{"id":"7dff0a46-818a-47b3-86ff-6b29ffefbcb8","arxiv_id":"2605.25768","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Full end-to-end hybrid training decouples trainability from PQC expressibility, unlike pure PQCs which show only a weak regime-dependent trade-off.","lead":"This paper examines the expressibility-trainability trade-off in hybrid quantum neural networks across pure quantum circuits and different hybrid training setups. It finds that full end-to-end training with classical layers can eliminate the trade-off observed in pure parameterized quantum circuits.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Claim that full hybrid training eliminates the trade-off rests on unvalidated metric and simulation choices","rationale":"The reader’s weakest assumption directly identifies the same load-bearing point. Because the full manuscript was not supplied in the query, no additional internal inconsistency or stronger objection could be located; the concern remains the untested fidelity of the chosen metrics and setups to real hybrid dynamics.","tokens_in":1788,"tokens_out":328,"duration_ms":23377,"concrete_test":"Re-run the full-hybrid-training experiments of §4 using an alternative trainability proxy (e.g., the variance of the full-model gradient with respect to a cross-entropy loss instead of the paper’s chosen metric) on the same qubit counts and entanglement topologies; if the reported elimination of the expressibility–trainability correlation disappears or reverses sign, the headline claim is metric-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result—that classical components can decouple trainability from PQC expressibility—depends on the specific definitions of expressibility (likely a fidelity or covering number on the PQC unitary) and trainability (gradient variance or loss curvature), plus the exact circuit families, depths, topologies, and end-to-end training protocol. The abstract notes consideration of “different trainability definitions,” indicating the observed elimination is sensitive to these choices. If the metrics do not track the effective optimization landscape once classical layers are present (e.g., because classical gradients dominate or because expressibility is measured only on the isolated PQC), the decoupling conclusion does not follow for practical HQNNs.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that pure parameterized quantum circuits (PQCs) exhibit only a weak, regime-dependent expressibility-trainability trade-off, whereas hybrid quantum neural networks (HQNNs) increasingly disrupt and can eliminate this relationship under full end-to-end training; classical components are said to reshape the optimization landscape and decouple trainability from PQC expressibility. The work also introduces a multi-objective neural architecture search (NAS) framework that jointly optimizes expressibility, trainability, and task performance, revealing different Pareto fronts for quantum-only versus full hybrid training.","tokens_in":1896,"tokens_out":604,"duration_ms":27401,"significance":"If the empirical results are robust to metric choice and simulation details, the finding that hybridization can eliminate the conventional trade-off would be significant for quantum machine learning, as it suggests that classical layers can mitigate barren-plateau issues even for expressive PQCs and motivates architecture search over combined classical-quantum spaces. The consideration of multiple trainability definitions is a positive step toward generality.","major_comments":[{"comment":"The central claim that full hybrid training eliminates the expressibility-trainability trade-off rests on the specific operational definitions of expressibility (likely a fidelity or covering-number measure on the isolated PQC unitary) and trainability (gradient variance or loss curvature). These definitions are not shown to remain valid once classical layers are inserted and trained end-to-end; if classical gradients dominate or if expressibility is never re-measured on the composite model, the decoupling conclusion does not follow for practical HQNNs.","section":"§4 (Results) and abstract"},{"comment":"No error bars, number of independent runs, or statistical tests are reported for the key figures or tables that demonstrate elimination of the trade-off under full hybrid training. Without these, it is impossible to determine whether the observed disruption is statistically reliable or sensitive to the chosen circuit depths, qubit counts, entanglement topologies, and random seeds.","section":"§4 and Methods"},{"comment":"The multi-objective NAS framework is presented as revealing different Pareto-optimal solutions, yet the paper provides no ablation on the weighting of the three objectives or comparison against single-objective baselines or random search; this weakens the claim that the framework is practically useful for hybrid design.","section":"§5 (NAS framework)"}],"minor_comments":[{"comment":"The abstract states that 'different trainability definitions' are considered but does not list them explicitly; a short enumerated list in the introduction or methods would improve clarity.","section":"Abstract / §1"},{"comment":"Notation for the hybrid loss function and the interface between classical and quantum layers should be defined once in a dedicated subsection rather than introduced piecemeal.","section":"§3 (Model definition)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments on our manuscript arXiv:2605.25768. We address each of the major comments point-by-point below, providing clarifications and indicating where revisions will be made to strengthen the paper.","responses":[{"response":"We