{"id":"9294d076-2fe9-4f2f-9209-d3f4e916b27c","arxiv_id":"2607.23211","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Q-BRIDGE, a graph-transformer denoiser conditioned on backend features, reconstructs ideal outcome distributions from noisy quantum executions and markedly improves oracle-based bug detection.","lead":"A new machine-learning tool called Q-BRIDGE cleans up the noisy outputs of quantum programs so that software tests can tell real bugs apart from hardware noise. Tested on 23 simulated IBM quantum devices, it spots bug-induced failures with about 94% precision and 83% recall, well above the previous best method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference clips outputs but does not normalize them, so the 'denoised distribution' may not sum to 1; Hellinger distances and oracle thresholds may be computed on invalid probability vectors.","rationale":"The reader's weakest assumption concerns external validity: simulated Fake-backend noise models may not represent real device drift and crosstalk. That is a legitimate threat to generalization, but it does not undermine the correctness of the measurements on the simulated backends themselves. The probability-simplex issue is more load-bearing because it threatens the internal validity of every reported number: if the model output is not normalized, the Hellinger distances, the WODF oracle decisions, and the precision/recall/F1 values are not well-defined for a probability measure. The paper even claims to enforce a valid probability simplex, yet the described inference does not do so. This is a concrete, checkable flaw that could overturn the central claim even on the exact backends evaluated. I therefore disagree with the reader's selection of the weakest assumption, while agreeing that the verdict should remain conditional pending verification. The concrete test would settle whether the outputs are actually normalized in practice or by the code; if they are not, the paper would need to re-run all evaluations with proper normalization or explicitly describe it.","tokens_in":19234,"tokens_out":4124,"duration_ms":37232,"concrete_test":"Take a random sample of 100 test instances (circuit-family × backend × seed) from the evaluation suite. For each, compute the sum of the model's final outputs \\hat p_t over all basis states after the clipping step. If the sum deviates from 1 by more than 1% for any instance, and especially if the mean absolute deviation across instances is nontrivial, then the evaluation metrics are computed on non-normalized outputs. Also inspect the release code to confirm whether renormalization is applied outside the described pipeline; if it is, the paper's description is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper states in Section II.C that it 'enforce[s] a valid probability simplex at inference,' but Section III.E describes only clipping each output to [0,1] with no normalization step. Section III-A's notation \\hat p_t ∝ clip(\\tilde p_t,0,1) suggests normalization, yet the actual inference procedure in Section III-E does not divide by the sum. The training objective (Section III-D) is a per-state MSE with no sum-to-one constraint, so nothing forces the model to emit a normalized vector. Consequently, the model's output may not be a probability distribution. The evaluation metrics (Hellinger distance, UOF/WODF oracles) and the headline precision/recall/F1 numbers presuppose a valid probability vector; computing them on non-normalized outputs could make the reported improvements over QOIN meaningless or artifactual. This is an internal inconsistency that directly affects the validity of the empirical claims, independent of whether the noise models transfer to real hardware.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Q-BRIDGE is a graph-transformer-based denoiser for quantum program testing. The method encodes the transpiled circuit as a DAG with node/edge features derived from gate types, calibration data, and scheduling, and uses a FiLM-style head to condition on per-state observation statistics extracted from noisy executions. A general model also conditions on a learned backend embedding. The model is trained by per-state MSE to match ideal-simulator distributions. Experiments cover 23 Qiskit Aer fake IBM noise models and 6 circuit families, with two deployment scenarios (backend-wise and general). The paper reports lower Hellinger distance than the QOIN baseline on all backends/families and improved precision/recall under UOF/WODF oracles.","tokens_in":19533,"tokens_out":5333,"duration_ms":56449,"significance":"If the empirical claims are sustained, Q-BRIDGE would be a useful contribution to noise-aware quantum software testing: it incorporates circuit structure and backend information more explicitly than prior summary-based methods, and it addresses the practical scenario of distribution-level denoising rather than single-observable mitigation. The paper has clear strengths: a public replication package, held-out CFUT inputs, five random seeds, consistent baselines under the same pretrain/fine-tune protocol, and explicit ablation