{"id":"d7306044-4f94-4c52-97eb-d63d111c1c1d","arxiv_id":"2507.01462","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"D-Wave's NL-Hybrid solver achieves 0.83-0.94 approximation ratios in 4-6 seconds on five real-world 3D inspection TSP instances, beating OR-Tools on quality and GUROBI on runtime.","lead":"This paper tests two D-Wave hybrid quantum-classical solvers on robot inspection path planning problems modeled as 3D traveling salesman variants. The NL-Hybrid solver returns good solutions in seconds, outperforming GUROBI's hours-long runtimes and improving on OR-Tools' solution quality.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantum attribution is untested: the reported 0.83–0.94 approximation ratios in 4.5–5.4 s are properties of the full D-Wave NL-Hybrid service, not of its QPU, and no classical-only control is reported.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing point: the paper interprets end-to-end results of a hybrid vendor service as evidence of quantum potential without ablating the quantum component. This is the single most important unresolved condition for the central claim. The paper has genuine strengths: source code is provided, 15 independent runs are performed, two D-Wave solvers and two classical baselines are compared, and the conclusion is appropriately modest about not claiming quantum advantage. However, Table I cannot discriminate between the vendor service's classical heuristics and any quantum contribution, so the headline speed/quality trade-off is not yet attributable to the QPU. I also considered the absence of error bars and the ambiguous definition of approximation ratio, but those affect confidence and reporting precision rather than the causal attribution that the title and framing depend on. The proposed ablation is the one concrete check that would settle whether the 'quantum-assisted' framing is justified. Because the existing CONDITIONAL verdict already encodes this caveat, I see no reason to change the verdict; it remains appropriate pending the control experiment.","tokens_in":3337,"tokens_out":3272,"duration_ms":40765,"concrete_test":"Run a classical-only control on the same five benchmark instances using the same NL-Hybrid model and permutation encoding, but replace the Quantum Module's responses with random perturbations or disable QPU access if the D-Wave API permits it. Keep the same Classical Heuristic Module, time budget, and 15-run protocol. Alternatively, independently reimplement the Classical Heuristic Module described in [5] without any QPU calls. If the classical-only control reproduces NL-Hybrid's approximation ratios and runtimes within a small margin (e.g., 2%), the quantum-guidance attribution is not supported and the paper's title and abstract should be weakened. If the classical-only control is substantially worse, that would provide direct evidence for the quantum module's contribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central framing — the title, the abstract's reference to \"quantum approaches,\" and Section II's description of NL-Hybrid as a hybrid of a Classical Heuristic Module and a Quantum Module whose queries \"guide the heuristic module\" — all invite the inference that the reported speed/quality trade-off evidences a quantum contribution. Nothing in the experiment supports that inference. Table I reports only end-to-end results of a proprietary vendor service. There is no run of the same model with the Quantum Module disabled, no replacement of QPU queries with random or noiseless perturbations, and no comparison against a classical heuristic using the same permutation encoding and time budget. The NL-Hybrid solver's default behavior includes substantial embedded classical search, so the observed average approximation ratios of 0.83–0.94 in 4.5–5.4 seconds are fully consistent with purely classical heuristics. If the Classical Heuristic Module alone accounts for all observed performance, the \"Quantum-Assisted\" claim is unverified and the contribution reduces to a vendor benchmark. The paper does not overstate its conclusions as a demonstration of quantum advantage, but its central claim of quantum potential still depends on this untested separation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a benchmark of hybrid quantum-classical solvers on an industrial robotic inspection path-planning problem modeled as a 3D open-route TSP with incomplete graphs. Five real-world instances (106–194 nodes) are solved with D-Wave's CQM-Hybrid and NL-Hybrid solvers, and results are compared against GUROBI and Google OR-Tools. The main reported finding is that NL-Hybrid achieves approximation ratios of 0.83–0.94 in 4.5–5.4 seconds, while GUROBI requires 17.4 to 68,400 seconds and OR-Tools achieves ratios of 0.69–0.83 in 6.8–13.5 seconds. The paper concludes that this demonstrates the potential of quantum approaches for industrial optimization.","tokens_in":3571,"tokens_out":2909,"duration_ms":36181,"significance":"If the results are taken at face value, the paper provides a useful data point for a specific industrial TSP variant: a proprietary hybrid solver can return competitive-quality solutions much faster than an exact solver on medium-size instances. The authors publicly release code and data, and they report 15 runs per solver, which is a reasonable effort for an exploratory benchmark. However, the significance of the work for the quantum computing community is currently limited by the absence of any