{"id":"8a8dc03f-e61e-4dda-91e7-53fa48809514","arxiv_id":"2502.10233","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-agent reinforcement learning model with hierarchical parallel decoding achieves state-of-the-art min-max travel distance on mixed-shelves picker routing and generalizes to larger instances.","lead":"Researchers trained a neural network to plan routes for multiple warehouse pickers at once, minimizing the longest picker route instead of total distance. It outperforms existing heuristic and learned solvers, especially on large warehouses it has not seen during training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's MDP allows pickers to start new tours after returning to the packing station, but the MIP and HAM baselines are restricted to one tour per picker, so the reported SOTA may compare different problems.","rationale":"The reader's weakest assumption is that M=ceil(sum d_p/kappa) ties workforce to demand and limits real fixed-workforce applicability. I agree that is a practical limitation, but the more damaging issue is internal: the paper's own MDP explicitly allows a picker to revisit the packing station, restoring capacity, and to continue picking. Each such revisit creates a new tour, so a feasible MAHAM policy can use more than M tours. The Appendix A MIP, however, models B as a set of independent tours and minimizes max tour length; it has no decision variables linking tours to pickers, so it cannot represent the multi-tour-per-picker case. HAM is adapted under the statement 'there are exactly as many pickers as there are tours' (Appendix B), i.e., one tour per picker, and 2d-Ptr/Equity Transformer/PARCO likewise generate one route per agent. The neural-baseline comparison is therefore asymmetrical: MAHAM searches a larger solution space. Because extra tours can reduce the min-max objective by splitting a long assignment across pickers (at the cost of extra total travel, which the min-max objective does not penalize), the growing margins in Table 2 could be explained by this flexibility rather than by the hierarchical/parallel decoding contribution. For the same reason, the Gurobi comparison in Table 1 is not a comparison on the same problem. This is not a matter of external applicability; it is a correctness risk in the central empirical claim. The paper can be repaired either by constraining MAHAM to one tour per picker (making it consistent with the MIP and baselines) or by reformulating all baselines and the MIP to the multi-tour setting with per-picker distance aggregation; until one of these is done, the headline claim is unsubstantiated. Hence my verdict moves from the reader's CONDITIONAL toward REJECT for the current version, because the reported comparisons do not establish the claimed superiority.","tokens_in":15532,"tokens_out":19110,"duration_ms":223367,"concrete_test":"Instrument the released code's transition function to count, for every MAHAM solution in Tables 1 and 2, how many times each picker leaves the packing station after having already returned to it. If any picker performs more than one tour in a test instance, re-run the comparison with a 'one tour per picker' constraint (agents terminate at first station return) or, equivalently, reformulate the MIP to aggregate tour distances per picker and allow multi-tour assignment. If MAHAM's advantage over HAM, PARCO, and Gurobi shrinks or reverses, the reported SOTA is an artifact of comparing different problem formulations.","verdict_should_be":"REJECT","load_bearing_attack":"Section 3.1 defines an action that lets a picker visit the packing station, unload, restore capacity, and then continue; each agent's reward is total distance over its entire route. Thus a feasible MAHAM trajectory can contain more than M=ceil(sum_p d_p/kappa) station-to-station tours, and splitting long assignments across extra tours can lower the min-max objective. In contrast, the Appendix A MIP models exactly one independent station-to-station tour per element of B (constraints (16)-(18)) and minimizes the maximum tour length, with no per-picker aggregation, and the HAM baseline is adapted 'thanks to our assumption, that there are exactly as many pickers as there are tours' (Appendix B). The other neural baselines also construct one route per agent. Consequently Table 1 and Table 2 may measure a different, more constrained problem for the baselines, so the claimed consistency and growing margins may be an artifact of this extra degree of freedom rather than of the proposed hierarchical/parallel decoding. This is an internal inconsistency in the problem definition, not merely a practical limitation of the M=ceil assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAHAM, a multi-agent hierarchical attention model for the min-max Mixed-Shelves Picker Routing Problem (MSPRP). MAHAM extends the Hierarchical Attention Model (HAM) with a joint logit-space decoder that generates actions for all pickers in parallel, followed by a sequential action-selection scheme to guarantee feasibility. The model is trained with a self-improvement cross-entropy method. Experiments compare MAHAM against Gurobi, a greedy heuristic, HAM, Equity Transformer, 2d-Ptr, and PARCO on MSPRP instances with 10, 25, 40, and 50 shelves, reporting better solution quality and speed, especially on large out-of-distribution instances.","tokens_in":15694,"tokens_out":7393,"duration_ms":79916,"significance":"If the results were established, the