{"id":"e8646f51-95bf-4876-a8ee-b3b01b0a3ffd","arxiv_id":"2606.10081","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors introduce 75 benchmark instances for railway crew scheduling and demonstrate that their branch-and-price algorithm achieves average optimality gaps below 1%, driven mainly by primal heuristics.","lead":"This paper creates 75 new public benchmark instances for railway crew scheduling that mimic real planning data and tests a branch-and-price algorithm on them, reporting average optimality gaps below 1%. A smart generalist might read it to see how optimization methods perform on large transportation scheduling tasks and what drives their success.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Representativeness of the 75 generated instances for real railway crew scheduling difficulty remains the load-bearing assumption.","rationale":"The reader's weakest_assumption directly identifies the same point; the full-text description of instance generation does not add independent evidence (such as statistical matching tables or hardness benchmarks) that would remove the risk.","tokens_in":1685,"tokens_out":272,"duration_ms":16370,"concrete_test":"Run the published algorithm and solution checker on any real (anonymized) NS instance set of comparable size; if average optimality gaps rise above 2% or the relative gap closure from heuristics drops below 70% of total, the representativeness assumption fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline performance claim (average gaps <1% driven by primal heuristics, with branching/reduced-cost fixing contributing little) and the broader conclusion that 'improving the dual bound remains the main challenge' both require that the synthetic instances exhibit the same structure and hardness as real NS instances. The paper states the instances are generated by mimicking the full planning process and validated against NS data, yet provides no quantitative comparison (e.g., duty network density, task clustering statistics, or dual-bound tightness distributions) that would confirm the generated set is not systematically easier for column-generation heuristics.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents a branch-and-price algorithm for railway crew scheduling and introduces a publicly available library of 75 benchmark instances (450–3,016 tasks) generated by mimicking the full planning process and validated against Netherlands Railways data. It reports a computational study of algorithmic components (branching rules, primal heuristics, reduced-cost fixing) on these instances, claiming average optimality gaps well below 1% driven primarily by the primal heuristics, with branching and reduced-cost fixing contributing little, and concludes that improving the dual bound remains the main challenge for exact methods.","tokens_in":1796,"tokens_out":546,"duration_ms":15619,"significance":"If the instances are representative of real railway crew scheduling difficulty, the work supplies a much-needed public benchmark set with best-known solutions and a solution checker, together with empirical evidence that column-generation primal heuristics are effective while dual-bound improvement is the bottleneck. These contributions would be useful for the operations research community working on crew scheduling and column generation.","major_comments":[{"comment":"Instance generation section: The statement that the 75 instances are 'validated against data from Netherlands Railways' is presented without quantitative metrics (e.g., comparisons of duty-network density, task-clustering statistics, or dual-bound tightness distributions between generated and real instances). Because the headline performance claims (gaps <1 %, attribution to primal heuristics, and the conclusion that dual bounds are the main challenge) rest on the assumption that the synthetic instances exhibit the same structure and hardness as real NS problems, the absence of such statistics is load-bearing.","section":"Instance generation section"},{"comment":"Computational results section (tables reporting gaps and component contributions): The attribution that 'branching and reduced cost fixing contribute little' is based on aggregate averages; without per-instance or per-size-class breakdowns showing the marginal effect of each component (or ablation runs), it is difficult to confirm that the observed gaps are not simply an artifact of the instance set being easier for the primal heuristics than for the dual-side techniques.","section":"Computational results section"}],"minor_comments":[{"comment":"Abstract: the phrase 'well below 1%' should be replaced by the precise average gap value (or range) once the tables are finalized.","section":"Abstract"},{"comment":"The paper states that instances and a solution checker are publicly available; the repository URL and exact license should be given in the main text (not only in a footnote) to facilitate immediate use by readers.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments and the recommendation for major revision. We address each major comment point by point below, providing clarifications and indicating where revisions will be made to strengthen the manuscript.","responses":[{"response":"We agree that providing quantitative validation metrics would better support the representativeness of the instances. In the revised version, we will expand the instance generation section to include comparisons of key structural statistics, such as duty-network density and task-clustering measures, between the generated instances and real data from Netherlands Railways. This will address the concern regarding the assumption of similar structure and hardness.","revision_made":"yes","referee_comment":"[Instance generation section] Instance generation section: The statement that the 75 instances are 'validated against data from Netherlands Railways' is presented without quantitative metrics (e.g., comparisons of duty-network density, task-clustering statistics, or dual-bound tightness distributions between generated and real instances). Because the headline performance claims (gaps <1 %, attribution to primal heuristics, and the conclusion that dual bounds are the main challenge) rest on the assumption that the synthetic instances exhibit the same structure and hardness as real NS problems, the absence of such statistics is load-bearing."},{"response":"We acknowledge that aggregate averages alone may not fully demonstrate the marginal contributions. In the revision, we will include per-size-class breakdowns of the results, showing the impact of each algorithmic component (branching, primal heuristics, reduced-cost fixing) on the optimality gaps. This will provide clearer evidence that the gaps are indeed driven primarily by the primal heuristics across different instance sizes.","revision_made":"yes","referee_comment":"[Computational results section] Computational results section (tables reporting gaps and component contributions): The attribution that 'branching and reduced cost fixing contribute little' is based on aggregate averages; without per-instance or per-size-class breakdowns showing the marginal effect of each component (or ablation runs), it is difficult to confirm that the observed gaps are not simply an artifact of the instance set being easier for the primal heuristics than for the dual-side techniques."