asserts the surrogate 'can fit through sparse samples that detect slight differences in L' without any scaling analysis, and Section II-D concedes that in practice a fixed surrogate 'may introduce some bias ε_t.' That concession is exactly the unproven part: the proof needs ε_t to be controllable, but the protocol gives no bound linking sample count, circuit size, and ε_t. Thus Theorem 2 does not prove convergence or barren-plateau avoidance; it assumes the desired property. The empirical section also undercuts 'superior accuracy': Table I reports 89.90% for QNN (Direct Grad) on CIFAR-10 while the text reports 86.31%, and on CIFAR-100 the classical CNN (63.63%) beats the surrogate QNN (58.65%).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid quantum-classical image classifier in which a classical neural network surrogate is trained to predict measurement outcomes of a parameterized quantum circuit with mid-circuit measurement and reset of ancillas, and the surrogate's classical gradient is used to update the circuit parameters without direct quantum gradient evaluation. The architecture also compresses input images via a CNN and projects features into circuit angles. The paper claims that this gradient-free approach circumvents barren plateaus, that the mid-circuit measurement-reset introduces useful nonlinearity, and that experiments on MNIST, CIFAR-10, and CIFAR-100 show superior accuracy relative to direct-gradient QNNs and classical baselines.","tokens_in":9932,"tokens_out":3992,"duration_ms":39421,"significance":"If the central claims were correct, the paper would offer a broadly applicable training framework for deep QNNs that avoids vanishing gradients while improving accuracy, which would be a meaningful contribution to variational quantum machine learning. The manuscript provides a concrete 15-qubit, 6-layer circuit design, a detailed experimental setup, and comparisons on three benchmark datasets, which is a useful empirical data point. However, the theoretical argument is not valid as stated: the key convergence theorem assumes the very property the paper claims to prove, and the empirical results partly contradict the abstract's superiority claim. The nonunitarity result is standard quantum channel theory and does not by itself establish the claimed nonlinear activation mechanism. The manuscript therefore does not currently substantiate its main claims.","major_comments":[{"comment":"The convergence theorem assumes that at each iteration the surrogate gradient error ||∇S(t)(θ(t)) − ∇L(θ(t))|| can be made arbitrarily small, but this assumption is exactly the property the paper claims to achieve. In a barren-plateau landscape the true gradient is exponentially small, so achieving ε_t ≪ ||∇L|| requires resolving exponentially small differences in the loss, which in turn requires exponentially many circuit evaluations. The manuscript provides no sample-complexity bound linking the number of circuit executions, the circuit size, and the attainable ε_t, and Section II-D itself concedes that a fixed surrogate 'may introduce some bias ε_t.' Consequently Theorem 2 does not prove convergence or plateau avoidance; it assumes the desired outcome.","section":"II-D, Theorem 2 assumption (2); II-C, Eq. (11)-(13)"},{"comment":"Lemma 2 assumes S(θ(t)) = L(θ(t)) and ||∇S(θ) − ∇L(θ)|| ≤ ε, whereas the surrogate is defined in Section II-C as approximating the measurement outcome vector m(θ,x), not the loss L. No argument shows that minimizing the MSE between S and m implies the required closeness of the surrogate's loss gradient to the true loss gradient. The descent lemma therefore does not apply to the procedure as actually described.","section":"II-C, Lemma 2 and Eq. (11)"},{"comment":"The text reports the QNN (Direct Grad) CIFAR-10 test accuracy as 86.31%, but Table I lists 89.90% for the same entry. In addition, the abstract claims 'superior accuracy compared to direct-gradient QNNs and classical baselines,' yet on CIFAR-100 the classical CNN achieves 63.63% while the surrogate QNN achieves 58.65%. These inconsistencies and the selective reporting materially weaken the empirical support for the paper's central claim.","section":"Table I and Section III-B, CIFAR-10 paragraph"},{"comment":"The claim that measuring and resetting ancillas induces 'nonlinear transformations' on the main qubits is not supported: the derived map Φ(ρ) = Σ_k V_k ρ V_k^† is a