Pith. sign in

REVIEW 4 major objections 5 minor 46 references

A multi-round warm-start QAOA protocol improves Pareto-front hypervolume over single-pass weighted-sum QAOA under matched shot budgets across three benchmark stages.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 14:08 UTC pith:ARXP3TOO

load-bearing objection A plausible modular multi-round QAOA framework for multi-objective optimization, but the headline gain may ride on an MPS truncation artifact and in-sample model selection. the 4 major comments →

arxiv 2607.18848 v1 pith:ARXP3TOO submitted 2026-07-21 quant-ph

Quantum-Enhanced Multi-Objective Optimization

Maolin Luo , Jiapei Zhuang , Zuoheng Zou , Man-Hong Yung This is my paper
classification quant-ph
keywords multi-objective combinatorial optimizationQAOAwarm-startPareto fronthypervolumeparameter transferPBI scalarizationIsing model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a fixed quantum sampling budget yields a better Pareto front if it is spent in several guided rounds instead of one blind pass. QEMOO is a modular framework that takes the existing weighted-sum QAOA approach, adds a classical feedback loop that selects elite non-dominated bitstrings, and feeds them back as warm-start biases into later QAOA rounds. Across three benchmarks—small grids, a strongly conflicting engineered set, and larger grids simulated with tensor networks—the multi-round protocol improves hypervolume over the single-pass baseline at matched shot counts, and matched random-sampling controls indicate the quantum sampler itself matters. The point matters because quantum sampling is expensive, so any protocol that extracts more front coverage per shot is a practical step toward useful quantum-assisted multi-objective optimization.

Core claim

The paper's central claim is that a fixed quantum sampling budget buys a higher-quality Pareto approximation when it is distributed over multiple feedback rounds rather than spent in one pass. After an initial cold-start round of QAOA sampling over a pool of weight directions, non-dominated sorting selects elite bitstrings; a seed-selection rule (Balanced Coverage, hypervolume-ranking, or kNN sparsity) picks warm-start seeds, and an adaptive weight-direction rule (Inherit, Forward, or a PBI-inspired local linearization) generates the scalarized target for the next round. The same transferred QAOA angles are reused across all rounds; only the initial state bias changes. Under a fixed total bu

What carries the argument

The central mechanism is the multi-round feedback loop: a Direction-Seed Matrix pairs an adaptive weight-direction method with a seed-selection method, yielding nine schemes (S1–S9). The load-bearing component is the warm-start QAOA circuit that interpolates between the uniform superposition and a product state biased toward an elite seed (Eq. 9), with a rotated mixer (Eq. 10) that keeps the biased state as a fixed point; combined with transferred QAOA angles (depth p=3, trained on q_target=2 instances), this converts earlier elites into sampling priors at no additional variational cost. For strong-conflict regimes, the PBI-inspired update locally linearizes the non-linear PBI scalarization

Load-bearing premise

The framework assumes the pre-computed QAOA angles stay near-optimal for every problem size and every weight direction even though it never retunes them per instance or per direction; if that transfer assumption degrades, the benchmark compares two mistuned samplers and the reported gains would not transfer to hardware or other problem classes.

