PaperlayerView exact revision on arXiv

Source-linked AI summary

QROB: Quantifying Realization Overhead in Quantum Compilation via Reverse Construction

Jintao Li, Kaiqi Li, Rui Wang, Yilun Zhao, Kaixuan Huang, Ying Wang, Jialin Zhang, Zheng-An Wang, Xiaoming Sun, Heng Fan

arXiv:2609.07776v1quant-phcs.AR

TL;DR

Practical-scale quantum compilation evaluations lack calibrated references for realization overhead under hardware locality constraints. QROB constructs instances backward from directly realizable configurations and retains feasible, compiler-independent references across routing and scheduling settings. Evaluations report substantial overhead gaps, improved learned routing, and higher mirror-circuit survival for QROB references than Qiskit O3 compilations on real processors.

  • Problem

    At practical scales, compilation evaluations lack independently assessable reference solutions because optimal locality-constrained compilation is NP-hard and achievable optima are generally unknown.

    Source: D. Problem Statement · p. 4

  • Method

    QROB constructs compilation instances backward from directly realizable configurations, retaining feasible references for NISQ routing, lattice-surgery scheduling, and capacity-constrained memory-access scheduling.

    Source: §I · p. 2, §VI · p. 14

  • Results

    QROB references yield a 1.65× median mirror-circuit survival ratio versus Qiskit O3 across three IBM Heron-r2 processors, while trained routers outperform SABRE on 84.8% of real circuits.

    Source: H. Summary · p. 14

  • Takeaways & Limitations

    QROB provides a common framework for measuring realization overhead and a supervision source for learned compilation across geometric-routing settings and a memory-access extension.

    Source: §VI · p. 14

    QROB isolates architecture-induced realization overhead rather than modeling the full compilation and execution stack, including upstream synthesis and several fault-tolerant costs.

    Source: F. Scope and Limitations · p. 8

Abstract

from arXiv · show

Quantum compilation reconciles a program's idealized interaction topology with hardware locality constraints, yet evaluations at scale lack calibrated references for realization overhead. We present QROB, a scalable reverse-construction methodology that generates compilation instances backward from directly realizable configurations, retaining the inverse paths as feasible, compiler-independent references. QROB provides a common evaluation substrate for NISQ SWAP routing and fault-tolerant lattice-surgery scheduling, while extending its reference-preserving principle to capacity-constrained quantum memory-access scheduling. Across systems ranging from 9 to 156 qubits, evaluations highlight QROB's utility as both a diagnostic benchmark and a data source. First, for compiler characterization, QROB reveals substantial realization gaps in existing tools, with NISQ compilers incurring up to 24.1x the reference SWAP cost and fault-tolerant compilers requiring up to 7.0x the reference makespan. Second, as a supervision source for data-driven compilation, a router trained on QROB references outperforms Qiskit SABRE on 84.8% of real-world application circuits. Finally, on real hardware, QROB reference realizations achieve a median mirror-circuit survival rate 1.65x that of full Qiskit O3 compilations across three 156-qubit IBM Heron-r2 processors, demonstrating that closing algorithmic compilation gaps translates directly into physical fidelity gains.

I. INTRODUCTION

Quantum compilation adds realization costs when hardware locality conflicts with a program’s idealized interactions, but practical-scale evaluations lack calibrated references for achievable overhead. QROB constructs instances backward from directly realizable configurations, retaining feasible references for compiler evaluation across routing and scheduling settings.

I. INTRODUCTION

  • Compilers must reconcile hardware-agnostic interactions with locality constraints, which require SWAP routing on near-term devices or additional space–time overhead in lattice-surgery architectures.The problem is hard in general, and practical-scale solution quality is difficult to evaluate because achievable optima are generally unknown.Source: §I · pp. 1–2
  • Existing evaluations lack compiler-independent calibrated references for the minimum realization overhead needed to map ideal programs onto constrained hardware.Compiler-dependent ceilings based on existing workflows are circular as measures of compiler quality.Source: §I · p. 2
  • QROB generates harder instances from directly realizable configurations and retains the inverse transformations as feasible references without solving the forward optimization problem.The retained reference is reproducible and compiler-independent, but is not assumed globally optimal.Source: §I · p. 2
  • QROB applies its interaction-locality framework to NISQ routing and fault-tolerant lattice-surgery scheduling, and tests extension to capacity-constrained memory-access scheduling.Its stated benchmark scope is instances from 9 to 900 qubits generated in sub-second time.Source: §I · p. 2
  • 7.2× the QROB reference SWAP count on average, rising from 3× on 16-qubit systems to 24× on 156-qubit processors; lattice-surgery compilers incur 6.3–7.0× reference execution time.These comparisons characterize overhead in the evaluated NISQ and fault-tolerant settings.Source: §I · p. 2
  • Routers trained on QROB references outperform SABRE on 84.8% of real circuits, while QROB reference realizations achieve 1.65× median mirror-circuit survival versus Qiskit O3 compilations on three 156-qubit IBM Heron-r2 processors.The evaluation presents QROB as both a characterization benchmark and a supervision source for data-driven compilation.Source: §I · p. 2