and scalability studies. However, two load-bearing issues currently prevent acceptance: the inference procedure does not appear to output a normalized probability vector despite the paper's claims, and all 'backends' are vendor-supplied simulator noise models rather than real quantum hardware. Both issues affect the interpretation of the headline numbers.","major_comments":[{"comment":"There is a direct internal inconsistency in the definition of the denoised distribution. Section III.A writes \\hat p_t ∝ clip(\\tilde p_t,0,1), which implies renormalization, and Section II.C claims the model 'enforce[s] a valid probability simplex at inference.' However, Section III.E describes only clipping to [0,1] with no division by the sum. Since the training objective in Section III.D is a per-state MSE with no sum-to-one constraint, the model has no incentive to emit a normalized vector. Hellinger distance and the UOF/WODF oracles in Section IV.F/V all assume a valid probability distribution; computing them on an unnormalized vector can produce meaningless or even negative values inside the Hellinger square root. This is load-bearing for every quantitative claim in Tables III–V. Please clarify the actual inference code: if it normalizes, correct the text; if it does not, re-run th","section":"§III.A and §III.E"},{"comment":"All 23 'IBM noise backends' are Qiskit Aer device noise models (Fake* backends), not real hardware executions. Section VII correctly acknowledges that real-device transfer is future work, but the abstract and headline phrasing ('23 IBM noise backends') overstates the evidence. The paper's broader claim that considering 'physical characteristics of specific quantum backends' is a practical route to reliable testing is not yet supported for actual devices, since calibration drift, crosstalk, and simulator-model mismatch are explicitly named as open threats. At a minimum, revise the abstract and conclusions to say 'simulated IBM noise models', and temper the practical-route language. A real-hardware validation on even one or two backends would substantially strengthen the claim.","section":"§IV.B.2, §VII, Abstract"},{"comment":"The general model conditions on a learned backend embedding table E indexed by the 23 training backends. There is no described mechanism to compute an embedding for an unseen backend at inference time. Thus RQ2 demonstrates parameter sharing across the 23 seen backends, not generalization to new devices. The phrase 'cross-backend generalization' should be qualified accordingly, and the paper should state what happens when the model is applied to a backend not in the embedding table (e.g., zero vector, nearest neighbor, or retraining required).","section":"§III.C.2.d and RQ2"},{"comment":"The testing evaluation uses the same ideal simulator as the training target and as the oracle reference for ground-truth labeling. This is an acceptable denoising benchmark because QOIN is evaluated under the same protocol, but the phrase 'detecting bug-induced failures' should be qualified. The protocol measures whether denoising makes noisy outputs consistent with a known ideal reference distribution; it does not evaluate a situation where no ideal reference is available. Please state this explicitly in the RQ3 discussion so readers do not over-interpret the precision/recall numbers as evidence of fault detection without an ideal oracle.","section":"§V.C and Table V"}],"minor_comments":[{"comment":"Use 'simulated IBM noise models' or 'Qiskit Aer fake backends' instead of 'IBM noise backends' to avoid implying real-hardware evaluation.","section":"Abstract"},{"comment":"The notation '\\hat p_t ≜ p_ideal_t' is confusing: it suggests the output is defined as the target. Use a target variable such as y_t = p_ideal_t.","section":"§III.A"},{"comment":"The observation feature includes log(ODR_t), but ODR_t is undefined if POS_t is clipped to ε or 1-ε. The clipping choice ε=1e-6 is not analyzed; a sentence on sensitivity would help.","section":"§III.B.1"},{"comment":"The ablation study is only summarized in words and relegated to the replication package. Since the paper claims each architectural component contributes positively, include at least a compact table (e.g., Hellinger values with/without edge bias, FiLM, backend embedding) in the main text.","section":"§VI"},{"comment":"The WODF threshold is set to 0.01 following prior work, but no sensitivity analysis is reported. Given the large effect on precision/recall in Table V, a brief threshold sweep would strengthen the results.","section":"§IV.F"},{"comment":"Minor typographical issues: 'QISKIT' is inconsistently capitalized; 'we hypothesis' should be 'we hypothesize'; Table III uses many decimal places for standard deviations that are occasionally exactly zero, which may indicate rounding artifacts.