appreciate the referee raising this point on the validity of the metrics in the hybrid setting. Expressibility is measured on the PQC in isolation using standard fidelity-based metrics, as is conventional. Trainability is evaluated through the gradient variance of the PQC parameters under the hybrid loss. Our results indicate that end-to-end training with classical layers modifies the loss landscape such that the gradient variances for PQC parameters do not follow the expected trade-off with expressibility. This decoupling is observed empirically across configurations. We will add a new paragraph in Section 4 to explicitly discuss the application of these metrics to HQNNs and address potential concerns about classical gradient dominance by showing that the PQC parameter gradients are indeed influenced but remain the focus of our analysis.","revision_made":"partial","referee_comment":"[§4 (Results) and abstract] The central claim that full hybrid training eliminates the expressibility-trainability trade-off rests on the specific operational definitions of expressibility (likely a fidelity or covering-number measure on the isolated PQC unitary) and trainability (gradient variance or loss curvature). These definitions are not shown to remain valid once classical layers are inserted and trained end-to-end; if classical gradients dominate or if expressibility is never re-measured on the composite model, the decoupling conclusion does not follow for practical HQNNs."},{"response":"This is a valid observation. Although our simulations involved multiple independent runs with different random seeds to ensure reliability, these details and error bars were not included in the presented figures. We will revise the manuscript to report the number of runs (e.g., 20 per configuration), include error bars representing standard deviation in all plots, and add statistical analysis in the Methods section to confirm the significance of the observed differences.","revision_made":"yes","referee_comment":"[§4 and Methods] No error bars, number of independent runs, or statistical tests are reported for the key figures or tables that demonstrate elimination of the trade-off under full hybrid training. Without these, it is impossible to determine whether the observed disruption is statistically reliable or sensitive to the chosen circuit depths, qubit counts, entanglement topologies, and random seeds."},{"response":"We agree that additional experiments would enhance the NAS section. The primary goal was to show that different training regimes lead to distinct Pareto fronts. To address this, we will include ablations on objective weightings and comparisons with single-objective NAS and random search in the revised version, either in the main text or as supplementary material.","revision_made":"yes","referee_comment":"[§5 (NAS framework)] The multi-objective NAS framework is presented as revealing different Pareto-optimal solutions, yet the paper provides no ablation on the weighting of the three objectives or comparison against single-objective baselines or random search; this weakens the claim that the framework is practically useful for hybrid design."}],"tokens_in":1515,"tokens_out":669,"duration_ms":34811,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main observation is that pure PQCs show only a weak, regime-dependent link between expressibility and trainability, while adding classical layers and training the whole model end-to-end largely breaks that link. The authors reach this by comparing three setups—standalone PQCs, quantum-only training inside a hybrid model, and full hybrid training—across depths, qubit counts, and entanglement topologies. They also outline a multi-objective NAS that searches over the combined classical-quantum space.\n\nThe systematic sweep across configurations is the clearest contribution. It moves beyond the usual single-circuit or single-training-mode studies and directly tests how the trade-off behaves once classical components are present. The NAS proposal follows logically from the empirical findings and gives a concrete way to explore the design space.\n\nThe soft spots sit where the stress-test note points. Expressibility is still measured on the isolated PQC, and trainability uses definitions that may not capture the effective landscape once classical gradients are active. The abstract flags that different trainability definitions were considered, which already hints the elimination effect is not automatic. Without reported error bars, run counts, or explicit checks on whether classical layers dominate the gradients, it is hard to know how far the decoupling claim travels beyond the simulated cases. The paper is framed as an empirical study, so these gaps matter more than they would in a purely formal derivation.\n\nThis work is aimed at researchers already inside the hybrid quantum neural network literature who treat the expressibility-trainability trade-off as a settled design constraint. A reader who knows the barren-plateau and expressibility papers will see the value in the controlled comparisons. It is worth sending to peer review because the question is timely and the experimental design is broader than most prior work, even though the methods section will need tighter reporting and metric validation before the conclusions can be taken as general.","headline":"Full end-to-end hybrid training appears to weaken or remove the expressibility-trainability trade-off in these simulations, but the result tracks the chosen metrics and circuit families closely.","tokens_in":2379,"tokens_out":448,"would_cite":false,"duration_ms":26427,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Hybrid