II. BACKGROUND

Quantum compilation must reconcile program interaction requirements with architecture constraints; the background distinguishes geometric routing from capacity-constrained memory residency and defines a shared cost framework.

A. NISQ and FTQC Routing

  • On NISQ devices, nonadjacent logical qubits are brought together by inserted physical SWAPs, while static FTQC patches use temporary routing ancilla to form lattice-surgery channels.NISQ routing moves states; lattice surgery leaves data patches stationary and allocates temporary routing space.Source: A. NISQ and FTQC Routing · p. 3, Figure 1 · p. 3

B. Extension Setting: Quantum Memory-Access Scheduling

  • Quantum memory-access scheduling jointly chooses a legal gate order and a capacity-bounded compute-resident set to minimize compute–memory exchanges.A nonresident operand is admitted while one current resident is evicted; gates sharing a qubit retain their relative order.Source: B. Extension Setting: Quantum… · p. 3

C. Evaluation Metric

  • Implementation cost evaluates a valid realization under architecture constraints using a setting-specific metric, while optimal cost is the minimum over all valid realizations.The framework covers SWAP count, execution time, space–time volume, and compute–memory data movements.Source: C. Evaluation Metric · pp. 3–4
  • For NISQ, f counts inserted SWAPs; for FTQC, execution time in lattice-surgery steps is primary, with space–time volume secondary.The FTQC time metric captures serialization from routing conflicts; space–time volume counts occupied patches and routing ancilla over time.Source: C. Evaluation Metric · p. 4

D. Problem Statement

QROB addresses the lack of practical-scale calibrated references for architecture-constrained compilation by constructing benchmarks backward from feasible realizations. Its method builds realizable circuits, scrambles mappings, and retains inverse paths as references.

D. Problem Statement

  • Optimal implementation cost is NP-hard for core routing problems, so QROB targets practical-scale benchmarks with retained references, polynomial-time generation, and extension to capacity-constrained scheduling.The benchmark objective also calls for calibration against exact solvers on tractable instances and certification where restricted constructions apply.Source: D. Problem Statement · p. 4

III. METHODOLOGY

  • QROB first builds directly realizable circuits, then influence-guided scrambling perturbs mappings while retaining inverse exchanges as feasible references across both geometric-routing settings.The construction uses ideal-circuit strategies and configurable influence weighting; the retained inverse perturbation sequence supplies the reference realization.Source: §III · p. 4, Figure 2 · p. 5

A. Preliminaries

  • Mappings bind logical qubits bijectively to hardware sites, and reverse construction walks a restricted Cayley graph using exchanges only between connected sites, bounding spatial routing overhead.The exchange operators swap logical assignments at connected hardware sites.Source: A. Preliminaries · pp. 4–5

B. Ideal Circuit Construction

QROB builds zero-overhead circuits and then scrambles them backward, retaining the inverse exchanges as feasible references. Its influence-guided scoring and structural controls support varied routing workloads across NISQ and fault-tolerant settings.

B. Ideal Circuit Construction

  • QROB starts from circuits whose two-qubit gates are directly realizable on sampled hardware edges, giving them zero routing overhead by construction.The ideal circuit generator can also accept custom interaction strategies.Source: B. Ideal Circuit Construction · p. 5
  • The Snake, Partition Snake, and random-walk strategies create distinct interaction structures, including dependency chains, concurrency, and localized bottlenecks.Source: B. Ideal Circuit Construction · p. 5
  • Extended Variational Strategy emulates repeated, bounded-degree entanglement layers, and adding its instances to GNN training improves performance on real-world variational benchmarks.The authors describe this transfer as circumstantial rather than a proof of structural equivalence.Source: B. Ideal Circuit Construction · p. 5

