{"id":"9da74f6f-0630-4ae8-86cb-219e44a5b7f8","arxiv_id":"2601.20871","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A benchmarking study reports gate cancellation as the best optimization pass and pulse duration as the strongest fidelity predictor, but the main correlation is built into its exponential-decay model.","lead":"This preprint benchmarks four standard quantum-circuit optimization passes by simulating fidelity with a Lindblad model and running eight jobs on IQM Garnet hardware. It reports that gate cancellation helps most and pulse duration is the main fidelity predictor, a result useful mainly to quantum compiler engineers if the artifacts can be verified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pulse-duration correlation is an artifact of the exponential-decay model, not an empirical result.","rationale":"The reader's weakest assumption correctly identifies the exponential-decay approximation as the source of the pulse-duration correlation. This is the most load-bearing concern because the paper's central, actionable claim—that compiler designers should minimize pulse duration—rests on this correlation. The manuscript itself states the approximation in Section VII.B, confirming the model's built-in time dependence. I also considered a second serious issue: Eq. (2) defines process fidelity via a trace over a unitary U_impl, which is ill-defined for a non-unitary Lindblad process. However, the exponential-decay artifact is more directly tied to the headline correlation and is explicitly self-admitted, making it the cleaner target. The abstract/body inconsistencies (72% vs. 68%, missing Qiskit comparison, absent upper-bound analysis) further weaken the paper but are secondary; even if those were fixed, the modeling artifact would remain. A full-Lindblad re-computation is a concrete, feasible test because the paper claims cross-validation against qiskit-dynamics. If the correlation persists, the claim would gain credibility; if it vanishes, the rejection stands.","tokens_in":5827,"tokens_out":5287,"duration_ms":63264,"concrete_test":"Run the same 371-circuit corpus through a full Lindblad master-equation simulation using qiskit-dynamics (the tool the paper says it cross-validated against) with IQM Garnet noise parameters, and recompute the Pearson r and R² between pulse duration and process fidelity. If |r| drops well below 0.74 or R² falls below 0.55, the headline correlation is an artifact of the exponential-decay approximation. As an additional control, compute the partial correlation of pulse duration with fidelity after controlling for total gate count; a near-zero partial correlation would indicate that pulse duration is only a proxy for circuit size.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim—that pulse duration is the strongest negative predictor of process fidelity (Table IV: r=-0.74, R²=0.55)—is effectively guaranteed by the fidelity model. Section VII.B explicitly admits: 'The pulse simulation uses an exponential decay model approximating the full Lindblad dynamics.' Under an exponential-decay ansatz, process fidelity is a decreasing function of total evolution time, so any regression of fidelity against ideal pulse duration will produce a strong negative correlation regardless of the actual noise structure. The paper provides no evidence that this correlation survives when non-exponential errors (crosstalk, leakage, calibration drift, per-gate coherent errors) are included. The hardware validation cannot settle this because only 4 distinct circuits were run with 160 shots each, and the hardware fidelity estimator is not described. Moreover, the abstract's claimed upper-bound analysis (mean overestimate 0.49) is absent from Section VI, so it cannot independently support the correlation. The actionable guidance 'minimize pulse duration' is therefore a restatement of the modeling assumption, not a robust empirical optimization result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents qco-integration, a modular framework connecting a C++ circuit optimizer to a Python pulse-level simulation, and claims an end-to-end fidelity analysis across 371 benchmark circuits (abstract says 4,452 experiment runs) on IQM Garnet parameters. The central claims are: gate cancellation is the dominant optimization pass (68% of circuits improved, 14,024 gates removed), pulse duration is the strongest negative predictor of process fidelity (r = -0.74, R² = 0.55 in the body; r = -0.73, R² = 0.53 in the abstract), the optimal pass ordering is cancel→commute→rotate, and hardware execution on IQM Resonance Garnet validates the model as a consistent upper bound with mean overestimate 0.49. The abstract also claims a Qiskit transpilation comparison and a formal pass-ordering ablation (Kruskal–Wallis