{"id":"495c1310-070a-4ea7-b2bc-832585e4e65e","arxiv_id":"2412.04844","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Cutting a quantum circuit into smaller trainable subcircuits allows hybrid quantum-classical neural networks to run on devices with fewer qubits while roughly preserving accuracy.","lead":"This paper applies a known quantum-circuit technique, called circuit cutting, to hybrid quantum-classical neural networks so they can run on much smaller quantum chips. The authors show on two image datasets that the cut networks reach similar accuracy to the original networks while needing fewer qubits, but the approach is an approximation and has not been rigorously validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cut replacement in §III-B is a lossy measurement-plus-angle-encoding channel, not a circuit cut; the cut HQNN is a different model, so 'preserves accuracy' is unestablished.","rationale":"The reader's weakest assumption identifies the same lossy cut approximation, but the more fundamental problem is that the paper never reconstructs the original circuit: it trains a different, autoencoded model. This makes the abstract's wording misleading. However, the experimental results are a plausible proof-of-concept, the code is promised, and the paper itself acknowledges the 2-qubit exception. The conditional verdict is appropriate: acceptance should require either a theoretical characterization of the cut-channel error or a clearly restricted empirical claim with proper statistics. No change to the reader's conditional verdict is needed.","tokens_in":12099,"tokens_out":7419,"duration_ms":78502,"concrete_test":"Re-run the 8-3 and 10-3 cut experiments on Digits and MNIST with 30 independent seeds and report mean final validation accuracy with 95% confidence intervals for both original and cut circuits, using a pre-registered equivalence margin of 2 percentage points. If the cut circuit's CI lies outside the original's CI by more than the margin, the accuracy-preservation claim for 3-qubit cuts fails; if the intervals overlap within the margin, the empirical claim survives but must still be paired with an analytic bound on the cut channel before the phrase 'preserves the accuracy of the original circuits' is justified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the definition of a cut in Section III-B: 'Subcircuits are generated by replacing cuts with measurement operations and re-encoding quantum states (using angle embedding).' This is not a circuit cut in the standard sense. A computational-basis measurement followed by angle re-encoding induces the channel Γ(ρ) = Σ_b ⟨b|ρ|b⟩ |ψ_b⟩⟨ψ_b|, which destroys all coherence across the cut and transmits at most one classical scalar per wire. Standard wire cutting reconstructs expectation values by sampling over a quasi-probability decomposition (for example Pauli frames); no such reconstruction is implemented or analyzed here. Consequently the cut HQNN is a different, information-limited quantum-classical model, not an execution of the original n-qubit circuit on m qubits. The accuracy curves in Figures 8 and 9 therefore cannot establish the abstract's claim that the method 'preserves the accuracy of the original circuits'; at best they show that this new autoencoded model can be trained to comparable accuracy on two small datasets. The 2-qubit degradations in Figure 8 (labels 1-3) are the expected signature of the lossy channel, and with only five runs and no reported variance it is unclear whether the 3-qubit agreements exceed run-to-run noise. Without a bound on the cut-channel error, or a restriction of the claim to empirical training equivalence with error bars, the central claim is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a framework for quantum circuit cutting of hybrid quantum-classical neural networks (HQNNs) so that an n-qubit quantum layer can be trained and executed on an m-qubit device with m < n. The framework consists of a greedy cutting-point design algorithm (Algorithm 1, Section III-A) and an implementation in which each subcircuit becomes an independent QNode and is converted to a trainable Keras layer (Section III-B). The authors evaluate the approach on Digits and MNIST for original circuits of 4 to 10 qubits cut to 2 to 4 qubits, and additionally report 20- and 50-qubit cases. The central claim, repeated in the abstract, Section IV-H, and the conclusion, is that the cut circuits preserve the accuracy of the original circuits while supporting training of quantum parameters in all subcircuits. The empirical component reports five-run average training and validation accuracy curves for the cut and uncut circuits, FLOPs measurements, and a small ablation on embedding strategies.","tokens_in":12331,"tokens_out":4582,"duration_ms":51522,"significance":"If the central claim were established, the paper would address a real NISQ-era bottleneck: executing and training quantum layers that exceed the qubit count of available hardware. The concrete engineering contribution is a multi-QNode implementation that preserves end-to-end differentiability through cut subcircuits, which is genuinely different from prior circuit-cutting work focused on expectation-value reconstruction. The authors also ship code and report repeated runs, which supports