What would settle it

Fix the shot budget and compare QEMOO against a version that re-optimizes (or per-direction rescales) the QAOA angles on the n=42 Grid Large instances at the same total shot count; if per-instance angle optimization closes the hypervolume gap or reverses the ranking, the multi-round feedback gain is an artifact of a mistuned single-pass baseline. A concrete run: same weight pool, same 50,000 shots, compare transferred-angle QEMOO with angle-optimized QEMOO and angle-optimized single-pass QAOA, and report final hypervolume.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under a fixed 50,000-shot budget, the best QEMOO scheme improves mean hypervolume over the single-pass weighted-sum QAOA baseline on every instance in all three benchmark stages (10/10 positive cases per stage).
  • Replacing the QAOA sampler with uniformly random bitstrings at matched budget causes a large drop in front quality, so the gain is not solely from the classical selection logic.
  • The dominant adaptive direction rule is stage-dependent: the PBI-inspired update dominates the strongly conflicting benchmark, while Inherit and Forward rules stay competitive on the grid benchmarks.
  • The warm-start advantage is largest at the smallest per-weight budget tested (1,000 shots) and shrinks at 5,000 shots, suggesting the protocol is most valuable in shot-constrained settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the QAOA angles are never retuned per direction or per instance, a natural extension is to test whether per-direction angle optimization at the same shot budget closes the gap; if it does, the framework's gains reflect a mistuned baseline rather than a fundamentally better search.
  • The PBI-inspired 'linearize around the elite' trick is not specific to PBI: the same local-gradient construction could synthesize effective directions for other non-linear scalarizations (e.g., Chebyshev), extending geometry-aware updates to any QAOA workflow.
  • The appendix's weight-pool analysis suggests that the choice of weight distribution trades off absolute final quality against warm-start amplification; one untested design axis is to choose the pool specifically to maximize round-over-round gains.
  • A hardware prediction follows from the rotated-mixer warm start: since the extra gates are single-qubit, the scheme's overhead on real devices is small, but whether the transferred angles remain near-optimal under noise is untested and would decide the practical value.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. QEMOO is presented as a multi-round, budget-matched extension of weighted-sum QAOA for multi-objective Ising problems. It splits a fixed shot budget into three rounds, uses elite seeds from non-dominated sorting to construct warm-start QAOA circuits, and couples this with three direction-update rules (Inherit, Forward, and PBI-inspired local linearization) and three seed selectors. The main evidence is an empirical comparison on three Ising benchmark suites (Grid Small n=20, Strong Conflict n=18, Grid Large n=42 with MPS χ=30) against a single-pass QAMOO baseline and uniform random sampling. The reported results select a best scheme per dataset (S6, S9, S2) and claim HV improvements in all ten instances per dataset, with the largest relative gain on Grid Large. Appendices provide parameter scans and a derivation of the PBI local linearization.

Significance. The modular separation of adaptive direction updates and seed selection is conceptually useful, and the PBI linearization in Appendix F is a clear, checkable derivation. The paper also has strengths in matching total shot budgets, including random-sampling controls, per-instance gains, and code/data release. However, the empirical claims rest on deterministic single-seed runs, in-sample selection of the best of several schemes, and a tensor-network backend with fixed bond dimension for the largest and most important gain. These issues currently prevent me from treating the 'across three benchmark stages' claim as established, although they are addressable with additional reruns and analyses.

major comments (4)
  1. [Sec. IV.D and Tables I-III] All experiments use fixed random seeds, so the reported means over 10 instances are single deterministic trajectories and 'positive cases 10/10' is not a statistical win rate. No confidence intervals or hypothesis tests are given. Please rerun with multiple random seeds/weight-pool draws and report paired confidence intervals and tests (e.g., Wilcoxon across instances and seeds). Without this, one cannot separate systematic improvement from seed luck.
  2. [Sec. V.B-V.D and Fig. 2] The headline schemes S6/S9/S2 are chosen as the best among the evaluated combinations on the same datasets used for the final tables; the manuscript itself describes a 'rerun benchmark' with stage-appropriate rows. No multiplicity control or holdout validation is applied. Since the abstract claims QEMOO 'improves' HV, please report the distribution over all schemes, or a pre-registered/holdout choice, or at least adjusted comparisons.
  3. [Sec. IV.D, Table III, App. C.2] The Grid-Large result is evaluated with mqmps at χ=30 on n=42; a vertical cut of the 6x7 grid requires up to χ=64 for exact representation. Warm-start and standard QAOA states can have different entanglement, so a fixed χ can differentially truncate baseline vs QEMOO. App. C.2 reports only final HV stability, not the baseline-vs-QEMOO gap or discarded weight. The largest single gain (7.5% on Grid Large) is on this backend, so the internal validity of the central claim depends on ruling out a truncation artifact. Please report ΔHV and per-scheme discarded weight vs χ (including χ=64 or larger if feasible).
  4. [Sec. II.C and IV.B] Transferred p=3 angles from q_target=2 are used for all directions and sizes without checking their quality on the scalarized Hamiltonians. If the transferred angles are far from optimal on n=20/18/42, the comparison may be between two variants of an ineffective sampler; this would not change the relative conclusion but would undermine the title-level 'quantum-enhanced' claim and transferability. Please provide a sanity check, e.g., approximation ratio or energy of the transferred-angle QAOA vs classically optimized angles on a few representative scalarized instances.
minor comments (5)
  1. [Throughout] Typos and spacing errors: 'perplayers' in Sec. II.C; 'can offers' in the first paragraph; 'aQuantum' in the contribution list; 'researchs' in Sec. I.
  2. [Appendix E, Table V] Table V uses normalized HV values around 1.04/0.98, while main-text tables report raw HV (e.g., 10.94e9 for Grid Large). Please clarify the normalization and state which dataset the table refers to.
  3. [References] References [12] and [35] appear to be the same paper (Farhi, Goldstone, Gutmann, and Zhou, Quantum 6, 759 (2022)); please consolidate.
  4. [Fig. 2 caption] The caption's definition of 'HV_best' and the '+1' offset is hard to parse; please specify whether HV_best is the stage-best mean HV and why the offset is used.
  5. [Sec. III.C] The phrase 'a cap of d_max seeds per weight direction' is ambiguous; it should be 'at most d_max seeds assigned to the same parent weight direction'.

