{"id":"f0242bab-0563-42e3-b870-7c5b87e59e83","arxiv_id":"2512.10670","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A pulse-level data re-uploading classifier outperforms its gate-based counterpart in noisy superconducting-qubit simulation.","lead":"This paper replaces the trainable gates of a quantum machine-learning classifier with microwave control pulses and simulates the result on a noisy transmon model. In the simulation, the pulse version reaches higher test accuracy and tolerates more noise than the same classifier built from standard gates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Noise-model ambiguity in pulse simulation (footnote 56 vs Appendix B.3) leaves the central noise-resilience claim unverified.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the noise model's treatment of Trotterized pulses is under-specified, and the central advantage claimed in Fig. 6(b) could be an artifact of where noise is inserted. I agree with this assessment. The paper's other potential weaknesses (e.g., learnable target states, parameter-count differences) are secondary to the noise-model ambiguity because they affect the interpretation of generalization but not the fundamental validity of the noise-resilience comparison. The concrete test I propose is direct and actionable: inspect the public codebase to determine the actual noise insertion points, and if necessary rerun the key experiment with corrected noise placement. This would either validate or refute the central claim. Since the issue is fixable and the reader's verdict is already CONDITIONAL, I see no reason to change that verdict.","tokens_in":39762,"tokens_out":5815,"duration_ms":58606,"concrete_test":"Inspect the public code at https://github.com/nacedob/Pulsed-Data-Reuploading-Quantum-Models to locate where noise channels are added relative to the Trotterized pulse. If noise is applied after each Trotter gate, rerun the p-sweep of Fig. 6(b) with noise applied once per physical pulse (after the full Trotterized evolution), using the same pulse duration and error parameters. If the pulsed model's test accuracy drops to the gate-based level, the central claim is an artifact of the noise placement. Alternatively, model the pulse with a Lindblad master equation (QuTiP) with T1, T2 and a depolarizing rate fitted to the native gate error, and compare the same gate-based model under the same setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section IV C, Fig. 6(b)) is that the pulsed model maintains ~80% test accuracy for depolarizing probability p up to 0.1 while the gate-based model degrades sooner. This hinges on how noise is inserted in the pulse simulation. Appendix B.3 states that depolarizing, amplitude damping, and phase damping channels are applied after each electromagnetic control pulse or single-qubit gate. However, footnote 56 says each pulse is simulated via Trotter-Suzuki decomposition into 'dozens of gates.' If noise channels are inserted after each Trotter slice, a single pulse receives dozens of error events, making the pulsed model a deep noisy circuit; its observed robustness to p≈0.1 would be highly implausible. If instead noise is applied once per physical pulse, the error strength for a pulse is not calibrated to native gate errors the way the gate-based model is, and the comparison is not 'under equivalent noise conditions.' The manuscript provides no information on which convention is used. This is load-bearing because either resolution could reverse the main conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pulse-level formulation of data re-uploading quantum machine learning, replacing parameterized gate-based variational layers with native control pulses on a simulated transmon processor. Trainable parameters are moved into the pulse amplitudes, phases, and detunings, while the encoding blocks remain gate-based. The authors benchmark the pulse-based model against a gate-based counterpart on binary MNIST classification (digits 0 vs 8), using a two-qubit data re-uploading architecture with warm-start initialization. The central empirical claims are that the pulsed model achieves substantially higher test accuracy (~80% vs ~60%) and maintains this accuracy for depolarizing noise probability p up to about 0.1, whereas the gate-based model degrades sooner. Noise is modeled with amplitude damping, phase damping, and depolarizing channels parameterized from IBM Brisbane calibration data, with a separate depolarizing sweep in Fig. 6(b).","tokens_in":40050,"tokens_out":4400,"duration_ms":49265,"significance":"If the central claims hold, the paper would provide a concrete, hardware-aligned alternative to gate-based variational QML, with meaningful implications for noise-resilient model design on superconducting processors. The manuscript has several strengths: it uses publicly available device parameters, provides a code repository, includes a warm-start training procedure, and systematically studies both layer-count and noise-strength dependence. The generalization gap and the noise-resilience plateau are striking and potentially important. However, the validity of the empirical benchmark depends on whether the pulse and gate models are compared under genuinely equivalent noise conditions. The main weakness is that the noise insertion convention for the Trotterized pulse simulation is unspecified, and this ambiguity directly affects the paper's central conclusion. If the noise insertion is per Trotter slice, the reported pulse robustness is very surprising and may be an artifact; if it is per physical pulse, the comparison may not be at equivalent error strength. The learnable target states and the different two-qubit parameterizations also introduce possible confounds that need to be a","major_comments":[{"comment":"The central comparison in Fig. 6(b) requires that the pulse and gate models incur comparable error events. Appendix B.3 states that depolarizing, amplitude damping, and phase damping channels are applied after each electromagnetic control pulse or single-qubit gate, while footnote 56 states that each pulse is simulated via a Trotter-Suzuki