evidence that the quantum component contributes to the observed performance; the reported behavior is a property of the full vendor service, which contains substantial embedded classical heuristics. The paper is honest in speaking of 'potential' rather than proven advantage, but its title and framing nevertheless invite a quantum attribution that the experiment does not support.","major_comments":[{"comment":"The central claim of 'quantum-assisted' performance is not supported by the experiments. The reported runtimes and approximation ratios are end-to-end results of the commercial D-Wave NL-Hybrid service, which combines a Classical Heuristic Module with a Quantum Module. No ablation is provided: there is no run with the Quantum Module disabled, no replacement of QPU queries by random or noiseless perturbations, and no comparison against a classical heuristic using the same permutation encoding and time budget. Consequently, the observed 0.83–0.94 approximation ratios in 4.5–5.4 seconds are fully consistent with a purely classical heuristic embedded in the service. To justify the title and abstract's 'quantum' framing, the authors must either provide such an ablation or explicitly re-frame the contribution as a benchmark of a hybrid commercial service without any claim about the quantum component's role.","section":"§II and Table I"},{"comment":"The statistical reporting is insufficient for the main comparative claims. The text states that results are based on 15 independent runs and Table I reports averages, but no standard deviations, confidence intervals, or significance tests are given. Without these, the reader cannot assess, for example, whether the difference between NL-Hybrid's 0.86 and OR-Tools' 0.69 on the Aircraft instance is meaningful or within run-to-run variability. Additionally, the GUROBI runtime appears to be a single deterministic value; the authors should specify how many runs were performed for each solver and whether variability was observed for the classical solvers.","section":"§III and Table I"},{"comment":"The experimental setup is under-specified to the point of non-reproducibility. The patch segmentation thresholds (maximum patch area and maximum surface normal variation) and the robot cost function between points are not defined; the edge structure of the incomplete graph and the cost distribution are direct consequences of these choices and can strongly affect the relative performance of TSP heuristics. Likewise, the OR-Tools configuration is described only as 'a variation of the default configuration' with no details, and no GUROBI parameters (time limit, MIP gap tolerance, threads) are reported. The authors should provide these definitions or cite the public code repository with explicit parameter values.","section":"§III, instance generation and cost function"},{"comment":"The paper's focus on NL-Hybrid is the result of a post-hoc selection: the text states that 'extensive testing revealed that NL-Hybrid consistently delivered superior results' and then proceeds with only that solver. While CQM-Hybrid results appear in Table I and are therefore not hidden, the narrative emphasis on the better-performing solver, selected on the same benchmark, introduces a selection effect that should be acknowledged. The authors should either present the pre-registered comparison or explicitly discuss the possibility that the reported superiority of NL-Hybrid over CQM-Hybrid is inflated by the selection procedure.","section":"§I and §II, solver selection"}],"minor_comments":[{"comment":"The phrasing 'Five references has been selected' is ungrammatical and 'references' should be 'instances'; also 'results are benchmarked against two classical methods' would be clearer as 'compared with'.","section":"Abstract and §I"},{"comment":"The mathematical expression for the objective function is typeset incorrectly: 'f (x) =PN −1 i=1 cost matrixxi,xi+1' should be 'f(x) = sum_{i=1}^{N-1} c_{x_i, x_{i+1}}'.","section":"§II, objective function"},{"comment":"The table reports no approximation ratio for GUROBI; the authors should state explicitly that AR=1 by definition for the baseline. Also, the runtime units are inconsistent (e.g., '1.2K', '2.4K'), and the authors should use a uniform unit such as seconds with scientific notation.","section":"§III, Table I"},{"comment":"Figure 1 would be more informative if the path's cost or the approximation ratio of the shown solution were reported, so the reader can connect the visualization to the quantitative results.","section":"§III, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a vendor benchmark for a commercial hybrid solver, and the quantum attribution rests on the vendor's description of the service rather than on any controlled experiment. The authors should be encouraged to consider whether the journal's readership would be better served by a re-framed manuscript that positions the work as an evaluation of a hybrid solver on an industrial TSP variant, with an explicit disclaimer that no quantum advantage is claimed. The reliance on the authors' own prior publication [5] for the description of NL-Hybrid is also a concern; the present paper should contain enough detail to be self-contained on the mechanism being tested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, honest benchmark of D-Wave's NL-Hybrid solver on five real CAD-derived inspection TSP instances, and the headline result—approximation ratios around 0.83–0.94 in 4.5–5.4 seconds, while GUROBI needs 17 seconds to 19 hours—is believable and worth knowing. But the title and abstract's talk of \"quantum approaches\" invites the inference that the QPU is doing something, and nothing in the paper tests that. The NL-Hybrid is a proprietary hybrid service with substantial embedded classical search; there is no run with the quantum module disabled, no replacement of QPU queries with random perturbations, and no classical heuristic using the same encoding and time budget. The authors are careful not to claim quantum advantage, but the \"quantum-assisted\" language still outruns the evidence.