{"id":"255c9301-2a00-4213-8f9d-4808860373a9","arxiv_id":"2505.11464","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multilevel QAOA can transfer p=1 parameters across coarsening levels, and warm-starting Burer-Monteiro with these solutions improves many Max-Cut benchmarks.","lead":"The paper tests a hybrid solver that runs a shallow quantum circuit on progressively smaller versions of large graph problems, then carries the best angles back up to the original problem. It reports that using these solutions to warm-start a classical solver improves results on most benchmark graphs and lowers runtime.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parameter-transfer claim rests on a 3-sigma Jaccard landscape proxy that does not directly test whether transferred QAOA angles are near-optimal at finer levels.","rationale":"The reader's weakest_assumption identifies exactly the same gap: the z-score-based landscape overlap is a coarse proxy and the paper does not validate that transferred parameters are near-optimal on each finer graph. This is the most load-bearing concern because the paper's headline contribution is the parameter-transfer mechanism, not the multilevel structure itself; without reliable transfer, the method would require per-level optimization, eliminating the advertised reduction in quantum resources. The concern is empirical, not a logical contradiction: the authors may be correct, but the current evidence does not yet establish the claim. The paper also acknowledges in Section VI that an analytic proof of transfer is open, which reinforces that the empirical proxy is doing all the work. The missing hyperparameter values and lack of code/data release are secondary but compound the difficulty of checking the claim. Because the reader already assigned CONDITIONAL with moderate confidence and this concern matches the reader's weakest assumption, no verdict adjustment is needed.","tokens_in":11155,"tokens_out":4006,"duration_ms":44689,"concrete_test":"Take one Gset graph (e.g., G1) and, for every hierarchy level, compute the true p=1 QAOA optimal parameters by dense grid search with local refinement. Then compare the quality, measured by final objective/approximation ratio, of (a) parameters transferred from the coarsest level according to Algorithm 1 and (b) level-specific optimized parameters. Also compute the parameter-space distance between the coarsest-level optimum and each finer-level optimum. If transferred parameters consistently achieve an approximation ratio within, say, 0.5% of the level-specific optimum across all levels, the transfer claim is supported; if not, the 3-sigma Jaccard similarity does not establish near-optimal transfer and the central claim loses its main evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is parameter transfer from the coarsest level to finer levels (Section IV-B), supported by the claim that relaxation-based coarsening preserves QUBO structure needed for QAOA parametrization. The only quantitative support is Section V-A: grid points with expectation above 3 sigma are detected, and the Jaccard overlap between consecutive levels is reported as roughly 95% on one example (Figure 3). This proxy is not load-bearing for the transfer claim. A high overlap of thresholded high-energy regions does not imply that the parameter set that is optimal at the coarsest level remains near-optimal at finer levels; broad regions of the landscape can be similar while the optimum shifts or narrows. The transfer mechanism in Algorithm 1 prunes the parameter grid to the top s solutions from the previous level without re-optimization, so if the coarsest-level parameters are not near-optimal downstream, the claimed reduction in quantum optimization resources fails. The authors explicitly list an analytic proof of parameter transfer as open (Section VI), so the empirical proxy is the only support. Additionally, the threshold-merging hyperparameters Delta_1 and Delta_2 (Section IV-A) are never specified, which affects which hierarchy is built and therefore which landscape is being compared.