decomposition into 'dozens of gates.' It is not specified whether noise channels are inserted per Trotter slice or per physical pulse. If noise is applied per slice, a single pulse suffers many error events, making the pulsed model a deep noisy circuit and the robustness at p≈0.1 implausible. If noise is applied per pulse, the per-event error strength and pulse duration are not calibrated to native gate errors. Please state the noise insertion points explicitly (ideally with a code snippet) and report the number of Trotter steps per pulse.","section":"Appendix B.3 and footnote 56"},{"comment":"The noise-sweep protocol is incompletely defined. The text says p is the probability of collapsing to the maximally mixed state and that all other noise parameters are held constant, but it does not say how p is applied in the pulsed simulation: per Trotter slice, per physical pulse, or per native gate. It also does not define 'equivalent noise conditions' operationally. If the pulsed model applies p once per physical pulse while the gate model applies it per gate, the comparison is not at matched error counts. Please specify the exact noise schedules for both models and, if necessary, rerun the sweep under matched error conventions.","section":"Section IV.C / Fig. 6(b)"},{"comment":"The learnable target states |s0(θ,φ)> and |s1(θ,φ)> are introduced in the experimental setup and are used in the fidelity-based cost. It is not stated explicitly whether these target-state parameters are trained for both the gate-based and pulse-based models. If they are used only in the pulsed model, the test-accuracy gap in Fig. 6(a) could be explained by this extra trainable flexibility rather than by pulse-level training. If they are used in both models, this should be stated when the models are defined in Section III.","section":"Section IV.A / Eqs. (8)-(9)"}],"minor_comments":[{"comment":"Please clarify what the error bars represent and how they are computed over the five random seeds. The caption and text should state whether the plotted intervals are standard deviations, standard errors, or min/max ranges.","section":"Fig. 6"},{"comment":"Reference [22] lists 'H. T. el al.' and is incomplete; reference [37] is missing a title; reference [55] is missing a title. Please fix these bibliographic entries.","section":"References"},{"comment":"Typo: 'quantum channelss' should be 'quantum channels.' Also, the notation in Eqs. (B2)-(B4) uses p, γ, and λ without explicitly connecting the depolarizing probability p to the device error rates described later; a brief summary of the mapping would improve readability.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is interesting and potentially publishable, but the main 'equivalent noise conditions' assertion is not verifiable from the manuscript as written. The noise insertion convention for Trotterized pulses must be clarified and the experiments possibly rerun. I am not recommending rejection because this is a fixable methodological ambiguity rather than an internally inconsistent derivation. Please ensure the authors address the noise schedule and the target-state confound in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea—train a data re-uploading classifier at the pulse level rather than the gate level—is sensible, and the two-qubit CR-inspired pulse ansatz with tunable amplitude, phase, and detuning is a reasonable piece of new design. The authors honestly note that their single-qubit pulse block is just an arbitrary SU(2) reparameterization, and they provide code and data. The warm-start initialization for the two-qubit network is also a nice practical touch.\n\nThe trouble is the main empirical claim. The abstract and Section IV C say the pulsed model keeps ~80% test accuracy for depolarizing probability p up to 0.1, while the gate-based model degrades faster. That conclusion rests on how noise is inserted in the pulse simulation, and the manuscript doesn't say clearly. Footnote 56 says each pulse is simulated via Trotter-Suzuki into 'dozens of gates.' Appendix B.3 says noise channels are applied 'after each electromagnetic control pulse or single-qubit gate.' If noise is applied after each Trotter slice, the pulsed model is a deep noisy circuit and would likely be destroyed by p=0.1 across hundreds of error events—so the claimed robustness would be implausible. If noise is applied once per physical pulse, then the per-pulse error strength is not calibrated against the per-gate errors in the gate-based model, so the comparison is not 'under equivalent noise conditions.' Either way, the central plot in Fig. 6(b) is not a trustworthy benchmark as reported.\n\nTwo smaller issues: the learnable target states in Eqs. (8)-(9) are introduced as a relaxation, but it's never stated whether the gate-based baseline also gets this flexibility. If it doesn't, the generalization comparison is unfair. And the abstract says 'higher fidelity,' but the results report accuracy; that's a wording mismatch, not a substantive problem.\n\nThe paper is worth refereeing because the question matters for QML practice and the fix is straightforward: specify the noise insertion convention, run both models with the same per-operation error budget, and give the gate baseline the same target-state freedom. But as written, I can't trust the headline claim. Send it to peer review with the expectation of a major revision.","headline":"Plausible idea, but the main noise-resilience claim is unverified because the noise model's insertion rule is ambiguous.","tokens_in":40516,"tokens_out":4141,"would_cite":false,"duration_ms":44482,"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":"Training a data re-uploading quantum classifier with native control pulses instead of gate rotations yields ~80% test accuracy and far greater noise tolerance than the equivalent gate-based circuit, in simulations of a superconducting trans","keywords":["quantum machine learning","pulse-level control","data re-uploading","transmon qubits","noise resilience","generalization","variational quantum circuits","cross-resonance"],"falsifier":"Read the released simulation