{"id":"6fab660b-650a-4a1d-b69d-dd82d89c9bc6","arxiv_id":"2506.06291","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A GNN trained with behavior cloning and RL provides MILP warm starts that reduce solve time for multi-agent task allocation, but quality trails deadline-based heuristics and the evaluation is limited.","lead":"The authors train graph neural networks with behavior cloning and reinforcement learning to generate starting solutions for a mixed-integer linear program that schedules construction robots. In a small simulated setting, these learned starting points cut solver time compared with starting from scratch, though simple deadline heuristics still produced better-quality schedules.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quality metric and optimality protocol contradict: if solvers run to optimal, final quality must be equal across methods; Figure 4's differences imply a fixed budget or initial-solution quality, leaving 'maintains quality' undefined.","rationale":"The paper's central claim is that the BC+RL warm-start reduces optimization time and variance while maintaining quality and feasibility. The reader's weakest assumption concerned instance representativeness and solver determinism, which are legitimate but secondary. My review identifies a more direct problem: the evaluation protocol appears internally inconsistent. The text claims solvers are run to optimality and time to optimal is recorded, yet Figure 4 reports quality differences across methods, which cannot occur if all runs reach the same optimal solution. This contradiction means the reported quality scores are either computed on a fixed-budget incumbent or on the initial warm-start solution, and the paper does not state which. In neither case is the headline claim well-defined: if fixed-budget, then 'time to optimal' is not actually measured; if initial-solution quality, then 'maintaining quality' refers to the learned schedule, not the solver output, and needs a different evaluation framing. Additionally, Section IV.B's statement that BC RL 'showed the most variability' directly contradicts the abstract's variance-reduction claim and Figure 3's caption. While the qualitative direction is plausible and consistent with prior warm-starting literature, the reported evidence as written cannot support the claim without clarifying the metric and reporting raw per-instance data. The reader's conditional verdict is appropriate, but the deeper metric ambiguity pushes toward unverdictable pending clarification and artifact release.","tokens_in":5534,"tokens_out":3528,"duration_ms":39882,"concrete_test":"Obtain or regenerate the 100 instances and record for each of the 10 repeats: initial warm-start quality (Equation 2 on the initial assignment), final incumbent quality after the solver terminates, solve time to proven optimality, and whether any instance hit a time limit. Then verify that final quality is identical across methods on every instance, and recompute Figures 3 and 4 from these logs. If final quality differs across methods, the paper's quality metric is not final optimal quality, and the 'maintaining quality' claim is undefined. If no time limit was used, the reported quality differences indicate a metric inconsistency that invalidates the stated comparison.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing weakness is that the evaluation protocol does not define what 'quality score' measures, and the two reported metrics are mutually inconsistent. Section III.E says the MILP solver is run with a warm start 'to achieve the optimal solution' and optimization time is recorded. If every run terminates at the same proven optimum, the final schedule quality (which depends only on makespan and feasibility) must be identical across methods; the large quality differences in Figure 4 would then be impossible. The only consistent readings are (a) runs hit a time limit and quality is measured on the incumbent, or (b) 'quality score' is computed on the initial warm-start schedule before solving. In reading (a), the reported 'time to optimal' is not the actual metric, so the time comparison is incommensurable; in reading (b), the headline claim of 'maintaining solution quality and feasibility' refers to the GNN's raw schedule, not the solver output, and no text or equation states this. Either way, the central claim cannot be checked from the paper as written. Supporting evidence of the same measurement problem: Section IV.B says BC RL 'showed the most variability,' directly contradicting Figure 3's 'stable performance' and the abstract's variance-reduction claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a learning-based warm-starting framework for MILPs in a multi-agent task allocation and scheduling domain. A GNN is trained first by behavior cloning on solver-generated solutions and then fine-tuned with reinforcement learning; the resulting schedules are fed as initial solutions to an off-the-shelf MILP solver. The authors report experiments on 100 randomly generated instances with 10 agents and 20 tasks, comparing a baseline solver without warm start, EDF, constraint-aware EDF, BC, and BC+RL. They claim that the learning-based warm starts reduce optimization time and variance while maintaining solution quality and feasibility.","tokens_in":5724,"tokens_out":3301,"duration_ms":36835,"significance":"If the reported results are correct, the paper would provide a practically relevant demonstration that a GNN-based warm start can accelerate a MILP solver on temporally constrained task allocation problems while remaining compatible with off-the-shelf solvers. The focus on a construction-inspired environment with heterogeneous agents and precomputed motion paths is timely, and the comparison against EDF-style heuristics is sensible. However, the current manuscript does not provide sufficient numerical detail or a consistent evaluation