Circularity Check

2 steps flagged

Headline HV gains are in-sample best-of-scheme/hyperparameter selections: the winning QEMOO scheme (S6/S9/S2) and Grid-Large settings (c=0.4, χ=30) are chosen from the same benchmark data used to report the gains, so the central empirical claim is partly selected rather than independently predicted.

specific steps
  1. fitted input called prediction [Sec. V.A (Fig. 2), Tables I–III; scheme selection described in Sec. IV.B]
    "the plot keeps only the single best QEMOO scheme from the rerun benchmark ... Specifically, the displayed QEMOO schemes are S6 for DataSet 1: Grid Small, S9 for DataSet 2: Strong Conflict, and S2 for DataSet 3: Grid Large."

    S6, S9, and S2 are not fixed methods chosen before evaluation; they are the highest-HV entries among the S1–S9 method grid scored on the same datasets (Sec. IV.B and App. A heatmaps). The headline gains and the 10/10 positive-case counts in Tables I–III are then reported for those post-hoc winners. Thus the central empirical claim 'QEMOO improves across three benchmark stages' is a max over in-sample configurations: the scheme identity is read from the same HV outcomes that serve as evidence, so the reported advantage is partly constructed by selection rather than independently predicted.

  2. fitted input called prediction [Appendix C.1/C.2, Sec. IV.D, Sec. V.D, Table III]
    "the resulting Hypervolume follows a clear unimodal distribution peaking at c≈0.4 ... the highest final mean HV appearing near the latest χ=30 rerun setting ... the main Grid-Large rerun uses bond dimension χ=30."

    The Grid-Large headline result (S2, mean gain 7.64e8, Table III) is produced with c=0.4 and χ=30, the hyperparameter values identified by HV scans performed on the same Grid-Large benchmark (App. C.1/C.2; App. B fixes 'the Grid-Large setting ... χ=30, and c=0.4'). Therefore the main-text Grid-Large gain is reported for hyperparameters optimized on the evaluation data itself; the appendix 'stability' scans are refits on the same data, not out-of-sample validation. The paper's own limitation statement—'settings such as χ=30, c=0.4 ... are the strongest tested choices in our benchmark suite, not universal optima'—confirms this in-sample status.

full rationale

The mathematical derivation chain is not inherently circular: Appendix F gives a genuine first-order linearization of PBI and shows how λeff maps back to a two-body Ising Hamiltonian; parameter transfer is cited to the external result [16]; warm-starting is cited to the external result [40]; and the random-sampling control is a matched control rather than a fitted outcome. However, the paper's strongest empirical conclusion—that QEMOO improves Pareto-front hypervolume across three benchmark stages—relies on stage winners S6/S9/S2 being selected as the best of the compared schemes on the same datasets used for the headline tables, and on c=0.4 and χ=30 being chosen from Grid-Large HV scans before being reused in the Grid-Large main result. This is in-sample model selection: the 'best' configuration and the reported gain come from the same evaluation, so the headline numbers are partially selected rather than independent predictions. The MPS χ=30 truncation concern raised externally is a serious internal-validity threat for the Grid-Large and quantum-vs-random evidence, but it is an experimental artifact concern, not a definitional circularity, so it does not by itself raise the circularity score. No load-bearing self-citation chain was found. Overall, the core derivation is self-contained, but the central empirical claim is partly fitted, warranting a partial-circularity score of 6 rather than 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

