{"id":"a2d49433-44b5-4441-8461-3f0025136c14","arxiv_id":"2608.06426","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A contextual bandit that adaptively selects ZNE folding levels reduces quantum circuit execution round trips by up to 40% and exchanged bytes by 35% while slightly improving estimator fidelity in simulated VQC training.","lead":"This paper introduces CMAB-ZNE, a system that uses a contextual multi-armed bandit to choose which noise-amplification levels to run when applying zero-noise extrapolation to variational quantum circuits. The adaptive selection cuts quantum circuit execution round trips by up to 40% and bytes exchanged by up to 35% in simulations, while slightly improving estimator fidelity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline cost savings are computed from an offline counterfactual replay in which all folding scales are precomputed (Sec. III.A.4–5), and the real-hardware experiment does not run the bandit online; the central claim of 40%/35%/30% reductions is therefore not yet validated in a live deployment.","rationale":"The paper is transparent about the offline-replay methodology and the theoretical analysis is a standard LinUCB argument, so the concern is not internal inconsistency or bad faith. Rather, the gap is between the headline communication-savings claim and the evidence: the savings are counterfactual projections from a replay, and the only real-hardware experiment does not run the bandit online. This is exactly the weakest assumption the reader identified, and the conditional verdict already reflects it. My stress test does not find a reason to move the verdict; the same load-bearing concern remains and would be settled by an online single-arm-feedback experiment. I therefore recommend no change to the reader's verdict.","tokens_in":23902,"tokens_out":6070,"duration_ms":61473,"concrete_test":"Run an online single-arm-feedback version of the simulation: at each epoch, execute only the scales in the selected arm, update the bandit from that arm's reward, and apply the budgeted-stopping rule without access to unselected scales; measure actual round trips, bytes transferred, and end-to-end cost over 50 epochs under the 10 Mbps budget, and compare against fixed {1,3,5}. If the observed reductions fall below the claimed 40%/35%/30% levels, or if the fidelity gain disappears, the headline claim is not supported. As a second check, recompute Table V round-trip reductions from the logged arm choices alone: without budgeted stopping or cached reuse, reductions from the 2/3/4-scale arm set cannot exceed 33.3% relative to {1,3,5}.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and Table V, is that CMAB-ZNE cuts QPU-CPU round trips by ~40%, bytes by ~35%, and end-to-end cost by ~30% versus the strongest fixed-ZNE arm {1,3,5}. The evidence for these numbers comes entirely from an offline replay. Sec. III.A.4 says that for every epoch and scale lambda in {0,1,3,5,7}, the expectation vectors are stored as dumps; Sec. III.A.5 then says the bandit can reconstruct the ZNE output and compute rewards for all unselected arms, calling these counterfactual rewards used for offline replay. Thus the bandit's arm choice does not control which circuits are actually executed: all scales are always computed, and the reported round-trip reduction is an accounting measure that counts only the scales in the selected arm, not a measured reduction in quantum executions. The real-hardware validation (Sec. III.C) also does not exercise the bandit online: it evaluates all three arms from a shared scale set {1,3,5,7} and uses the QPU only for frozen inference. Consequently, the operational claim that the adaptive policy cuts quantum circuit execution round trips has not been demonstrated under live partial feedback, where only the selected arm's reward is observed, exploration has real cost, and the context-reward relationship may differ from the counterfactual replay. A secondary arithmetic check strengthens this concern: with arms of 2, 3, and 4 scales, a >33% reduction in round trips relative to {1,3,5} requires the budgeted-stopping rule (Eq. 20) or caching to make some epochs use fewer than two new executions; that mechanism is not validated online.