protocol for the central claims to be verified; the significance is therefore conditional on a substantive revision of the experimental reporting.","major_comments":[{"comment":"The evaluation protocol is internally inconsistent. Section III.E states that the MILP solver is initialized with a warm start 'to achieve the optimal solution' and that 'optimization time required to achieve the optimal solution is recorded.' If every run terminates at the same proven optimum, then the final schedule quality—feasibility and makespan—should be identical across methods, yet Figure 4 reports clear quality-score differences between Baseline, EDF, Constraint-Aware EDF, BC, and BC RL. The only consistent readings are that runs hit a time limit and quality is measured on the incumbent, or that the 'quality score' is computed on the initial warm-start schedule before solving. Either interpretation must be stated explicitly, and the corresponding time metric (time-to-optimal versus time-to-budget) must be defined accordingly. As written, the central claim that the method 'maintains solution quality and feasibility' cannot be checked.","section":"III.E and Figure 4"},{"comment":"The experimental reporting lacks the numerical detail and statistical validation needed to support the variance-reduction claim. No actual means, standard deviations, confidence intervals, or hypothesis tests are reported; the reader only sees bar charts. Moreover, the text and captions directly contradict each other: the Figure 3 caption describes BC RL as achieving 'stable performance,' while Section IV.B states that 'BC RL showed the most variability.' Since variance reduction is an explicit part of the abstract's claim, this contradiction must be resolved with concrete numbers and appropriate statistical tests (e.g., paired tests across the same 100 instances).","section":"IV.A and Figures 3-4"},{"comment":"The RL fine-tuning component, which is a key part of the proposed framework, is not specified to a reproducible level. Equation (2) defines a quality reward and Equation (3) defines a time reward, but no combination weight or scalarization rule is given, no RL algorithm is named (e.g., PPO, policy gradient), and no training hyperparameters are reported. The authors state in Section III.D that 'the design of this reward structure is still being explored,' which further weakens the claim that BC RL is a complete proposed method rather than a preliminary experiment. Since Section V concedes that RL fine-tuning provides only 'modest gains' and 'lacks consistency,' the paper should either report a fully specified training pipeline and ablation over the reward combination, or frame BC RL as a preliminary study.","section":"III.D, Equations (2)-(3)"},{"comment":"The generality of the conclusions is not supported by the experimental scope. All results are on a single problem size (10 agents, 20 tasks), with 100 randomly generated instances whose generation distribution is not described (e.g., how obstacles, time windows, and order constraints are sampled, and whether the test instances are drawn from the same distribution as the training instances). The authors themselves list 20-agent/100-task scalability as future work, so the abstract's claim about 'large-scale, real-time scenarios' is not established. The paper should either restrict the claims to the evaluated setting or add a larger-scale experiment.","section":"IV.A and V"}],"minor_comments":[{"comment":"The MILP formulation is incomplete as written: the cost function f(A, S_1, ..., S_NA) is never defined, and several symbols in the constraints (e.g., A_ij, S^i_jk, t^T_ijk, O_jk, W_jk, s_k, e_k, tE_ik) are not formally introduced in the text. Adding a notation table or clear definitions would greatly improve reproducibility.","section":"III.C, Equation (1)"},{"comment":"The quality score is described as ranging from 1 to 21 and as 'combining feasibility and makespan relative to deadlines,' but the relationship between this description and Equation (2) is not made explicit; in particular, it is unclear how infeasible assignments are penalized in the reported score. Please reconcile the metric definition used in the experiments with the reward formula.","section":"IV.A"},{"comment":"The sentence 'The lack of significant improvement in the optimization time for BC and BC+RL may be due to this' uses the word 'significant' without any statistical test being reported; either report significance tests or rephrase as 'substantial improvement.'","section":"IV.B"},{"comment":"The figures do not show units for optimization time, and there are no numerical labels on the bars or standard-deviation whiskers. Adding a table with the exact mean and standard deviation for each method would make the results auditable.","section":"Figures 3-4"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work (e.g., references [5], [6], [9], and [19]); this is not by itself a problem, but the novelty relative to [9] in particular should be clarified explicitly in the revision. There is also no statement about code or data availability, which would be valuable given the current lack of numerical tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one-sentence take: a BC+RL-trained GNN warm-starts a MILP solver for a 10-agent/20-task construction scheduling problem and cuts solve time; the direction is unsurprising given Nair et al. and related work, and the novelty is the specific application setting and the EDF comparison, not the mechanism.\n\nWhat is good: the environment and MILP are described with enough detail to follow; comparing against no-warm-start exact solver, EDF, and constraint-aware EDF is sensible; all warm-start variants reduce solve time, which is a consistent qualitative result. The authors are also candid that reward design is unfinished and scalability to 20 agents/100 tasks is future work. Credit for that.