p = 0.302) that do not appear in the body. The paper releases code and data as open source.","tokens_in":6051,"tokens_out":3464,"duration_ms":40750,"significance":"If the claims were supported, the paper would provide useful compiler guidance: prioritize cancellation and minimize pulse duration to improve NISQ fidelity. The open-source framework, actual IQM hardware executions, and per-pass statistics are concrete contributions. However, the headline empirical claims are undermined by internal inconsistencies and, more fundamentally, by the fact that the pulse-duration correlation is built into the fidelity model through the exponential-decay approximation. The claimed hardware validation and Qiskit comparison are not actually reported in the body. The significance of the paper therefore rests on claims that the presented evidence does not substantiate.","major_comments":[{"comment":"The central claim that pulse duration is the strongest negative predictor of process fidelity (Table IV, r = -0.74, R² = 0.55) is an artifact of the modeling assumption. Section VII.B states: 'The pulse simulation uses an exponential decay model approximating the full Lindblad dynamics.' Under exponential decay, process fidelity is a decreasing function of total evolution time by construction, so any regression of fidelity against pulse duration will produce a strong negative correlation regardless of actual noise structure. The paper provides no test with non-exponential errors (crosstalk, leakage, calibration drift, coherent per-gate errors), so the actionable conclusion 'minimize pulse duration' in Section VII.A is a restatement of the model, not an empirical result.","section":"Section VII.B and Eq. (1)"},{"comment":"The abstract and body report irreconcilable numbers for the same campaign: abstract says 4,452 experiment runs over 371 circuits while Section V.A/Table II say 'Total circuit runs 371'; gate-cancellation improvement is 72% in the abstract vs. 68% in Table III; pulse-duration correlation is r = -0.73, R² = 0.53 in the abstract vs. r = -0.74, R² = 0.55 in Table IV; input-gate correlation is r = -0.78 in the abstract but Table IV gives r = -0.606. These are not minor typos; they affect the headline numbers and prevent reproducibility assessment.","section":"Abstract vs. body (Table II, Table III, Table IV)"},{"comment":"The abstract claims: 'Comparing against Qiskit transpilation levels...' and 'A formal ablation study shows that pass ordering has no significant effect... Kruskal–Wallis p = 0.302.' Neither analysis appears in the body. Section V reports per-pass effectiveness and Figure 5 baseline vs. optimized, but no Qiskit comparison and no statistical test of pass ordering. These are load-bearing claims for the paper's stated contributions and must either be reported with data or removed.","section":"Abstract vs. body: missing Qiskit and pass-ordering analyses"},{"comment":"The hardware validation does not support the abstract's claim that the model is a 'consistent upper bound' with mean overestimate 0.49. That number is absent from the body. No process-fidelity estimator is described (e.g., randomized benchmarking, tomography, or readout correction), so it is unclear what Table V's 'Fidelity' numbers mean. Only 4 distinct circuits are reported with 160 shots each, and GHZ circuits show no optimization benefit; QFT shows 0.088 vs. 0.100, a difference that is not statistically assessed. This is insufficient to validate the upper-bound claim or the predictive strength of pulse duration.","section":"Section VI and Table V"},{"comment":"Even setting aside the exponential-decay artifact, Table IV reports only univariate correlations. Pulse duration, gate count, depth, and qubit count are strongly inter-correlated, so the claim that pulse duration has the 'strongest predictive power' requires a multivariate or partial-correlation analysis, or at least a demonstration that the R² difference is not confounded. Without such an analysis, Table IV does not establish that pulse duration is the dominant factor.","section":"Table IV"}],"minor_comments":[{"comment":"The paper contains many small inconsistencies, e.g., 'SW AP' for 'SWAP' in Section III.A, and the abstract says 8 circuits while Section VI.A says 8 jobs with 4 distinct circuit types. Please standardize terminology and numbers.","section":"Throughout"},{"comment":"The error bars in Figure 2 are described as standard deviation across the corpus, but the bar for 'CommutationPass' shows zero gates removed with no explanation of how a standard deviation is shown for a zero-variance quantity.","section":"Figure 2"},{"comment":"The limitation that hardware validation and simulation use different circuit sets is appropriately acknowledged, but this limitation directly