{"id":"30753671-d9a1-4777-9e40-95d8baaeb87b","arxiv_id":"2501.02977","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CAMP is a new attention-based multi-agent RL solver for vehicle routing with per-client profiles, outperforming prior neural baselines on both preference and zone-constrained variants.","lead":"This paper introduces CAMP, a multi-agent reinforcement learning model for profiled vehicle routing, a problem in which each vehicle has preferences or zone restrictions for individual clients. In tests, CAMP runs in under a second and beats other neural solvers, while remaining a few percent behind the leading classical heuristic.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-seed comparisons make the 'consistently outperforms all neural baselines' claim unsupported; the observed margins are small and no variance is reported.","rationale":"The reader identified reward indexing as the weakest assumption, but that is not the load-bearing issue. In Eq. (12) and Eq. (21), p_{ik} is attached to the tail node i of arc (i,j). In any feasible route starting and ending at the depot, every client appears exactly once as a tail and once as a head, so summing p_{ik} over outgoing arcs yields exactly the total preference of all served clients, identical in value to a destination-based indexing. The formal objective defined in Eq. (1) uses tail indexing, so the reward is internally consistent with the stated problem, provided the code follows the equations. The real weak point is statistical: the headline superiority is a universal claim over many sizes and configurations, but every number in Table 1 is a single-run point estimate. Differences of 1-3% against strong neural baselines are plausible training noise for one-seed NCO runs; the paper itself shows CAMP can lose to CAMP(-EC) on some cells, indicating the evaluated improvements are not large compared to run-to-run variation. Therefore the claim should remain conditionally accepted, pending multi-seed validation. This matches the reader's CONDITIONAL verdict, though for a different primary reason; I would not require a change in the reward equation unless code inspection reveals a mismatch with Eq. (21).","tokens_in":18269,"tokens_out":10400,"duration_ms":94407,"concrete_test":"Retrain CAMP and the strongest neural baselines, at least PARCO and 2D-Ptr, with 5-10 different seeds on PVRP-P and PVRP-ZC at N=100 with m=5 and m=7, keeping the same instance set and evaluation protocol. Report mean and standard deviation of the adjusted cost over seeds, and run a paired test such as Wilcoxon signed-rank over the 1280 test instances comparing CAMP to each baseline. If any baseline is statistically indistinguishable from or better than CAMP, replace 'consistently outperforms' with 'competitive with' in the abstract and Section 5.2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (Section 5.2) is that CAMP consistently outperforms all other neural solver baselines across every client/vehicle size for both PVRP-P and PVRP-ZC. The evidence is a single table of point estimates: each model is trained once, and each cell reports one greedy/sampling number with no standard deviation, no number of seeds, and no significance test. The margins over the strongest neural baselines are often small, for example PVRP-P N=80, m=5: CAMP(s.) 8.57 vs PARCO(s.) 8.66, about 1%, and PVRP-ZC N=100, m=5: CAMP(s.) 19.26 vs 2D-Ptr(s.) 19.72, about 2.4%. At N=60, m=3 in PVRP-ZC, CAMP(s.) is 12.54, slightly worse than its own ablation CAMP(-EC)(s.) at 12.49, showing that even the value of the full architecture is not stable at the reported precision. Because the universal superiority claim is the central empirical contribution, a single seed per baseline is insufficient; the apparent consistency could be within training noise. This is a correctness risk, not a disagreement with community consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Profiled Vehicle Routing Problem (PVRP), a generalization of the heterogeneous capacitated VRP in which each vehicle k carries a per-client profile p_i^k that either adds a soft preference term (PVRP-P) or imposes a hard zone constraint (PVRP-ZC). The authors give an ILP formulation (Eqs. 1-11), an MDP reformulation (Section 4.1), and propose CAMP, an attention-based construction method that maintains a separate client embedding per vehicle profile, refines them with bipartite-graph message passing in the encoder, lets agents exchange messages at each decoding step through a transformer communication layer, and emits actions for all vehicles in parallel via a batched pointer with priority-based conflict resolution. Training uses REINFORCE with a SymNCO-style shared baseline and a reward-balancing normalization across preference distributions. The method is benchmarked on PVRP-P and PVRP-ZC at N=60/80/100 with m=3/5/7 vehicles against OR-Tools, HGS-PyVRP, and four neural baselines (ET, DPN, 2D-Ptr, PARCO). Section 5.2 claims CAMP consistently outperforms all neural baselines at every configuration; the reported inference times are sub-second (0.17-0.54 s), with a roughly 3-5% gap to HGS-PyVRP.","tokens_in":18577,"tokens_out":26441,"duration_ms":242461,"significance":"If the headline empirical claim holds, CAMP is a solid first learner for a practical VRP variant: it is, to the best of the paper's own survey, the first construction method designed for per-client vehicle profiles, it reaches about 3-5% gap to a state-of-the-art heuristic (HGS-PyVRP) in sub-second inference while beating OR-Tools under time limits, and the four ablations give credible evidence that the profile-specific embeddings, encoder communication, and reward balancing each contribute. The paper earns credit for matched training budgets across baselines, an external anchor in OR-Tools/HGS-PyVRP (which prevents the benchmark from being entirely self-defined), honest reporting that HGS remains superior, and a public code release. The principal weakness is statistical: the central claim is supported only by single point estimates per model, with the decisive margins often near 1-2.4%; two table cells even show the full model slightly worse than its own ablation. The contribution's value for the community depends on whether the headline margin exceeds training noise, so the empirical claim needs variance or significance evidence before the comparison can be taken at face value.","major_comments":[{"comment":"The central claim that 'CAMP consistently outperforms all other neural solver baselines in experiments across all client and vehicle sizes' is based on single point estimates: no number of training seeds, no standard deviation, and no significance test are reported anywhere in Section 5. I verified that at the point-estimate level CAMP (both greedy and sampling) beats ET, DPN, 2D-Ptr, and PARCO in every one of the 18 size/vehicle configurations, but several margins are small (e.g., PVRP-P N=80 m=5 sampling: 8.57 vs 8.66, about 1%; PVRP-ZC N=100 m=5 sampling: 19.26 vs 19.72, about 2.4%). Two cells in Table 1 go the other way against the paper's own ablation CAMP(-EC): PVRP-ZC N=60 m=3 sampling (CAMP 12.54 vs CAMP(-EC) 12.49) and PVRP-ZC N=60 m=7 greedy (CAMP 13.14 vs CAMP(-EC) 13.05), which shows that differences at the reported precision are comparable to training noise. Since the headline claim is the paper's main empirical contribution, the authors should report the number of seeds and mean plus/minus standard deviation for the main models (or provide paired significance tests), or explicitly soften the claim to the level of evidence actually provided.","section":"Table 1 / Section 5.2"},{"comment":"The aggregation that produces the PVRP-P numbers in Table 1 and Figure 3 is under-specified. The text says results are averaged over 'settings of alpha ranging from 0.0 to 0.2' but does not state the alpha values or their count, does not say which preference distributions (Random, Angle, Cluster, and/or Zone) enter the average, and does not state whether the tabulated 'cost' is the combined objective (duration minus alpha times preference) or the duration component alone. Because the ILP objective (Eq. 1) and the reward (Eq. 21) are linear in alpha for a fixed route, the reported gaps to HGS-PyVRP and the Pareto curves in Figure 3 depend on this grid, and different plausible choices give different averages. The same paragraph must specify how OR-Tools and HGS-PyVRP were given the preference-dependent objective for each alpha (e.g., per-vehicle cost matrices) and how the reported 'Gap(%)' was computed for the PVRP-P rows.","section":"Section 5.1"},{"comment":"Several load-bearing pieces of the architecture are not defined precisely enough to reimplement from the text. The message-passing function Phi in the bipartite encoder is never defined, even though Eqs. (26)-(28) describe the rest of the encoder in detail. In Eq. (27), h^k_k and h^k_i are used without defining which index is the vehicle and which is the current node, and the notation is not time-indexed. In Eq. (30), the matrix L is said to be a projection of the node embeddings h, but h is defined as a list of m profile-specific tensors h^k in R^{(m+n) x d_h}; the dimensions do not line up for the batched pointer as written. Finally, the factorization in Eq. (25) conditions each vehicle's action only on its own history, while the decoder (Eqs. 27-28) uses cross-vehicle communication; the factorization and the procedure should be made consistent. The open-source release mitigates this, but a methods paper should be self-contained.","section":"Section 4.2.2-4.2.3"}],"minor_comments":[{"comment":"The reward in Eq. (21) (and the objective in Eqs. (1) and (12)) credits p_i^k on edge (i,j), i.e., the preference of the departure node, although p_i^k is defined as the preference for serving client i. I checked whether this is a substantive error and concluded it is not: in a closed route the sum of departure-node preferences over all edges equals the sum of the served-client preferences over all visited clients (the two sums contain the same terms, since each client appears exactly once as a departure node), provided p_0^k = 0 for