reproducibility. However, the significance as stated is conditional: because the cut operation is a lossy measurement-and-re-encoding channel rather than a standard circuit cut, the paper does not establish that the original circuit's accuracy is preserved, and the experimental evidence lacks the statistical detail needed to support the 'comparable accuracy' claim beyond selected configurations.","major_comments":[{"comment":"The cut operation is not a circuit cut in the standard sense and does not preserve the original circuit. Replacing a cut with a computational-basis measurement followed by angle embedding implements the channel Γ(ρ)=Σ_b ⟨b|ρ|b⟩ |ψ_b⟩⟨ψ_b⟩, which destroys all coherence across the cut and transmits only one classical bit per cut wire. Standard wire cutting, as in the cited works of Peng et al. [26] and Lowe et al. [28], reconstructs expectation values by sampling over a quasi-probability decomposition; no such reconstruction is implemented or analyzed here. Consequently the cut circuit is a different, information-limited quantum-classical model, and the abstract's claim that the method 'preserves the accuracy of the original circuits' is not supported. The authors should either implement and analyze a proper reconstruction with error bounds, or explicitly reframe the contribution as empirical training of an approximate autoencoded variant and restrict the claims accordingly.","section":"Section III-B and Figure 7; Abstract; Section V"},{"comment":"The experimental evidence for 'comparable accuracy' is not statistically supported. The paper reports only the average over five runs, with no error bars, no per-run spread, and no final numeric accuracy values for any n-m configuration. In several subplots of Figure 8 the cut and original curves overlap or cross, so it is unclear whether the claimed agreement for 3- and 4-qubit targets exceeds run-to-run noise. Conversely, the large 2-qubit gaps in Figure 8 (labels 1-3) are the expected signature of the lossy cut channel, and the longer-training results in Figure 11 also lack variance information. The authors should report mean ± standard deviation over seeds, and ideally effect sizes or confidence intervals, for every configuration discussed in Sections IV-B and IV-C.","section":"Section IV-A and Figures 8, 9, 11"},{"comment":"The cutting-point design algorithm is underspecified, which makes the methodology difficult to reproduce and its correctness impossible to assess. The priority order in 'Sort Gates in Priority Order' is not defined, the wire selection criterion in 'Select m wires based on priority' is informal, and the dependency-unavailable case is described only by example. No invariant, termination argument, or proof that the resulting subcircuits are valid trainable HQNN components is provided. Since every downstream claim depends on this partition, the algorithm needs a precise specification of its inputs, priority metric, wire-selection rule, and a correctness statement.","section":"Section III-A, Algorithm 1"},{"comment":"The scalability evidence is incomplete. For the 50-qubit experiment only the cut circuit is reported, with no original 50-qubit baseline to compare against, so the claim that the cut circuit 'remains effective' for larger circuits is not supported. In addition, the section title refers to '20-5 Cut & 50-5 Cut' while the text and Figure 10 describe 20-4 and 50-4 cuts and state that all cut circuits use 4 qubits; this inconsistency must be corrected before the results can be evaluated.","section":"Section IV-E and Figure 10"}],"minor_comments":[{"comment":"The in-text references to labels 8 and 9 are unclear because those labels are not associated with particular subplots in the caption; please tie each label to a named configuration.","section":"Section IV-B, Figure 8"},{"comment":"The FLOPs measurement is not described: the authors should state how forward and backward FLOPs are profiled and whether the reported values are deterministic or averaged.","section":"Table II"},{"comment":"The sentence 'All cut circuits are implemented using 4 qubits' conflicts with the section title's '50-5 Cut'; please harmonize the nomenclature throughout.","section":"Section IV-E"},{"comment":"The accuracy axes have no tick labels or scale, which makes it difficult for the reader to judge the magnitude of the observed gaps; adding labeled axes and perhaps final-accuracy annotations would materially improve interpretability.","section":"Figures 8 and 9"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper has a useful engineering contribution—trainable multi-QNode HQNN subcircuits—and the public code is a plus. My main concern is that the central claim of 'preserving accuracy' is conceptually unsupported because the proposed cut is a lossy measurement-and-re-encoding channel rather than standard wire cutting with reconstruction. This is fixable by reframing the contribution as an approximate autoencoding method and by strengthening the empirical evidence with error bars and final accuracy tables, but it is not merely a presentation issue. I would not recommend rejection, because the implementation and empirical trends could still support a meaningful narrower claim if the authors are willing to revise the framing and statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The genuinely new thing is the PennyLane/Keras integration: each subcircuit becomes its own QNode and is wrapped as a KerasLayer, so gradients flow across the cut during training. That solves a real engineering pain point and isn't in the earlier cutting work. The greedy cutting-point search is simple but fine for the sizes tested, and the FLOPs overhead table is a useful, honest addition.