undermines the claim of 'consistency' between simulation and hardware. Please discuss the implications more fully.","section":"Section VII.B"}],"recommendation":"reject","confidential_remarks":"The manuscript has load-bearing internal contradictions and a central correlation that is an artifact of the exponential-decay model. Even if the authors removed the unsupported Qiskit and pass-ordering claims, the main result about pulse duration would require a fundamentally different noise model and a coherent experimental campaign. I do not see a route to acceptance within the scope of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The pulse-duration correlation is not discovered, it's assumed. Once you model dynamics as exponential decay, fidelity must fall with total evolution time, so Table IV's r = -0.74 is a restatement of Eq. (1) plus the Section VII.B approximation. The hardware data can't rescue it: four circuits, 160 shots, and no described fidelity estimator. And the abstract keeps promising things the body doesn't deliver: the Qiskit comparison, the pass-ordering ablation, and the \"consistent upper bound\" analysis (mean overestimate 0.49) never appear in Section VI.\n\nThe paper does have real value. The qco-integration harness is a sensible decoupled pipeline, and the per-pass ranking—cancellation first, rotation merging second, identity elimination negligible, commutation only enabling—is likely correct and consistent with existing compiler experience. Running eight jobs on a real Garnet device with released job IDs is honest evidence that the toolchain works, even if the statistical weight is small. The 70% QFT gate reduction is concrete.\n\nThe soft spots are not minor. The abstract/body disagreements (72% vs 68% improved, r = -0.73 vs -0.74, R² = 0.53 vs 0.55) look like a sloppy or inflated abstract. The missing promised analyses mean a reader can't verify the strongest claims. And the central correlation is an artifact of the modeling premise. The paper also claims cross-validation against qiskit-dynamics, but I don't see that experiment in the body. The citation list is appropriate; no red flags there.\n\nI'd send it to a referee, not because the current claims hold, but because the integration framework is worth salvaging and a referee could push the author to strip the abstract down to what is actually done: a modular benchmark of standard passes with a known decoherence model, plus a small hardware demonstration. As written, no.\n\nDo not cite it yet. Maybe bring it to the reading group as a case study in model-induced correlations.","headline":"The headline pulse-duration result is written into the model; the abstract describes analyses the body doesn't include.","tokens_in":6493,"tokens_out":2430,"would_cite":false,"duration_ms":28153,"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":"Total pulse duration is the strongest fidelity predictor in compiled quantum circuits, and gate cancellation is the most effective optimizer pass.","keywords":["quantum circuit optimization","process fidelity","Lindblad master equation","pulse-level control","gate cancellation","compiler passes","NISQ hardware validation","decoherence modeling"],"falsifier":"Run the same logical circuit on real hardware with two pulse schedules that have identical gate counts and depths but different total durations, for example by inserting calibrated idle time. If the longer schedule does not show lower fidelity, the pulse-duration claim fails. Alternatively, measure fidelity under simultaneous-gate crosstalk: if crosstalk cancels or reverses the duration effect, the exponential-decay model is incomplete.","tokens_in":5713,"feed_emoji":"⚛️","tokens_out":5329,"duration_ms":54562,"temperature":0.7,"pith_summary":"This paper tries to establish which stage of quantum compilation actually matters for near-term hardware fidelity. It connects a gate-level circuit optimizer to a Lindblad-based pulse-level noise simulation and runs 371 benchmark circuits through the pipeline. Its central findings: gate cancellation is the dominant optimization pass (removing roughly 14,000 gates and improving most circuits), and total pulse duration is the strongest negative predictor of process fidelity (R²≈0.55), ahead of gate count, depth, and qubit count. Because it also runs a small set of circuits on a real 20-qubit processor, it claims the model preserves relative difficulty ordering, though it overestimates absolute fidelity by about 0.49. A sympathetic reader would take away that compiler design should target execution time and cancellation, not just gate-count reduction.","feed_headline":"Pulse duration beats gate count as quantum fidelity predictor","feed_subtitle":"A 371-circuit campaign finds cancellation is the key compiler pass and