the depot. The authors should still fix the notation, defining p_0^k = 0 and either re-indexing to the destination client or adding a sentence noting the equivalence, because as written the reward appears to credit the wrong node, and Eq. (19) correctly uses the destination index p_j^k for the zone constraint, making the inconsistency visible.","section":"Eq. (21)"},{"comment":"Notation slips in the ILP: Eq. (13) writes y_ik without the trip and vehicle indices used in the definition y^r_ik in Section 3.1, and Eq. (9) uses a single subtour-elimination variable w_i per node although the constraint is imposed separately for each trip r; w should be indexed by r (or the trips removed) for the formulation to be correct.","section":"Eqs. (13) and (9)"},{"comment":"The abstract's claim that 'no learning method exists to solve the more practical and challenging PVRP' is stronger than what the paper demonstrates: Section 5 trains and benchmarks four learning baselines on PVRP instances, so the intended claim must be that no learning method has been proposed specifically for PVRP. The related-work section also omits the closest problem neighbors (e.g., colored TSP variants and other per-agent-node-constrained routing problems), which should be discussed to scope the novelty claim.","section":"Abstract"},{"comment":"Section 5.3 states that the '- Vehicle-specific Profile Embedding' ablation 'corresponds to PARCO,' but Table 2 reports a duration gap of 6.85% for this ablation on PVRP-P N=100, while Table 1 lists PARCO's average PVRP-P gaps as 13.21% (greedy) and 6.44% (sampling); neither value matches, so the basis of Table 2's gap numbers (and the definition of the 'preference gap' column) must be stated.","section":"Table 2"},{"comment":"In the REINFORCE estimator, B is defined but the summation index L (presumably the number of augmented solutions or samples per instance) is not; the notation should be defined in the text.","section":"Eq. (32)"},{"comment":"Minor wording issues: the communication layer in Eq. (28) is said to capture 'intra-vehicle dynamic relationships' but it operates between agents and should read 'inter-vehicle'; the encoder text uses 'client' and 'customer' interchangeably; and Eq. (12) drops the trip superscripts on x relative to Eq. (1).","section":"Section 4.2.3"},{"comment":"The Pareto plot does not state the set of alpha values used, and the panel labeled 'Average Preference Distribution' is not defined; the axis quantities ('Duration Gap', 'Preference Gap') also need explicit definitions, as neither gap measure is defined in the text.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper comes from the same group that developed PARCO, SymNCO, and RouteFinder; PARCO is both a baseline and the architectural predecessor of CAMP (parallel autoregressive decoding, communication layer, priority-based conflict handler), and the incremental contribution consists of profile-specific client embeddings, bipartite encoder message passing, and the PVRP benchmark itself. That is a reasonable conference-scale increment, but the abstract's 'no learning method exists' novelty claim should be substantially scoped. My main reservation is statistical: the headline comparison rests on single seeds with small margins, and I would want to see seed variance or significance evidence before taking the 'consistently outperforms' claim at face value. If the authors supply that evidence (or temper the claim), I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading this paper. First, the new problem formulation (PVRP) and the model are reasonable: per-vehicle profile embeddings, bipartite message passing, a communication layer, and an ablation study that gives some credit to each component. Second, the headline claim that CAMP consistently beats every neural baseline is not actually supported by the table, because every number is a single run with no variance, and some margins are tiny.\n\nWhat's new: they name the Profiled VRP, give two variants (preferences and zone constraints), and propose CAMP, built directly on PARCO. They compare against ET, DPN, 2D-Ptr, PARCO, plus OR-Tools and HGS-PyVRP. The code is open, the sweep across N and m is broad, and the limitations section honestly admits they do not beat HGS. That is real engineering, credit where due.\n\nThe soft spots are real, in increasing order of severity. Minor: the aggregation over alpha (0.0 to 0.2) is underspecified—each cell reports one 'cost' number, but how the preference objective is folded in is not clear. More serious: Eq. (21) credits preference p_i k on edge (i,j), i.e., the departure node, rather than p_j k for the client being served. If the code follows the equations, the reward signal is misaligned with the stated objective. This might be a typo, but the authors need to clarify.