{"id":"3af7ed59-616b-4a21-8386-67fb40866903","arxiv_id":"2506.22159","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid quantum-classical framework that tunes variational circuit parameters with NSGA-II and augments hypervolume with coverage indicators improves Pareto front quality and diversity on small simulated problems.","lead":"This paper proposes two improvements to a variational quantum algorithm for multicriteria optimization: using a classical multiobjective optimizer (NSGA-II) to tune the quantum circuit parameters, and adding coverage metrics to the hypervolume cost function. On small noiseless benchmarks, the authors report better average performance and up to 40% better Pareto front coverage with minimal loss in quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QMOOC cost function in Section 5 has a sign inconsistency: minimizing p·I + (1−p)·(−HV) cannot increase extent-type coverage indicators such as PS, OD, and M*3, yet Section 6.2 reports exactly such increases, making the 40% coverage claim unreproducible as written.","rationale":"I focus on the cost-function sign error rather than the Section 4.1 equivalence. The equivalence issue is real but secondary: it concerns the correctness of a theoretical framing, and QMOOM's empirical behavior could survive even if the equivalence fails. The sign error, by contrast, makes the paper's second headline result impossible under its own formal specification. Table 3 states 'Function to minimize: pI(Y(θ)) + (1−p)(−HV(Y(θ)))'; under minimization, a positive p coefficient on an extent-type indicator is a penalty, so the optimizer should shrink PS, OD, and M*3. Section 6.2 says these indicators grew by up to 10–40% as p rose. The only readings are that the text's sign is wrong relative to the code, or that the figures and data are incompatible with the stated algorithm; either way, the 40% coverage claim cannot be accepted as stated. I therefore keep the REJECT verdict. The reader's weakest_assumption names a different point, but the reader's rationale already flags the cost-function issue, so my agreement is partial.","tokens_in":24908,"tokens_out":6632,"duration_ms":74037,"concrete_test":"Re-run the Section 6.2 experiment for (UMOCO-2, n=10, COBYLA) with QMOOC(PS, p=0.5) implemented literally as minimizing p·PS + (1−p)·(−HV), over the same 20 instances and 40 seeds. If the text is correct, δ2(PS,0.5) will be non-positive; any positive value proves that the cost function that produced the reported numbers is not the one written in the paper. Independently, inspect the repository to confirm whether the QMOOC objective returns +I or −I for these indicators.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 defines QMOOC(I,p) as minimizing p·I(Y(θ)) + (1−p)·(−HV(Y(θ))), with the hypervolume term rewarding good hypervolume. Because the optimizer minimizes, the p·I term forces I downward for any p > 0. This is coherent for the uniformity-type indicators DM, Δ*, and EV, where smaller values are better. But the paper also uses PS, OD, and M*3 as coverage indicators, and these are extent-type metrics: PS is a product of normalized objective ranges, OD is the maximum extent, and M*3 is a sum of extents. For these, larger values mean better coverage. Under the stated cost function, increasing p must drive PS, OD, and M*3 down, yet Section 6.2 reports the opposite: 'coverage can be increased by 40%' for PS, up to 20% for OD, and up to 10% for M*3, with positive δ2(I,p). The reported improvements therefore cannot be produced by the algorithm as written. The only resolution is that the implemented cost used the opposite sign for these indicators (or maximized instead of minimized), in which case the text and the experiments disagree, and the 40% coverage claim must be re-derived under the published convention. This is an internal inconsistency, not merely a departure from prior consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two modifications to the variational quantum multicriteria optimization algorithm QMOO of Ekström et al. First, it reformulates the variational parameter search as a multicriteria problem by mapping a parameter vector θ to the objective-value vector of the P most frequent basis states sampled from the ansatz, and then applies NSGA-II to this vector (QMOOM). Second, it augments the hypervolume-based cost with coverage indicators (PS, OD, M*3, DM, Δ*, EV) through a weighted sum pI + (1-p)(-HV) (QMOOC). Experiments on four bicriteria benchmark families with n=10 and n=13 report average