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CMAB-ZNE, a contextual multi-armed bandit framework that selects ZNE folding-scale sets adaptively during variational quantum circuit training. The method defines arms as sets of folding scales, uses a cost-aware LinUCB-type policy with a normalized cost proxy, and provides a regret bound for the cost-aware selection problem. The authors evaluate the approach in PennyLane simulations on CIFAR-10 and EuroSAT under a time-varying noise proxy, and report a real-hardware validation on Rigetti Cepheus-1-108Q. The headline claims are that CMAB-guided folding reduces QPU-CPU round trips by up to 40%, exchanged bytes by up to 35%, and end-to-end cost by up to 30% under a 10 Mbps budget, with up to 6.9% higher estimator fidelity, compared with fixed-fold and grid-search ZNE.","tokens_in":24393,"tokens_out":4817,"duration_ms":51039,"significance":"If the central claims were fully validated, the paper would make a useful contribution by connecting ZNE fold selection to communication cost and dynamic noise. The strengths include a publicly released code repository, a real-hardware execution with 240 successful QPU tasks on Rigetti hardware, and a standard, clearly presented regret analysis for the cost-aware LinUCB selection rule. However, the headline savings are computed from an offline counterfactual replay rather than a live deployment, the hardware experiment does not run the bandit online, and the estimator-fidelity metric is undefined and inconsistent with the accuracy tables. These issues affect the paper's central operational claims, so the significance of the results as currently stated cannot be fully assessed.","major_comments":[{"comment":"The headline round-trip, byte, and end-to-end cost reductions are computed from an offline counterfactual replay on stored expectation-value dumps for all scales in {1,3,5,7}. Because every scale is always executed and dumped each epoch, the bandit's arm choice does not reduce actual quantum executions; the 40%/35%/30% figures are an accounting of which scales the selected arm would have used, not a measured reduction from a live system in which only the selected arm runs and unselected rewards are unavailable. The real-hardware experiment in Section III.C also does not exercise the bandit online: all three arms are reconstructed from the shared scale set {1,3,5,7} and the QPU runs only frozen inference. The authors' own limitation statement in Section III.D confirms this. I request either reframing these numbers as offline counterfactual savings with an explicit caveat, or adding an online deployment experiment with partial feedback and real exploration cost.","section":"Section III.A.4-III.A.5, Table V"},{"comment":"The 'estimator fidelity' metric is never defined. If it means test accuracy, Table IV shows that the best arm [1,3,5,7] reaches 85.58% on CIFAR-10 while the fixed baseline [1,3,5] reaches 85.34%, a difference of 0.24 percentage points, not the 6.9% reported in Table V. If it is a fidelity of the extrapolated expectation values relative to the noiseless run, no formula is given in the paper. The headline improvement therefore cannot be checked. Please define the metric precisely, state how it is computed, and reconcile it with the accuracy and loss numbers in Table IV.","section":"Section III.B, Tables IV-V"},{"comment":"The regret bound in Theorem 1 relies on Assumption 2, E[r_{t,k}|x_t] = x_t^T theta*_k, but the implemented reward in Eq. (16) is a sum of the cross-entropy loss of a nonlinear VQC classifier, the variance of the extrapolated estimate, and a discrete depth penalty. There is no reason to expect this reward to be linear in the context vector, which includes epoch index, loss, accuracy at lambda=1, and variance across {1,3}. As written, the theory applies to a stylized linear-reward model rather than to the actual reward used in the experiments. Either restrict the theoretical claim to that stylized model, or provide a validation of the linearity assumption and a regret bound for the actual reward structure.","section":"Section II.C, Eq. (16), and Section II.D, Assumption 2"},{"comment":"The method depends on many free parameters: alpha, beta, gamma_c in Eq. (15); alpha, beta, d0 in Eq. (16); eta, lambda_c, rho, lambda_0 in Algorithm 1; and sigma^2_target and epsilon in Eq. (20). No sensitivity analysis is reported, and the specific values used in the experiments are not given in the section. Since the communication-savings claim is driven primarily by the cost penalty lambda_c and the depth penalty alpha, the reader cannot tell whether the reported 40%/35%/30% reductions are robust or are artifacts of the chosen cost weights. Please report all parameter values and add an ablation study over the main cost and exploration parameters.","section":"Section II.C and Algorithm 1"}],"minor_comments":[{"comment":"The sentence 'As Assgrows, performance benefits decline (curves fors!