\n\nThe problem: Section III.E says the solver is run with a warm start \"to achieve the optimal solution\" and optimization time is recorded. If every run terminates at proven optimum, final schedule quality is identical across methods. Figure 4 shows differences in quality. So the protocol must be either (a) time-limited runs with quality computed on the incumbent, or (b) quality computed on the initial warm-start schedule before solving. The paper never says which. As written, the claim that the method \"maintains solution quality and feasibility\" cannot be checked. That is the load-bearing soft spot, and it is real.\n\nThere is also a smaller but annoying inconsistency: the abstract and Figure 3 say BC RL has stable performance / low variance, while Section IV.B says BC RL showed the most variability. That needs a fix. And the evaluation would be much stronger with tables, statistical tests, and a comparison against a learned warm-start baseline like Nair et al.; currently it only compares to heuristics and the exact solver.\n\nWho is this for: people working on learned warm-starts or construction-oriented multi-agent scheduling. Not a theory paper, not a benchmark-scale study. It deserves a serious referee rather than a desk reject, because the central empirical claim is checkable and the domain is reasonably motivated. I would not cite it until the metric ambiguity is resolved and code/data are released; but if the authors address those, it could become a solid workshop or short-conference contribution.","headline":"A cleanly written small empirical study showing learned warm-starts cut MILP solve time in a 10-agent/20-task construction scheduling domain; the speedup is plausible and unsurprising, but the paper never defines what its quality score is measured on, which blocks the central claim as written.","tokens_in":6311,"tokens_out":2469,"would_cite":false,"duration_ms":25080,"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":"The paper claims that a graph neural network trained by behavior cloning and fine-tuned with reinforcement learning can warm-start a MILP solver for multi-agent task allocation and scheduling, reducing optimization time and variance while…","keywords":["mixed integer linear programming","warm-starting","graph neural networks","behavior cloning","reinforcement learning","multi-agent task allocation and scheduling","branch-and-bound acceleration"],"falsifier":"Run the same warm-start comparison on 20-agent/100-task instances with the solver's default settings, and separately log search time excluding validation time; if the BC+RL warm start no longer beats the cold-start baseline on optimization time while maintaining feasibility, the central claim is falsified.","tokens_in":5301,"feed_emoji":"⏱️","tokens_out":12313,"duration_ms":102952,"temperature":0.7,"pith_summary":"The paper is trying to establish that a learned model—a graph neural network trained first by behavior cloning on expert solver outputs and then fine-tuned by reinforcement learning—can generate high-quality initial schedules that warm-start a MILP solver on a multi-agent task allocation and scheduling problem. If true, this gives a practical way to cut the long computation times that limit MILP use in time-sensitive settings such as construction, manufacturing, and logistics, without modifying the solver itself. On 100 random instances with 10 agents and 20 tasks, the trained warm starts report lower optimization time and variance than a cold-start baseline and than earliest-deadline-first heuristics, while keeping feasibility. The paper is also explicit that learned warm starts do not reach the peak schedule quality of a constraint-aware EDF heuristic and that scaling to larger instances remains future work, so the contribution is a speed-and-stability result on small instances rather than a claim of better schedules.","feed_headline":"Learned warm starts speed up MILP solvers, keep plans feasible","feed_subtitle":"Behavior cloning plus reinforcement learning warms the solver; run-to-run time variance drops on scheduling tasks.","key_machinery":"The load-bearing object is a graph neural network trained in two stages: behavior cloning to imitate expert MILP solutions, then reinforcement learning fine-tuning whose reward combines schedule quality—feasibility and makespan $t_{\\mathrm{ms}}$ relative to the deadline $t_{\\mathrm{ddl}}$—with optimization time $t_{\\mathrm{optimization}}$. The trained GNN outputs a task-assignment schedule that is passed to an off-the-shelf MILP solver as a warm start, so branch-and-bound search begins from that initial solution instead of from scratch. Supporting machinery includes a MILP formulation with order-constraint and time-window constraints (C1–C10) and a multi-agent rapidly-exploring random graph module that precomputes collision-free travel times between task locations. The mechanism's work is to convert a learned guess about assignments into a feasible starting point that shortens solver convergence.","core_discovery":"On the paper's own terms, the central discovery is that warm-starting a MILP solver with a GNN policy trained by behavior cloning and fine-tuned by reinforcement learning reduces the time to reach the optimal solution in a construction-inspired multi-agent task allocation and scheduling domain. Across 100 random 10-agent/20-task instances, all warm-start methods beat the cold-start baseline on optimization time, and the BC+RL policy recorded the lowest average time with stable performance, while behavior cloning alone beat both EDF and Constraint-Aware EDF on time. Quality scores remained feasible and competitive, although the EDF-based heuristics achieved the highest average quality and the BC+RL method showed the largest quality variance. The paper is candid that the reported time metric mixes search time and validation time, that the RL gains over BC are modest and inconsistent, and that scalability