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The normalization issue is the most serious technical concern and is fixable only if the authors clarify and re-run the experiments. The simulated-backend limitation is a framing/scope issue, not a fatal flaw, provided the claims are honestly scoped. The empirical scaffolding is above average for the area, and if the corrected numbers remain similar, I would be supportive of publication. I would not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine new combination — a graph transformer over transpiled circuit DAGs plus FiLM conditioning on shot statistics and backend identity — evaluated broadly and carefully against QOIN. The empirical protocol is solid: held-out inputs, five seeds, consistent baselines, and a good spread of backends and circuit families. Credit where earned: the encoding of scheduling and calibration features into the graph is thoughtful, and the cross-backend general model is a real step beyond per-device training.\n\nThe problem is an internal contradiction that cuts to the metric. Section II.C says the model 'enforce[s] a valid probability simplex at inference,' and Section III.A writes \\hat p_t ∝ clip(...). But Section III.E only clips each output to [0,1] with no normalization, and the training objective is per-state MSE with no sum-to-one constraint. So the 'denoised distribution' may not sum to 1. Hellinger distance is only defined for probability vectors; WODF thresholds also assume a normalized distribution. If Q-BRIDGE's outputs aren't normalized, the reported relative improvements and the 93–95% precision could be artifacts of scale, not of better denoising. This isn't a minor typo — it directly affects every number in Tables III–V.\n\nThe abstract also says '23 IBM noise backends' when all runs are Qiskit Aer FakeBackends, i.e., vendor-supplied noise models, not real hardware. The paper admits this in the threats section, but the abstract overstates it. That's a softer issue, more about external validity.\n\nThe ablation study and some scalability details are in the replication package rather than the paper; that's a minor transparency concern, though the code link is there.\n\nOverall: the idea is interesting and the engineering is honest, and the code is available, so this deserves a serious referee. But the normalization contradiction is load-bearing. If the authors can show that clipped outputs are effectively normalized, or add renormalization and the results persist, this is a useful paper for quantum software testing. As is, I would not cite it, but I'd bring it to a reading group to discuss exactly this kind of metric trap.","headline":"A real new architecture and a careful evaluation, but the denoised vectors aren't actually normalized despite the claim, so the headline metrics may not mean what they appear to.","tokens_in":19923,"tokens_out":4029,"would_cite":false,"duration_ms":40115,"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":"Encoding a quantum circuit's structure and its backend's physics lets a learned denoiser catch faults that earlier filtering almost entirely misses.","keywords":["quantum program testing","noise mitigation","graph transformer","backend-aware machine learning","outcome distribution denoising","oracle-based testing","Hellinger distance","fault detection"],"falsifier":"Run the same six circuit families and the same fault-injected mutants on real quantum hardware backends, apply the trained denoiser, and compare denoised distributions to the ideal simulator distributions over several days. If the Hellinger reduction over the unfiltered output shrinks toward zero, or recall of bug-induced failures drops toward the prior baseline, the central claim is refuted; if calibration drift changes results from day to day, the model's backend conditioning is incomplete.","tokens_in":19172,"feed_emoji":"⚛️","tokens_out":4963,"duration_ms":46183,"temperature":0.7,"pith_summary":"Quantum program testing on noisy hardware is unreliable because device noise shifts measured distributions and can flip pass/fail verdicts. This paper tries to establish that a graph transformer which encodes the transpiled circuit's gate structure, timing, and backend calibration, and then conditions that encoding on observed shot statistics, can undo enough of the noise to make oracle-based tests trustworthy. Across 23 simulated device noise models and six circuit families, the denoiser moves output distributions closer to the noise-free ideal than the strongest prior learned filter on every device. In fault-injection testing it detects over 80 percent of bug-induced failures while keeping precision near 94 percent, where the prior approach detects under 10 percent. A single model shared across all backends preserves most of this benefit, so per-device retraining is not strictly necessary.","feed_headline":"Graph denoiser lifts quantum test recall from 9.6% to 83.5%","feed_subtitle":"Encoding the compiled circuit's structure and backend physics lets oracles tell real bugs from device noise.","key_machinery":"The load-bearing mechanism is a graph transformer with edge-biased attention combined with a state-wise FiLM (feature-wise linear modulation) conditioning head. The encoder treats each gate as a node and each dependency as an edge, with edge attributes encoding how much a qubit's state decays during the idle gap between gates via exp(-Δt/T1) and