>=!256 are near)' is garbled and should be rewritten, and the notation 's!>=!256' in the caption is nonstandard.","section":"Section III.B, Fig. 7"},{"comment":"The symbols alpha and beta are reused for cost weights in Eq. (15) and for reward weights in Eq. (16), which is confusing; please rename one set of parameters.","section":"Eqs. (15)-(16)"},{"comment":"Reference [20], 'Noise pollution disrupts freshwater cetaceans,' appears unrelated to the NISQ noise statement it supports; please verify and replace with a relevant reference.","section":"References [19]-[20]"},{"comment":"The contributions list claims 'sample complexity' analysis, but Section II.D provides only a regret bound for the bandit selection rule, not a sample-complexity result for learning the ZNE estimate or the noise model; please align the claims with the actual theoretical content.","section":"Section I.B and Section II.D"},{"comment":"The reward in Algorithm 1 is written as CE_{t,k_t}^{ZNE} without an explicit definition; although Eq. (16) refers to cross-entropy of the ZNE-corrected prediction, the formula should be stated clearly for reproducibility.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the new thing is real—using a contextual bandit to pick ZNE folding sets with a cost-aware reward—and there is actual engineering substance behind it: public code, a simulator, and a small but real Rigetti run. The headline savings, though, are not yet demonstrated in a live deployment.\n\nWhat is genuinely new here is the arm formulation: the agent chooses among folding-scale sets, not just a single fold factor, and the reward explicitly trades extrapolation quality against round trips, bytes, and runtime. That combination is absent from the prior work they cite, and their Table I positioning is fair. The theory is standard LinUCB with a known cost penalty; the regret proof is routine but correct as far as I can tell. The simulation study is broad: two datasets, multiple widths and depths, shot-count ablations, and a dynamic noise schedule. The real-hardware experiment, while limited to frozen inference, is still useful evidence—240 tasks on Rigetti, all executed, with a shared-scale protocol that cut 540 tasks to 240. I also give them credit for an honest limitations section and for releasing the code.\n\nThe main soft spot is exactly what the stress-test note flags, and it holds up on reading. The 40/35/30% savings in the abstract and Table V are computed from offline replay. In Section III.A.4-5, expectation-value dumps for all scales are stored every epoch; the bandit chooses which precomputed dumps to use, not which circuits actually run. So 'cutting round trips' is an accounting measure, not a measured reduction in QPU executions. The real-hardware section does not run the bandit online either; it evaluates all arms from a shared scale set. The central operational claim therefore remains unvalidated under live partial feedback, where exploration has real cost and the context-reward relationship could differ. This is fixable, but it needs to be fixed before the headline is taken at face value.\n\nTwo smaller issues. 'Estimator fidelity' is never defined, and the 6.9% figure does not match the accuracy deltas in Table IV, which are around 0.2-0.3 percentage points. The linear-reward assumption is unverified, but that is standard for LinUCB and not a real weakness beyond the usual caveat.\n\nWho is this for: people working on adaptive error mitigation, especially ZNE for hybrid VQC workloads where communication cost matters. It deserves a serious referee. My recommendation: send it to review, but ask for an online or partial-feedback experiment, a definition of the fidelity metric, and a rewrite of the savings claims to distinguish executed savings from replay savings.","headline":"Genuine CMAB-ZNE combination with real code and hardware work, but the headline savings come from offline replay, not live deployment—worth refereeing with revisions.","tokens_in":24855,"tokens_out":3849,"would_cite":true,"duration_ms":40432,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A bandit that chooses ZNE folding arms online cuts quantum runs by up to 40%","keywords":["quantum error mitigation","zero-noise