{"id":"ba4bed73-9e8d-40fa-8edb-1e83d1ae5eaf","arxiv_id":"2507.15385","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A transformer that predicts binary EV routing and scheduling decisions lets Gurobi solve a day-ahead MIP about 98% faster with negligible quality loss on a 20 to 100 EV test set.","lead":"This paper trains a transformer to predict the binary decisions in a day-ahead electric-vehicle routing and charging problem, then uses those predictions to shrink the search space that Gurobi must explore. The best tested model cuts solve time by about 98% with negligible quality loss on a 20 to 100 EV test bed, though that model was chosen after looking at the test results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 98.1% runtime-reduction headline is selected from four transformer variants on the same test set used to report it, so the number is an optimistic selection statistic; no held-out evaluation or confidence intervals are provided to show it is not a multiple-comparisons artifact.","rationale":"The central claim is an empirical engineering claim: the proposed transformer plus thresholding lets Gurobi solve the EVJRS MIP roughly 50x faster with negligible quality loss. The direct comparisons in Tables II and III are the strongest evidence, and the availability of the GitHub repository is a plus for reproducibility. I do not see an internal contradiction that falsifies the method; the reader's CONDITIONAL verdict is appropriate. My stress-test focuses on the evaluation protocol behind the headline number. Section V-B selects TF15 using the same test set on which the final metrics are reported, and the abstract/conclusion present TF15's numbers as the method's performance without a held-out model-selection split or uncertainty estimates. Since four models were compared and the reported r ranges from 89.00% to 98.18%, the difference between TF15 and TF20 is only 0.05 percentage points, well within what test-set noise could produce; the 98.1% figure is therefore not established as the expected performance of a single proposed method. The negative l entries in Tables II and III reinforce that the Gurobi reference is not certified optimal, so the tiny quality-loss claim should be treated cautiously. The reader's weakest_assumption about deterministic-to-stochastic transfer is plausible but less decisive: because the stochastic constraints and objective are scenario-wise separable (constraints 2-11 are all indexed by sc, and objective 1 is a probability-weighted sum), optimal binaries for each scenario can in principle be obtained from deterministic solves, so the transfer may be valid if each test scenario is fed separately. The absence of an explicit validation of this step and the low Acc1 (about 64%) still warrant attention, but they are secondary to the selection-bias issue for the headline quantitative claim. The 'any fleet size' wording also overstates the evidence, since tests cover only 20-100 EVs and the future-work section concedes that scaling to large fleets is untested; this is a scope limitation, not a fatal flaw. Overall, the correct disposition remains CONDITIONAL, pending a fresh held-out evaluation of the chosen model.","tokens_in":11439,"tokens_out":14237,"duration_ms":181080,"concrete_test":"Construct a fresh stochastic test set disjoint from the current test set (e.g., 100 instances with EV counts in [20,100] not divisible by 5), evaluate all four trained transformer models on it, and compute bootstrap 95% confidence intervals for r, l, and feas. If TF15 is no longer the best model or its fresh-set runtime reduction is significantly below 98.1%, the headline is a test-selection artifact; if TF15 remains best with overlapping CIs, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-B trains four transformer models (TF5, TF10, TF15, TF20), evaluates all four on the same testing dataset in Table II, and then declares that 'T F15 will be the best option to assist Gurobi'. The abstract and conclusion then report TF15's 98.13% runtime reduction, 100% feasibility, and 6.5e-4% quality loss as the performance of 'the proposed approach'. Because the same test set was used both for model selection and for the final performance claim, the headline numbers are maxima over four configurations and are likely optimistically biased. The paper does not state the test-set size, provide per-instance results, or report confidence intervals; the training/validation split is 90/10, implying only 80 validation samples per dataset, and the separate test set is unquantified. This matters because the four variants differ by only small margins in runtime reduction, so selection noise can change which model is 'best' and the associated headline. A second symptom of the same evaluation fragility is that Tables II and III contain negative quality losses (TF5: -2.4e-4%, CNN5: -0.01%), which cannot occur if the Gurobi baseline is a proven optimum; this suggests the reference objectives are not certified optimal, so the 'less than 0.0007% solution quality loss' is relative to an unverified baseline. The deterministic-to-stochastic transfer concern raised by the reader is weaker than it first appears because constraints (2)-(11) and objective (1) are scenario-separable, so per-scenario deterministic labels can in principle be combined; the test-set-selection issue is the load-bearing threat to the central quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a transformer-based deep learning model that predicts optimal binary decisions of a mixed-integer programming formulation of a day-ahead electric vehicle joint