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multilevel hybrid quantum-classical solver for large QUBO/MaxCut problems. It coarsens graphs via a relaxation-based sphere embedding with a new threshold-based pairwise merging rule, solves the coarsest level with depth-1 QAOA using analytic expectation values, feeds the resulting spin-correlation matrices into Quantum Relax & Round, and refines candidate solutions with a genetic algorithm as the hierarchy is uncoarsened. The central new claim is that p=1 QAOA parameters optimized at the coarsest level can be carried to finer levels without re-optimization, because the coarsening is said to preserve the structural information needed for QAOA parameterization. This claim is supported primarily by a z-score-based Jaccard similarity between QAOA energy landscapes at consecutive hierarchy levels, and the resulting solver is evaluated by warm-starting the Burer-Monteiro SDP heuristic on Gset and SuiteSparse instances up to 14,000 nodes. The paper states that an analytic proof of the parameter-transfer property remains open and that the current evidence is numerical.","tokens_in":11409,"tokens_out":5788,"duration_ms":57074,"significance":"If the parameter-transfer claim holds, the paper would reduce a major cost of multilevel QAOA, namely repeated variational optimization at every level, and would provide a practical NISQ-era strategy for large MaxCut/QUBO instances. The paper has genuine strengths: it uses the analytical p=1 QAOA correlation formulas to avoid expensive sampling, it combines Quantum Relax & Round with a genetic algorithm in a sensible way, it tests on standard large public benchmarks, and it honestly compares against strong classical solvers while acknowledging that the quantum-only solutions are not competitive. The significance is conditional, however, because the load-bearing transfer claim currently rests on a landscape-similarity proxy that is not directly validated, and the numerical evidence in Table I lacks statistical detail.","major_comments":[{"comment":"The 3-sigma Jaccard landscape-similarity metric does not directly establish that parameters optimal at the coarsest level are near-optimal at finer levels. The metric compares only the sets of grid points whose expectation exceeds a threshold, and a high overlap of broad high-energy regions can coexist with a shifted or narrowed optimum. Because Algorithm 1 prunes the parameter grid to the top s parameter sets from the previous level without re-optimization, this gap is load-bearing for the central claim. I request a direct test: for every benchmark graph and every hierarchy level, report the expectation value obtained by evaluating the coarsest-level optimal parameters at that level, the level-optimized optimum, and a random-parameter baseline, with medians and quantiles across the benchmark set. The reported ~75% similarity at the coarsest level is especially relevant, since that is exactly the level from which transfer originates.","section":"V-A, Figure 3, and Algorithm 1"},{"comment":"The coarsening thresholds Delta_1 and Delta_2, the number of retained parameter sets k, the candidate count s, and all genetic-algorithm hyperparameters are never specified, and the source code and data are currently only promised upon acceptance. These parameters determine the hierarchy that is built, the landscape that is compared, and the transfer search, so the similarity numbers in Section V-A and the results in Table I cannot be reproduced or assessed for sensitivity. Please report the actual values, a sensitivity study for Delta_1 and Delta_2, and a permanent code/data artifact or a detailed instance-level results table.","section":"IV-A and IV-B"},{"comment":"The performance claim that warm-starting Burer-Monteiro with multilevel solutions 'provides an advantage in the majority of the tested cases' is not statistically supported. The table contains single objective values per instance, and QAOA+BURER02 exceeds plain BURER02 in 10 of 17 rows, ties in one row, and is worse in six rows, often by small margins. There are no error bars, repeated trials, or per-instance runtime breakdowns. I request means and standard deviations over multiple runs, a statement of how often the warm-started variant beats the baseline by more than a meaningful tolerance, and a per-instance comparison of the total runtime including the multilevel preprocessing.","section":"V-B, Table I, and Figure 4"}],"minor_comments":[{"comment":"The constraint in Eq. (2) is garbled: 's.t: p_t_i - 1 = 0' should presumably be the unit-norm constraint ||p_t_i||_2^2 = 1, consistent with the preceding sentence.","section":"II-A, Eq. (2)"},{"comment":"The sentence 'QAOA withplayers alternatingly apply' should read 'QAOA with p layers'.","section":"II-A"},{"comment":"The phrase 'separating nodes from their neighbors connected by positive weight while bounding neighbors connected by positive weight' repeats 'positive weight'; the second occurrence should