\n\nThe load-bearing issue is the complete absence of error bars. Table 1 shows one number per cell, no seeds, no standard deviation. The margins over the strongest baselines are often around 1% (e.g., PVRP-P N=80, m=5: 8.57 vs 8.66 for PARCO sampling), and in PVRP-ZC N=60 m=3, the full CAMP(s.) at 12.54 is worse than its own ablation CAMP(-EC)(s.) at 12.49. That undercuts the universal superiority claim—the 'consistently' in Section 5.2 is not justified at this precision. This is a correctness risk, not a disagreement with community norms.\n\nAlso, the relationship to site-dependent VRP (SDVRP) is not discussed. PVRP-ZC with hard constraints is essentially a site-dependent VRP, and the paper should position itself against that literature before asserting that no learning method exists.\n\nWho this helps: researchers in neural combinatorial optimization who want a baseline for multi-agent routing with heterogeneous client preferences. It deserves a serious referee, but the referee should ask for multi-seed statistics, a clarified reward indexing, and a discussion of SDVRP. My recommendation: send to peer review, but with expectation of revision.","headline":"CAMP is a plausible and well-engineered solution to a newly named problem, but the central 'consistently outperforms' claim rests on single-seed numbers and the reward indexing looks wrong.","tokens_in":19087,"tokens_out":3022,"would_cite":true,"duration_ms":77897,"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":"CAMP, a collaborative attention model with profiles, is claimed to be the first learning-based solver for the Profiled Vehicle Routing Problem, beating existing neural baselines on both preference and zone-constrained variants.","keywords":["profiled vehicle routing problem","multi-agent reinforcement learning","attention model","neural combinatorial optimization","heterogeneous fleet routing","preference constraints","parallel decoding","route construction"],"falsifier":"Inspect the released code's reward calculation: if the preference term $\\alpha p_{ik}$ in Eq. (21) is computed with $i$ equal to the current node on each route edge rather than the destination client $j$, then the reported PVRP-P rewards do not match the stated objective, and a corrected implementation would change both training and evaluation results.","tokens_in":18095,"feed_emoji":"🚚","tokens_out":4684,"duration_ms":45488,"temperature":0.7,"pith_summary":"This paper introduces the Profiled Vehicle Routing Problem (PVRP), where each vehicle has client-specific preferences or constraints, and proposes CAMP, a multi-agent reinforcement learning solver that constructs routes for all vehicles in parallel. The paper claims CAMP is the first learning method for PVRP and that it consistently outperforms all neural solver baselines on both PVRP with preferences and PVRP with zone constraints, across all tested client and vehicle counts, while running in well under a second. A sympathetic reader would care because real fleets are heterogeneous: vehicles differ in access rights, driver familiarity, and regulatory permissions, and a learned solver that respects those profiles in real time could replace hand-tuned heuristics for routing.","feed_headline":"Profile-aware neural solver beats prior learners on PVRP","feed_subtitle":"CAMP builds routes for every vehicle in parallel, respecting client preferences, and runs in under a second.","key_machinery":"CAMP's core mechanism is a profile embedding that concatenates vehicle, client, and preference features into a combined representation for every vehicle-client pair, processed by multi-head attention. A bipartite graph message-passing step then lets vehicle and client embeddings exchange information across all profiles, so each vehicle's representation reflects the whole instance. During decoding, a transformer-based communication layer performs self-attention among vehicle queries at every step, and a multiple pointer mechanism produces logits for every vehicle-node pair in parallel; conflicts are resolved by prioritizing the vehicle with the highest action probability. This machinery is what allows heterogeneous vehicle profiles to be encoded and decoded jointly rather than sequentially.","core_discovery":"The central claim is that CAMP consistently outperforms all other neural solver baselines in experiments across all client and vehicle sizes for both PVRP-P and PVRP-ZC, while also beating a classical constraint-programming solver in solution quality and achieving inference times under a second. The paper further claims that no learning method existed for the PVRP before CAMP, making it the first learned approach to handle vehicle profiles on a per-client basis. The gains are attributed to three design choices: vehicle-specific profile embeddings in the encoder, a communication layer that lets agents share decisions during decoding, and a batched pointer mechanism that evaluates all vehicle actions in parallel.","pith_inferences":["A natural extension implied by the paper is testing whether the profile-embedding communication layer transfers to other heterogeneous combinatorial optimization problems, such as vehicle routing with time windows, skills, or compatibility constraints.","The reported gains are shown on synthetic preference distributions; a stronger test would evaluate zero-shot