C. Influence-Guided Scrambling

  • Scrambling proceeds backward from the final layer, applying exchanges that can break adjacency while retaining the reverse exchange sequence as a feasible restoration path.Source: C. Influence-Guided Scrambling · pp. 5–6
  • The influence score aggregates each candidate exchange’s layer-level degradation using weights that control how strongly effects couple across layers.Cut-Off, Decay, and Global weighting choices encode finite, diminishing, or execution-wide lookahead.Source: C. Influence-Guided Scrambling · p. 6
  • For NISQ routing, the score uses weighted executability loss; larger values indicate exchanges that disrupt direct realizability and require more SWAP routing.Source: C. Influence-Guided Scrambling · p. 6
  • For lattice surgery, the score measures conflict pressure on retained shortest paths, with larger values indicating less accessible conflict-free restoration.Source: C. Influence-Guided Scrambling · p. 6
  • Weighting controls whether overhead disruption stays localized or becomes more globally coupled, while the construction runs in O(D · |E| · d) time.Source: C. Influence-Guided Scrambling · p. 6

D. Extension to Memory-Access Scheduling

QROB-Mem applies reference-preserving reverse construction to capacity-constrained memory access by planting a legal trajectory of resident sets and recording its exchange cost. Evaluation hides that trajectory from schedulers, which must jointly find a gate order and residency schedule.

E. Reference Quality

QROB assesses reference quality through solver calibration for influence-guided instances and construction-based certification for controlled-distance instances. The former closely tracks exact solutions on tractable cases, while the latter provides certified references under restricted structural constraints.

  • IR-mode references are heuristic baselines, not guaranteed global optima, and their close tracking of OLSQ2 is established on tractable small instances.Source: E. Reference Quality · p. 7
  • QROB’s quality claims therefore combine solver-based calibration for IR mode with construction-based certification for CDD mode, whose guarantees apply only under restricted structures.Source: E. Reference Quality · p. 7
  • On a 3 × 3 grid, QROB IR references closely track OLSQ2’s exact optimum across increasing circuit complexity, while Qiskit degrades.The figure’s brown bars show QROB’s gap relative to OLSQ2.Source: E. Reference Quality · p. 7, Figure 3 · p. 7
  • CDD mode certifies references at scale by restricting exchanges to distance-increasing, qubit-disjoint, collateral-free chains, making reversal costs additive.Source: E. Reference Quality · pp. 7–8
  • For a lower bound, QROB selects qubit-disjoint disrupted interactions so their minimum exchange requirements cannot be shared.Source: E. Reference Quality · p. 8

F. Scope and Limitations

QROB benchmarks architecture-induced routing and data-movement overhead rather than the full compilation and execution stack. Across coupling-map systems, compiler overhead grows with scale; the evaluation spans multiple hardware settings and compiler families, with important scope limits on modeled system behavior.

F. Scope and Limitations

  • QROB isolates SWAP routing, lattice-surgery channel allocation, and compute–memory data movement, so it is a benchmark for that layer rather than a full-system model.Source: F. Scope and Limitations · p. 8
  • The benchmark omits upstream logical synthesis and several FTQC costs, including code-distance-dependent errors, decoder latency, factory footprint, distillation throughput, and magic-state queuing.Source: F. Scope and Limitations · p. 8
  • Magic-state routing results use abundant boundary supply and should be read as routing-level headroom, not end-to-end T-state throughput under other factory organizations.Source: F. Scope and Limitations · p. 8

A. Experimental Setup

  • The evaluation covers seven compilers across 18 hardware configurations, including 12 coupling topologies from 9–156 qubits and six surface-code lattices from 16–256 logical qubits.Source: §IV · p. 8, A. Experimental Setup · p. 8
  • The frozen suite records each instance’s reference, cost, configuration, and seed; reporting rules require per-instance and aggregate costs, dispersion, completion rates, and timeouts.Generator-suite results are reported separately from frozen-suite compiler rankings.Source: A. Experimental Setup · p. 9

B. Compilation Gap Characterization

  • 7.2×: Qiskit SABRE incurs this reference-cost ratio on average, while Tket, Cirq, and QMAP incur 12.9–23.0×.Source: B. Compilation Gap Characterization · p. 9
  • 24.1×: Qiskit’s overhead grows from 3× at 16 qubits to 24.1× at 156 qubits, an 8× ratio increase over a 10× increase in system size.With full Qiskit O3, the ratio still reaches ~10× at 119 qubits and corresponds to a measured hardware fidelity loss.Source: B. Compilation Gap Characterization · p. 9
  • Across coupling maps and surface-code lattices, overhead ratios grow with system size and resist heuristic optimization; the authors attribute this pattern to mapping-search-space growth relative to compiler planning horizons.Source: B. Compilation Gap Characterization · p. 9

C. Lattice-Surgery Compilation

QROB references expose substantial lattice-surgery realization overhead, driven largely by placement quality, and support capacity-aware evaluation of memory-access scheduling. The reported overhead persists when T-gate retrieval is included.

