{"id":"9e3255f9-420d-45a8-92e7-b8abf654074a","arxiv_id":"2412.05665","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural cost predictor embedded in a mixed-integer program chooses depots and customer assignments for location-routing, reaching near best-known solutions on large benchmarks with seconds of allocation time.","lead":"This paper trains a neural network to estimate vehicle routing costs and embeds it inside a mixed-integer optimization model to choose warehouse locations and customer assignments for the capacitated location-routing problem. The appeal is that large location decisions can be made in seconds after a one-time training cost, though the actual vehicle routes still have to be computed afterward.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'within 1% for large problems' claim relies on training data subsampled from the very Prodhon benchmark instances used for testing, so out-of-the-box generalization is not established.","rationale":"The central contribution of the paper is a pre-trained neural surrogate that can be embedded in a MIP and used out-of-the-box. The abstract explicitly promises training data 'entirely separate' from test instances to avoid information leakage. However, the main experimental results use RSCC, which subsamples training VRP instances from the same 30 Prodhon CLRP instances used for testing (Section 5.4 and Algorithm 1). This directly contradicts the abstract. The paper's own ablation (Section 5.5.1) shows that GVS, a truly independent sampling method, yields a median gap of 4.29% versus 1.84% for RSCC, demonstrating that the near-1% performance on 200-customer instances likely depends on training on the test-instance geometry. This is the single most load-bearing concern because the headline empirical claim—and the 'out-of-the-box' framing—collapses if the training data overlaps the test set. The proposed check, retabulating the 200-customer results with the GVS-trained model, directly settles whether the claim survives without this overlap. Other components, such as the Deep Sets representation (Theorem 1) and the MIP encoding, are not the main risk; the weakness is empirical validity of the generalization claim.","tokens_in":22807,"tokens_out":6513,"duration_ms":58817,"concrete_test":"Recompute Table 1 for the 200-customer instances using the GVS-trained single neural network from Section 5.5.1 (which is independent of the Prodhon instances) in place of the RSCC-trained network, reporting the average Egap_BKS and TLA over the same five runs. If the average gap exceeds 1% or the per-instance gaps rise materially above the RSCC-based values, the 'within 1%' claim is not a valid measure of generalization to new instances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the neural surrogate is trained on a dataset 'entirely separate from any CLRP test instances,' but the main experiments (Section 5.4) generate the 10,000 training VRP instances 'across all 30 CLRP instances using the RSCC data sampling method.' Algorithm 1 (Appendix B) explicitly selects a CLRP instance from the same benchmark set and then subsamples customers and a depot from that instance. Thus, the training instances share depot coordinates, customer coordinates, demands, capacities, and vehicle parameters with the test CLRP instances. The surrogate can therefore memorize the cost-to-geometry mapping for the specific test instances, and the resulting location-allocation decisions may not reflect performance on genuinely new instances. The paper's own ablation (Section 5.5.1) shows that training with GVS, which is independent of the Prodhon set, yields a median gap of 4.29% versus 1.84% for RSCC, indicating that the choice of sampling is consequential. The headline claim of 0.80% average gap for 200-customer instances in Table 1 is thus not evidence of out-of-the-box generalization; it reflects overlap between the training and test distributions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NEO-LRP, a framework that approximates the optimal vehicle routing cost of serving any customer subset from a depot by a permutation-invariant neural network (a Deep Sets architecture), embeds this ReLU network as a mixed-integer surrogate in a CLRP location-allocation MIP, solves the MIP with an off-the-shelf solver, and then evaluates the true routing cost of the resulting allocation with an exact VRP solver. The authors report experiments on the Prodhon benchmark set, comparing against GRASP/VNS and an FLP-VRP baseline, and provide ablations on sampling methods, sample size, routing label solver, problem size, and single versus customized networks. The central technical claims are that the surrogate MIP is easy to solve, that a single pre-trained model can be used out of the box, and that NEO-LRP achieves high-quality location-allocation decisions for large instances in seconds.","tokens_in":22934,"tokens_out":2260,"duration_ms":22806,"significance":"If the generalization claim holds, NEO-LRP offers a genuinely modular and simple alternative to routing-specific heuristics: it requires no routing-specific algorithm engineering, handles side constraints within the MIP, and leverages mature MIP and VRP solvers. The paper has notable strengths: Theorem 1 is a sound application of an existing set-function representation theorem; the final location-allocation solution is evaluated by an exact VRP solver, so the reported objective values are not based on the surrogate's own predictions; the ablation studies are broad and informative; and the authors release code and data. However, the main evidence for out-of-the-box generalization is currently compromised by the RSCC sampling procedure, which draws training VRP instances from the very Prodhon