generalization to instance sizes and preference patterns not seen during training.","The reward function in Eq. (21) credits the departure node's preference on each edge; if the intended preference is for the served client, the trained policy may be optimizing a slightly different objective, and re-running with the destination-client preference would settle whether the quality comparison is robust.","The conflict-resolution rule that chooses the highest-probability vehicle is a fixed heuristic; learning a non-greedy conflict handler or adding an improvement step after construction could close the remaining gap to the best heuristic solver."],"forward_implications":["CAMP constructs routes for all vehicles in parallel, so inference time stays near-constant as fleet size grows, enabling real-time re-routing in dynamic settings.","Because the encoder embeds each vehicle profile separately, the same trained model can handle fleets of different sizes and client-specific preference patterns without retraining on fixed vehicle counts.","On the zone-constrained variant, invalid actions are masked in the environment, so the learned policy never proposes a route where a vehicle serves a prohibited client.","The reward-balancing scheme lets one policy handle multiple preference distributions, such as random, angle-based, cluster-based, and zone-based preferences, without favoring one reward scale.","If the empirical results hold, a learned construction method can outperform a classical constraint-programming solver in quality while running in fractions of a second, though a stronger heuristic solver that takes minutes per instance still has the best reported solutions."],"supporting_citations":[{"why":"Supplies the parallel autoregressive decoding scheme, communication layer, and conflict-resolution strategy that CAMP builds on and compares against.","marker":"[4]"},{"why":"Provides the attention-based encoder-decoder architecture that CAMP adapts to multi-profile vehicle routing.","marker":"[32]"},{"why":"Provides the symmetric-transform shared baseline used for the REINFORCE policy-gradient training.","marker":"[29]"},{"why":"Supplies the REINFORCE gradient estimator that is the training algorithm of CAMP.","marker":"[63]"},{"why":"One of the sequential multi-agent neural baselines that CAMP claims to outperform.","marker":"[54]"},{"why":"Another sequential neural baseline that CAMP compares against and claims to beat.","marker":"[72]"},{"why":"A dual-decoder neural method for the heterogeneous capacitated VRP that is a key baseline in the comparison.","marker":"[44]"},{"why":"The heuristic solver whose reported solutions define the target gap in the benchmark tables.","marker":"[64]"},{"why":"The classical constraint-programming baseline that CAMP claims to outperform in solution quality.","marker":"[51]"}],"fun_headline_variants":["CAMP: first learned solver for profiled vehicle routing","Attention-based RL beats prior neural baselines on PVRP","Profile-aware vehicle routing solved in parallel by CAMP","CAMP solver improves PVRP solutions with multi-agent RL","Fast neural routing with per-client profiles: CAMP wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reward function used to train and evaluate the model is assumed to implement the PVRP objective correctly; if the preference term $p_{ik}$ on each edge $(i,j)$ rewards the departure node rather than the served client, the learned policy optimizes a different objective.","fun_headline_variants_meta":{"raw":{"variants":["CAMP: first learned solver for profiled vehicle routing","Attention-based RL beats prior neural baselines on PVRP","Profile-aware vehicle routing solved in parallel by CAMP","CAMP solver improves PVRP solutions with multi-agent RL","Fast neural routing with per-client profiles: CAMP wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3185,"prompt_tokens":918,"completion_tokens":2267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2185}},"tokens_in":534,"tokens_out":2267,"duration_ms":17205,"temperature":1.0,"reasoning_tokens":2185,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:59:20.687734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code's reward calculation: if the preference term $\\alpha p_{ik}$ in Eq. (21) is computed with $i$ equal to the current node on each route edge rather than the destination client $j$, then the reported PVRP-P rewards do not match the stated objective, and a corrected implementation would change both training and evaluation results.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"One of the sequential multi-agent neural baselines that CAMP claims to outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another sequential neural baseline that CAMP compares against and claims to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The heuristic solver whose reported solutions define the target gap in the benchmark tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The classical constraint-programming baseline that CAMP claims to outperform in solution quality."}],"review_version":1}