routing and scheduling (EVJRS) problem. The predicted binaries are threshold-filtered and used to prune the search space for Gurobi. The model is designed to handle varying EV fleet sizes without retraining, and is trained on deterministic instances while tested on the full stochastic problem. On a test set derived from the IEEE 33-bus distribution network coupled with the Nguyen-Dupuis transportation network, the best transformer variant (TF15) is reported to reduce Gurobi solution time by 98.1% on average, achieve 100% feasibility, and lose less than 0.0007% solution quality, compared with a CNN baseline using padding.","tokens_in":11760,"tokens_out":4111,"duration_ms":49079,"significance":"If the reported results hold, the paper addresses a practically relevant problem: accelerating the solution of stochastic EV routing and scheduling with variable fleet sizes, while preserving feasibility and solution quality. The paper makes its code publicly available, provides a formal MIP formulation with explicit constraints, and compares against a CNN baseline. The authors are also honest about a limitation of their attention design. However, the evaluation protocol has load-bearing weaknesses: the headline numbers are selected from four transformer variants evaluated on the same test set, the baseline Gurobi solutions appear not to be certified optimal, and the deterministic-to-stochastic transfer is asserted rather than validated. These issues need to be addressed before the central claims can be accepted.","major_comments":[{"comment":"This is the central evaluation issue and affects the main quantitative claims.","section":"Section V-B, Table II"},{"comment":"This assumption underpins the entire supervised-learning pipeline and is currently unverified.","section":"Section IV-A, Section V-B"},{"comment":"This affects the interpretation of both the quality and the runtime comparison.","section":"Equation (22), Tables II and III"}],"minor_comments":[{"comment":"The motivation for using a transformer mentions its ability 'to capture long term dependencies in sequential data', but the implementation repurposes the sequence dimension to be the feature axis and the feature dimension to be time; attention therefore operates across features rather than time. The authors do acknowledge this limitation, but the framing should be adjusted to avoid overstating the model's temporal modelling capability, or a temporal-attention variant should be considered.","section":"Section IV-B"},{"comment":"The description of the threshold-based post-processing is underspecified: it is not clear whether the mean prediction probability thresholds are computed per class, per instance, or globally over the dataset, nor how the thresholds interact with the class imbalance in a multi-label setting. Please clarify the exact threshold computation and selection rule.","section":"Section IV-D"},{"comment":"The notation for the test set, 'Et ∈ [20, 100] / E5', is difficult to parse. Please explicitly state which EV counts are in the test set (e.g., all integers in [20,100] that are not multiples of 5) and report the number of test instances used in Tables II and III.","section":"Table I"},{"comment":"There is a typo in the definition of the multipliers: '{5, 10, 5, 20}' should presumably be '{5, 10, 15, 20}'. Please correct this.","section":"Section V-B"},{"comment":"There are several language issues, including 'combinatorial challenging' in the abstract and 'we proposed' (past tense) in the abstract and introduction. These should be corrected in a final language pass.","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":"The central idea is timely and the code availability is a plus, but the evaluation protocol needs substantial strengthening before publication. The model-selection-on-the-test-set issue and the non-certified baseline are the main risks; both are fixable within the scope of a revision. I do not see evidence of inappropriate citation or novelty concealment, but the deterministic-to-stochastic transfer should be handled carefully in revision, as it is the methodological hinge of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The actual contribution here is narrower and more credible than the abstract suggests: a transformer encoder that puts attention on the feature axis instead of the time axis, so the model can accept variable EV fleet sizes without padding. That is a genuine architectural adaptation, and the paper shows it beats the authors' own prior CNN-with-padding approach on feasibility and speed. The EVJRS formulation is clearly laid out, the comparison to a CNN baseline is fair, and the code is public. That is real work and worth engaging with.\n\nThe soft spots are mostly about the headline number. The 98.13% runtime reduction is TF15, selected because it had the best balance of metrics in Table II, and Table II is computed on the same test set used to report the headline. With four variants and no confidence intervals or stated test-set size, the 98% is a selection statistic and likely optimistic. The negative quality loss for TF5 in the same table is a separate symptom: it implies the Gurobi reference objectives are not certified optimal, so the \"less than 0.0007% loss\" is relative to an unverified baseline. That should be easy to fix by reporting MIP gaps or using a longer time limit.\n\nI agree with the stress-test note that the deterministic-to-stochastic training mismatch is a weaker concern than it first looks: the scenario-separable structure means per-scenario deterministic labels can in principle be combined, though the paper does not validate that transfer directly. The \"any fleet size\" claim is also overbroad; the test range is 20–100 EVs, so \"a wide range of fleet sizes\" would be accurate, \"any\" is not.