QEMOO introduces no new physical entities. The load-bearing extra assumptions are algorithmic: parameter transfer, the PBI local-linearization heuristic, MPS fidelity for the large stage, and seed-independence of the fixed-seed runs. The free parameters are a mix of fitted values (c, chi) and unreported constants (theta_p, eta, tau, k) that a re-implementer would need.

free parameters (6)
  • Warm-start bias coefficient c = 0.4
    Chosen as the peak of mean HV in a scan on the Grid-Large benchmark (App. C, Fig. 8) and then used for all stages; no held-out validation.
  • MPS bond dimension chi = 30
    Selected in App. C (Fig. 9) as the value near the highest final HV for Grid-Large; approximation error could influence the large 7.5% gain.
  • PBI penalty theta_p = not reported
    Appears in Eq. (F3) for the PBI-inspired direction lambda_eff; no value is given, so Strong-Conflict results depend on an unstated constant.
  • Forward deformation eta = not reported
    lambda_fwd = Pi_Delta((1-eta)lambda + eta e_d*) in Sec. III B; eta is never specified.
  • Shot schedule (s1,s2,s3) = (300,100,100)
    Default R=3, 3:1:1 per-weight allocation chosen ad hoc; no schedule comparison appears in the main text.
  • BC separation threshold tau and kNN k = not reported
    Balanced-Coverage selector uses a distance threshold tau and kNN-sparsity uses k, but neither is specified, hampering exact reproduction.
axioms (6)
  • domain assumption QAOA parameters optimized on 2-qubit instances transfer to all n<=42 scalarized subproblems at depth p=3
    Invoked in Sec. II C and IV B; eliminates per-direction variational optimization and underlies every QAOA sample.
  • ad hoc to paper The local linearization lambda_eff = u + theta_p d2/||d2|| preserves the ability of PBI-style scalarization to target useful trade-off regions
    App. F, Eq. F3; no proof that a single weighted-sum round reaches unsupported Pareto points; this is the core mechanism credited for Strong-Conflict gains.
  • domain assumption MPS with bond dimension chi=30 faithfully reproduces exact QAOA output statistics and HV differences at n=42
    Sec. IV A and App. C; if approximation error correlates with warm-starting, the Grid-Large 7.5% gain may be an artifact.
  • standard math Weighted-sum scalarization of multi-objective Ising Hamiltonians yields a single Ising Hamiltonian with projected couplings and fields
    Eq. (7); this is linear algebra and not contentious.
  • domain assumption Hypervolume with fixed raw-energy reference points is an appropriate comparable measure across different instance scales
    Sec. IV C; absolute HV depends on objective scaling, which the paper acknowledges in Sec. VI.
  • domain assumption Fixed random seed 2026 is representative; results would not change qualitatively under other seeds
    Sec. IV D; no repeated-seed statistics are provided, so the 10/10 positive-case claims assume seed independence.

pith-pipeline@v1.3.0-alltime-deepseek · 18732 in / 17129 out tokens · 156703 ms · 2026-08-01T14:08:13.238958+00:00 · methodology

0 comments
read the original abstract

Multi-objective combinatorial optimization requires identifying Pareto-optimal trade-off solutions among conflicting objectives, often making it more demanding than its single-objective counterpart. Although quantum multi-objective optimization methods have begun to emerge, most existing quantum optimization workflows are still built around single-objective or fixed-scalarization settings. Building on existing weighted-sum QAOA approaches to quantum multi-objective optimization, we propose QEMOO, a quantum-enhanced multi-objective optimization framework that combines Pareto-based selection and warm-started QAOA sampling in a multi-round protocol under the same total shot budget. We further introduce a PBI-inspired adaptive direction-update scheme to improve coverage in strongly conflicting benchmark regimes. Across three benchmark stages, QEMOO improves Pareto-front hypervolume over the single-pass weighted-sum QAOA baseline under matched shot budgets, suggesting a practical route toward shot-efficient quantum-assisted multi-objective optimization and its future applications.

Figures

Figures reproduced from arXiv: 2607.18848 by Jiapei Zhuang, Man-Hong Yung, Maolin Luo, Zuoheng Zou.