likely be 'negative weight' or otherwise be clarified.","section":"IV-A"},{"comment":"The column header 'wE' is unexplained; please define it as the set of edge weights used in the instance.","section":"Table I"},{"comment":"Lines 7-8 are ambiguous: 'Select top n-th from k x m best candidates' and 'Initialize {(gamma', beta')} with parameters from top n-th candidates' do not specify how n and m are chosen and how the pruned parameter sets are paired with the selected candidate bitstrings.","section":"Algorithm 1"},{"comment":"The Index Terms use 'Quantum Relax & Rounds' with an inconsistent plural; the main text consistently uses 'Round', so the index term should be singular.","section":"Index Terms"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a quantum-computing venue and advances the authors' previous multilevel QAOA line with a concrete parameter-transfer mechanism. The main risk is not circularity but validation strength: the landscape proxy is too weak to carry the central claim, and the benchmarking lacks the statistical detail needed to judge whether the warm-start gains are robust. I would recommend major revision rather than rejection because the requested direct validation experiments and hyperparameter disclosure appear feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this one. First, it is an honest empirical paper about a hybrid quantum-classical pipeline: it never claims the QAOA part beats classical heuristics, and it says so plainly in Section V. Second, the headline result is real but narrow — warm-starting Burer-Monteiro with seeds from a multilevel QAOA+QRR+GA pipeline improves cut values on many large graphs from Gset and SuiteSparse, with reported runtime savings of at least 800s compared to a non-multilevel BM run.\n\nWhat's actually new: (1) parameter transfer between hierarchy levels in multilevel QAOA — the coarsest-level QAOA angles are reused at finer levels instead of re-optimized; (2) a threshold pairwise merging rule that only merges nodes with low pairwise distance, which can produce more levels but preserves QUBO structure; (3) combining Quantum Relax & Round with a genetic algorithm on the measurement outcomes. Using the analytic p=1 QAOA expectation values from Ozaeta et al. to build correlation matrices is a sensible way to scale to tens of thousands of variables without simulating the full circuit.\n\nCredit where due: the baseline comparison is wide (Burer-Monteiro, FES02GVP, Duarte05, Glover10) and the paper is transparent that \"the solutions obtained from the multilevel QAOA p=1 solver are not of sufficient quality to compete with state-of-the-art classical solvers.\" There is no hand-waving about quantum advantage.\n\nSoft spots, in proportion. The load-bearing novelty — parameter transfer — is supported by a landscape-similarity proxy in Section V-A: they threshold the p=1 energy landscape at 3 sigma and compute Jaccard overlap between consecutive levels (~95% except coarsest at ~75%). But that is not a direct test of whether the coarsest-level optimum stays near-optimal at finer levels. Broad overlap of high-energy regions can coexist with a shifted optimum. The paper itself lists an analytic proof as open (Section VI), so the empirical proxy is the only support. There is no direct comparison of transferred parameters vs re-optimized parameters at each level. Fig 4 does compare QRR-only, GA-only, and both, on one graph (G1), and the both version wins, which is indirect evidence the transfer helps — but it's one instance. Also, the threshold-merging hyperparameters Delta_1 and Delta_2 are never specified, so the hierarchy itself is not reproducible. No code or data is released (link \"will be provided upon acceptance\"), no error bars, and no classical warm-start control (e.g., warm-starting BM with a simple greedy or random seed) to isolate the value of the QAOA seeds. Some of these are minor individually, but together they make the central empirical claims hard to verify.\n\nWho this is for: people working on hybrid quantum-classical solvers for QUBO/MaxCut, especially multilevel or parameter-transfer methods. It's a useful data point that low-depth QAOA seeds can help a classical solver, but it's not a demonstration of quantum advantage.