paper would make a useful empirical contribution to neural combinatorial optimization for min-max multi-agent routing: the proposed joint logits plus sequential masking is a plausible way to obtain coordination without fully autoregressive decoding, and the ablation of ranking strategies is informative. The public code and the detailed description of the architecture are strengths. However, the central empirical claim is currently compromised by a mismatch between the problem solved by MAHAM and the problem solved by the MIP and the baseline methods, and by the lack of statistical reporting. With the problem definition aligned and repeated experimental runs, the contribution could be significant.","major_comments":[{"comment":"The MDP in Section 3.1 permits a picker to visit the packing station, unload, restore capacity, and continue, so the per-picker distance used in the reward can span multiple station-to-station tours. In contrast, the MIP in Appendix A models each element of B as exactly one independent tour (constraints (15)-(18), (20)) and minimizes the maximum tour length, with no aggregation of tours per picker. Appendix B states that HAM is adapted \"thanks to our assumption, that there are exactly as many pickers as there are tours.\" Consequently, Tables 1 and 2 compare MAHAM on a less constrained problem than the MIP and the one-tour-per-picker baselines, and the reported margins may reflect this extra degree of freedom rather than the proposed hierarchical and parallel decoding. Please align the problem definitions: either make the packing station a terminal action that ends an agent's episode (one tour per picker), or reformulate the MIP and the baselines to allow multiple tours per picker and compare against the same min-max total-distance objective.","section":"Section 3.1, Appendix A, Appendix B"},{"comment":"No repeated runs or error bars are reported. The Gurobi rows are described as solving a single test instance per configuration, and the neural baseline entries report only the best objective among 1280 sampled solutions, without standard deviations over instances or training seeds. Given that several margins in Table 1 are below 0.5% (e.g., MAHAM 0.01% vs PARCO 0.13% for MSPRP10 with |P|=3), the paper should report means and standard deviations over multiple test instances and multiple seeds, or otherwise the claim of consistent superiority is not statistically supported.","section":"Section 5.1, Table 1"},{"comment":"The large-scale generalization claims rest on a single comparison protocol in which each neural baseline is evaluated with 1280 sampled solutions. It is not stated how many test instances are used, whether the same instances are used for all methods, or whether the reported times are for the full sampling procedure or for a single greedy rollout. Please specify the evaluation protocol, including the number of test instances and the precise timing convention, so that the speed comparisons in Table 2 are reproducible.","section":"Table 2, Section 5.2"}],"minor_comments":[{"comment":"Decimal commas appear inconsistently: entries such as \"0,30s\" and \"0,25s\" use a comma, while other entries use a period; please use a single decimal separator throughout.","section":"Table 1"},{"comment":"The header for the second block reads \"SKUs\" while the first block uses \"|P|\"; please use a consistent label for the number of SKUs.","section":"Table 1"},{"comment":"The normalization expression for P_ma is typeset incorrectly (the fraction bar is missing); it should read P_ma = exp(L'_ma / beta) / suma over all agent-action pairs of exp(L'_ij / beta).","section":"Algorithm 1, Section 4.3"},{"comment":"The scale parameter C is introduced but its value or training scheme is never specified; please state how C is set in the experiments.","section":"Section 4.1, Equation (10)"},{"comment":"References [15] and [16] are duplicate entries for the same Matrix Encoding Networks paper; please merge them.","section":"References"},{"comment":"There are minor formatting issues: \"Figure3a\" and \"Figure3b\" lack spaces, and \"table 2\" is lowercase; please fix these for journal style.","section":"Section 5.3, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the problem mismatch between the MAHAM MDP and the MIP/baseline problem. If the authors restrict the MDP to one tour per picker, the experimental comparison becomes meaningful; alternatively, they must extend the baselines to the multi-tour setting. The lack of any repeated-run statistics also needs to be addressed for the empirical claims to be credible. The paper is otherwise a decent candidate after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is the first neural solver for the min-max MSPRP, and the architecture is a sensible combination of HAM's hierarchical decoder with PARCO-style parallel decoding, plus a rank-based agent encoder and a sequential action selection over joint logits. The empirical story is initially plausible: MAHAM beats the other neural baselines on all tested sizes and matches or beats Gurobi on larger instances where Gurobi times out. The out-of-distribution results on MSPRP50 are new. Code is public.