{"id":"ff56bbc2-b6b9-406d-9739-725b52591a71","arxiv_id":"2505.05119","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A unified transformer-based reinforcement-learning policy, USPR, encodes arbitrary vehicle-client profile scores and profile weights and outperforms prior neural PVRP solvers on synthetic, large-scale, and CVRPLib-derived benchmarks.","lead":"USPR is a single reinforcement-learning model that plans delivery routes when different trucks have different permissions or customer preferences, removing the need to retrain a neural solver for each new preference setting. It reports better solution quality than prior learning-based routing solvers and keeps working when tested on instances up to ten times larger than its training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'arbitrary profile types' claim is untested: every experiment reuses the same Uniform(0,1) profile generator, so the unified-solver evidence is conditional on an unvaried distribution family.","rationale":"The reader's weakest-assumption identification is correct: the profile-distribution gap is the most load-bearing weakness in the central claim. The SOTA and efficiency comparisons are internally coherent, and the unified-model training savings are real conditional on the experiments as run. My read does not move the verdict: acceptance should remain conditional on demonstrating profile-family robustness and on releasing the code so that the PSR formula and the baseline-adaptation protocol can be checked. I additionally note that Eq. (5) appears to have a sign/indexing issue, since writing -log(dist_ij + p_ij) makes higher preference scores decrease the logit, which is opposite to the stated objective, and the term is not scaled by alpha; this should be verified against the actual implementation. Missing error bars, the unavailable code artifact, and the unspecified adaptation of ET, DPN, 2D-Ptr, and PARCO all support keeping the verdict conditional rather than accepting as-is.","tokens_in":22023,"tokens_out":7096,"duration_ms":71002,"concrete_test":"Take the released USPR checkpoint and run the same in-distribution test protocol (N=60-100, M=3-7, the alpha grid, greedy and sampling) on profiles drawn from at least two structurally different families: a categorical preference scale such as {0,1,2,3}, a Beta(0.5,0.5) continuous distribution, spatially correlated scores that increase with proximity to each client's zone, and sector-based hard zone constraints as described in the supplementary. Compare the average gap to HGS-PyVRP against the Uniform(0,1) baseline in Table 1. If the gap degrades by more than roughly two percentage points on any family, or if retraining is required to restore performance, the 'arbitrary profile types' claim should be narrowed to independent real-valued profiles.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a single USPR model handles arbitrary vehicle-client profile matrices and profile weights without retraining. For that claim to hold, the model must remain effective as the profile distribution changes. The experiments never test this. In 'Experiments: Data Generation' and the supplementary data-generation paragraph, profile scores are always drawn independently from Uniform(0,1), with 0% to 10% of entries randomly set to plus or minus infinity. The PVRPLib set retains real coordinates, demands, and vehicle counts from CVRPLib but regenerates profiles using the same synthetic protocol, so it does not provide a different profile distribution. The zone-constrained variant described in the supplementary is described but not evaluated in any reported table. Consequently, every headline gap—Table 1, Tables 3/8/9, and the PVRPLib results in Table 4—is produced under a single profile family. This does not mean the method fails; the architecture's linear embeddings of raw scalars could plausibly transfer to other profile distributions. But the 'arbitrary' generalization claim is precisely the load-bearing scope condition of the paper, and the evidence leaves that condition untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes USPR, a unified neural solver for the Profiled Vehicle Routing Problem (PVRP), in which a single model trained once handles arbitrary vehicle-client profile matrices and profile weights without retraining. The method combines Profile Embeddings (PE) to encode profiles and weights, Multi-Head Profiled Attention (MHPA) for bidirectional vehicle-client interactions, and Profile-aware Score Reshaping (PSR) that injects explicit distance and profile information into decoder logits. The authors report in-distribution results on N=60-100, out-of-distribution scale generalization up to N=1000, and results on a newly constructed PVRPLib dataset; they also report an ablation and an efficiency comparison showing one USPR model replaces ten CAMP models. The central empirical claim is state-of-the-art performance among learning-based PVRP solvers, with a single model generalizing to instances up to 10x larger than training data.","tokens_in":22145,"tokens_out":4418,"duration_ms":40791,"significance":"The paper addresses a real limitation of existing neural PVRP solvers, which typically require separate models for different profile distributions and weights. The architectural ideas (PE, MHPA, PSR) are sensible and the experimental results consistently show USPR outperforming the neural baselines, including CAMP, across all tested scales, while being substantially cheaper to train. The scale-generalization numbers (e.g., Table 3, N=1000) are notable and support the method's practical potential. If the unified-solver claim could be substantiated across genuinely varied profile