C. Lattice-Surgery Compilation

  • At L=256, WISQ and EDPC require 6.3× and 7.0× the reference makespan, respectively, with overhead growing across L=64–256.The trend also holds for space–time volume and under a tunable gate-density knob.Source: C. Lattice-Surgery Compilation · p. 9, Figure 6 · p. 10
  • On DASCOT’s compact layout, the makespan ratio rises from 9.2× at L=64 to 43.8× at L=256 as limited routing space heightens placement sensitivity.The reference still runs most merges in parallel, while a weaker placement is forced onto a single shared corridor.Source: C. Lattice-Surgery Compilation · p. 9
  • For one fully shuffled L=64 instance, the QROB reference completes interactions in 4 steps while WISQ takes 10, a 2.5× ratio determined by placement quality.With patches fixed, remaining routing freedom is limited; the reference executes 24–32 merges per step versus 4–17 for WISQ.Source: C. Lattice-Surgery Compilation · p. 9, Figure 5 · p. 10
  • At L=256 and full T density, WISQ and EDPC require 5.4× and 6.7× the reference makespan, respectively.T-state retrieval is charged to the same ancilla fabric; the layer count remains unchanged.Source: C. Lattice-Surgery Compilation · p. 10, Figure 6 · p. 10

D. Extension Case Study: Memory-Access Scheduling

  • QROB-Mem evaluates legal execution orders and resident-set trajectories for a 200-logical-qubit system, counting compute–memory exchanges while abstracting away access latency and transfer paths.The tested configurations use capacities K=4–24 and an active pool A=2K; the evaluated memory policies include FIFO, LRU, bounded next-use, and max-unlock.Source: D. Extension Case Study: Memory-Access… · pp. 10–11, Figure 7 · p. 11
  • The evaluated 2K-gate next-use policy stays within 1.01–1.16× retained-reference cost, while the K-gate policy rises from 1.14× at K=4 to 2.24× at K=24.An eight-gate horizon is sufficient at K=4 but not K=8; max-unlock rises from 1.48× to 2.00×, while FIFO and LRU remain substantially above reference.Source: D. Extension Case Study: Memory-Access… · p. 11

E. Learning-Based Compilation

QROB references provide calibrated supervision that transfers to real routing workloads, while structural and hardware checks assess the practical validity of its generated instances. Against full Qiskit O3, QROB reference realizations also show higher mirror-circuit survival on three IBM Heron-r2 processors.

E. Learning-Based Compilation

  • With the model architecture fixed, QROB supervision raises SABRE beat rate from 57.8% to 82.2% on 750 test circuits and lowers reference-cost ratio from 1.93 to 1.29.The calibration is reproducible across five seeds with standard deviation below 0.5%.Source: E. Learning-Based Compilation · p. 11
  • On 125 real circuits from 10 algorithm families, the QROB-trained model beats SABRE on 84.8% with 100% completion, while an adversarial-label model fails to complete 75%.The adversarial-label model reaches 94.7% validation accuracy, indicating that label correctness alone is insufficient when benchmark structure diverges from real workloads.Source: E. Learning-Based Compilation · p. 11
  • Generic QROB supervision yields aggregate SWAP ratios of 0.737 on QASMBench and 0.762 on MQT Bench against SABRE-Basic.Family-conditioned supervision further improves main-suite ratios to 0.701 and 0.733, respectively, and lowers QAOA/VQE-style probe ratios.Source: E. Learning-Based Compilation · p. 11
  • QROB-supervised routing remains below SABRE-Basic in aggregate SWAP count across 3×3 to 10×10 grids and reaches near-parity on 155-qubit Baihua without retraining.The figure reports aggregate SWAP ratio and per-circuit beat rate because a few large circuits can dominate total routing cost.Source: E. Learning-Based Compilation · p. 11, Figure 8 · p. 12