\n\nThe real soft spot is not the missing error bars or the single-instance Gurobi comparison, though those also matter. The problem is that MAHAM and the baselines may not be solving the same problem. Section 3.1 lets a picker return to the packing station, unload, restore capacity, and keep going; an agent's total route is the sum of its tours. The MIP in Appendix A and the neural baselines (HAM, 2d-Ptr, ET, PARCO) construct exactly one tour per picker (or per element of B). Since M is set to the minimum number of tours required, MAHAM has the extra freedom to split a long assignment into multiple trips, which directly reduces the min-max objective. That extra degree of freedom could explain the growing margins in Tables 1 and 2 just as well as the proposed decoding scheme. The paper should either restrict MAHAM to one tour per agent or extend the MIP and baselines to allow multiple tours per picker. As written, the SOTA claim is not yet established.\n\nMinor issues: no variance statistics over seeds, and the Gurobi comparison uses a single test instance per configuration. The MIP's exponential subtour constraints are standard; the concern above is the mismatch.\n\nWho is this for? Researchers in neural combinatorial optimization and warehouse logistics. The method is a reasonable engineering contribution, and the problem is practical. But the load-bearing comparison needs to be redone.\n\nMy recommendation: send it to peer review, but the authors need to address the tour-count mismatch and report variance. If they fix that, it's a solid paper; as is, I'd be skeptical of the headline numbers.","headline":"First neural solver for the min-max MSPRP, but the headline comparison is suspect because MAHAM can split a picker's route into multiple tours while the MIP and neural baselines are restricted to one tour per picker.","tokens_in":16309,"tokens_out":3722,"would_cite":false,"duration_ms":37192,"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":"This paper claims that a multi-agent hierarchical attention model with parallel decoding and sequential action selection is the first neural solver to handle the min-max Mixed-Shelves Picker Routing Problem effectively, matching exact…","keywords":["mixed-shelves picker routing problem","min-max routing","multi-agent reinforcement learning","neural combinatorial optimization","parallel decoding","hierarchical decoder","warehouse logistics","sequential action selection"],"falsifier":"Run MAHAM on instances where the number of pickers is fixed below ceil(sum(dp)/kappa); if the min-max objective degrades sharply or feasibility breaks, the claimed practical relevance to real warehouses with fixed picker counts would fail. A second check would be finding a large instance where Gurobi provably reaches an optimal solution within its time budget and beats MAHAM.","tokens_in":15238,"feed_emoji":"📦","tokens_out":4202,"duration_ms":47037,"temperature":0.7,"pith_summary":"The paper proposes a learned solver, MAHAM, for the min-max Mixed-Shelves Picker Routing Problem, where several pickers in a mixed-shelves warehouse retrieve items and the goal is to minimize the longest tour. The authors claim that MAHAM coordinates pickers by generating a joint distribution over all agents' next actions in parallel, then resolving conflicts with a sequential action selection mechanism. They report that MAHAM matches Gurobi's optimal solutions on small instances and beats Gurobi on larger instances where the exact solver cannot find an optimum within an hour. If true, this would be the first neural method for the min-max variant and a practical option for real-time, large-scale warehouse routing.","feed_headline":"Neural picker-routing solver beats Gurobi on large warehouses","feed_subtitle":"MAHAM coordinates multiple pickers in parallel and solves large, unseen instances in seconds.","key_machinery":"The load-bearing components are a heterogeneous-graph encoder with shared cross-attention parameters, an agent context encoder with ranking-based positional encoding and multi-head self-attention among pickers, and a hierarchical decoder that first selects shelf actions and then SKU actions. The sequential action selection samples one agent-action pair at a time from a single normalized distribution over the joint logit space, so the policy learns which picker should act first and avoids two agents selecting the same shelf-SKU combination. Training uses a self-improvement scheme that samples many candidate solutions from the current best policy, keeps the best one as a pseudo-optimal demonstration, and trains with cross-entropy loss.","core_discovery":"The central claim is that MAHAM solves the min-max MSPRP by combining a hierarchical decoder with parallel decoding over all pickers, so that a single shared policy can balance workload between agents without the high latency of purely autoregressive construction. The joint logits over shelf and SKU choices are decoded through a learned sequential action selection, which masks infeasible actions and implicitly ranks agents by priority. The paper reports state-of-the-art objective values among neural baselines across all tested warehouse sizes, with gaps to the best known solutions below 1.7% on in-distribution instances, and shows large-scale generalization to 50-shelf instances with up to 500 SKUs that Gurobi cannot solve within an hour.","pith_inferences":["Editorial inference: the joint-logit sequential selection strategy could be ported to other multi-agent problems with composite action spaces and hard collision constraints, such as multi-robot task allocation, where parallel decoding currently relies on hand-built conflict handlers.","Editorial inference: because the number of pickers is set equal to