distributions, this would be a meaningful contribution to the NCO literature. However, the current evidence is narrower than the claim: all experiments use one synthetic profile generation protocol, and the absence of variance estimates and code availability limits the strength of the empirical conclusions.","major_comments":[{"comment":"The central claim of a unified solver for 'arbitrary profile types' is not tested. In every experiment, including PVRPLib, profile scores are drawn independently from Uniform(0,1) with between 0% and 10% of entries set to ±infinity. No experiment varies the profile distribution family (e.g., categorical, correlated, negative-valued, or real driver-customer preferences), and the zone-constrained variant described in the supplementary is not evaluated in any reported table. Since the claim of handling arbitrary profile types is the paper's key selling point, the evidence currently supports only the narrow case of i.i.d. uniform soft scores with sparse hard constraints. The authors should either add experiments with substantially different profile distributions or explicitly restrict the scope claim in the abstract and introduction.","section":"Experiments: Data Generation / Supplementary Data Generation"},{"comment":"No table reports variance or multiple-seed results. All neural methods are evaluated on a single run, so the observed gaps between USPR and the strongest baseline CAMP (e.g., Table 1 greedy gap 9.20% vs. 10.50%; Table 3 greedy gap 19.35% vs. 23.72%) could be within run-to-run noise. The paper's central empirical claim is that USPR achieves state-of-the-art results among learning-based methods; this claim requires at least a few independent training runs with reported means and standard deviations, or statistical significance tests.","section":"Tables 1, 3, 4, 5, 8, 9"},{"comment":"The abstract states that source code is publicly available, but the manuscript contains no link or repository identifier. The supplementary says only that code 'will be released upon acceptance.' This breaks the reproducibility expectation for a machine learning paper and makes it impossible to verify the reported results or baseline adaptations. The authors should provide an anonymous or working link to code and trained models, or clearly state in the abstract that code will be released upon publication.","section":"Abstract and Supplementary 'Code Implementation and Hardware'"},{"comment":"The statement 'We follow CAMP to adapt ET, DPN, 2D-Ptr, and PARCO to PVRP' is not substantiated. The paper gives no details on how profiles are incorporated into these architectures, whether hyperparameters were retuned, or whether the adaptation is equivalent across baselines. Without this information, the fairness of the comparison cannot be assessed. The authors should specify the adaptation protocol, ideally in an appendix.","section":"Experiments: Neural Baselines"},{"comment":"The PSR formula uses log(dist_ij + p_ij). Since the paper's problem definition allows p_ik in R∪{±infinity}, this expression is undefined for p_ij = -infinity and for any p_ij ≤ -dist_ij, and it is also potentially ill-defined for +infinity. The paper does not explain how hard constraints are handled in the logit computation before masking, nor how the formula generalizes to negative or arbitrary real-valued profiles. This is a technical gap directly related to the 'arbitrary profile types' claim.","section":"Eq. (5), Profile-Aware Score Reshaping"}],"minor_comments":[{"comment":"There is an inconsistency in component naming: the Introduction contribution list says 'PAR to ensure robust generalization', while the Methodology section and all experiments refer to PSR (Profile-aware Score Reshaping). Please correct the typo.","section":"Introduction / Methodology"},{"comment":"The caption claims results for instances 'up to 10× the number of agents M and 10× the number of nodes N'. While N=1000 is 10x the training maximum of N=100, the agent number M=35 is only 5x the training maximum of M=7. Please correct the factor for M.","section":"Table 3 caption"},{"comment":"The ablation labels '-PSR & SR' and '-PSR, SR & MHPA' are unclear; 'SR' is not defined in the main text. Please state clearly what SR (distance-only score reshaping?) denotes and what each ablation removes.","section":"Table 5"},{"comment":"The conclusion states that USPR is 'matching or exceeding classical solvers', but the reported gaps to HGS-PyVRP are consistently positive (e.g., 9.20% in Table 1 and 19.35% in Table 3). The paper only outperforms OR-Tools in some settings, not HGS-PyVRP. Please rephrase to avoid overclaiming.","section":"Conclusion"},{"comment":"PVRPLib is described as 'real-world' data, but only the coordinates, demands, and vehicle counts come from CVRPLib; the profile scores are generated with the same synthetic Uniform(0,1) protocol. The terminology 'real-world' is misleading and should be clarified, e.g., 'real-world geometry with synthetic profiles'.