hypervolume gains of up to 10 percentage points for QMOOM and coverage increases of up to 40% with at most 2% hypervolume loss for QMOOC.","tokens_in":25265,"tokens_out":10660,"duration_ms":119754,"significance":"If correct, the paper would provide a practical plug-and-play interface between classical multicriteria metaheuristics and variational quantum circuits, and would show that diversity-oriented costs can substantially improve Pareto-front coverage with little hypervolume sacrifice. The experimental study is reasonably broad: four problem families, two problem sizes, 20 instances, and 40 seeds per instance, with hypervolume measured against the true Pareto front and no fitted free parameters used for prediction. However, two load-bearing issues—the sign of the coverage term in the QMOOC cost function and the unproven and in fact false set-to-vector equivalence in Section 4.1—must be resolved before the paper's central claims can be accepted.","major_comments":[{"comment":"The QMOOC cost function is defined as minimizing p·I(Y(θ)) + (1−p)·(−HV(Y(θ))). Because the optimizer minimizes, the I term rewards small values of I for every p>0. This is coherent for the uniformity-type indicators DM, Δ*, and EV, where smaller values are better. It is not coherent for the extent-type indicators PS, OD, and M*3, which Section 5 itself describes as measuring extent and boundary coverage, so larger values denote better coverage. The reported results in Section 6.2 and Figure 6 nevertheless show positive δ2(I,p) for exactly these indicators, including the statement that Pareto Spread coverage can be increased by 40% with no significant hypervolume loss. A positively weighted I in a minimization cannot produce an increase of a larger-is-better indicator. Thus the experiments as described cannot realize the reported coverage improvements: either the implementation minimized p·(−I)+(1−p)·(−HV) (or equivalently maximized p·I+(1−p)·HV), in which case the published cost function is wrong, or the reported δ2 values are inconsistent with the stated algorithm. The formula, the implementation, and the empirical claims must be reconciled and the tradeoff curves re-derived under one consistent convention.","section":"Section 5 / Section 6.2"},{"comment":"The claimed equivalence 'X(θ) is minimal in P_P(F_2^n) w.r.t. ⊴_f if and only if F(θ) is minimal in F(R^{KP}) w.r.t. ≦' is stated without proof and is false as stated. A concrete counterexample with K=1 and P=2: take four decision vectors a,b,c,d with objective values f(a)=5, f(b)=0, f(c)=0, f(d)=5. Suppose for parameter θ the two most frequent states are a then b, so F(θ)=(5,0), and for θ' the two most frequent states are c then d, so F(θ')=(0,5). As sets, X(θ')={c,d} dominates X(θ)={a,b} under ⊴_f via the permutation matching c with b and d with a, because 0≤0 and 5≤5. Yet F(θ')=(0,5) is not ≤ F(θ)=(5,0) in R^2, since the second coordinate gives 5≤0. The structural reason is that F fixes the coordinate order by the frequency ranking inside each θ, whereas ⊴_f allows an arbitrary matching permutation. Therefore the reduction of MCO3 to ordinary vector dominance is not established, and the theoretical justification for QMOOM as formulated collapses. The authors should either prove the equivalence under additional assumptions or restate QMOOM as a heuristic whose objectives are the frequency-ordered objective vectors, without claiming the equivalence.","section":"Section 4.1"},{"comment":"The comparison between QMOOM and QMOO does not control for computational effort. QMOOM is given an explicit budget of at most 200 generations and 4000 evaluations, as stated in Section 4.2. No analogous termination criterion or maximum number of function evaluations is reported for COBYLA, Nelder-Mead, or Powell inside QMOO, and the number of quantum circuit executions per run is not stated. The observed average gains of 1 to 10 percentage points, and especially the claims about improved worst-case behavior, could therefore be an artifact of unequal evaluation budgets. The authors should report the total number of cost evaluations or wall-clock time for each configuration and, ideally, run all methods under matched budgets.","section":"Section 6.1 / Section 4.2"}],"minor_comments":[{"comment":"The appendix plots have axes labeled only by the numbers '1' and '2', with no axis titles, so it is not possible to verify the sign conventions of δ1 and δ2 from the figures; please add axis labels and a legend explaining the plotted quantities.","section":"Appendix