\n\nWhat is genuinely new: five instances derived from CAD surface segmentation (car door, bumper, toy bear, aircraft, half-sphere), a permutation-based encoding for NL-Hybrid that avoids QUBO overhead, and a publicly available dataset and code. The comparison against GUROBI and OR-Tools is external and the approximation-ratio table is internally consistent. This is a legitimate proof-of-concept application, not a demo of quantum advantage, and the paper does not overstate it beyond the speed/quality trade-off.\n\nSoft spots, in rough order of severity. First, the missing control for the quantum component is the load-bearing gap; the central claim of \"quantum potential\" collapses to a vendor benchmark until that is addressed. Second, no standard deviations, confidence intervals, or significance tests are reported despite 15 runs—minor for a proof-of-concept, but easy to fix. Third, the authors selected NL-Hybrid after their own tests because it was superior, which is post-hoc selection; reporting CQM-Hybrid's numbers only in the table softens this a little. Fourth, the patch segmentation thresholds and the robot cost function are described only in vague terms, and the OR-Tools configuration is \"a variation of the default\" without details. Fifth, the paper leans on the authors' own prior paper for NL-Hybrid's performance credibility; that is not disqualifying, but independent confirmation would help. The instance sizes (106–194 nodes) are small, so the industrial-claims are modest but not misleading.\n\nWho is this for? Practitioners who want to know what a commercial hybrid solver can do on routing-style problems, and quantum-optimization researchers looking for concrete benchmarks. It deserves a serious referee because the data and code are real and the claims are mostly scoped properly. Send it to review with major comments: demand a classical-only baseline or a reframed claim, plus basic statistics. Don't desk-reject.","headline":"A credible vendor-benchmark result on five real CAD-derived TSP instances, but the 'quantum-assisted' framing outruns what the experiment actually shows: no control run isolates the QPU's contribution.","tokens_in":4112,"tokens_out":1823,"would_cite":false,"duration_ms":23008,"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 hybrid quantum-classical solver plans robotic inspection routes on real factory parts with 83–94% of exact solution quality, in about five seconds where a classical exact solver needs up to nineteen hours.","keywords":["quantum optimization","hybrid quantum-classical solver","traveling salesman problem","robot path planning","quality inspection","D-Wave","Industry 4.0","open-route TSP"],"falsifier":"An ablation experiment that runs NL-Hybrid on the same five instances with the quantum module disabled or with quantum guidance replaced by random noise would settle the causal role of the quantum component; if the ablation produces statistically indistinguishable runtime and approximation ratios, the central claim of quantum-assisted advantage is falsified.","tokens_in":3163,"feed_emoji":"⚛️","tokens_out":9121,"duration_ms":87005,"temperature":0.7,"pith_summary":"The paper tries to establish that a commercially available hybrid quantum-classical solver, D-Wave's NL-Hybrid, is a practical tool for planning robot inspection trajectories in industrial quality control. The problem is modeled as a three-dimensional traveling-salesman variant with an incomplete graph and an open route, derived from CAD models of five real parts. Across instances of 106–194 inspection points, NL-Hybrid returns routes within 83–94% of the optimal baseline in roughly five seconds, while the exact classical solver GUROBI takes anywhere from 17 seconds to more than 19 hours. The authors conclude that this competitive quality at much lower runtime shows quantum-assisted optimization moving toward real industrial use.","feed_headline":"Quantum hybrid solver plans robot inspection routes in seconds","feed_subtitle":"On five real factory parts, the solver hit 83–94% of exact quality in ~5 seconds; the classical exact solver took hours.","key_machinery":"The load-bearing component is D-Wave's NL-Hybrid solver, a two-phase hybrid algorithm that runs parallel threads combining a classical heuristic module with a quantum module that sends queries to the Advantage_system6.4 QPU. The problem is encoded natively as a permutation: a decision variable is a list of node indices, and the objective is the sum of transition costs along the path, $\\sum_{i=1}^{N-1} C_{x_i,x_{i+1}}$, so no extra QUBO constraints are needed. The solver alternates classical exploration with quantum-guided refinement; this specific division of labor is what the authors credit for the speed-quality tradeoff they report.","core_discovery":"The central claim is that NL-Hybrid, a hybrid quantum-classical solver from D-Wave's Hybrid Solver