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim (unified handling of arbitrary profile types) is the main selling point, and the current experiments do not vary the profile distribution family at all. This is not a case of disagreeing with consensus; it is a mismatch between the claim and the evidence. The lack of variance reporting and missing code link would also need to be addressed before publication. The self-citation pattern is concentrated on the authors' own prior work (CAMP, PARCO, RL4CO) but does not appear to suppress alternative baselines; the main missing consideration is a stronger non-uniform profile baseline distribution. I would lean toward major revision rather than rejection because the architecture and in-distribution results are credible and the generalization claim could be fixed with additional experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: USPR is a real step forward for neural PVRP. Training one policy over randomized profile weights and hard-constraint rates, with profile embeddings and bidirectional vehicle-client attention, is a genuine departure from CAMP's per-alpha models, and the reported gains hold up reasonably. Tables 1, 3, 8, and 9 consistently show USPR beating CAMP and other neural baselines at similar or lower inference cost, and the 10x scale generalization is impressive if it reproduces. The ablation shows each component earns its place; MHPA removal hurts most. PSR is a hand-designed heuristic, not fit to test sets, so I don't see circularity in the benchmark comparisons.\n\nThe soft spots are real but mostly about scope, not the core numbers. No table reports variance or seeds, so I would want at least a couple of seeds before trusting the margins; that is standard for the area and fixable. The bigger issue is the 'arbitrary profile types' claim. Every experiment, including PVRPLib, generates profiles as independent Uniform(0,1) draws with 0-10% infinities. The OOD tests change instance scale, not the profile distribution family. Zone-constrained profiles are described in the appendix but never evaluated in any table, yet the conclusion claims success on 'zone-constrained routing problems.' That overclaim should be corrected. Also, the abstract promises public code, but the appendix says code will be released upon acceptance, and no link is present now. Baseline adaptation for ET, DPN, 2D-Ptr, and PARCO is 'we follow CAMP,' which is plausible but unverifiable without code. There's also a minor indexing bug: Eq. (1) uses p_ik with x^k_ij where the reward should attach to the served client j as in Eq. (11); it is cosmetic but should be fixed.\n\nThe citation pattern is normal; self-cites to CAMP and PARCO are expected since those are the main neural baselines. The memory limitation is stated honestly.\n\nWho is this for? Researchers in neural combinatorial optimization, especially multi-agent and profiled variants. It deserves serious peer review: the central claim is interesting, the empirical core is plausible, and the gaps are addressable in revision. I would recommend engaging rather than desk-rejecting, with the expectation that the authors either soften the 'arbitrary profile types' language or actually vary the profile distribution, add seed variance, and release code.","headline":"A credible single-model PVRP solver with solid in-distribution and scale gains, but the 'arbitrary profile types' headline claim is untested — every experiment uses one Uniform(0,1) profile generator.","tokens_in":22808,"tokens_out":2430,"would_cite":true,"duration_ms":23709,"reading_group":"yes","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 reinforcement-learning-trained transformer, USPR, handles arbitrary vehicle-client preference profiles and preference weights without retraining while reporting better solution quality than prior learning-based PVRP solvers on…","keywords":["profiled vehicle routing problem","unified neural solver","multi-head profiled attention","profile embeddings","score reshaping","reinforcement learning","out-of-distribution generalization","vehicle-client preferences"],"falsifier":"Generate a test set of PVRP instances whose profile scores come from a different distribution family, such as correlated preferences, ordinal categories, or real driver-customer ratings, and measure USPR's greedy gap to HGS-PyVRP; if the gap degrades sharply or falls behind CAMP on the same instances, the 'arbitrary profile types' unification claim would be refuted.","tokens_in":21680,"feed_emoji":"🚚","tokens_out":7530,"duration_ms":67462,"temperature":0.7,"pith_summary":"USPR is a single transformer-based policy trained once to solve the Profiled Vehicle Routing Problem (PVRP), where each vehicle-client pair has a preference score that can also encode hard constraints. The paper claims that one USPR model replaces the previous practice of training ten separate models on the same profile distribution, because its profile embeddings make the policy continuous in the profile matrix and in the preference weight $\\alpha$. Across in-distribution instances with 60-100 clients and out-of-distribution instances up to 1,000 clients, USPR reports a lower average gap to the HGS-PyVRP reference than all compared learning-based solvers, at a fraction of the training cost. The paper's contribution is a unified neural construction policy for preference-aware routing, not a new classical algorithm.","feed_headline":"One model, trained once, solves profiled vehicle routing","feed_subtitle":"USPR replaces ten per-profile models, cuts training time by an order of magnitude, and scales to 10x larger instances.","key_machinery":"Three components carry the argument. Profile Embeddings (PE) project clients, vehicles, profile scores (including $\\pm\\infty$ hard constraints) and the scalar weight $\\alpha$ into a shared latent space so the model can condition on any profile matrix. Multi-Head Profiled Attention (MHPA) runs four attention directions (client-client, vehicle-vehicle, vehicle-client, client-vehicle) to build bidirectional vehicle-client representations. Profile-aware Score Reshaping (PSR) adds an explicit distance-plus-preference term to the decoder logits, so the policy's outputs stay calibrated when profile distributions or instance sizes shift. Together they let a single autoregressive, parallel-vehicle decoder produce feasible routes without