Figures 8-22"},{"comment":"The reference [SE24] is given as 'https://github.com/username/repo', which is a placeholder; the actual repository link is needed for reproducibility of the benchmark data.","section":"References"},{"comment":"The Pareto Spread formula uses ~ y^I and ~ y^N in the denominator, but Section 2 defines only the true ideal y^I and the approximate nadir ~ y^N; the approximate ideal ~ y^I is never defined and should be specified or removed.","section":"Section 5 / Section 2"},{"comment":"The statement that 'the other problem types and solvers behave similarly' is too strong: in the appendix, DM shows a positive δ2 in some configurations (e.g., Figure 17), so the sign and magnitude of the effect are indicator- and setting-dependent; the main text should qualify this claim.","section":"Section 6.2 and Appendix"},{"comment":"The manuscript contains numerous typographical errors, including 'apporach', 'calld', 'euclidiean', 'intervall', 'NSGSA-II', and 'opzimizer'; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for acceptance: the QMOOC cost function as written cannot produce the reported 40% coverage improvement, and the Section 4.1 equivalence that underpins QMOOM is false as stated. Both issues are substantive but potentially repairable—the sign error by correcting the cost convention and re-deriving the tradeoff curves, and the equivalence by either proving it under restrictive assumptions or explicitly downgrading QMOOM to a heuristic. Given the broad benchmark study and the practical relevance of the framework, I would be willing to look at a thoroughly revised version rather than recommending outright rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the set-based reformulation (MCO3) and the vector-valued encoding of the ansatz's top-P bitstrings. That genuinely opens the door to using classical multiobjective metaheuristics for variational parameter search, and the QMOOM results, while confined to noiseless 10- and 13-qubit simulations, are consistent and reasonably thorough (20 instances x 40 seeds, four problem families, three classical baselines). The authors also deserve credit for stating their limitations plainly rather than overselling the hardware readiness.\n\nThe load-bearing problem is the QMOOC cost function. Section 5 defines it as minimizing p·I(Y(θ)) + (1-p)·(-HV(Y(θ))). For the extent-type indicators PS, OD, and M*3, larger values mean better coverage. Minimizing p·I for any p>0 must therefore drive those indicators down, yet Section 6.2 reports coverage increases of up to 40% for PS, 20% for OD, and 10% for M*3. The appendix plots confirm positive δ2 values for those indicators. As written, the algorithm cannot do that. Either the implemented cost used the opposite sign for these indicators, or the text misstates the convention. Either way, the headline coverage claim is not reproducible from the published description. The uniformity indicators (DM, Δ*, EV) are lower-better, so the sign convention is coherent for those, but the paper does not separate the two cases.\n\nThe second soft spot is Section 4.1's claimed equivalence: X(θ) minimal under set dominance iff F(θ) minimal under vector dominance. This is stated as an observation, not proved, and it is not obviously true because the set dominance relation allows arbitrary permutation while F(θ) fixes the order by frequency. If that equivalence fails, the theoretical underpinning of QMOOM is shaky, although the algorithm might still work empirically.\n\nA smaller concern: the QMOOM vs QMOO comparison may not be apples-to-apples. QMOOM gets 4000 evaluations through NSGA-II with a population of 5; the QMOO baselines use single-point optimizers, and the paper never states how many cost-function evaluations those solvers were allowed. That could explain part of the performance gap.\n\nThis paper has a promising core idea and a reasonably careful experimental setup, but the sign inconsistency is central and must be fixed (either by correcting the cost function or re-deriving the results under the published convention). The equivalence also needs a real proof or a honest statement of it as a heuristic assumption. I would send it to review, not desk reject it, because the flaws look fixable and the underlying approach is worth engaging with. The authors need to re-run or re-report the QMOOC experiments and tighten the theory.","headline":"New framing for VQA parameter search as multicriteria optimization, but the QMOOC cost function as written cannot produce the reported