CLRP instances used for testing. The paper's own GVS ablation shows that changing the sampling method substantially changes the optimization gap, so this issue is load-bearing for the headline claims.","major_comments":[{"comment":"The abstract states that the neural network is trained on a dataset 'entirely separate from any CLRP test instances', but this is contradicted by the experimental protocol. Section 5.4 states that the 10,000 training instances are generated 'across all 30 CLRP instances using the RSCC data sampling method', and Algorithm 1 in Appendix B explicitly selects a CLRP instance from the benchmark set and subsamples customers and a depot from that instance. Consequently, the training VRP instances share depot coordinates, customer coordinates, demands, capacities, and vehicle parameters with the test CLRP instances. The headline result of a 0.80% average gap for 200-customer instances in Table 1 is therefore not evidence of generalization to unseen CLRP instances. I request either retraining with a data-generation procedure that is independent of the Prodhon set (e.g., GVS) and reporting the corresponding Table 1, or substantially revising the abstract and the out-of-the-box generalization claims.","section":"Abstract and Section 4.4; Appendix B, Algorithm 1"},{"comment":"The ablation on sampling methods shows that the choice of sampling is highly consequential: RSCC achieves a median optimization gap of 1.84%, while GVS, which is independent of the Prodhon instances, achieves 4.29%. The paper attributes this difference to 'alignment between the training data and the customer assignments Ji(x) during optimization', but an equally natural explanation is that RSCC training instances are subsampled from the same benchmark instances used for testing. Because the main experiments in Table 1 use RSCC, the reported gaps may be optimistic for unseen instances. This does not invalidate the framework, but it means the paper currently establishes in-distribution performance on Prodhon rather than out-of-the-box generalization.","section":"Section 5.5.1, Figure 3"},{"comment":"The single-versus-customized comparison does not resolve the generalization question. Both the 'single' and 'customized' networks are trained with RSCC on the same Prodhon benchmark set, so the comparison measures the effect of aggregating training data across instances, not the effect of testing on instances that were never used in training. To support the claim that a single pre-trained model generalizes across CLRP instances, the authors should evaluate a model trained without any instances from the test benchmark, or on a separate benchmark family.","section":"Section 5.5.5, Table 2"}],"minor_comments":[{"comment":"In Section 4.2, the normalization constant P is defined as the maximum range of customer coordinates after centering the depot at (0,0), whereas Theorem 1 treats P as any fixed positive constant. Please clarify that the specific choice of P in the implementation is one admissible instantiation of the theorem's constant.","section":"Section 4.2"},{"comment":"Equation (17) uses the index m both as the output of the final layer and in the notation w^K_{\\ell m}; please adjust the notation to avoid confusion with the hidden-layer index m used in equations (14)-(16).","section":"Appendix A, Eq. (17)"},{"comment":"In the row for instance 200-10-3b in Table 2, the BKS value is printed as '362, 320' with an extra space; please correct the formatting.","section":"Table 2"},{"comment":"The phrase 'Tone-time' should be 'one-time' or 'T_one-time'; the subscript appears to be a typographical artifact.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core methodological idea is sound and well-presented, and the derivation-level reasoning is not circular. The problem is empirical: the paper's headline generalization claim is not supported by the current RSCC-based experiments, because the training data are subsampled from the test benchmark. The GVS ablation in Section 5.5.1 already provides a concrete path for the revision: re-running the main comparison with an independent sampling scheme would either restore the claim or force a more modest interpretation. I therefore see this as a major-revision situation rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful, honest framework paper whose headline empirical claim doesn't survive contact with its own data-generation procedure. The method relies on a training set subsampled from the same Prodhon instances it is tested on, so the sub-1% gaps for 200-customer cases are not evidence of out-of-the-box generalization.\n\nWhat's actually new: first neural-embedded MIP for the capacitated location-routing problem. Theorem 1 is a clean application of the Deep Sets sum-decomposition to depot-relative VRP costs; the MIP encoding of the ReLU regressor is standard but well-executed; and the modular pipeline (off-the-shelf MIP + a posteriori exact VRP solving) is a genuinely useful engineering contribution. The ablations on sampling methods, sample sizes, and exact-vs-heuristic labeling are thoughtful, and the code and data are public. The paper is also transparent that NEO-LRP does not produce routes.\n\nThe soft spot is load-bearing. The abstract claims training is 'entirely separate from any CLRP test instances,' but the main experiments (Sec 5.4) generate 10,000 RSCC training instances by subsampling across all 30 Prodhon instances, which are exactly the test instances. Appendix B confirms the sampling draws a CLRP instance from the benchmark, then subsamples customers and a depot. So the surrogate has seen the test geometries. The paper's own ablation (Sec 5.5.1) shows GVS, which is independent, gives a 4.29% median gap versus 1.84% for RSCC. That gap is the difference between the headline claim and a merely good result. The paper should either retrain on clearly disjoint instance families or report RSCC and GVS results on all benchmarks, and qualify the abstract accordingly.