\n\nThe soft spot is the definition of a 'cut.' In Section III-B they replace a cut with a computational-basis measurement followed by angle re-encoding. That is not a circuit cut in the standard sense — no quasi-probability reconstruction, no Pauli frames. It's a lossy channel that discards coherence and transmits one classical scalar per wire. So the cut HQNN is a different, information-limited model, not an execution of the original n-qubit circuit on m qubits. The abstract's 'preserves the accuracy of the original circuits' is therefore unestablished. What the experiments show is that this autoencoded model can be trained to comparable accuracy on Digits and MNIST for 3- and 4-qubit targets. That is still interesting, but it's a weaker claim.\n\nAlso, there are no error bars. The paper says 'average over five runs' but doesn't report variance, so the 3-qubit agreements might be within noise. The 2-qubit failures are acknowledged, but they're framed as an exception rather than the expected signature of the lossy channel. And the comparison with the original HQNN isn't apples-to-apples: the cut circuit has extra classical layers doing the re-encoding, so it's a different architecture that could do better or worse for reasons unrelated to cutting.\n\nThe 50-qubit result is just a 50-4 cut with no uncut counterpart — it shows convergence but not preservation. The FLOPs overhead is real but not analyzed against the cost of running many sequential subcircuits on actual hardware.\n\nThis is a useful proof-of-concept for a practical problem, and the code availability is a point in its favor. But the central claim needs to be downgraded, and the statistics need work. I'd send it to peer review with a request for major revision: report per-seed results, characterize the cut error, and stop calling it accuracy preservation.","headline":"A useful engineering recipe for training cut HQNNs, but the lossy re-encoding 'cut' doesn't support the accuracy-preservation claim.","tokens_in":12863,"tokens_out":2937,"would_cite":false,"duration_ms":30253,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"Quantum circuit cutting lets large hybrid networks train on few-qubit hardware without losing accuracy.","keywords":["quantum circuit cutting","hybrid quantum-classical neural networks","NISQ","quantum machine learning","trainable subcircuits","greedy cutting algorithm","MNIST"],"falsifier":"Train the same HQNN architecture as both an n-qubit original and an n-to-m cut for m = 2 and n >= 6, running enough epochs for learning curves to plateau; if the validation accuracy gap between cut and original never closes, the claim that cutting preserves accuracy fails in that regime. More directly, compare the quantum state on the cut wire before the cut with the state re-encoded after measurement; a fidelity well below 1 would show the information loss behind the accuracy drop.","tokens_in":11870,"feed_emoji":"✂️","tokens_out":7570,"duration_ms":64698,"temperature":0.7,"pith_summary":"This paper argues that quantum circuit cutting can make hybrid quantum-classical neural networks (HQNNs) trainable on NISQ devices that have far fewer qubits than the network's original circuit. The authors propose a concrete methodology: a greedy algorithm chooses where to cut the circuit, and each resulting subcircuit becomes its own differentiable layer, so gradients still flow through the whole network during backpropagation. On standard benchmarks like MNIST and handwritten Digits, they report that cut circuits using three or more qubits match or sometimes exceed the accuracy of the original circuits, with the exception of two-qubit cuts, which converge slowly and lose accuracy. The payoff is a practical path to running larger quantum machine learning models on current, qubit-limited hardware.","feed_headline":"Circuit cutting lets big quantum networks train on tiny devices","feed_subtitle":"New method preserves accuracy of hybrid quantum-classical networks while cutting qubit needs by up to ten times.","key_machinery":"The load-bearing mechanism is replacing a wire cut with a computational-basis measurement and an angle-embedding re-encoding. This turns one big quantum node into several smaller quantum nodes, each packaged as a trainable layer so gradients flow through the classical backend between subcircuits. The greedy cutting-point algorithm sorts gates by data dependency and places cuts before gates whose dependencies lie on wires that cannot be mapped to the m-qubit device. The measurement-plus-re-encoding is the lossy step that dominates the accuracy results.","core_discovery":"The central claim is that a hybrid quantum-classical neural network whose quantum circuit uses n qubits can be cut into a set of subcircuits, each using m < n qubits, and trained end-to-end without an accuracy penalty. The authors show that the cut wires can be replaced by