Figure 1
Figure 1. Figure 1: FIG. 1. Overview of the QEMOO framework. Panel a shows how the multi-objective Ising model is scalarized into a pool [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Budget-normalized performance comparison and instance-level gains of the best QEMOO scheme on each benchmark [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The full stage-wise heatmaps of method combi [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: The table lists the mean raw Hypervolume (HV) gain [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Marginal Hypervolume gains of the Seed-Selection Methods in the rerun raw-HV benchmark. The three panels [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Performance heatmaps detailing the hypervolume improvements across the rerun raw-HV benchmark suite: (a) Grid [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Simplex-style visualization of the nine weight-pool schemes used in the latest all-9 Grid-Large analysis. Each panel is [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Relationship between weight-pool geometry and warm-start gain in the latest all-9 Grid-Large rerun. Panel (a) plots [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Round-wise decomposition of warm-start gains for the nine weight-pool schemes in the latest all-9 Grid-Large rerun. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Mean HV vs. warm-start coefficient [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Bond-dimension scaling for the best Grid-Large case. The left panel shows the HV-vs-shots trajectories for different [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Weight-count scaling for the best Grid-Large case. The left panel shows the HV-vs-shots trajectories for different [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11. Hypervolume scaling behavior as a function of the per-weight sampling budget [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12. Analysis of mean ∆HV gain (left y-axis, red line) and positive instance counts (right y-axis, transparent bars) across [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 1 canonical work pages

  1. [1]

    Ehrgott,Multicriteria Optimization, 2nd ed

    M. Ehrgott,Multicriteria Optimization, 2nd ed. (Springer, 2005)

    2005

  2. [2]

    •Whenc→1:θ q →π·b q, preparing a state close to the computational basis state|b⟩(greedy exploita- tion)

    =π/2 for allq, recovering|+⟩ ⊗n (standard QAOA). •Whenc→1:θ q →π·b q, preparing a state close to the computational basis state|b⟩(greedy exploita- tion). The mixing Hamiltonian must also be adapted to pre- serve the biased initial state as a fixed point. Follow- ing [40], we replace the standardR X (2β) mixer with a rotatedmixer: U (warm) M (β) = nY q=1 R...

    2026

  3. [3]

    Aguilera, R

    E. Aguilera, R. de Santiago, B. L´ opez-Prado, C. Motz, and G. Olague, Mathematics12, 1291 (2024)

    2024

  4. [4]

    hijacking

    Repeat Steps 2–3 until allRrounds are complete. This design introduces afeedback loop: each round’s non-dominated front informs the next round’s initial quantum states, enabling adaptive reallocation of quan- tum resources toward the most promising regions of ob- jective space. B. Adaptive W eight-Direction Strategies The first design axis of QEMOO is the...

  5. [5]

    Heese, T

    R. Heese, T. Leleu, S. Reifenstein, C. Nietner, and Y. Yamamoto, Hybrid quantum-classical optimiza- tion for multi-objective supply chain logistics (2026), 12 arXiv:2602.05364v1, arXiv:2602.05364 [math.OC]

    arXiv 2026

  6. [6]

    P. Xu, Y. Ma, W. Lu, M. Li, W. Zhao, and Z. Dai, Jour- nal of Materials Informatics5(2025)

    2025

  7. [7]

    J. R. Figueira, C. M. Fonseca, P. Halffmann, K. Klam- roth, L. Paquete, S. Ruzika, B. Schulze, M. Stiglmayr, and D. Willems, Technical Report (2016)

    2016

  8. [8]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, IEEE Transactions on Evolutionary Computation6, 182 (2002)

    2002

  9. [9]

    Zhang and H

    Q. Zhang and H. Li, IEEE Transactions on Evolutionary Computation11, 712 (2007)

    2007

  10. [10]

    Boland, H

    N. Boland, H. Charkhgard, and M. Savelsbergh, Euro- pean journal of operational research260, 904 (2017)

    2017

  11. [11]

    D¨ achert, T

    K. D¨ achert, T. Fleuren, and K. Klamroth, Mathematical Methods of Operations Research100, 351 (2024)

    2024

  12. [12]

    Farhi, J

    E. Farhi, J. Goldstone, and S. Gutmann, arXiv preprint arXiv:1411.4028 (2014)

    Pith/arXiv arXiv 2014

  13. [13]

    M. P. Harrigan, K. J. Sung, M. Neeley, K. J. Satzinger, F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends,et al., Nature Physics17, 332 (2021)

    2021

  14. [15]