extrapolation","contextual multi-armed bandits","variational quantum circuits","dynamic noise","unitary circuit folding","communication overhead","LinUCB"],"falsifier":"Run the full CMAB-ZNE loop online on a noisy simulator or QPU where each epoch executes only the selected arm and its scales, then compare actual round trips, bytes, and estimator fidelity against the offline-replay numbers; if the gap to fixed folding is much smaller than 40%/35%/30% or fidelity no longer exceeds fixed ZNE, the central claim is refuted.","tokens_in":23703,"feed_emoji":"⚛️","tokens_out":5140,"duration_ms":47065,"temperature":0.7,"pith_summary":"Zero-noise extrapolation (ZNE) needs a set of noise-amplifying folding scales, and choosing that set poorly wastes quantum executions or biases the estimate. This paper claims that the choice can be treated as a contextual multi-armed bandit problem: at each epoch a policy observes circuit and noise features, picks one of several folding-scale sets (arms), and receives a reward that penalizes both poor extrapolation and excessive circuit depth. Under a simulated time-varying noise schedule, the adaptive picker is claimed to cut quantum-circuit execution round trips by up to 40%, bytes transferred by up to 35%, and end-to-end cost by up to 30%, while raising estimator fidelity by up to 6.9% against fixed-fold and grid-search ZNE baselines. A cost-limited run on a superconducting processor showed the shallow arm matching deeper arms in accuracy at lower cost, supporting the idea that adaptivity saves overhead without sacrificing fidelity.","feed_headline":"Adaptive circuit-folding choices cut quantum runs by up to 40%","feed_subtitle":"Under time-varying noise, a bandit policy trims bytes and cost while improving extrapolated estimator fidelity.","key_machinery":"The central object is a cost-aware contextual multi-armed bandit whose arms are ZNE folding-scale sets: arm k is a set $\\Lambda_k$ like $\\{1,3\\}$, $\\{1,3,5\\}$, or $\\{1,3,5,7\\}$, each evaluated by folding the circuit at those scales, estimating expectation values, and extrapolating to zero noise. Selection uses a LinUCB rule: for each arm the policy keeps a ridge-regression estimate of reward as a linear function of context, adds an exploration bonus, and subtracts a normalized execution cost that combines round trips, transferred bytes, and runtime. The reward balances the cross-entropy loss of the ZNE-corrected prediction, the variance of the extrapolated estimate, and a penalty for folding depth, so the bandit is rewarded for staying shallow when shallow arms suffice. The same structure yields the paper's sublinear regret bound: with a bounded linear reward model and known arm costs, the cost-aware UCB selection converges to the optimal arm in hindsight.","core_discovery":"The paper's central claim is that adaptive, context-aware selection of ZNE folding arms outperforms fixed schedules under time-varying noise. Concretely, the authors assert that a cost-aware LinUCB bandit, using a context vector built from epoch, loss, accuracy, and variance signals, learns to prefer shallow folding sets such as {1,3} early in training and deeper sets only when extrapolation gain justifies the extra executions. In simulations with dynamic noise on CIFAR-10 and EuroSAT, the adaptive policy is reported to reduce quantum-circuit execution round trips by up to 40%, bytes exchanged by up to 35%, and end-to-end cost by up to 30% under a 10 Mbps link budget, with up to 6.9% higher estimator fidelity versus fixed-fold and grid-search ZNE. On Rigetti hardware, all three ZNE arms gave the same accuracy, and the shallow arm had the best cost-aware reward, which the paper presents as evidence that unnecessary folding is waste.","pith_inferences":["The headline round-trip and byte savings are measured by offline replay over saved expectation-value dumps, not by an online run where only the selected arm is executed; whether the savings transfer to live deployment is an open question the paper's own hardware experiment does not close.","If the cost-aware reward were weighted differently, for example prioritizing latency over byte count, the arm selection would shift toward even shallower sets; the framework could be tuned to different network budgets without redesign.","The same arm-selection machinery could be applied to other ZNE variants, such as noise-aware or global-folding schemes, by changing only the arm definition and cost model.","A strong testable