code to check where depolarizing, amplitude-damping, and phase-damping channels are applied relative to Trotter-Suzuki slices; then rerun the two-qubit 20-layer experiment with noise applied continuously over pulse duration. If the gate-based model then reaches about 80% test accuracy, the claimed pulse advantage is an artifact of the noise discretization.","tokens_in":39663,"feed_emoji":"⚛️","tokens_out":4835,"duration_ms":48426,"temperature":0.7,"pith_summary":"The paper aims to show that quantum machine learning models can be trained directly at the pulse level by leaving pulse amplitude, phase, and detuning as free parameters in the transmon's driven Hamiltonian, and that this hardware-native formulation outperforms the standard gate-based data re-uploading model. On a simulated two-qubit transmon processor with realistic amplitude damping, phase damping, depolarizing, and readout errors, the pulse-trained model reaches about 80% test accuracy on an MNIST 0-versus-8 task, while the gate-based model plateaus near 60%. Under increasing depolarizing noise, the pulse model keeps its accuracy up to depolarizing probability p≈0.1, whereas the gate model degrades much sooner. If correct, this establishes pulse-level training as a viable route to noise-resilient, generalizing QML on near-term hardware.","feed_headline":"Pulse control keeps quantum classifier accurate under noise","feed_subtitle":"On a simulated transmon, pulse-trained model holds ~80% test accuracy where gate circuits drop toward random guessing.","key_machinery":"The key object is the pulse-level replacement of the trainable unitary blocks. Each single-qubit SU(2) becomes two virtual-Z rotations (error-free phase updates) around a resonant constant-amplitude pulse whose amplitude and phase are trainable. Each two-qubit entangling block is a single cross-resonance-style pulse with trainable amplitude, phase, and frequency detuning, able to modulate or suppress entanglement. A fidelity-based loss and a warm-started two-qubit initialization (starting from the trained one-qubit parameters) carry the training.","core_discovery":"The central claim is that replacing each parameterized gate in a data re-uploading QNN with a physically motivated pulse schedule—resonant single-qubit pulses flanked by virtual-Z rotations, and a single cross-resonance-inspired entangling pulse with trainable amplitude, phase, and detuning—yields a model that is at least as expressive and substantially more noise-tolerant than the gate-based original. In numerical simulation of a superconducting transmon processor using device-like coherence times and error rates, the pulse-based model achieves approximately 80% test accuracy across layer counts from 5 to 40, while the gate-based model saturates near 60%; and the pulse model retains roughly","pith_inferences":["The fairness of the gate-versus-pulse comparison hinges on a simulation detail flagged in the paper's own footnote: pulses are Trotterized into 'dozens of gates,' and if noise is charged per Trotter slice rather than per physical pulse, the pulsed model is effectively a deeper noisy circuit and the reported advantage may be an artifact.","A direct hardware experiment with real pulse schedules, where T1/T2 decay during each pulse duration is physically integrated, would be the cleanest test of whether the advantage survives outside the simplified Markovian noise model.","The authors leave pulse-based data encoding to future work; making the encoding itself trainable could either widen the generalization gap or add noise-sensitive parameters, and is an obvious next step."],"forward_implications":["Pulse-native training reduces the number of concatenated error-prone operations, so learning models can exploit more layers without immediately succumbing to noise.","The roughly 80% test accuracy sustained up to depolarizing probability p≈0.1 gives near-term QML a larger usable noise budget than gate-based circuits.","Because the methodology replaces arbitrary gates by native pulses, it can be applied to other variational QML architectures, not just data re-uploading.","Pulse-level models overfit later as depth grows, indicating better generalization from the same amount of training data.","The approach moves QML closer to the native control language of hardware, reducing calibration overhead in principle."],"fun_headline_variants":["Pulse-based quantum model outlasts gate circuits under noise","Quantum machine learning gets noise boost from pulse control","Pulse-level training beats gates for quantum classifiers","Quantum data re-uploading: pulses triumph over gates in noise","Pulse-driven QML retains accuracy where gate circuits fail"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison assumes the simplified noise model charges each pulse a single error event with probabilities set by native gate errors; if noise is applied per Trotter-Suzuki slice instead of per physical pulse, the claimed pulse advantage could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Pulse-based quantum model outlasts gate circuits under noise","Quantum machine learning gets noise boost from pulse control","Pulse-level training beats gates for quantum classifiers","Quantum data re-uploading: pulses triumph over gates in noise","Pulse-driven QML retains accuracy where gate circuits fail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":1001,"prompt_tokens":722,"completion_tokens":279,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":199}},"tokens_in":466,"tokens_out":279,"duration_ms":3324,"temperature":1.0,"reasoning_tokens":199,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:03:36.832625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Read the released simulation code to check where depolarizing, amplitude-damping, and phase-damping channels are applied relative to Trotter-Suzuki slices; then rerun the two-qubit 20-layer experiment with noise applied continuously over pulse duration. If the gate-based model then reaches about 80% test accuracy, the claimed pulse advantage is an artifact of the noise discretization.","supporting_citations":[],"review_version":1}