{"id":"c9c2f65b-6059-41b0-99fa-9c081317574e","arxiv_id":"2512.19321","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A learning-assisted variable neighborhood search jointly selects substation connections and road-constrained cable routes, cutting modeled construction costs by 30–50% on synthetic urban grids.","lead":"An optimization paper combines genetic search, A*, and variable neighborhood search with deep-reinforcement-learning agents to plan underground cable routes in cities, claiming 30–50% cost savings over simpler planning approaches. The savings come mainly from choosing routes that share a single excavated trench for multiple cables, and from jointly optimizing which substations to connect and exactly which roads to use.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 30–50% cost reduction is largely an artifact of comparing against baselines that optimize a different cost model; fair path-aware comparisons show only 1–8% gains.","rationale":"The reader correctly identified that external validity is unsupported (only four synthetic regular lattices appear despite the abstract promising 12 benchmarks and 3 GIS-derived cases). However, the single most load-bearing concern about the central claim is more fundamental: the reported 30–50% reduction is not a fair algorithmic comparison. MCWS and HGS are evaluated with a cost model that never allows two cables to share a trench, so merely introducing trench sharing into the objective can produce large cost differences that have nothing to do with the quality of the search algorithm. The paper itself provides the fair comparison: SNS-1/2/3, which use the same path-aware cost model, are only 0.12–8.00% worse than MVNS/L-MVNS (Table 2). The algorithm's genuine contribution over a strong single-operator search is therefore modest, and the 'compared with representative 2024 approaches' phrasing overstates it. This is not an internal inconsistency or a fraudulent claim; it is an experimental-design issue. The fix is straightforward: report the cost gaps under the same objective, and include the promised GIS-based instances to test external validity. The algorithmic core—MVNS with three operators and learned neighborhood sampling—is plausible and the ablation/sensitivity studies are a positive sign. I would keep the verdict at CONDITIONAL: the direction is credible, but the headline claim and evidence need revision before acceptance. My concern differs from the reader's weakest assumption, so I mark disagreement, though both concerns point to the need for more rigorous evaluation.","tokens_in":16591,"tokens_out":7308,"duration_ms":76409,"concrete_test":"Recompute MCWS and HGS solutions under the same objective as L-MVNS: after obtaining each baseline's feeder-connection plan, route every feeder on the road graph with A*, then evaluate with Eq. (6) counting trenching cost once per edge regardless of the number of cables (i.e., shared trenches). If the 30–50% gap shrinks to below ~10%, the headline improvement is an artifact of objective mismatch rather than algorithmic superiority. The authors should also report such recomputed gaps in Table 2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that MVNS/L-MVNS cut costs by 30–50% versus representative 2024 approaches rests on comparisons with MCWS and HGS, which the paper classifies as 'relation-only baselines' that ignore explicit trench sharing and 'thereby equat[e] trench length with cable length' (§5.2). These baselines optimize the Stage-I objective F1 (total cable length), not the full cost model F2 in Eq. (6) where trenching cost is shared across parallel cables. The gap between the proposed methods and HGS is 27–42%, but when compared against SNS-1/2/3—path-aware methods that use the same F2 cost model—the advantage of MVNS/L-MVNS is only 0.12–8.00% (Table 2). Thus the large reduction is a modeling effect (permitting shared trenches) rather than an algorithmic improvement over 2024 methods; the paper's own fair algorithmic comparison shows only a few percent gain. The central claim, as worded in the abstract and conclusion, is therefore not supported by the experimental evidence as presented.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates urban cable routing as a bilevel connectivity–path co-optimization problem on a road graph, with costs for trenching and cable laying where parallel cables can share a trench. It proposes MVNS, a multi-operator variable neighborhood search with three destruction operators and A* repair, and L-MVNS, a DRL-guided variant in which three agents are trained to bias neighborhood sampling. Experiments are run on four synthetic regular-lattice benchmark cases (Section 3.5) with 10 independent runs per method. The paper claims that MVNS and L-MVNS reduce total construction costs by approximately 30–50% relative to representative 2024 approaches, and that L-MVNS further improves accuracy and variance on larger instances.","tokens_in":16959,"tokens_out":3442,"duration_ms":35787,"significance":"The practical problem is relevant: accounting for road-constrained routes and trench sharing can substantially change planning outcomes. If the algorithmic claims were cleanly established, the DRL-guided neighborhood selection in L-MVNS would be a