quantum neural networks can eliminate the expressibility-trainability trade-off of pure quantum circuits.","keywords":["hybrid quantum neural networks","expressibility","trainability","parameterized quantum circuits","barren plateaus","neural architecture search","quantum machine learning","optimization landscape"],"falsifier":"A persistent strong negative correlation between the expressibility and trainability metrics that remains even after full end-to-end training of the hybrid model on standard classification benchmarks would falsify the decoupling result.","tokens_in":2664,"feed_emoji":"⚛️","tokens_out":753,"duration_ms":20549,"temperature":0.7,"pith_summary":"The paper tests the widespread assumption that highly expressive parameterized quantum circuits suffer worse trainability due to barren plateaus, and checks whether this holds once those circuits sit inside larger hybrid classical-quantum models. It compares three regimes: training a pure quantum circuit alone, training only the quantum part inside a hybrid model, and training the entire hybrid model end-to-end. Pure circuits show only a weak, regime-dependent version of the expected trade-off, but full hybrid training increasingly removes the correlation, indicating that the classical layers change the loss landscape. This matters for model design because it suggests expressibility and trainability need not be traded off when classical and quantum components are optimized together. The authors also introduce a multi-objective search procedure that jointly tunes expressibility, trainability, and task accuracy over the combined design space.","feed_headline":"Hybrid training removes expressibility-trainability trade-off","feed_subtitle":"Classical components in hybrid quantum neural networks reshape the loss landscape and decouple trainability from quantum circuit expressibil","key_machinery":"Systematic comparison of expressibility-trainability correlation across pure PQC training, quantum-only training inside hybrids, and full end-to-end hybrid training, performed over varying circuit depths, qubit counts, and entanglement topologies, plus a multi-objective neural architecture search that optimizes the three objectives jointly.","core_discovery":"Pure parameterized quantum circuits exhibit only a weak and regime-dependent negative correlation between expressibility and trainability, while hybrid quantum neural networks increasingly disrupt and can eliminate this relationship once the full model is trained end-to-end; classical components reshape the optimization landscape and thereby decouple trainability from the expressibility of the embedded quantum circuit.","pith_inferences":["The result suggests that scaling quantum machine learning may benefit more from careful hybrid architecture choices than from further engineering of the quantum circuit alone.","Similar decoupling could appear in other hybrid quantum-classical settings whenever classical gradient steps dominate the loss surface.","A direct test would compare the same quantum circuit inside a hybrid model versus as a standalone circuit on identical data, measuring whether the correlation reappears when the classical layers are removed."],"forward_implications":["Designers of hybrid models can select more expressive parameterized quantum circuits without incurring the expected trainability penalty once classical layers participate in training.","The dominant factor controlling trainability shifts from the quantum circuit properties to the overall hybrid architecture and training mode.","Pareto fronts for expressibility, trainability, and task performance differ markedly between quantum-only and full end-to-end training regimes.","Standard guidelines that treat expressibility as a direct proxy for trainability apply only to isolated quantum circuits, not to embedded ones inside hybrids."],"fun_headline_variants":["Hybrid training eliminates expressibility-trainability trade-off","Classical components decouple trainability from quantum expressibility","Pure PQCs weak trade-off disrupted in full hybrid models","End-to-end training reshapes quantum optimization landscape","Hybrid architectures remove PQC expressibility-trainability link"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The chosen numerical metrics for expressibility and trainability, together with the simulated circuit configurations and training procedures, accurately reflect the optimization behavior that would appear in practical hybrid quantum neural network applications.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid training eliminates expressibility-trainability trade-off","Classical components decouple trainability from quantum expressibility","Pure PQCs weak trade-off disrupted in full hybrid models","End-to-end training reshapes quantum optimization landscape","Hybrid architectures remove PQC expressibility-trainability link"]},"model":"grok-4.3","cost_usd":0.003305,"raw_usage":{"total_tokens":1776,"prompt_tokens":693,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":33049500,"prompt_tokens_details":{"text_tokens":693,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1010,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":693,"tokens_out":73,"duration_ms":8788,"temperature":1.0,"reasoning_tokens":1010,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T21:41:38.805276+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A persistent strong negative correlation between the expressibility and trainability metrics that remains even after full end-to-end training of the hybrid model on standard classification benchmarks would falsify the decoupling result.","supporting_citations":[],"review_version":1}