\n\nThe citation pattern and arithmetic are fine. The paper positions itself honestly within the ML-assisted MIP literature, and the self-referential aspect of training on Gurobi labels and then benchmarking against Gurobi is inherent to this line of work, not a flaw.\n\nWho is this for? Researchers working on ML-guided MIP for power-system operations, especially anyone dealing with varying problem dimensions. It is a solid engineering contribution whose quantitative claims need sharpening, not a paper with a load-bearing logical hole.\n\nRecommendation: send it to peer review. A serious referee should ask for a proper held-out model-selection protocol, error bars or per-instance results, and an explicit statement of the Gurobi optimality gap on the baseline solutions. Those are addressable, and the core method is worth publishing after that.","headline":"A legitimate fleet-size-agnostic transformer adaptation for predicting MIP binary variables, but the 98% speedup headline is a selected best-of-four statistic and needs a proper held-out evaluation before I'd trust the number.","tokens_in":12310,"tokens_out":1085,"would_cite":true,"duration_ms":15103,"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":"A transformer trained on deterministic EV routing problems predicts the optimal binary decisions well enough to prune the mixed-integer search space, cutting Gurobi runtime by 98.1% on average with 100% feasibility and less than 0.0007%…","keywords":["Electric Vehicles","Power system simulation","Optimization","Neural networks","Vehicle routing","Transformer","Mixed-integer programming","Search space pruning"],"falsifier":"Solve enough stochastic EVJRS test instances to proven optimality, apply the transformer's thresholded predictions as fixed binary variables, and check whether the true optimal solution of any instance violates a fixed variable; if it does, the pruning step has cut away the optimum and the near-zero quality-loss claim collapses.","tokens_in":11224,"feed_emoji":"⚡","tokens_out":8225,"duration_ms":74973,"temperature":0.7,"pith_summary":"This paper tries to show that a transformer-based neural network can make a hard day-ahead optimization problem—scheduling and routing electric vehicles across a power grid and a traffic network—tractable by predicting the optimal binary decisions before a commercial solver starts. The authors argue that fixing the predicted high-confidence binary variables prunes the mixed-integer search space enough that Gurobi reaches near-optimal solutions much faster, and that the model's attention-based design lets it handle any fleet size without retraining. On simulations of an IEEE 33-bus distribution network coupled to the Nguyen-Dupuis transportation network, the assisted solver is claimed to cut runtime by 98.1% on average while keeping every tested instance feasible and losing less than 0.0007% of solution quality. If correct, the method points to a general recipe: use learned structure to shrink the combinatorial core of MIPs, while leaving constraint satisfaction to the solver.","feed_headline":"Transformer cuts EV routing solve time 98 percent, no retraining","feed_subtitle":"A fleet-size-agnostic transformer keeps every solution feasible while Gurobi runs 98 percent faster.","key_machinery":"The load-bearing object is a two-layer transformer encoder whose sequence dimension is repurposed from time to features (bus loads, PV output, and EV job schedules), with time kept as the feature dimension, so that the number of EV-related tokens can vary from instance to instance. After embedding, three feed-forward classification layers with a sigmoid output produce a probability for each binary decision variable, and a thresholding filter derived from the mean prediction probability of each class selects the bits used to fix variables in Gurobi. The same variable-size FFN mechanism lets the model produce predictions for fleets of sizes never seen in training, which is what makes the fleet-size-agnostic claim work.","core_discovery":"The central claim is that the optimal binary variables of the EVJRS stochastic mixed-integer program—EV routing choices and charging/discharging statuses—can be predicted well enough by a transformer trained on deterministic instances that the predictions can be used to prune the MIP's search space before Gurobi solves it. The authors report that on a held-out set containing stochastic instances and EV counts unseen during training, the best transformer model achieves 100% feasibility, an average runtime reduction of 98.1% over bare Gurobi, and an average objective loss of only $6.5 \\times 10^{-4}\\%$ (below 0.0007%), without retraining. They also report that models trained on coarser fleet-size granularity (e.g., EV counts in steps of 15) balance feasibility and solution quality better than finer-grained training. The mechanism is not direct solution: the network emits binary bit probabilities, a thresholding filter keeps only confident bits, and Gurobi completes the problem with the remaining variables.","pith_inferences":["The deterministic-to-stochastic transfer is the least tested link: the paper trains on deterministic instances and tests on stochastic ones, but does not compare the predicted binary pattern against the true stochastic optimum, so a targeted optimality check on small solvable instances would settle whether the pruning ever