Service, delivers near-optimal solutions for a 3D open-route TSP on incomplete graphs with a large reduction in computation time relative to exact classical solving. In the authors' experiments, NL-Hybrid achieved average approximation ratios of 0.83–0.94 against the GUROBI baseline across five real-world instances (106–194 nodes), with runtimes of 4.5–5.4 seconds; GUROBI required 17.4 to 68,400 seconds, and OR-Tools achieved ratios of 0.69–0.83 with runtimes of 6.8–13.5 seconds. The authors interpret this as evidence that quantum-assisted optimization can address industrial-scale problems, not just abstract benchmarks.","pith_inferences":["The paper does not test whether the quantum module is the cause of the speedup; without an ablation that disables quantum guidance, a classical-heuristics explanation remains plausible, so the 'quantum-assisted' attribution is an interpretation rather than an established mechanism.","A fairer classical comparison would give a purpose-built local-search heuristic the same five-second budget as NL-Hybrid; if that heuristic matched or beat the reported ratios, the practical advantage would be a property of heuristic design, not of quantum computing.","The five instances are modest in size (106–194 nodes), and the largest instance (H-Sphere) has the lowest ratio (0.83); whether the quality advantage persists at thousands of inspection points is an open scaling question.","The reported runtimes are wall-clock times for the full hybrid service, which includes cloud communication and classical preprocessing; the marginal cost of the quantum queries themselves is not isolated."],"forward_implications":["For a fixed product, the inspection route can be re-optimized in seconds, which matters when production lines switch between customized parts.","On all five instances, NL-Hybrid dominates OR-Tools in both solution quality and runtime, so the paper identifies it as the stronger practical heuristic of the two.","The permutation-based formulation avoids QUBO constraint overhead, which the authors credit for the solver's ease of use on this routing problem.","The study adds a real-world benchmark set (five CAD-derived parts) that future quantum-optimization evaluations can reuse."],"supporting_citations":[{"why":"It defines the CAD-based automatic trajectory generation method whose output (inspection points and cost matrix) constitutes the five benchmark instances.","marker":"[3]"},{"why":"It describes D-Wave's Hybrid Solver Service, the environment in which NL-Hybrid and CQM-Hybrid run.","marker":"[4]"},{"why":"It provides the description and performance analysis of NL-Hybrid, the solver whose two-phase structure is central to the paper's approach.","marker":"[5]"},{"why":"It is a survey of OR-Tools and machine-learning path planning cited to justify OR-Tools as a recognized classical baseline.","marker":"[6]"}],"fun_headline_variants":["Quantum hybrid plans robot inspection routes in seconds","Hybrid quantum solver beats classical on 5 factory cases","Robot inspection routing: quantum hits 94% quality in 5s","D-Wave hybrid tops OR-Tools for 3D inspection paths","Quantum-assisted path planning: near-optimal in ~5 seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The interpretation of the results as evidence of quantum potential rests on the assumption that NL-Hybrid's speed and solution quality come from its quantum-guided search rather than from its built-in classical heuristics or from the particular structure of the five test instances, since no ablation of the quantum component is provided.","fun_headline_variants_meta":{"raw":{"variants":["Quantum hybrid plans robot inspection routes in seconds","Hybrid quantum solver beats classical on 5 factory cases","Robot inspection routing: quantum hits 94% quality in 5s","D-Wave hybrid tops OR-Tools for 3D inspection paths","Quantum-assisted path planning: near-optimal in ~5 seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000103,"raw_usage":{"total_tokens":962,"prompt_tokens":811,"completion_tokens":151,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":66}},"tokens_in":427,"tokens_out":151,"duration_ms":2766,"temperature":1.0,"reasoning_tokens":66,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:50:03.663566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation experiment that runs NL-Hybrid on the same five instances with the quantum module disabled or with quantum guidance replaced by random noise would settle the causal role of the quantum component; if the ablation produces statistically indistinguishable runtime and approximation ratios, the central claim of quantum-assisted advantage is falsified.","supporting_citations":[{"cited_title":"Garrote, G","cited_arxiv_id":null,"evidence_quote":"It defines the CAD-based automatic trajectory generation method whose output (inspection points and cost matrix) constitutes the five benchmark instances."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It describes D-Wave's Hybrid Solver Service, the environment in which NL-Hybrid and CQM-Hybrid run."},{"cited_title":"Osaba and P","cited_arxiv_id":null,"evidence_quote":"It provides the description and performance analysis of NL-Hybrid, the solver whose two-phase structure is central to the paper's approach."},{"cited_title":"Benoit and P","cited_arxiv_id":null,"evidence_quote":"It is a survey of OR-Tools and machine-learning path planning cited to justify OR-Tools as a recognized classical baseline."}],"review_version":1}