measurement operations followed by angle re-encoding of the measured values into the next subcircuit, and that by wrapping each subcircuit in its own trainable layer, standard backpropagation still updates every quantum parameter. Their experiments across 4- to 50-qubit circuits on the Digits and MNIST datasets show the cut model's training and validation accuracy closely tracks the uncut model whenever the target device has at least three qubits. The two-qubit cuts are the documented failure mode: subcircuits with only two qubits have low correlation and entanglement capacity, so accuracy and convergence degrade.","pith_inferences":["The measurement-plus-angle-re-encoding approximation is effectively a classical bottleneck; replacing it with quasi-probability reconstruction or other circuit-knitting estimators could close the two-qubit gap while keeping the method differentiable.","The faster convergence of cut circuits may be an implicit regularization effect: cutting reduces entanglement capacity, which can act as a regularizer on small datasets.","The greedy partitioning becomes combinatorially more expensive as the number of cuts grows, so scaling to much larger circuits will likely need a smarter search over cutting points."],"forward_implications":["HQNNs with 8, 10, 20, and 50-qubit circuits can be trained on devices with as few as 3-4 qubits, with accuracy comparable to the uncut circuit.","Training remains differentiable across all subcircuits, so quantum parameters in every subcircuit receive gradient updates.","The method introduces a FLOP overhead that grows as target qubit count drops, but the qubit savings outweigh this cost for NISQ execution.","Cut circuits with 3+ qubits sometimes converge faster and generalize slightly better than the original circuit, as observed for 8-3 and 10-3 cuts.","Two-qubit cuts are the known limitation: they need many more epochs and still lag in accuracy."],"supporting_citations":[{"why":"Provides the underlying wire-cutting technique this work adapts for hybrid quantum-classical neural networks.","marker":"[26]"},{"why":"Supplies the automatic differentiation machinery that lets each subcircuit be trained like a standard neural network layer.","marker":"[41]"},{"why":"Digits dataset used for the main accuracy comparisons between original and cut circuits.","marker":"[42]"},{"why":"MNIST dataset used to test scalability and generalization to larger circuits.","marker":"[65]"},{"why":"Defines the angle and amplitude embeddings used for re-encoding measured values into the next subcircuit.","marker":"[43]"}],"fun_headline_variants":["Cutting quantum circuits trains large neural nets on few qubits","Quantum circuit cutting enables big QNN training on tiny devices","Cut large quantum networks into trainable subcircuits for small qubit devices","Greedy circuit cutting runs large QNNs on limited qubits","Training large quantum networks via circuit cutting on limited hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that replacing a cut wire with a measurement followed by angle re-encoding loses no information the network needs for training, an assumption the paper's own two-qubit results show can break down.","fun_headline_variants_meta":{"raw":{"variants":["Cutting quantum circuits trains large neural nets on few qubits","Quantum circuit cutting enables big QNN training on tiny devices","Cut large quantum networks into trainable subcircuits for small qubit devices","Greedy circuit cutting runs large QNNs on limited qubits","Training large quantum networks via circuit cutting on limited hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2866,"prompt_tokens":991,"completion_tokens":1875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1787}},"tokens_in":607,"tokens_out":1875,"duration_ms":14752,"temperature":1.0,"reasoning_tokens":1787,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:12:28.576041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same HQNN architecture as both an n-qubit original and an n-to-m cut for m = 2 and n >= 6, running enough epochs for learning curves to plateau; if the validation accuracy gap between cut and original never closes, the claim that cutting preserves accuracy fails in that regime. More directly, compare the quantum state on the cut wire before the cut with the state re-encoded after measurement; a fidelity well below 1 would show the information loss behind the accuracy drop.","supporting_citations":[{"cited_title":"Simulating large quantum circuits on a small quantum computer,","cited_arxiv_id":null,"evidence_quote":"Provides the underlying wire-cutting technique this work adapts for hybrid quantum-classical neural networks."},{"cited_title":"Pen-Based Recognition of Handwritten Digits,","cited_arxiv_id":null,"evidence_quote":"Digits dataset used for the main accuracy comparisons between original and cut circuits."},{"cited_title":"The mnist database of handwritten digit images for machine learning research [best of the web],","cited_arxiv_id":null,"evidence_quote":"MNIST dataset used to test scalability and generalization to larger circuits."},{"cited_title":"Schuld and F","cited_arxiv_id":null,"evidence_quote":"Defines the angle and amplitude embeddings used for re-encoding measured values into the next subcircuit."}],"review_version":1}