to 20 agents and 100 tasks is left to future work.","pith_inferences":["Because the GNN consumes only the instance graph, the same warm-start pipeline could be lifted to other assignment-style MILPs such as vehicle routing or crew scheduling, where the speedup would depend on how tightly clustered the optimal solutions are; the paper does not test any of these.","The quality gap between learned warm starts and Constraint-Aware EDF suggests the current reward may over-weight optimization time; reshaping the reward to penalize infeasible or long-makespan assignments more heavily could recover quality without giving up the time gain.","If the reported time reduction comes mainly from the search phase, the method's value should grow with instance size; if it comes from validation time, the practical gain will shrink on solver configurations with cheap validation. The paper does not separate the two.","The learned policy's ceiling may be set by the small instances it imitates, so testing transfer to 20-agent/100-task problems would reveal whether the warm-start advantage is a property of the learned representation or an artifact of the test distribution."],"forward_implications":["Any MILP solver that accepts an initial feasible solution can be accelerated by a learned warm start, so the speedup is available without changing solver internals.","Time-sensitive deployments gain more predictable runtimes because the learned warm start lowers variance across repeated solves.","A learned warm start can beat simple greedy heuristics on time even when the start itself has lower schedule quality, because the solver begins the search closer to the optimum.","Feasibility under order constraints and time windows is preserved, which matters when infeasible assignments are unacceptable.","The balance between learning and heuristics depends on the reward weight on optimization time; the paper's BC-only versus BC+RL comparison shows this tradeoff can be tuned."],"supporting_citations":[{"why":"Supplies the multi-agent task-assignment and scheduling MILP formulation that this paper adapts with heterogeneous travel times.","marker":"[5]"},{"why":"Provides the graph-neural-network scheduling-policy approach that behavior cloning builds on.","marker":"[6]"},{"why":"Supports fine-tuning a behavior-cloned initializer with reinforcement learning using solver feedback.","marker":"[9]"},{"why":"Demonstrates that neural-network-generated solutions can warm-start MILP solvers, the premise this work extends.","marker":"[15]"},{"why":"Provides the multi-agent rapidly-exploring random graph method used to precompute collision-free travel times.","marker":"[16]"},{"why":"Defines the Earliest Deadline First heuristic used as a time-optimization baseline.","marker":"[17]"},{"why":"Defines the Constraint-Aware EDF heuristic that achieves the highest quality scores in the evaluation.","marker":"[18]"},{"why":"Supports the explanation for why behavior-cloned policies can yield near-optimal warm starts despite lower direct quality scores.","marker":"[19]"}],"fun_headline_variants":["GNN warm starts cut MILP solve time in scheduling tests","Learned warm starts reduce MILP solve time and variance","BC and RL warm starts beat cold-start MILP on time","Warm-starting MILP with GNNs cuts solve time in allocation","Reinforced GNN warm starts trim MILP solve time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured speedup rests on the assumption that the 100 randomly generated 10-agent/20-task instances capture the real construction scenarios the paper targets; if larger or more realistic instances erase the learned warm start's advantage, the central claim would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["GNN warm starts cut MILP solve time in scheduling tests","Learned warm starts reduce MILP solve time and variance","BC and RL warm starts beat cold-start MILP on time","Warm-starting MILP with GNNs cuts solve time in allocation","Reinforced GNN warm starts trim MILP solve time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1483,"prompt_tokens":830,"completion_tokens":653,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":565}},"tokens_in":446,"tokens_out":653,"duration_ms":6390,"temperature":1.0,"reasoning_tokens":565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:47:42.125947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same warm-start comparison on 20-agent/100-task instances with the solver's default settings, and separately log search time excluding validation time; if the BC+RL warm start no longer beats the cold-start baseline on optimization time while maintaining feasibility, the central claim is falsified.","supporting_citations":[{"cited_title":"Learning Coordination Policies over Heterogeneous Graphs for Human-Robot Teams via Recurrent Neural Schedule Propagation,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-agent task-assignment and scheduling MILP formulation that this paper adapts with heterogeneous travel times."},{"cited_title":"Planning and scheduling under uncertainty: a re- view across multiple sectors,","cited_arxiv_id":null,"evidence_quote":"Provides the graph-neural-network scheduling-policy approach that behavior cloning builds on."},{"cited_title":"Human-Robot Teaming: Grand Challenges,","cited_arxiv_id":null,"evidence_quote":"Supports fine-tuning a behavior-cloned initializer with reinforcement learning using solver feedback."},{"cited_title":"Learn- ing primal heuristics for mixed integer programs,","cited_arxiv_id":null,"evidence_quote":"Demonstrates that neural-network-generated solutions can warm-start MILP solvers, the premise this work extends."},{"cited_title":"Rapidly exploring random graphs: mo- tion planning of multiple mobile robots,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-agent rapidly-exploring random graph method used to precompute collision-free travel times."}],"review_version":1}