coverage gains, and the key equivalence is unproven.","tokens_in":25733,"tokens_out":1845,"would_cite":false,"duration_ms":22813,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12","90C29"],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating a variational quantum algorithm's parameter search as a multiobjective problem, solved by population search, improves hypervolume up to 10 points and can lift Pareto-front coverage 40% with minimal loss.","keywords":["variational quantum algorithms","multiobjective optimization","Pareto front approximation","hypervolume indicator","NSGA-II","coverage indicators","parameter search reformulation","hybrid quantum-classical optimization"],"falsifier":"Search for a counterexample to the Section 4.1 equivalence: on a small instance, enumerate all parameter vectors, compute $X(\\theta)$ and $F(\\theta)$, and check whether every pair with $X(\\theta_1) \\trianglelefteq_f X(\\theta_2)$ also satisfies $F(\\theta_1) \\leq F(\\theta_2)$ and conversely; one violated pair refutes the stated iff. Empirically, running QMOOM against QMOO on a noisy device with the same benchmarks would also settle whether the reported gains survive realistic noise.","tokens_in":24773,"feed_emoji":"⚛️","tokens_out":7237,"duration_ms":74632,"temperature":0.7,"pith_summary":"The paper tries to establish that the parameter-search loop of a variational quantum algorithm for multiobjective problems can itself be treated as a multiobjective problem, and that classical population-based optimizers can then drive the quantum circuit better than standard single-objective optimizers. It reports that this hybrid design, called QMOOM, beats the baseline QMOO in average and worst-case hypervolume across benchmark problems by up to 10 percentage points. It further claims that adding a coverage indicator to the hypervolume cost, in a variant called QMOOC, raises Pareto-front coverage by up to 40% while sacrificing at most about 2% hypervolume. If these effects hold outside the noiseless simulations tested, they would give practitioners a relatively cheap way to get more diverse solution sets from near-term quantum hardware.","feed_headline":"Multiobjective tuning lifts quantum optimizer up to 10%","feed_subtitle":"Coverage term also raises Pareto-front coverage by 40% with almost no hypervolume loss.","key_machinery":"The load-bearing object is the set-valued map $\\theta \\mapsto X(\\theta)$, the $P$ most frequent computational-basis states in the ansatz output, and its vectorization $F(\\theta) \\in \\mathbb{R}^{KP}$ that sorts those states by frequency and then evaluates all objectives. The identity that carries the argument is the claimed equivalence that $X(\\theta)$ is minimal under set dominance $\\trianglelefteq_f$ if and only if $F(\\theta)$ is minimal under ordinary vector dominance; this is what lets NSGA-II treat $\\theta$ as a population member and rank parameters by the nondomination of their extracted solution sets. The second mechanism is the averaged cost function $p\\,I(Y(\\theta)) + (1-p)(-\\mathrm{HV}(Y(\\theta)))$, where $I$ is one of six coverage indicators, which the paper uses to trade a small amount of hypervolume for a large increase in spread, uniformity, or boundary coverage.","core_discovery":"The paper's central claim is that variational parameter optimization for multiobjective VQAs need not be scalarized into a single cost function. By mapping each parameter vector $\\theta$ to the $P$ most probable measurement outcomes and arranging their objective values into a vector $F(\\theta) \\in \\mathbb{R}^{KP}$, the search over $\\theta$ becomes an ordinary multicriteria problem whose nondominated points correspond to nondominated $P$-element subsets of the decision space. Feeding that vector-valued objective to NSGA-II yields better hypervolume and stability than optimizing a single hypervolume scalar with COBYLA, Nelder-Mead, or Powell. Separately, replacing the pure hypervolume cost with a weighted average of hypervolume and a coverage indicator such as Pareto spread or outer diameter can improve front coverage by up to 40% with near-zero loss, because the tradeoff curve is concave with a steep initial gain.","pith_inferences":["An inference beyond the paper: if the Section 4.1 equivalence is later shown to be false, QMOOM could still work empirically, but its status would change from a formal reduction to a heuristic analogy, and the experimental comparison would become the main evidence.","The concave tradeoff curves suggest a practical