useful contribution. The paper also provides a reproducible benchmark suite, though only four synthetic cases are actually used. However, the headline 30–50% cost reduction is an artifact of comparing against baselines that optimize a different cost model; the fair path-aware comparison shows only 0.12–8% gains. The abstract also promises 12 benchmark and 3 GIS instances that do not appear in the body. These issues undermine the central claims as currently stated.","major_comments":[{"comment":"The claim that MVNS/L-MVNS reduce total construction costs by approximately 30–50% is not supported by the paper's own fair algorithmic comparisons. Table 2 shows that this large gap appears only against MCWS and HGS, which optimize the Stage-I objective (Eq. (2)) and explicitly ignore trench sharing. Against path-aware baselines SNS-1/2/3, which use the same cost model (Eq. (6)), MVNS/L-MVNS are only 0.12–8.00% better (Case-2: 0.12%, Case-4: 0.18%). The 30–50% figure is therefore a modeling effect of allowing shared trenches, not an algorithmic improvement. The abstract and conclusion should be reframed to separate the modeling benefit from the algorithmic advantage.","section":"Abstract, §5.2, Table 2"},{"comment":"The abstract states 'Comprehensive experiments on 12 benchmark instances and 3 GIS-derived instances' and claims reductions of 0.92% to 73.72%. The body constructs only four synthetic lattice cases (Section 3.5) and reports results only for those cases (Table 2, Fig. 8). No GIS-derived instances or 12-instance suite appears anywhere in the experimental section. This is a major discrepancy between claimed and actual evidence. Either supply the promised experiments or remove/correct the abstract statement.","section":"Abstract vs. Sections 3.5 and 5"},{"comment":"The feasibility constraints are only specified as abstract sets X and Y, with the statement that the 'full algebraic specification' is deliberately omitted. The network topology constraints (ring/interconnected, capacity, connectivity realization, trenching-cabling consistency) are central to the problem and to the feasibility of the repair operator. Without a precise statement of these constraints, the problem is not fully defined and the implementation cannot be independently verified. This is not a presentation nicety: the claim that the proposed method finds feasible optimal solutions depends on these sets being correctly encoded and enforced.","section":"§3.2–3.3, Eqs. (3) and (7)"},{"comment":"L-MVNS is pretrained on Case-0 and then 'online fine-tuned' on the first 200 iterations of the target instance before being evaluated. This means the DRL policy has already seen part of the test instance's search trajectory, so the comparison between L-MVNS and MVNS is not out-of-sample. The claim that L-MVNS 'delivers additional gains on larger instances' is therefore at risk of in-sample tuning artifact. To support generalization, report L-MVNS without online fine-tuning (pretrained only), or use a train/test instance split and show that fine-tuning on the test instance does not materially change the conclusions.","section":"§5.1, L-MVNS evaluation protocol"},{"comment":"The comparisons are based on only 10 independent runs per case. Several gaps are small (e.g., Case-2 MVNS 5295.26 vs. SNS-1 5325.11, 0.68%; Case-1 L-MVNS 2489.23 vs. MVNS 2481.42, L-MVNS is actually worse but the text only notes Case-1 is an exception). No statistical tests (e.g., Wilcoxon signed-rank, Mann–Whitney U) or confidence intervals are given. Claims of 'consistently yields the best' and 'higher stability' need quantitative support, especially for differences of a few percent.","section":"Table 2, §5.2.4"}],"minor_comments":[{"comment":"The benchmark generation says 'with random seed 42, we generate four scales' but the specific seeds for K-means and demand draws are not documented, which hampers exact reproducibility.","section":"§3.5"},{"comment":"Typo: 'CNY0.5 million CNY/km' should be '0.5 million CNY/km' or 'CNY 0.5 million/km'.","section":"§3.5"},{"comment":"The notation for the number of feeders, arcs, and road edges is introduced as |F|, |A|, |R|, but the earlier notation uses script letters and the symbol F is not formally defined; please align the notation.","section":"§3.4"},{"comment":"The text says SNS-1/2/3 use a neighborhood size of 30, while MVNS/L-MVNS use 10 per operator (total 30). This is fair, but the per-operator budget difference (30 vs. 10) should be explicitly acknowledged when interpreting operator-frequency statistics in Fig. 10.","section":"§5.2"},{"comment":"The reward is defined only on improvement; the baseline function b(s_i) is described as a three-layer fully connected network but its inputs and training target are not specified. A short clarification would help reproducibility.","section":"§5.1, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper's main contribution is potentially interesting, but the current framing is misleading. The 30–50% cost reduction is a model-based comparison, not an algorithmic result, and the promised 12-instance/GIS evaluation is absent from the body. The in-sample fine-tuning of the DRL component further weakens the 'additional gains on larger instances' claim. I recommend major revision: require the authors to correct the abstract/conclusion, add or remove the missing benchmark/GIS experiments, formalize the constraint sets, and add out-of-sample DRL evaluation plus statistical tests. If these cannot be done, the paper may be better positioned as a problem-formulation and modeling study rather than an algorithmic superiority claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper is a reasonable engineering contribution with an overstated headline. The 30–50% reduction over \"representative 2024 approaches\" is largely a modeling artifact: the baselines (MCWS, HGS) don't account for shared trenching, so they're optimizing a different cost model. In the paper's own fair comparison against path-aware SNS methods, MVNS/L-MVNS gain only 0.12–8.00%. That's the number to focus on.\n\nWhat's genuinely new: the integrated formulation coupling substation connectivity, road-constrained routing, and trench sharing; the four synthetic lattice benchmarks; and the MVNS framework with three destruction operators plus a DRL-guided neighborhood selector. The experiments are internally consistent, and the ablation showing multi-operator search beats single-operator search is meaningful. The stability improvement from L-MVNS (lower variance) is also credible, though modest.\n\nSoft spots, in order of severity. First, the abstract promises 12 benchmark instances and 3 GIS-derived instances; the body reports only four synthetic cases. That's not a small omission—it's a mismatch between the advertised and actual evaluation. Second, the headline claim. As written, it invites the reader to think the algorithm beats 2024 methods by 30–50%, but the comparison is against baselines with a different objective. The fair comparison is 1–8%, and even that is within a small number of runs with no significance tests. Third, the L-MVNS advantage comes partly from online fine-tuning on the first 200 iterations of the very instance it's evaluated on. That's in-sample tuning, which inflates the reported gains relative to MVNS. Fourth, the constraints are abstracted into sets (Eqs. 3 and 7) rather than given explicitly, which makes reproduction harder; and no code is available.\n\nNone of this invalidates the core direction. The problem is practically important, and the integrated treatment of trench sharing is the right thing to model. The paper deserves a serious referee, but it needs a major revision: either provide the GIS experiments and a fair baseline set, or rewrite the abstract and conclusion to state the actual comparison. The benchmark suite and formulation alone are worth publishing, even if the algorithmic novelty is thinner than claimed.","headline":"Solid engineering paper with an overstated headline: the 30–50% cost saving is a modeling effect, not an algorithmic breakthrough; fair comparisons show 1–8% gains.","tokens_in":17359,"tokens_out":2760,"would_cite":true,"duration_ms":27373,"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":"Co-optimizing substation connectivity and road-constrained cable paths, with shared-trench costs internalized, cuts urban cable construction cost by roughly 30–50% compared with representative 2024 approaches.","keywords":["urban cable routing","connectivity–path co-optimization","variable neighborhood search","deep reinforcement learning","multi-agent","trench sharing","distribution network planning","benchmark suite"],"falsifier":"Run L-MVNS and a relation-only baseline (e.g., MCWS) on a real GIS road network with actual trenching and cable cost data for a mid-size city; if the cost gap is well below 30%, the headline savings do not generalize. Alternatively, on the paper's own Case-1 instance, set the trench-to-cable cost ratio to 1:1 and check whether the reported gap between path-aware and relation-only methods collapses—which would show the savings are tied to the cost assumption.","tokens_in":16496,"feed_emoji":"⚡","tokens_out":7913,"duration_ms":65827,"temperature":0.7,"pith_summary":"Urban underground cable networks are extremely expensive, and traditional planning algorithms typically decide which substations to connect using straight-line distances, only afterwards drawing cable paths along roads. This paper argues that this sequential, relation-only approach costs cities 30–50% more than necessary, because in cities cables must follow roads and running several cables in the same trench sharply reduces construction cost. It reformulates the problem as a two-level co-optimization of substation connectivity and road-constrained routing, and solves it with a multi-operator variable neighborhood search whose neighborhood choices are guided by a multi-agent deep-reinforcement-learning module. On four synthetic grid benchmarks, the proposed L-MVNS is reported to cut total construction cost by roughly 30–50% versus representative 2024 baselines, with lower run-to-run variance on larger instances. If the result transfers to real street layouts, it gives planners a practical way to save a large fraction of an expensive infrastructure budget.","feed_headline":"Shared-trench routing cuts urban cable cost by 30–50%","feed_subtitle":"Planners who route cables along roads and share excavation costs could save a third to half of construction