    Boulebnane, A

    S. Boulebnane, A. Khan, M. Liu, J. Larson, D. Herman, R. Shaydulin, and M. Pistoia, Evidence that the quan- tum approximate optimization algorithm optimizes the sherrington-kirkpatrick model efficiently in the average case (2025), arXiv:2505.07929 [quant-ph]

    Pith/arXiv arXiv 2025

  15. [16]

    Lykov, J

    D. Lykov, J. Wurtz, C. Poole, M. Saffman, T. Noel, and Y. Alexeev, npj Quantum Information9, 10.1038/s41534-023-00718-4 (2023)

    doi:10.1038/s41534-023-00718-4 2023

  16. [17]

    Shaydulin, C

    R. Shaydulin, C. Li, S. Chakrabarti, M. DeCross, D. Her- man, N. Kumar, J. Larson, D. Lykov, P. Minssen, Y. Sun, Y. Alexeev, J. M. Dreiling, J. P. Gaebler, T. M. Gatter- man, J. A. Gerber, K. Gilmore, D. Gresh, N. Hewitt, C. V. Horst, S. Hu, J. Johansen, M. Matheny, T. Men- gle, M. Mills, S. A. Moses, B. Neyenhuis, P. Siegfried, R. Yalovetzky, and M. Pist...

    doi:10.1126/sciadv.adm6761 2024

  17. [18]

    S. H. Sureshbabu, D. Herman, R. Shaydulin, J. Basso, S. Chakrabarti, Y. Sun, and M. Pistoia, Quantum8, 1231 (2024)

    2024

  18. [19]

    Shaydulin, S

    R. Shaydulin, S. Hadfield, T. Hogg, and M. Pistoia, ACM Transactions on Quantum Computing4, 1 (2023)

    2023

  19. [20]

    X. Zhu, Z. Zou, F. Jin, P. Mosharev, M. Luo, Y. Wu, J. Chen, C. Zhang, Y. Gao, N. Wang, Y. Zou, A. Zhang, F. Shen, Z. Bao, Z. Zhu, J. Zhong, Z. Cui, Y. Han, Y. He, H. Wang, J.-N. Yang, Y. Wang, J. Shen, G. Liu, Z. Song, J. Deng, H. Dong, P. Zhang, C. Song, Z. Wang, H. Li, Q. Guo, M.-H. Yung, and H. Wang, National Science Review 10.1093/nsr/nwag124 (2026)

    work page doi:10.1093/nsr/nwag124 2026

  20. [21]

    Chandarana, S

    P. Chandarana, S. V. Romero, A. G. Cadavid, A. Simen, E. Solano, and N. N. Hegade, Hybrid sequential quantum computing (2025), arXiv:2510.05851 [quant-ph]

    arXiv 2025

  21. [22]

    Kotil, A

    A. Kotil, A. Banerjee, S. Ghosh, A. Yakaryilmaz, D. Her- man, S. H. Sureshbabu, R. Shaydulin, Y. Sun, M. Pis- toia, and S. Chakrabarti, Nature Computational Science , 1 (2025)

    2025

  22. [23]

    Das and J

    I. Das and J. E. Dennis, SIAM Journal on Optimization 8, 631 (1998)

    1998

  23. [24]

    Ekstr¨ om, H

    L. Ekstr¨ om, H. Wang, and S. Schmitt, Physical Review Research7, 023141 (2025)

    2025

  24. [25]

    Ekstrøm, H

    L. Ekstrøm, H. Wang, S. Schmitt, and N. Awasthi, arXiv preprint arXiv:2602.10952 (2026)

    arXiv 2026

  25. [26]

    A. D. King, arXiv preprint arXiv:2511.01762 (2025)

    arXiv 2025

  26. [27]

    Ayodele, R

    M. Ayodele, R. Allmendinger, M. L´ opez-Ib´ a˜ nez, and M. Parizy, inOperations Research Proceedings 2022, Lec- ture Notes in Operations Research, edited by O. Grothe, S. Nickel, S. Rebennack, and O. Stein (Springer Nature Netherlands, 2023) pp. 393–399

    2022

  27. [28]

    Schworm, X

    P. Schworm, X. Wu, M. Klar, M. Glatt, and J. C. Aurich, Journal of Manufacturing Systems72, 142 (2024)

    2024

  28. [29]