exp(-Δt/T2) survival factors. The FiLM head modulates the graph embedding separately for every computational basis state using that state's observed count statistics, allowing the model to decide per-outcome whether to amplify or suppress probability mass. This lets the model reconstruct a full output distribution instead of predicting a single expec","core_discovery":"The central claim is that noise mitigation for quantum program testing is best done as structure-aware distribution reconstruction rather than summary-based reweighting. Q-BRIDGE converts the transpiled circuit into a directed acyclic graph whose node features carry gate type, per-qubit calibration (T1, T2, gate error, readout error), timing, and gate parameters, and whose edge features carry idle-time decoherence survival factors computed from scheduling gaps. A graph transformer with edge-biased attention encodes this graph, and a feature-wise linear modulation (FiLM) head combines the graph embedding with per-outcome observation features—empirical probability, its complement, log odds, an","pith_inferences":["Editorial extension: because the experiments use vendor-supplied simulated noise models rather than live hardware, the most direct testable prediction is that the same gains will appear on real devices; if calibration drift is severe, the general model's backend embedding may need to be refreshed continuously.","Editorial extension: the per-state observation features (empirical probability, complement, log odds, shot count) are essentially sufficient statistics of a multinomial observation, so the method may extend naturally to other output alphabets, including continuous or hybrid observables, by swapping the embedding layer.","Editorial extension: the failure-mode analysis suggests a cheap diagnostic—backends whose readout-error heterogeneity or two-qubit gate-error heterogeneity is high will be the ones where learned denoising helps least, so a testing team could use those indicators to decide when mitigation is worth running.","Editorial extension: one could test the general model on a backend never seen in training by initializing its embedding and fine-tuning only that vector, turning cross-backend generalization into a few-shot adaptation problem."],"forward_implications":["Distribution-level denoising can be framed as supervised graph-to-distribution regression trained on paired noisy and ideal executions, and this framing transfers across devices.","A shared backend-conditioned model recovers most of the accuracy of per-device models, lowering the cost of deploying noise mitigation to many backends.","Oracle-based test verdicts become substantially more reliable: far fewer correct programs are rejected, while most bug-induced failures are still caught.","On larger circuits (8–15 qubits) the same approach still beats the prior learned filter, though absolute error grows, suggesting scalability but with diminishing returns.","Residual error concentrates on backends with highly heterogeneous readout errors and two-qubit gate errors, so calibration quality, not just circuit structure, bounds what any learned denoiser can recover."],"fun_headline_variants":["Graph denoiser lifts quantum test recall to 83.5%","Backend-aware graph model catches quantum bugs with 94% precision","Q-BRIDGE: Backend-aware graph learning for quantum test denoising","One graph model denoises across 23 IBM quantum backends","Quantum test reliability boosted by graph-based noise denoising"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline results rest on the assumption that the 23 vendor-supplied simulated device noise models faithfully capture what real quantum hardware does—including calibration drift and crosstalk—so the measured Hellinger improvements and test metrics may not transfer to actual devices.","fun_headline_variants_meta":{"raw":{"variants":["Graph denoiser lifts quantum test recall to 83.5%","Backend-aware graph model catches quantum bugs with 94% precision","Q-BRIDGE: Backend-aware graph learning for quantum test denoising","One graph model denoises across 23 IBM quantum backends","Quantum test reliability boosted by graph-based noise denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000993,"raw_usage":{"total_tokens":4078,"prompt_tokens":810,"completion_tokens":3268,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":3175}},"tokens_in":554,"tokens_out":3268,"duration_ms":20911,"temperature":1.0,"reasoning_tokens":3175,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T03:16:39.904154+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same six circuit families and the same fault-injected mutants on real quantum hardware backends, apply the trained denoiser, and compare denoised distributions to the ideal simulator distributions over several days. If the Hellinger reduction over the unfiltered output shrinks toward zero, or recall of bug-induced failures drops toward the prior baseline, the central claim is refuted; if calibration drift changes results from day to day, the model's backend conditioning is incomplete.","supporting_citations":[],"review_version":1}