completely positive trace-preserving map and therefore linear on density operators. The paper does not establish that this channel yields the type of nonlinearity needed for enhanced expressivity; the output probabilities depend nonlinearly on the input state, but that is a generic property of quantum measurements rather than a new nonlinear activation mechanism.","section":"Lemma 1 and Theorem 1"}],"minor_comments":[{"comment":"The caption of Figure 5 says 'MNIST dataset,' but the plot and the surrounding text describe CIFAR-100; the caption should be corrected.","section":"Figure 5 caption"},{"comment":"The text states the CNN has 1.6M parameters and the NN has 2.1M parameters, while Table I reports 1.8M and 2.3M respectively; these numbers should be reconciled.","section":"Section III-B and Table I"},{"comment":"Reference [12] is cited for ReLU activation, but the listed paper by Householder (1941) is not a source for ReLU; the proper citation is Nair and Hinton (2010), which appears as [13], and the citation order should be corrected.","section":"References [12] and [13]"}],"recommendation":"reject","confidential_remarks":"The mismatch between the abstract's superiority claim and the CIFAR-100 result, together with the internal inconsistency in the CIFAR-10 direct-gradient accuracy, suggests the empirical claims require careful re-auditing before any resubmission. The theoretical gap is fundamental: the convergence proof assumes the vanishing-gradient problem is already solved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the load-bearing claim is not established. Theorem 2 assumes that at each iteration the surrogate gradient error can be made arbitrarily small, and in a barren-plateau landscape that is exactly what you cannot do with polynomial samples. The paper does not provide any sample-complexity bound connecting sample count, circuit size, and the surrogate error epsilon. So the conclusion that this scheme 'circumvents barren plateaus' is an assumption, not a result. The paper even concedes in Section II-D that a fixed surrogate 'may introduce some bias epsilon_t,' which is the unproven part dressed up as a theorem.\n\nWhat the paper does well: the Kraus operator derivation in Lemma 1 is standard but correct, and the writing is clear. The combination of classical surrogate gradients with mid-circuit measurement and reset is a natural package, and the idea of fitting a surrogate to measurement outcomes rather than to the loss is sensible. For a high-school student, this is a serious effort.\n\nWhere it falls apart: the internal inconsistency in the numbers is hard to overlook. Table I lists the direct-gradient QNN on CIFAR-10 as 89.90% test accuracy, but the text says 86.31%. On CIFAR-100, the classical CNN beats the surrogate QNN (63.63% vs. 58.65%), which directly contradicts the abstract's claim of 'superior accuracy compared to ... classical baselines.' Figure 5 is captioned as MNIST but the surrounding text describes CIFAR-100. No code, surrogate architecture, or training hyperparameters are provided, so the results are not reproducible.\n\nThe theoretical gap is not a minor fix. Lemma 2 requires the surrogate to approximate the true gradient to within epsilon, and Theorem 2 requires epsilon_t to be arbitrarily small. But in a barren-plateau landscape the true gradient is exponentially small, so detecting it requires exponentially many circuit evaluations. The paper's argument that the surrogate 'can fit through sparse samples that detect slight differences in L' has no scaling analysis behind it. This is a circularity, not a missing lemma.\n\nWho is this for? Readers interested in QML training tricks might find the experimental setup worth a quick look, but the central claim is unsupported. I would not cite it in its current form.