the minimal number of tours implied by demand and capacity, the method implicitly assumes flexible staffing; adapting it to a fixed workforce would require treating the number of pickers as an input and likely a different reward design.","Editorial inference: the reported out-of-distribution robustness suggests that learning an agent ordering from logits is a transferable skill, and a testable prediction is that the learned ranking would generalize across different warehouse layouts even if the rest of the decoder is retrained.","Editorial inference: the self-improvement training loop selects the best of 100 sampled trajectories, so the reported margins over baselines may depend on the quality of the initial reference policy and the sampling budget."],"forward_implications":["MAHAM matches Gurobi's optimal solutions on MSPRP10 and beats Gurobi's one-hour results on MSPRP25 and MSPRP40, with gaps below 1% on most in-distribution settings.","Large-scale generalization tests show MAHAM finding solutions in seconds for 50-shelf, 500-SKU instances where Gurobi finds nothing in an hour, with objective gaps of 4% or less against the best known solutions.","Parameter sharing in the cross-attention layers cuts model size by about 20% and improves both inference speed and out-of-distribution generalization.","The sequential action selection avoids the conflicts that PARCO incurs, so MAHAM needs fewer decoding steps and trains faster than the alternative parallel method.","The first neural solver for the min-max variant of the MSPRP provides a concrete alternative to hand-crafted heuristics for mixed-shelves warehouse operations."],"supporting_citations":[{"why":"Supplies the hierarchical decoder and heterogeneous-graph encoder architecture that MAHAM extends from single-picker to multi-picker settings.","marker":"[18]"},{"why":"Provides the PARCO baseline and the concept of parallel solution construction with conflict handling that MAHAM compares against and improves upon.","marker":"[2]"},{"why":"The attention model decoder whose cross-attention mechanism MAHAM adapts to generate joint logits over agent actions.","marker":"[11]"},{"why":"Supplies the self-improvement training method that MAHAM uses to learn from pseudo-optimal sampled solutions.","marker":"[19]"},{"why":"Provides the 2d-Ptr baseline, a purely autoregressive multi-agent pointer network that MAHAM outperforms in speed and quality.","marker":"[17]"},{"why":"Provides the Equity Transformer baseline for min-max routing problems, which MAHAM outperforms.","marker":"[20]"}],"fun_headline_variants":["Parallel decoding makes neural picker routing scale and beat Gurobi","Hierarchical and parallel decoding: neural picker routing beats Gurobi","MAHAM: state-of-the-art neural solver for mixed-shelves picker routing","Parallel decoder handles 500 SKUs, beats Gurobi in seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The number of pickers is assumed to equal the minimum number of tours required by the demand and capacity, so the method only applies when staffing can be flexibly matched to the workload rather than being fixed in advance.","fun_headline_variants_meta":{"raw":{"variants":["Parallel decoding makes neural picker routing scale and beat Gurobi","Hierarchical and parallel decoding: neural picker routing beats Gurobi","MAHAM: state-of-the-art neural solver for mixed-shelves picker routing","Parallel decoder handles 500 SKUs, beats Gurobi in seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001107,"raw_usage":{"total_tokens":4569,"prompt_tokens":851,"completion_tokens":3718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":3637}},"tokens_in":467,"tokens_out":3718,"duration_ms":25184,"temperature":1.0,"reasoning_tokens":3637,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T18:51:20.245763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MAHAM on instances where the number of pickers is fixed below ceil(sum(dp)/kappa); if the min-max objective degrades sharply or feasibility breaks, the claimed practical relevance to real warehouses with fixed picker counts would fail. A second check would be finding a large instance where Gurobi provably reaches an optimal solution within its time budget and beats MAHAM.","supporting_citations":[{"cited_title":"In: Pro- ceedings of the International Conference on Automated Planning and Scheduling","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical decoder and heterogeneous-graph encoder architecture that MAHAM extends from single-picker to multi-picker settings."},{"cited_title":"net/forum?id=ByxBFsRqYm Hierarchical and Parallel Decoding for Picker-Routing 15","cited_arxiv_id":null,"evidence_quote":"The attention model decoder whose cross-attention mechanism MAHAM adapts to generate joint logits over agent actions."},{"cited_title":"Transactions on Machine Learning Research (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the self-improvement training method that MAHAM uses to learn from pseudo-optimal sampled solutions."},{"cited_title":"In: Proceedings of the 23rd International Conference on Autonomous Agents and Mul- tiagent Systems","cited_arxiv_id":null,"evidence_quote":"Provides the 2d-Ptr baseline, a purely autoregressive multi-agent pointer network that MAHAM outperforms in speed and quality."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"Provides the Equity Transformer baseline for min-max routing problems, which MAHAM outperforms."}],"review_version":1}