budgets.","key_machinery":"The carrying mechanism is a two-level formulation: a substation-level graph encodes which feeders connect which substations, and a road-level graph encodes which road segments are trenched and how many parallel cables each carries. The algorithm iteratively destroys and repairs both levels: three destruction operators (path-related removal, intra-feeder 2-opt, and inter-feeder 2-opt) alter the candidate network, a modified A* repair re-plans affected paths while counting trench-sharing costs, and an adaptive perturbation size balances exploration and exploitation. In L-MVNS, three lightweight agents (each an LSTM plus multi-head attention) read the current solution and output a probability f","core_discovery":"The paper's central claim is that co-optimizing which substations to connect and exactly which road segments to excavate, with shared-trench savings internalized, reduces total construction cost by approximately 30–50% against representative 2024 planning approaches. The evidence comes from four synthetic regular-lattice benchmarks, where path-aware methods consistently beat relation-only baselines (MCWS and HGS) that use precomputed shortest paths and ignore trench sharing; the multi-operator variable neighborhood search (MVNS) and its learning-assisted version (L-MVNS) give the lowest costs, and L-MVNS adds moderate gains and markedly lower variance on larger cases.","pith_inferences":["The 30–50% figure is measured against MCWS and HGS on synthetic grids; whether it transfers to real street networks with irregular geometry and non-uniform costs is untested in the body of the paper, even though the abstract promises GIS-derived instances that do not appear in the reported experiments.","The benefit likely depends on the trench-to-cable cost ratio: if the ratio approaches 1:1, the incentive to share trenches weakens and the gap between path-aware and relation-only planning should shrink—the paper's sensitivity analysis varies only algorithmic parameters, not the cost ratio.","The same connectivity–path co-optimization pattern, with trench sharing priced in, could be applied to other shared-infrastructure networks such as district heating, fiber optics, or water mains, where routing constraints and joint excavation costs dominate.","Because the DRL module's gains over the non-learning MVNS are modest (sub-2% on reported cases), a sharper test would be to evaluate whether a learned destruction policy generalizes across instance scales and cost structures without per-instance fine-tuning."],"forward_implications":["Using path-aware, trench-sharing planning instead of relation-only planning is reported to cut total construction cost by roughly 30–50% on the benchmark suite.","The multi-agent DRL guidance reduces run-to-run variance, making cost estimates from the algorithm more predictable on large instances.","The four regular-lattice instances form a standardized, scalable benchmark for future urban cable routing algorithms.","The auxiliary-task initialization (hybrid genetic search for connections, A* for paths) produces feasible starting points that accelerate the subsequent neighborhood search without special tuning.","The method's gap to the best single-operator search is small but consistent, indicating that combining complementary destruction operators is more effective than any one operator alone."],"fun_headline_variants":["AI-driven cable routing cuts costs by 50%","Smart routing shares trenches, slashes cable costs","Cable planning algorithm saves up to 50% on urban builds","Learning to sample: better cable routes, lower costs","Urban cable routing optimization cuts construction costs 30–50%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that planning with explicit roads and shared trenches saves 30–50% rests on four synthetic regular-grid benchmarks with a fixed 3:1 trench-to-cable cost ratio standing in for real urban road networks; if those benchmarks are not representative, the savings figure may not transfer to actual cities.","fun_headline_variants_meta":{"raw":{"variants":["AI-driven cable routing cuts costs by 50%","Smart routing shares trenches, slashes cable costs","Cable planning algorithm saves up to 50% on urban builds","Learning to sample: better cable routes, lower costs","Urban cable routing optimization cuts construction costs 30–50%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001002,"raw_usage":{"total_tokens":4088,"prompt_tokens":771,"completion_tokens":3317,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":3237}},"tokens_in":515,"tokens_out":3317,"duration_ms":27141,"temperature":1.0,"reasoning_tokens":3237,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:43:52.587979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run L-MVNS and a relation-only baseline (e.g., MCWS) on a real GIS road network with actual trenching and cable cost data for a mid-size city; if the cost gap is well below 30%, the headline savings do not generalize. Alternatively, on the paper's own Case-1 instance, set the trench-to-cable cost ratio to 1:1 and check whether the reported gap between path-aware and relation-only methods collapses—which would show the savings are tied to the cost assumption.","supporting_citations":[],"review_version":1}