\n\nRecommendation: yes, send it to peer review. It's a serious, coherent contribution with real novelty, and the gaps are fixable with code release, hyperparameter reporting, error bars, a direct transfer-vs-reoptimization test, and a classical warm-start baseline. It's not desk-reject material.","headline":"Plausible hybrid multilevel QAOA+QRR+GA pipeline with an honest classical comparison; the parameter-transfer claim is promising but only indirectly supported, and the paper lacks reproducibility basics.","tokens_in":11927,"tokens_out":4192,"would_cite":true,"duration_ms":39474,"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":"A single low-depth QAOA parameter search can drive a multilevel solver for large QUBO problems, because coarsening preserves the structure QAOA depends on.","keywords":["QAOA","multilevel optimization","parameter transfer","QUBO","MaxCut","Quantum Relax and Round","genetic algorithm","near-term quantum devices"],"falsifier":"On a fresh graph, compute the best fine-level p=1 QAOA angles by an exhaustive fine-grid search and compare the objective value they produce with the objective produced by the transferred coarse-level angles; if the transferred angles are substantially worse while the landscape overlap is still near 95%, the similarity measure is not a valid proxy.","tokens_in":10923,"feed_emoji":"⚛️","tokens_out":8267,"duration_ms":77668,"temperature":0.7,"pith_summary":"The paper argues that a depth-1 QAOA circuit does not need to be re-optimized at each scale of a multilevel solver for large QUBO and MaxCut problems. Its central claim is that relaxation-based coarsening preserves the structural information that fixes QAOA's energy landscape, so the angles found on the coarsest graph remain near-optimal at every finer level. If this is right, a single low-depth quantum parameter search, followed by classical refinement, can seed high-quality solutions for graphs with thousands to tens of thousands of nodes. The paper also reports that warm-starting a classical rank-two SDP solver with those multilevel solutions improves the best objective on most benchmark graphs while using less total runtime.","feed_headline":"QAOA angles survive coarsening: one search handles every scale","feed_subtitle":"A low-depth quantum optimization circuit only needs tuning on the coarsest graph; finer levels reuse those angles.","key_machinery":"The carrying mechanism is the V-cycle multilevel hierarchy with threshold-based coarsening. Nodes are embedded on a sphere and repositioned to maximize weighted distances from their neighbors; pairs are merged only when their sphere distance is below a threshold and their edge weight is non-positive, which keeps the coarse graph faithful to the original QUBO. At the coarsest level, p=1 QAOA expectation values and two-point correlations are obtained analytically, Quantum Relax & Round turns the correlation matrices into candidate bitstrings, and a genetic algorithm selects the best solution. The selected QAOA parameters are then transferred unchanged to each finer level, where the parameter grid is pruned to the top candidates from the previous level.","core_discovery":"The paper's central claim, stated on its own terms, is that QAOA parameter transfer across a multilevel hierarchy works because the coarsening phase is information-preserving. For p=1 QAOA, the paper reports that the energy landscapes at consecutive hierarchy levels overlap by roughly 95% in their interesting regions, and that the optimized parameters from the coarsest level, carried to finer levels with a pruned parameter grid, yield the best final solution without further variational optimization. The solver combines analytically computed single-layer QAOA correlations with Quantum Relax & Round to produce candidate bitstrings, uses a genetic algorithm to refine them, and uses the resulting solutions as warm starts for a classical rank-two SDP relaxation. On benchmark graphs with 3,400 to 14,000 nodes, the warm-started solver matches or beats the bare classical solver on most instances and runs at least 800 seconds faster.","pith_inferences":["A stronger test of the transfer claim would compare transferred-parameter performance with per-level re-optimization on a held-out set of graphs; the paper's three-standard-deviation overlap measure is only an indirect proxy.","If the landscape similarity is generic, the same parameter-transfer scheme could be applied to weighted QUBO families beyond MaxCut, such as portfolio or scheduling problems, without per-instance quantum optimization.","The coarsening geometry is a design choice; using different embeddings or merging criteria might change how many levels are needed and how well the angles transfer.","The results hint that p=1 QAOA variational parameters are approximately scale-invariant