F. Practical Validity

  • The QROB convex hull covers 98.1% of real-circuit points in the 18-feature PCA space; eight of ten families are fully covered, with uncovered points only in ExcitationPreserving and QAOA.Family-conditioned initialization from a family’s interaction footprint recovers underrepresented structures.Source: Figure 9 · p. 12, F. Practical Validity · p. 12
  • On Quafu Baihua, CZ gate count explains 85.4% of XEB fidelity variance, and the cost–fidelity correlation also appears on IBM Brisbane, Origin Wukong, and Quafu Yudu.The passage reports per-gate CZ fidelity of 0.965 and an approximate SWAP fidelity of 0.898.Source: F. Practical Validity · p. 13, Figure 10 · p. 13

G. Fidelity Gap under Full Industrial Optimization

  • The hardware survival gap traces to O3’s higher two-qubit-gate count despite statistically indistinguishable CZ-edge fidelities, and the gate-count ratio grows from 1.14× at N=12 to 1.77× at N=40.On ibm marrakesh, the median survival ratio reaches 3.9× at N=30; full downstream O3 reduces but does not close the gap across circuit depth.Source: G. Fidelity Gap under Full Industrial… · p. 13, Figure 11 · p. 13
  • The QROB reference wins 57 of 65 hardware instance pairs, with a combined median mirror-circuit survival ratio of 1.65× against full Qiskit O3.The comparison uses three 156-qubit IBM Heron-r2 processors and the same downstream O3 passes after fixing QROB’s layout and routing.Source: Figure 11 · p. 13, G. Fidelity Gap under Full Industrial… · p. 13

H. Summary

QROB supplies retained feasible realizations for measuring compilation overhead across routing, lattice surgery, and memory-access scheduling. The paper concludes that calibrated references reveal gaps that grow with scale and can be evaluated alongside learning transfer and hardware outcomes.

H. Summary

  • QROB references enable learned compilers to beat SABRE on 84.8% of real circuits, support held-out workload transfer, and provide family-conditioned supervision without using benchmark circuits as routing-label sources.The summary also reports 98.1% routing-feature coverage and a median QROB/O3 mirror-circuit survival ratio of 1.65 across three IBM Heron-r2 processors.Source: H. Summary · p. 14

V. RELATED WORK

  • Exact mapping and routing formulations certify optimality only on small instances, while practical heuristics scale to hundreds of qubits but report costs without a reference.Learning-based routers trained on heuristic labels inherit label-quality constraints; QROB supplies feasible realizations and non-heuristic labels.Source: §V · p. 14
  • Application suites and their fault-tolerant counterparts provide realistic circuits without reference costs, whereas QROB’s scrambling spans structures and supports both coupling-map routing and lattice-surgery scheduling.Unlike the quantum hardware roofline, QROB references are not derived from compiled outputs.Source: §V · p. 14

VI. CONCLUSION

  • QROB constructs instances backward from retained feasible realizations, providing a common measurement framework for physical-qubit SWAP routing, lattice-surgery scheduling, and capacity-constrained memory-access scheduling.Its references are feasible rather than claimed global optima.Source: §VI · p. 14
  • Compilation overhead grows with system scale, persists across near-term and fault-tolerant layers, and translates into measured fidelity forfeiture under full industrial optimization.The evaluated 2K-gate next-use policy remains near the retained reference, unlike a fixed absolute horizon across compute-region capacities.Source: §VI · p. 14
  • The paper argues that compilation should be evaluated against calibrated reference costs as well as by relative comparisons among heuristics.It identifies additional architecture-constrained compilation settings as a possible direction for the reverse-construction perspective.Source: §VI · p. 14

APPENDIX

The artifact packages QROB and reproduces results through six self-contained sub-artifacts, with documented entry points and expected outputs. Its bundled benchmark data spans coupling maps, lattice-surgery instances, learning tests, and recorded hardware measurements.

A. Abstract

  • The artifact packages QROB’s reverse-construction framework and results as six Docker-containerized sub-artifacts, one per results section.Each includes a README, expected output for checking, and a one-command entry point.Source: A. Abstract · p. 15
  • The artifact documents exact-match checks for deterministic outputs and tolerances for stochastic or hardware-derived results; no live quantum-hardware access is required.Reproduction reanalyzes archived per-circuit and raw-device records, with collection scripts included and credentials removed.Source: A. Abstract · p. 15, B. Artifact check-list… · p. 15

B. Artifact check-list (meta-information) •

  • The provided environment uses pinned compilation stacks and Docker on Linux/x86-64, with at least 16 GB RAM and no GPU requirement.Quick paths take about 1–2 hours overall; disk use is approximately 10 GB.Source: B. Artifact check-list… · p. 15
  • Bundled data includes a 15-configuration coupling-map benchmark, lattice-surgery instances from 64–256 logical qubits, learning test sets through 155-qubit Baihua, and recorded IBM, Origin, and Quafu measurements.QASMBench data is shipped, while MQT Bench circuits are generated on the fly.Source: B. Artifact check-list… · p. 15

C. Description

The artifact is organized into six self-contained sub-artifacts, with bundled inputs and documented workflows for fast reproduction and deeper reruns.