\n\nA second, minor point: the 'less than 5 seconds' time is T_LA, the location-allocation phase only; total times in Table 1 are much larger (e.g., 431s average for 200-customer instances). The abstract does say 'location-allocation decisions,' so this is not a fatal misrepresentation, but it is easy to misread.\n\nVerdict: worth a serious referee. The framework is real, the math is sound, and the generalization question is fixable with honest reporting. I'd send it to review, expecting major revisions on the experiments.","headline":"A useful neural-embedded MIP framework for CLRP whose headline generalization claim is undermined by training-test overlap on the Prodhon benchmark.","tokens_in":23566,"tokens_out":2364,"would_cite":true,"duration_ms":20904,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C27","90B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural cost surrogate embedded in a MIP can make location-allocation decisions for large capacitated location-routing instances within about 1% of best-known solutions in seconds.","keywords":["location-routing","capacitated location-routing problem","vehicle routing","mixed-integer optimization","neural network surrogate","permutation-invariant set function","location-allocation","pretrained model"],"falsifier":"Hold out the six 200-customer benchmark instances entirely during training, train the same NEO-LRP pipeline on generic vehicle-routing instances or on subsamples from the other 24 instances, and measure the average optimization gap on the held-out instances; if the average gap is far above the reported 0.80%, the out-of-the-box generalization claim is falsified, whereas a gap near 1% would support it.","tokens_in":22516,"feed_emoji":"🚚","tokens_out":9096,"duration_ms":78963,"temperature":0.7,"pith_summary":"The paper is trying to establish that the capacitated location-routing problem (CLRP)—where one must choose depots to open and routes to serve customers—can be tackled by replacing the expensive vehicle-routing subproblem with a learned surrogate cost. It trains a permutation-invariant neural network on vehicle-routing instances to predict, for any depot and any subset of customers, the optimal routing cost, then embeds that network directly into a mixed-integer program whose solution gives the depot-opening and customer-assignment decisions. After a one-time training cost of about two hours, the method provides location-allocation decisions within 1% of the best-known solutions for 200-customer instances, in about 3.5 seconds on average, while remaining a simple, modular pipeline that an off-the-shelf MIP solver can handle. The paper is careful to note that it outputs locations and allocations, not the actual routes; those are computed afterwards with a standard vehicle-routing solver.","feed_headline":"Pretrained net inside a MIP picks depots within 1% of best known","feed_subtitle":"Trained once on routing costs, the model answers 200-customer location-allocation questions in about 3.5 seconds.","key_machinery":"The load-bearing object is the sum-decomposition identity for routing cost: the optimal vehicle-routing cost of a set of customers served by a depot is representable, up to normalization, as $\\rho$ applied to the sum of per-customer embeddings $\\varphi(s_{ij})$, with normalized features $s_{ij}=(P^{-1}(u_j-u_0^i),P^{-1}(v_j-v_0^i),Q^{-1}d_j)$. This converts a decision-dependent routing cost into a linear function of the assignment variables $x_{ij}$ once $\\varphi$ is precomputed, and makes the whole surrogate embeddable as mixed-integer linear constraints through the standard ReLU-network encoding. The design also lets one $\\varphi/\\rho$ pair serve every depot and any subset size, which is why a single pretrained network can be reused across instances.","core_discovery":"The paper's central claim is that the difficult part of the capacitated location-routing problem can be pushed into a learned, permutation-invariant cost model. Writing $g_i(S)$ for the optimal vehicle-routing cost of serving a customer set $S$ from depot $i$, the paper proves (Theorem 1) that after normalizing customer coordinates relative to the depot and demands by vehicle capacity, this set function admits a depot-independent decomposition $g_i(S)=\\rho\\left(\\sum_{j\\in S}\\varphi(s_{ij})\\right)$. The functions $\\varphi$ and $\\rho$ are approximated by ReLU feed-forward networks; because only $\\rho$ is embedded in the MIP and it is ReLU-activated, the trained surrogate becomes a system of linear constraints plus binary neuron variables, solvable by an off-the-shelf MIP solver. On the 30-instance benchmark, the learned surrogate's location-allocation decisions average 4.32%, 3.68%, 1.77%, and 0.80% above best known for 20, 50, 100, and 200 customers, with the 200-customer decisions taking about 3.5 seconds on average.","pith_inferences":["The framework's separation of location-allocation from routing means its practical ceiling is the quality of the learned cost ranking, not the accuracy of absolute cost predictions; a ranking-preserving loss could therefore improve the optimization gap more than reducing mean squared error, a direction the paper itself flags for future work.","Because the decomposition is depot-independent and feature-based, the same surrogate architecture could be lifted to other integrated problems with set-cost terms, such as two-echelon location-routing or combined production-distribution planning, by retraining the cost model