extension is to feed live calibration telemetry (T1/T2, gate and readout errors) directly into the context vector, which the paper lists as future work but does not implement."],"forward_implications":["A deployed CMAB-ZNE scheduler could replace fixed-fold or grid-search ZNE in VQC training loops, adapting fold choice per epoch without additional circuit executions for context.","Under the claimed savings, communication-constrained edge-cloud quantum machine learning becomes cheaper: fewer QPU-CPU round trips and fewer bytes for the same or better extrapolated accuracy.","The framework suggests a general pattern: any error-mitigation knob with a discrete set of cost-bearing choices can be cast as a contextual bandit with a cost-aware reward.","The budgeted stopping rule implies that when the extrapolated estimate is already stable, the policy can skip additional ZNE queries entirely, reducing latency in latency-sensitive workloads."],"supporting_citations":[{"why":"Introduces zero-noise extrapolation, the base mitigation method whose folding arms the bandit selects.","marker":"[9]"},{"why":"Formalizes digital ZNE with unitary folding, which defines the folding-scale arms used in the paper.","marker":"[10]"},{"why":"Provides best-practice ZNE with Pauli twirling/whirling, a baseline and motivation for regularizing noise before extrapolation.","marker":"[16]"},{"why":"Proposes adaptive folding with noise filtering (adaptive KIK), the closest adaptive baseline the paper compares against and improves on.","marker":"[6]"},{"why":"Gives the UCB1 bandit algorithm whose optimism principle underlies the LinUCB selection rule.","marker":"[21]"},{"why":"Supports the contextual bandit formulation for heterogeneous and drifting contexts.","marker":"[25]"},{"why":"Supplies the cryogenic temperature and attenuation model used in the time-varying thermal noise equation.","marker":"[31]"},{"why":"Supplies the metric-noise-resource gate fidelity model used to construct the noise context features.","marker":"[34]"}],"fun_headline_variants":["Bandit-tuned circuit folding saves 40% quantum runs","Adaptive ZNE cuts quantum runs by 40%","Time-varying noise? Bandit policy trims cost and boosts fidelity","Context-aware folding: 40% fewer runs, 6.9% better fidelity","Quantum noise: adaptive folding beats fixed schedules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 40% round-trip and 35% byte savings are computed by replaying selected arms over offline dumps where expectation values for all folding scales are already available; if live execution rewards differ from that replay, the savings may not materialize.","fun_headline_variants_meta":{"raw":{"variants":["Bandit-tuned circuit folding saves 40% quantum runs","Adaptive ZNE cuts quantum runs by 40%","Time-varying noise? Bandit policy trims cost and boosts fidelity","Context-aware folding: 40% fewer runs, 6.9% better fidelity","Quantum noise: adaptive folding beats fixed schedules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1473,"prompt_tokens":1076,"completion_tokens":397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":310}},"tokens_in":692,"tokens_out":397,"duration_ms":4210,"temperature":1.0,"reasoning_tokens":310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:25:38.481936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full CMAB-ZNE loop online on a noisy simulator or QPU where each epoch executes only the selected arm and its scales, then compare actual round trips, bytes, and estimator fidelity against the offline-replay numbers; if the gap to fixed folding is much smaller than 40%/35%/30% or fidelity no longer exceeds fixed ZNE, the central claim is refuted.","supporting_citations":[{"cited_title":"Adaptive folding and noise filtering for robust quantum error mitigation","cited_arxiv_id":"2505.04463","evidence_quote":"Proposes adaptive folding with noise filtering (adaptive KIK), the closest adaptive baseline the paper compares against and improves on."},{"cited_title":"Robust performance incen- tivizing algorithms for multi-armed bandits with strategic agents,","cited_arxiv_id":null,"evidence_quote":"Supports the contextual bandit formulation for heterogeneous and drifting contexts."},{"cited_title":"Engineering cryogenic setups for 100-qubit scale superconducting circuit systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the cryogenic temperature and attenuation model used in the time-varying thermal noise equation."}],"review_version":1}