recipe the paper does not explicitly test: run the baseline first and then post-select or re-weight with a coverage indicator, since most of the coverage gain appears at small weights $p$.","On noisy hardware, the ranking of most-frequent states that defines $X(\\theta)$ becomes noisy; a direct testable extension is whether QMOOC's 40% coverage gain survives finite shot noise with error mitigation.","For more than two objectives, coverage indicators that rely on pairwise neighbor distances may need renormalization, but the framework predicts the qualitative result, that coverage can be traded for hypervolume at a favorable rate, should still hold."],"forward_implications":["Any classical multiobjective metaheuristic can in principle replace NSGA-II inside QMOOM without modifying the quantum circuit, because the framework exposes a vector-valued objective to the classical solver.","The weighted-average cost function gives users a dial: with a small weight on a coverage indicator, Pareto Spread improves coverage by roughly 40% while hypervolume loss stays under about 2%, and returns diminish after roughly 10% of hypervolume is sacrificed.","QMOOM's advantage over QMOO is concentrated in average and worst-case behavior, meaning repeated runs are more consistent; the only noted exception is UMOCO-1 with Powell, where the means barely differ.","Coverage gains come without deeper circuits or more shot counts, so in applications where decision makers need a broad view of tradeoffs, the same quantum resources yield a more useful solution set.","The improvements appear across both linear and quadratic benchmark types and at both 10 and 13 variables, suggesting the mechanism is not tied to one problem structure."],"supporting_citations":[{"why":"Supplies the baseline QMOO algorithm: the multiobjective variational ansatz and the hypervolume cost that QMOOM and QMOOC modify.","marker":"[EWS24]"},{"why":"Defines NSGA-II, the population-based multiobjective optimizer used to search the variational parameters.","marker":"[DPAM02]"},{"why":"Introduces the hypervolume indicator and motivates it as the primary quality measure for Pareto-front approximations.","marker":"[ZT98]"},{"why":"Provides the QAOA ansatz template from which the multiobjective circuit used here derives.","marker":"[FGG14]"},{"why":"Is the implementation of NSGA-II that the paper modifies for the QMOOM loop.","marker":"[BD20a]"},{"why":"Introduces Pareto Spread, the coverage indicator whose weighting yields the reported 40% coverage increase.","marker":"[W A00]"},{"why":"Shows that maximizing hypervolume does not generally maximize coverage, justifying the coverage-augmented cost function.","marker":"[ABBZ09]"}],"fun_headline_variants":["Multiobjective tuning boosts quantum optimizer performance","VQA tuning as multicriteria search yields gains","Pareto front coverage improved in quantum optimization","Quantum variational algorithms enhanced via multiobjective methods","Coverage-aware costs lift Pareto fronts in VQAs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reformulation depends on the unproven claim that a set of candidate solutions is nondominated under elementwise set dominance exactly when its frequency-sorted objective vector is nondominated under ordinary vector dominance.","fun_headline_variants_meta":{"raw":{"variants":["Multiobjective tuning boosts quantum optimizer performance","VQA tuning as multicriteria search yields gains","Pareto front coverage improved in quantum optimization","Quantum variational algorithms enhanced via multiobjective methods","Coverage-aware costs lift Pareto fronts in VQAs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1429,"prompt_tokens":850,"completion_tokens":579,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":509}},"tokens_in":466,"tokens_out":579,"duration_ms":6422,"temperature":1.0,"reasoning_tokens":509,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:10:42.793257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search for a counterexample to the Section 4.1 equivalence: on a small instance, enumerate all parameter vectors, compute $X(\\theta)$ and $F(\\theta)$, and check whether every pair with $X(\\theta_1) \\trianglelefteq_f X(\\theta_2)$ also satisfies $F(\\theta_1) \\leq F(\\theta_2)$ and conversely; one violated pair refutes the stated iff. Empirically, running QMOOM against QMOO on a noisy device with the same benchmarks would also settle whether the reported gains survive realistic noise.","supporting_citations":[],"review_version":1}