per-distribution retraining.","core_discovery":"The paper's central claim is that a single model, USPR, can natively handle arbitrary vehicle-client profile matrices and profile weights without retraining, and that it does so while being the best learning-based PVRP solver in the comparisons. The authors report that USPR's greedy solutions have a 9.20% average gap to the HGS-PyVRP reference on in-distribution instances of 60-100 clients, against 10.50% for the strongest baseline CAMP, and a 19.35% gap at 1,000 clients against 23.72% for CAMP. In the same comparison, one USPR model replaces ten CAMP models trained at fixed profile weights, cutting parameter count, epochs, and wall-clock training time by roughly an order of magnitude. The paper therefore positions USPR as a unification result: profile-aware routing can be handled by a single construction policy rather than a zoo of task-specific ones.","pith_inferences":["A direct test the paper does not run is to change the profile distribution family at test time, for example correlated, categorical, or empirical driver-customer preferences; the architecture's embeddings should be distribution-agnostic, but no evidence yet shows this.","Because the reshaping term exposes distance and preference scales directly in the logits, the mechanism may also transfer across coordinate scales or non-Euclidean distance metrics, which the paper does not evaluate.","The memory limitation the paper reports, that profile matrix embeddings explode past roughly 10,000 clients, suggests a sparse or factorized profile-encoding variant as an obvious next step."],"forward_implications":["Deploying USPR removes the need to train and store a separate model for each preference weight or zone-constraint configuration.","On the paper's benchmarks, USPR's greedy solutions are closer to the HGS-PyVRP reference than those of the best prior neural baseline CAMP, with reported gaps of 9.20% versus 10.50% at N=60-100.","The reported gap advantage grows at out-of-distribution sizes, reaching 19.35% versus 23.72% at N=1000 in greedy mode.","One USPR model (1.5 million parameters, 100 epochs, about 11 hours) is reported to replace ten CAMP models (17.6 million parameters, 1000 epochs, about 4.6 days)."],"supporting_citations":[{"why":"It supplies the PVRP-specific neural baseline CAMP, whose per-\\alpha retraining protocol is the inefficiency USPR is designed to remove.","marker":"Hua et al. 2025"},{"why":"It provides the HGS-PyVRP reference solver whose solution costs define the gaps reported in every benchmark table.","marker":"Wouda et al. 2024"},{"why":"It supplies the instance-generation settings and the attention encoder-decoder paradigm that the USPR architecture extends.","marker":"Kool et al. 2018"},{"why":"It contributes the REINFORCE-style policy-gradient training and the logit scaling constant C=10 used by the score reshaping.","marker":"Bello et al. 2016"},{"why":"It provides the symmetric augmentation rollouts used to form the shared baseline for variance-reduced policy gradients.","marker":"Kim et al. 2022b"},{"why":"It supplies the distance-aware attention reshaping heuristic that profile-aware score reshaping generalizes by adding preference scores.","marker":"Wang et al. 2024"},{"why":"It provides the parallel multi-pointer decoding and communication layer that USPR adopts for multi-vehicle action generation.","marker":"Berto et al. 2024a"},{"why":"It is one of the sequential multi-agent baselines that USPR is compared against and reported to outperform.","marker":"Son et al. 2024"}],"fun_headline_variants":["One model, any profile, best-in-class routing","USPR: train once, solve all profiled routing","Single solver for arbitrary profiled vehicle routing","Replace ten models with one USPR solver","Unified profiled routing, no retraining needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every benchmark, including the real-world PVRPLib set, generates profile scores from the same synthetic protocol — independent Uniform(0,1) draws with zero to ten percent of entries randomly set to plus or minus infinity — and the baselines are adapted through an unspecified shared recipe; if genuinely different preference distributions or a different adaptation protocol change the ranking, the unified-solver claim is weaker than stated.","fun_headline_variants_meta":{"raw":{"variants":["One model, any profile, best-in-class routing","USPR: train once, solve all profiled routing","Single solver for arbitrary profiled vehicle routing","Replace ten models with one USPR solver","Unified profiled routing, no retraining needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1312,"prompt_tokens":934,"completion_tokens":378,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":305}},"tokens_in":550,"tokens_out":378,"duration_ms":3782,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:13:26.350932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test set of PVRP instances whose profile scores come from a different distribution family, such as correlated preferences, ordinal categories, or real driver-customer ratings, and measure USPR's greedy gap to HGS-PyVRP; if the gap degrades sharply or falls behind CAMP on the same instances, the 'arbitrary profile types' unification claim would be refuted.","supporting_citations":[{"cited_title":"A.; Lan, L.; and Kool, W","cited_arxiv_id":null,"evidence_quote":"It provides the HGS-PyVRP reference solver whose solution costs define the gaps reported in every benchmark table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is one of the sequential multi-agent baselines that USPR is compared against and reported to outperform."}],"review_version":1}