}],"tokens_in":1421,"tokens_out":452,"duration_ms":22351,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is a new library of 75 public benchmark instances for railway crew scheduling, sized from 450 to 3016 tasks, plus the result that primal heuristics account for most of the gap closure to under 1% on average while branching and reduced-cost fixing add little.\n\nThe paper does the field a service by releasing the instances with best-known solutions and a checker. Generating them from an end-to-end planning simulation and checking against Netherlands Railways data is a reasonable way to produce usable test cases. The component ablation is straightforward and gives a clear picture of what drives performance in their implementation.\n\nThe soft spot is the validation step. The abstract states the instances were validated against real data, but supplies no numbers on structural similarity such as task clustering, network density, or dual-bound behavior. Without that, the claim that improving the dual bound is the main remaining challenge rests on an unquantified assumption about instance difficulty.\n\nThis is for researchers who need reproducible test sets for column-generation or branch-and-price work in transportation scheduling. A reader looking for concrete performance numbers on a practical problem will find it useful.\n\nIt deserves peer review because the benchmarks are new and reusable even if the performance conclusions need closer scrutiny on the data.","headline":"New public benchmarks for railway crew scheduling plus evidence that primal heuristics close most gaps in their branch-and-price runs, though instance representativeness lacks quantitative backing.","tokens_in":2258,"tokens_out":328,"would_cite":true,"duration_ms":14823,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A branch-and-price algorithm solves railway crew scheduling instances to average optimality gaps below 1 percent, driven by primal heuristics.","keywords":["railway crew scheduling","branch and price","column generation","benchmark instances","primal heuristics","optimality gaps","reduced cost fixing"],"falsifier":"Running the same algorithm on proprietary instances from an actual operator and obtaining average gaps substantially larger than 1 percent would falsify the performance claim.","tokens_in":2583,"feed_emoji":"🚂","tokens_out":649,"duration_ms":21986,"temperature":0.7,"pith_summary":"The paper presents a branch-and-price algorithm for assigning tasks to crew duties at minimum cost and tests it on a new set of 75 benchmark instances that range from 450 to 3,016 tasks. These instances are generated to reproduce the full railway planning pipeline and are validated against data from Netherlands Railways before being released publicly with known solutions. Computational experiments show that the method routinely reaches gaps well below 1 percent across all sizes, with the largest contribution coming from column-generation-based primal heuristics; branching rules and reduced-cost fixing close little additional gap. The results therefore reinforce the practical value of heuristic column generation while identifying the dual bound as the main remaining obstacle to exact solutions.","feed_headline":"Branch-and-price reaches under 1% gaps on railway crew benchmarks","feed_subtitle":"New 75-instance library shows primal heuristics close most of the gap while branching adds little.","key_machinery":"Branch-and-price algorithm that generates duties via column generation, augmented by primal heuristics, on a library of 75 publicly released benchmark instances that mimic full railway planning.","core_discovery":"The branch-and-price procedure, built around column generation for feasible duties, produces average optimality gaps well below 1 percent on all 75 instances. Primal heuristics account for most of the progress toward optimality, whereas branching and reduced-cost fixing contribute little to closing the remaining gap. The instances themselves are constructed by simulating the entire planning process and are made available with a solution checker so that future methods can be compared directly against the reported bounds.","pith_inferences":["Similar benchmark-generation techniques could be applied to other transportation crew problems such as bus or airline scheduling.","The limited contribution of branching suggests that pure heuristic column generation may be the method of choice for operational use.","Public release of both instances and best-known solutions lowers the barrier for testing new dual-bound improvements."],"forward_implications":["Column-generation heuristics are sufficient to produce near-optimal crew schedules for problems of this scale.","Exact branch-and-price methods will require stronger dual bounds before they can routinely close the final gap.","The released instance library and checker provide a standard testbed for comparing future crew-scheduling algorithms.","Operators can adopt similar heuristic column-generation procedures for day-to-day planning with high confidence in solution quality."],"fun_headline_variants":["Branch-and-price hits sub-1% gaps on 75 railway crew instances","Primal heuristics drive branch-and-price below 1% on crew benchmarks","75 instances confirm sub-1% gaps from branch-and-price in railways","Branch-and-price reaches sub-1% optimality on new crew scheduling set"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The 75 generated instances accurately represent the structure and difficulty of real railway crew scheduling problems.","fun_headline_variants_meta":{"raw":{"variants":["Branch-and-price hits sub-1% gaps on 75 railway crew instances","Primal heuristics drive branch-and-price below 1% on crew benchmarks","75 instances confirm sub-1% gaps from branch-and-price in railways","Branch-and-price reaches sub-1% optimality on new crew scheduling set"]},"model":"grok-4.3","cost_usd":0.005146,"raw_usage":{"total_tokens":2498,"prompt_tokens":664,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":51462000,"prompt_tokens_details":{"text_tokens":664,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1762,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":664,"tokens_out":72,"duration_ms":14008,"temperature":1.0,"reasoning_tokens":1762,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T15:19:32.538349+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the same algorithm on proprietary instances from an actual operator and obtaining average gaps substantially larger than 1 percent would falsify the performance claim.","supporting_citations":[],"review_version":1}