discards an optimal solution.","The same architecture-pruning recipe should transfer to other MIPs whose binary variables scale with a population—e.g., unit commitment, fleet dispatch, or appliance scheduling—because the transformer's variable-length sequence removes the main obstacle of re-training per problem size.","A practical deployment would likely need an online calibration of the threshold, since the confidence filter is currently calibrated on the training data's mean prediction probabilities rather than on an operator-specified worst-case quality bound.","Large-fleet extrapolation remains an open question that the authors themselves flag: labels for very large instances are expensive, so a natural test is whether a model trained on small fleets can still prune well when applied to fleets far larger than 100 EVs."],"forward_implications":["Day-ahead EV coordination with fleets of 20–100 vehicles can be cleared in a fraction of the original solve time, making the approach usable in time-sensitive market settings where bare Gurobi becomes prohibitively slow.","A single trained model can be applied to unseen fleet sizes, so operators do not need to retrain whenever the EV fleet changes.","Feasibility is guaranteed by the solver rather than approximated by the network, because the network only fixes a subset of binary variables and Gurobi enforces all constraints on the remainder.","Coarser training granularity can dominate finer granularity on the metrics that matter: TF15 achieves 100% feasibility with nearly the same runtime reduction as TF20 and much lower objective loss than TF20.","The thresholding filter gives a tunable safety valve: raising the confidence threshold retains fewer but safer binary bits, trading a little speed for additional solution-quality insurance."],"supporting_citations":[{"why":"Supplies the EVJRS problem formulation, the dataset construction procedure, and the CNN-with-padding baseline used for comparison.","marker":"[26]"},{"why":"Provides the transformer encoder architecture and multi-head self-attention mechanism the model is built on.","marker":"[29]"},{"why":"Gurobi is the MIP solver whose runtime and objective values define the baseline and whose search space is pruned.","marker":"[15]"},{"why":"The Nguyen-Dupuis network is the transportation network used in the numerical evaluation.","marker":"[16]"},{"why":"Supplies the linearized DistFlow formulation used to model distribution network operational constraints.","marker":"[27]"},{"why":"Provides the prediction-probability/thresholding method used in the post-processing filter.","marker":"[30]"},{"why":"Supplies the original time-space network model that the modified TSN with virtual congestion nodes extends.","marker":"[28]"}],"fun_headline_variants":["Transformer speeds EV routing 98% without retraining","Fleet-size agnostic AI cuts EV routing solve time 98%","AI predicts EV routing decisions, MIP solves 98% faster","Transformer prunes MIP, slashes EV routing solve time 98%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that binary decisions learned from deterministic instances remain representative of the optimal binary decisions in the full stochastic problem, so that scenario-combined predictions can safely prune the stochastic MIP's search space.","fun_headline_variants_meta":{"raw":{"variants":["Transformer speeds EV routing 98% without retraining","Fleet-size agnostic AI cuts EV routing solve time 98%","AI predicts EV routing decisions, MIP solves 98% faster","Transformer prunes MIP, slashes EV routing solve time 98%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":3034,"prompt_tokens":1010,"completion_tokens":2024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1948}},"tokens_in":626,"tokens_out":2024,"duration_ms":13201,"temperature":1.0,"reasoning_tokens":1948,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:33:35.426800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve enough stochastic EVJRS test instances to proven optimality, apply the transformer's thresholded predictions as fixed binary variables, and check whether the true optimal solution of any instance violates a fixed variable; if it does, the pruning step has cut away the optimum and the near-zero quality-loss claim collapses.","supporting_citations":[{"cited_title":"Joint optimisation of electric vehicle routing and scheduling: A deep learning-driven approach for dynamic fleet sizes,","cited_arxiv_id":null,"evidence_quote":"Supplies the EVJRS problem formulation, the dataset construction procedure, and the CNN-with-padding baseline used for comparison."},{"cited_title":"An efficient method for computing traffic equilibria in networks with asymmetric transportation costs,","cited_arxiv_id":null,"evidence_quote":"The Nguyen-Dupuis network is the transportation network used in the numerical evaluation."},{"cited_title":"An interval power flow method based on linearized distflow equations for radial distribution systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the linearized DistFlow formulation used to model distribution network operational constraints."},{"cited_title":"Asymmetric loss for multi-label classification,","cited_arxiv_id":null,"evidence_quote":"Provides the prediction-probability/thresholding method used in the post-processing filter."},{"cited_title":"Battery-based energy stor- age transportation for enhancing power system economics and security,","cited_arxiv_id":null,"evidence_quote":"Supplies the original time-space network model that the modified TSN with virtual congestion nodes extends."}],"review_version":1}