    Sawamura, K

    K. Sawamura, K. Araki, N. Maruyama, R. Haba, and M. Ohzeki, Journal of the Physical Society of Japan95, 054002 (2026)

    2026

  29. [30]

    Zitzler, L

    E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. da Fonseca, IEEE Transactions on Evolutionary Computation7, 117 (2003)

    2003

  30. [31]

    C. M. Fonseca, L. Paquete, and M. L´ opez-Ib´ a˜ nez, in IEEE Congress on Evolutionary Computation(2006) pp. 1157–1163

    2006

  31. [32]

    Riquelme, C

    N. Riquelme, C. Von L¨ ucken, and B. Baran, inLatin American Computing Conference (CLEI)(2015) pp. 1– 11

    2015

  32. [33]

    Beume, C

    N. Beume, C. M. Fonseca, M. L´ opez-Ib´ a˜ nez, L. Paquete, and J. Vahrenhold, IEEE Transactions on Evolutionary Computation13, 1075 (2009)

    2009

  33. [34]

    A. P. Guerreiro, C. M. Fonseca, and L. Paquete, ACM Computing Surveys54, 1 (2022)

    2022

  34. [35]

    Lucas, Frontiers in Physics2, 74887 (2014)

    A. Lucas, Frontiers in Physics2, 74887 (2014)

    2014

  35. [36]

    Farhi, J

    E. Farhi, J. Goldstone, and S. Gutmann, arXiv preprint arXiv:1412.6062 (2015)

    Pith/arXiv arXiv 2015

  36. [37]

    Farhi, J

    E. Farhi, J. Goldstone, S. Gutmann, and L. Zhou, Quan- tum6, 759 (2022)

    2022

  37. [38]

    Akshay, H

    V. Akshay, H. Philathong, M. E. S. Morales, and J. D. Biamonte, Physical Review A104, 012403 (2021)

    2021

  38. [39]

    Wurtz and D

    J. Wurtz and D. Lykov, arXiv preprint arXiv:2107.00677 (2021)

    Pith/arXiv arXiv 2021

  39. [40]

    J. A. Monta˜ nez Barrera, D. Willsch, and K. Michielsen, arXiv preprint arXiv:2402.02833 (2024)

    Pith/arXiv arXiv 2024

  40. [41]

    Galda, X

    A. Galda, X. Liu, D. Lyou, Y. Feng, and Y. Alexeev, arXiv preprint arXiv:2106.07531 (2021)

    Pith/arXiv arXiv 2021

  41. [42]

    D. J. Egger, J. Mareˇ cek, and S. Woerner, Quantum5, 479 (2021)

    2021

  42. [43]

    Biscani and D

    F. Biscani and D. Izzo, Journal of Open Source Software 5, 2338 (2020)

    2020

  43. [44]

    smoother is always better

    MindQuantum Developer, Mindquantum, version 0.6.0, https://gitee.com/mindspore/mindquantum(2021), a hybrid quantum-classical framework. Appendix A: F ull Heatmaps of Method Combinations For completeness, we provide the full comparison of stage-wise method combinations for the rerun raw-HV bench- mark. This appendix is the detailed counterpart of the main-...

    2021

  44. [45]

    We explore its impact by scanning fromc= 0.1 (weak bias) toc= 0.8 (strong exploitation) across all test instances

    Selection Strategy for Mixing Coefficientc The warm-start bias coefficientc∈[0,1] serves as the primary control for the exploration–exploitation trade-off. We explore its impact by scanning fromc= 0.1 (weak bias) toc= 0.8 (strong exploitation) across all test instances. As shown in Figure 8, the resulting Hypervolume follows a clear unimodal distribution ...

  45. [46]

    The left panel in Fig

    Bond-Dimension Scaling for the Best Large-MPS Case To assess how tensor-network approximation quality interacts with the best Grid-Large configuration, we vary the bond dimensionχwhile keeping the same warm-start protocol and best observed direction policy. The left panel in Fig. 9 shows the HV-vs-shots trajectories for differentχvalues, and the right pan...

  46. [47]

    The left panel in Fig

    W eight-Count Scaling for the Best Large-MPS Case To complement the warm-start coefficient analysis, we also examine how the best Grid-Large configuration changes as the number of sampled weights varies while keeping the same transferred parameters and sampling protocol. The left panel in Fig. 10 shows the HV-vs-shots trajectories for different weight cou...

    1925