for structured optimization problems, a property that could be studied independently of the multilevel solver."],"forward_implications":["Only the coarsest level needs variational parameter optimization; every finer level reuses the transferred angles, so the quantum cost is essentially one p=1 QAOA parameter search plus correlation-matrix evaluations.","Warm-starting a classical rank-two SDP relaxation with multilevel solutions improves the best objective on most tested benchmark graphs and reduces total runtime by at least 800 seconds compared with the classical solver alone.","Combining Quantum Relax & Round samples with genetic-algorithm refinement during uncoarsening outperforms either interpolation strategy on its own.","The reported roughly 95% landscape similarity across levels suggests the transfer scheme is not limited to the single example shown and should carry over to other instances built with the same coarsening rules.","The authors expect that increasing QAOA circuit depth would improve solution quality further within the same framework."],"supporting_citations":[{"why":"Defines the QAOA circuit ansatz whose parameters are transferred between levels.","marker":"[1]"},{"why":"Supplies the analytic p=1 expectation values and spin correlations used to build Quantum Relax & Round candidates without sampling.","marker":"[9]"},{"why":"Introduces the Quantum Relax & Round procedure that converts correlation matrices into candidate bitstrings.","marker":"[10]"},{"why":"Establishes the multilevel QAOA scheme that this work extends with parameter transfer and threshold merging.","marker":"[7]"},{"why":"Prior multilevel QAOA formulation that motivates the hierarchy and refinement loop.","marker":"[8]"},{"why":"The rank-two SDP relaxation solver that is warm-started with the multilevel solutions in the numerical comparisons.","marker":"[40]"},{"why":"Supplies the benchmark graphs used for the performance comparison.","marker":"[41]"},{"why":"Supplies the larger real-world sparse graphs used in the experiments.","marker":"[42]"},{"why":"Introduces multilevel combinatorial optimization for quantum architectures, the framework's conceptual basis.","marker":"[5]"}],"fun_headline_variants":["One QAOA tuning covers all scales via multilevel transfer","95% landscape overlap lets one QAOA run solve large QUBOs","Multilevel QAOA: tune once, solve thousands of nodes","QAOA angles carry across scales, cutting runtime by 800s","Tune QAOA on the coarsest graph, reuse everywhere"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported overlap between the interesting regions of QAOA energy landscapes at neighboring levels, measured on grid points whose expectation value exceeds three standard deviations, really means the same parameters stay near-optimal at the finer level.","fun_headline_variants_meta":{"raw":{"variants":["One QAOA tuning covers all scales via multilevel transfer","95% landscape overlap lets one QAOA run solve large QUBOs","Multilevel QAOA: tune once, solve thousands of nodes","QAOA angles carry across scales, cutting runtime by 800s","Tune QAOA on the coarsest graph, reuse everywhere"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2937,"prompt_tokens":893,"completion_tokens":2044,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":1955}},"tokens_in":509,"tokens_out":2044,"duration_ms":14605,"temperature":1.0,"reasoning_tokens":1955,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:52:38.739561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fresh graph, compute the best fine-level p=1 QAOA angles by an exhaustive fine-grid search and compare the objective value they produce with the objective produced by the transferred coarse-level angles; if the transferred angles are substantially worse while the landscape overlap is still near 95%, the similarity measure is not a valid proxy.","supporting_citations":[{"cited_title":"Mlqaoa: Graph learning ac- celerated hybrid quantum-classical multilevel qaoa","cited_arxiv_id":null,"evidence_quote":"Establishes the multilevel QAOA scheme that this work extends with parameter transfer and threshold merging."},{"cited_title":"Gset - a suite-style benchmark for graph processing systems","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark graphs used for the performance comparison."},{"cited_title":"The university of florida sparse matrix collection.ACM TOMS, 38(1):1–25, 2011","cited_arxiv_id":null,"evidence_quote":"Supplies the larger real-world sparse graphs used in the experiments."}],"review_version":1}