1) How to access:

  • Six self-contained sub-artifacts are archived with README files, Dockerfiles, expected results, and run scripts.The artifact covers coupling-gap, OLSQ2, FTQC, NN, and two fidelity evaluations.Source: 1) How to access · p. 15

3) Software dependencies:

  • Docker is the recommended dependency; experiments run offline after images are built, while Python 3.8–3.12 suffices without Docker and Julia 1.9 is needed only for EDPC.The §IV-G reproduction path uses only the Python standard library.Source: 3) Software dependencies · p. 15

4) Data sets:

  • The archive includes benchmark instances and retained labels, full QROB learning test sets, and recorded hardware measurements.The learning test sets are approximately 1.4 GB and span small grids through the 155-qubit Baihua topology.Source: 4) Data sets · p. 15

E. Experiment workflow

  • Fast reproduction runs each sub-artifact as one Docker command, with deeper optional scripts documented in its README.The coupling-gap workflow ranges from sub-minute aggregation to seed regeneration and end-to-end runs.Source: E. Experiment workflow · p. 15
  • The workflows rerun routing, FTQC compilers, pretrained-model evaluations, PCA coverage, and fidelity calculations from recorded measurements.OLSQ2 also solves fresh IR-mode instances live.Source: E. Experiment workflow · p. 16

F. Evaluation and expected results

The evaluation artifact reproduces results across routing, fault-tolerant scheduling, learning, and hardware fidelity, and exposes configuration options for further experiments.

F. Evaluation and expected results

  • 7.2/12.9/16.0/23.0× SWAP overheads and Fig. 4 ratios from ×1.48 to ×24.12 are reproduced, alongside depth, time, and scale results.The artifact reports reproduction within ±0.1 for Table I and its 3× to 24× scale statement.Source: F. Evaluation and expected results · p. 16
  • Six deterministic routers reproduce per-circuit SWAP counts exactly, while seed-42 regeneration reproduces all 560 files and 280 labels.Stochastic-swap means agree within 2%.Source: F. Evaluation and expected results · p. 16
  • Fresh OLSQ2 instances yield QROB-reference depth exactly, with costs ordered OLSQ2 ≤ QROB ≤ SABRE ≤ Basic/Stochastic.At fully solved depths, OLSQ2 means are within approximately 0.3 mean SWAPs of the published data; high-depth timeouts reproduce the scalability claim.Source: F. Evaluation and expected results · p. 16
  • WISQ time overhead rises from 2.1× to 6.3× and EDPC from 3.4× to 7.0× as L increases from 64 to 256.At L=256, volume overhead is 6.4–6.5×; reruns reproduce per-L means within a few percent across eight instances per point.Source: F. Evaluation and expected results · p. 16
  • 84.8% of 125 real-circuit transfers are reported for QROB-supervised learning; the shipped-record verification rerun reports 82.4%.The QASMBench/MQT aggregate SWAP ratios are 0.737/0.762 for generic learning and 0.701/0.733 for family-conditioned learning.Source: F. Evaluation and expected results · p. 16
  • 98.1% PCA convex-hull coverage is reproduced, with 8 of 10 circuit families fully covered.The coverage is recomputed deterministically from the shipped feature files.Source: F. Evaluation and expected results · p. 16
  • 1.65× is the combined median survival ratio for QROB references versus Qiskit O3, with a 1.646× median on 57 of 65 clean pairs.The artifact also reports 18/27 platform×size×compiler cells reproduced exactly; a caveat is that the frozen 156-qubit headline uses a reference cell averaged over 208 Cirq-routed circuits, while all 280 circuits yield approximately 16×.Source: F. Evaluation and expected results · p. 16

G. Experiment customization

  • The generator, routers, OLSQ2 driver, FTQC sweep, and NN evaluations expose configuration options for customized experiments.Available controls include topology, seed, size, density, parallelism, timeout, depth range, and instance-size or model-selection settings.Source: G. Experiment customization · p. 16

Related papers

Loading 2609.07776v1…