on the corresponding vehicle-routing variant.","A natural stress test beyond the paper's experiments would be to hold out entire benchmark instances during training and evaluate on those held-out instances; this would directly measure whether the reported gaps reflect genuine out-of-the-box generalization or overlap between training subsamples and test instances."],"forward_implications":["When the paper is right, a single pretrained surrogate can be reused out of the box on a new CLRP instance of any size, without retraining for that instance.","Side constraints such as incompatible customer pairs or depot-specific customer restrictions reduce to adding linear inequalities to the MIP, so the framework extends without new routing logic.","The final evaluation is exact: after the MIP chooses depots and assignments, a VRP solver computes true routes and costs, so the reported gaps are against actual routing costs, not surrogate predictions.","For 200-customer instances the method beats the tailored heuristic baseline in both average gap (0.80% vs 1.24%) and total time; for 100-customer instances it is much faster but slightly less accurate.","Training-data quality matters more than volume: increasing training samples from 100 to 100,000 barely moves the optimization gap, while switching from generic VRP samples to benchmark-subsampled training samples halves it."],"supporting_citations":[{"why":"Supplies the state-of-the-art heuristic baseline and the best-known solution values that NEO-LRP is compared against.","marker":"[36]"},{"why":"Provides the mixed-integer linear representation of ReLU networks used to embed the neural surrogate in the MIP.","marker":"[25]"},{"why":"Gives the sum-decomposition architecture on which Theorem 1's depot-independent representation relies.","marker":"[61]"},{"why":"Its representation theorem provides the existence of the functions $\\varphi$ and $\\rho$ used in the proof of Theorem 1.","marker":"[60]"},{"why":"The exact vehicle-routing solver used to compute true routes and costs a posteriori and to produce exact training labels.","marker":"[22]"},{"why":"The heuristic vehicle-routing solver used to label the 10,000-sample training sets.","marker":"[26]"},{"why":"Defines the 30-instance benchmark set used for all experiments and best-known comparisons.","marker":"[44]"},{"why":"Documents the earlier neural-cost-predictor-plus-genetic-algorithm approach to CLRP that NEO-LRP is contrasted with as less modular.","marker":"[54]"}],"fun_headline_variants":["AI-learned routing costs MIP picks depots in seconds","Trained surrogate inside MIP for fast location-routing","Neural cost model embedded in MIP solves CLRP quickly","One-time training then 3.5s depot decisions via neural net","Hybrid neural-MIP approach finds near-optimal depots fast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise of the experiments is that training the surrogate on 10,000 vehicle-routing instances subsampled from the same 30 benchmark instances used for testing is a valid way to measure out-of-the-box generalization; if that overlap inflates the reported gaps, the principal claim of easy generalization to new instances is not established by this evidence.","fun_headline_variants_meta":{"raw":{"variants":["AI-learned routing costs MIP picks depots in seconds","Trained surrogate inside MIP for fast location-routing","Neural cost model embedded in MIP solves CLRP quickly","One-time training then 3.5s depot decisions via neural net","Hybrid neural-MIP approach finds near-optimal depots fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":2189,"prompt_tokens":1012,"completion_tokens":1177,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1089}},"tokens_in":628,"tokens_out":1177,"duration_ms":12599,"temperature":1.0,"reasoning_tokens":1089,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:28:42.898290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out the six 200-customer benchmark instances entirely during training, train the same NEO-LRP pipeline on generic vehicle-routing instances or on subsamples from the other 24 instances, and measure the average optimization gap on the held-out instances; if the average gap is far above the reported 0.80%, the out-of-the-box generalization claim is falsified, whereas a gap near 1% would support it.","supporting_citations":[{"cited_title":"EURO Journal on Computational Optimization 11, 100063 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the state-of-the-art heuristic baseline and the best-known solution values that NEO-LRP is compared against."},{"cited_title":"In: Proceedings of the 31st International Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Gives the sum-decomposition architecture on which Theorem 1's depot-independent representation relies."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Its representation theorem provides the existence of the functions $\\varphi$ and $\\rho$ used in the proof of Theorem 1."},{"cited_title":"INFORMS Journal on Computing 36(4), 956–965 (2024)","cited_arxiv_id":null,"evidence_quote":"The exact vehicle-routing solver used to compute true routes and costs a posteriori and to produce exact training labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The heuristic vehicle-routing solver used to label the 10,000-sample training sets."},{"cited_title":"In: MOSIM’04 (4` eme Conf","cited_arxiv_id":null,"evidence_quote":"Defines the 30-instance benchmark set used for all experiments and best-known comparisons."},{"cited_title":"Transportation Science (2024)","cited_arxiv_id":null,"evidence_quote":"Documents the earlier neural-cost-predictor-plus-genetic-algorithm approach to CLRP that NEO-LRP is contrasted with as less modular."}],"review_version":1}