\n\nRecommendation: the paper deserves a serious referee because it attacks an important problem, but it should be rejected in its current state. The author should either provide a real sample-complexity bound for the surrogate or, failing that, empirical evidence that surrogate gradients remain nonzero while true gradients vanish. With that, the direction could become publishable.","headline":"The paper's central claim that a classical surrogate bypasses barren plateaus is assumed in Theorem 2 rather than proved, and the experiments do not show consistent superiority over classical baselines.","tokens_in":10459,"tokens_out":2458,"would_cite":false,"duration_ms":26646,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a classical surrogate trained on circuit outputs replaces quantum gradients, letting deep quantum neural networks train past barren-plateau stalls and beat direct-gradient baselines on image benchmarks.","keywords":["quantum neural networks","barren plateaus","surrogate model","gradient-free optimization","mid-circuit measurement","ancilla reset","Kraus operators","image classification"],"falsifier":"Compute both the true parameter-shift gradient and the surrogate gradient for the 15-qubit, 6-layer circuit at a set of random parameter points, using a practical sample budget consistent with the paper's training setup; if the surrogate gradient is not substantially larger than the true gradient or does not point downhill on the true loss, the plateau-avoidance claim fails.","tokens_in":9348,"feed_emoji":"⚛️","tokens_out":15217,"duration_ms":128614,"temperature":0.7,"pith_summary":"This paper tries to establish that deep quantum neural networks for image classification can be trained without hitting barren-plateau stalls, by never computing quantum gradients directly. Instead, a classical neural network learns the local map from circuit parameters to measurement outcomes, and gradients for the classification loss are taken through that surrogate. To overcome the limited expressiveness of purely unitary circuits, four auxiliary qubits, called ancillas, are measured and reset after every layer, an operation the paper proves is a non-unitary but non-collapsing channel. Experiments on MNIST, CIFAR-10, and CIFAR-100 with 15-qubit, 6-layer circuits report higher accuracies than the same circuit trained by direct gradient computation, and accuracy close to or above classical CNNs with fewer parameters. If the claim is right, deep parameterized quantum circuits become trainable in practice and the recipe transfers to other quantum learning tasks.","feed_headline":"Train deep quantum nets without quantum gradients, dodge plateaus","feed_subtitle":"A classical surrogate learns the circuit, so training skips quantum gradients and beats direct-gradient baselines.","key_machinery":"The load-bearing machinery is the classical surrogate model $S_w(\\theta)$ trained by mean-squared error to match the circuit's measurement outcomes $m(\\theta)$, combined with the measure-and-reset channel. The reset map gives Kraus operators (the effective action of each measurement branch on the main qubits) $K_x = \\langle x|U$, so the effective channel $\\Phi(\\rho) = \\sum_x K_x \\rho K_x^\\dagger$ is non-unitary and, across layers, nonlinear. The surrogate supplies gradients $\\nabla_\\theta S_w$ in place of $\\nabla_\\theta L$, bypassing parameter-shift rules and the exponential decay of circuit gradients. The argument that this works rests on a descent lemma: if $\\|\\nabla S - \\nabla L\\| \\le \\epsilon$ near $\\theta_t$, then the update $\\theta_{t+1} = \\theta_t - \\eta \\nabla S$ decreases the true loss $L$ up to first order.","core_discovery":"The central claim is that a classical surrogate can replace quantum gradient computation in training a deep parameterized quantum circuit, and that interleaved ancilla measurement and reset turns each layer into a non-unitary but non-collapsing channel that preserves input distinctions while adding the nonlinear expressive power unitary circuits lack. Formally, the reset map is a completely positive trace-preserving (CPTP) map with Kraus operators $K_x = \\langle x|U$ on the main register, the ancillas are exactly reset to $|0\\rangle$ and reusable, and stacking such layers yields an effectively nonlinear feedforward map without deterministic collapse (Theorem 1). The descent results (Lemma 2 and Theorem 2) state that if the surrogate gradient stays within $\\epsilon$ of the true loss gradient in a neighborhood, surrogate updates decrease the true loss and converge to a stationary point. The experimental claim is that this scheme reaches 99.72% test accuracy on MNIST, 90.26% on CIFAR-10, and 58.65% on CIFAR-100 with under a million parameters, beating the same circuit trained by parameter-shift gradients on all three datasets and matching or beating a classical CNN on the first two.","pith_inferences":["A natural extension the paper does not test is applying surrogate-driven updates to variational quantum eigensolvers or quantum optimization