runtime, not gate count, sets decoherence limits.","key_machinery":"The load-bearing tool is a per-gate Lindblad master-equation fidelity model: each gate is executed as a control pulse with decoherence rates set by T1 and T2 times, and process fidelity is computed from the simulated density matrix. This lets the paper translate gate-level transformations into predicted process fidelity before hardware runs. The four optimizer passes—gate cancellation, commutation, rotation merging, and identity elimination—are the independent variables; cancellation is the workhorse, while commutation creates cancellation opportunities without removing gates itself.","core_discovery":"The paper's central claim is that end-to-end compiled-circuit fidelity is controlled less by how many gates remain and more by how long the whole pulse sequence takes, with gate cancellation as the dominant way to shorten it. In simulation over 371 circuits, cancellation alone removed about 14,000 gates and improved 68–72% of circuits, while pulse duration correlated with process fidelity at r≈−0.74, giving R²≈0.55; the next strongest predictor, input gate count, gave R²≈0.37. The paper also reports that pass ordering among cancellation, commutation, rotation merging, and identity elimination has no significant effect on two-qubit gate reduction, and that on a real 20-qubit processor the sim","pith_inferences":["The pulse-duration result is partly guaranteed by the exponential-decay Lindblad model, so the headline R² is not purely empirical; a hardware test that varies only pulse timing would separate model built-in from real effect.","Cancellation's dominance may not generalize to circuits that are already near-minimal, such as GHZ preparations; the gains concentrate in rotation-heavy circuits like QFT and QAOA.","The consistent 0.49 model-hardware gap suggests absolute fidelity is dominated by unmodeled noise sources; incorporating crosstalk, leakage, or calibration drift could close the gap.","If pulse duration is confirmed as the bottleneck, then optimizing routing to shorten execution time may matter more than minimizing swap count itself."],"forward_implications":["Compiler pipelines should run gate cancellation early and repeatedly; it is cheap and yields the largest fidelity gains.","Minimizing total pulse duration—through gate reduction and parallel scheduling—should be a primary NISQ optimization objective.","Since pass ordering does not significantly change two-qubit gate reduction, simple fixed orderings such as cancellation→commutation→rotation merging are sufficient.","Benchmarking should report two-qubit gate count and pulse duration, not just total gate count, because single-qubit consolidation can dominate total-gate metrics without hardware benefit.","The model can be used to rank circuits by difficulty even though its absolute fidelity predictions are consistently about 0.49 too high."],"fun_headline_variants":["Gate cancellation, not pass order, drives quantum fidelity","Pulse duration beats gate count for quantum fidelity","Quantum fidelity depends on pulse length, not gate count","Pulse time sets quantum fidelity more than gate count","Quantum fidelity predicted by pulse duration, not gate count"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fidelity model assumes decoherence follows exponential decay with fixed T1 and T2 times; under that assumption longer pulses mathematically must hurt fidelity, so the central pulse-duration correlation is substantially built into the model rather than independently observed.","fun_headline_variants_meta":{"raw":{"variants":["Gate cancellation, not pass order, drives quantum fidelity","Pulse duration beats gate count for quantum fidelity","Quantum fidelity depends on pulse length, not gate count","Pulse time sets quantum fidelity more than gate count","Quantum fidelity predicted by pulse duration, not gate count"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00046,"raw_usage":{"total_tokens":2213,"prompt_tokens":889,"completion_tokens":1324,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":1249}},"tokens_in":633,"tokens_out":1324,"duration_ms":10420,"temperature":1.0,"reasoning_tokens":1249,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:52:09.763859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same logical circuit on real hardware with two pulse schedules that have identical gate counts and depths but different total durations, for example by inserting calibrated idle time. If the longer schedule does not show lower fidelity, the pulse-duration claim fails. Alternatively, measure fidelity under simultaneous-gate crosstalk: if crosstalk cancels or reverses the duration effect, the exponential-decay model is incomplete.","supporting_citations":[],"review_version":1}