problems, where the true gradient is also subject to barren-plateau decay.","A direct scaling test would measure how the number of surrogate samples must grow with circuit depth for the accuracy gap to persist; if that growth is exponential, the method has moved the plateau from gradient size to sample cost.","Because the reported experiments are noiseless simulations, injecting realistic device noise into the surrogate fitting would show whether the accuracy advantage survives outside the simulator."],"forward_implications":["The same circuit architecture that stalls under parameter-shift gradients can be trained to higher accuracy, so depth and qubit count no longer automatically mean a vanishing update signal.","Because the surrogate is classical, training cost shifts from many quantum circuit evaluations per gradient step to classical backpropagation through the surrogate plus a fixed number of circuit samples.","Mid-circuit measurement and reset becomes a reusable template for adding nonlinearity to unitary layers without collapsing the state, since the ancillas are exactly returned to $|0\\rangle$.","The convergence result implies that any quantum neural network (QNN) training task with a smooth loss can use surrogate descent, provided a local surrogate with small gradient error can be maintained."],"supporting_citations":[{"why":"establishes the exponential gradient decay in deep unitary circuits that the method is built to avoid.","marker":"[1]"},{"why":"extends the plateau phenomenon to shallow circuits, motivating a gradient-free training route.","marker":"[2]"},{"why":"provides the parameter-shift gradient rule used by the direct-gradient baseline.","marker":"[3]"},{"why":"supports mid-circuit measurement and reset as a practical, resource-efficient technique.","marker":"[7]"},{"why":"supplies the quantum-circuit simulator used for the reported experiments.","marker":"[14]"},{"why":"is the optimizer used for all models in the comparisons.","marker":"[15]"}],"fun_headline_variants":["Surrogate-driven quantum classifier skips gradients, beats plateaus","Gradient-free quantum image classifier dodges barren plateaus","Classical surrogate trains deep QNN without quantum gradients","No quantum gradients needed: surrogate beats plateau limits","Surrogate predicts measurements to train deep quantum nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a practical classical surrogate can fit the circuit's local behavior well enough to supply a reliable gradient even where the true quantum gradient is tiny, and that this fit does not require exponentially many samples.","fun_headline_variants_meta":{"raw":{"variants":["Surrogate-driven quantum classifier skips gradients, beats plateaus","Gradient-free quantum image classifier dodges barren plateaus","Classical surrogate trains deep QNN without quantum gradients","No quantum gradients needed: surrogate beats plateau limits","Surrogate predicts measurements to train deep quantum nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2503,"prompt_tokens":930,"completion_tokens":1573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1496}},"tokens_in":546,"tokens_out":1573,"duration_ms":11260,"temperature":1.0,"reasoning_tokens":1496,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:09:18.649414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute both the true parameter-shift gradient and the surrogate gradient for the 15-qubit, 6-layer circuit at a set of random parameter points, using a practical sample budget consistent with the paper's training setup; if the surrogate gradient is not substantially larger than the true gradient or does not point downhill on the true loss, the plateau-avoidance claim fails.","supporting_citations":[{"cited_title":"Barren plateaus in quantum neural network training landscapes,","cited_arxiv_id":null,"evidence_quote":"establishes the exponential gradient decay in deep unitary circuits that the method is built to avoid."},{"cited_title":"Cost function dependent barren plateaus in shallow quantum circuits,","cited_arxiv_id":null,"evidence_quote":"extends the plateau phenomenon to shallow circuits, motivating a gradient-free training route."},{"cited_title":"Qubit-reuse compilation with mid-circuit measurement and reset,","cited_arxiv_id":null,"evidence_quote":"supports mid-circuit measurement and reset as a practical, resource-efficient technique."}],"review_version":1}