{"id":"0718b446-50c0-4675-9156-db88009c3d4f","arxiv_id":"2607.13263","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An ALNS heuristic with dynamic line generation and logit-based endogenous demand produces higher-frequency, higher-ridership bus networks for Odense, but the results depend on hand-set demand-sensitivity parameters.","lead":"The authors propose an adaptive large-neighborhood search heuristic that co-designs bus lines and frequencies while letting passenger demand respond to service quality through a logit mode-choice model. On Odense, Denmark's network, the heuristic produces denser, higher-frequency line plans that the model says would raise public-transport ridership share from 8% to 12-32%.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline ridership gains are outputs of a calibration, not findings: with β_d hand-set and the 5% floor unestimated, the 8%→12–32% claim can reverse within the plausible range the paper itself tests.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point, and I agree. The paper is otherwise carefully executed: the model and algorithm are described in detail, the ablation study isolates operator contributions, the SP/MNL re-evaluation checks route-choice sensitivity, and the sensitivity analysis is honest. None of these mitigates the calibration problem for the headline. In fact, §5.3 is the clearest evidence: a factor of two in β_d (0.05→0.10) more than doubles ridership and nearly doubles fleet size, while reducing β_d to 0.02 makes ridership fall below the observed baseline. Given that β_d and the 5% floor are not estimated from data, the quantitative claims in the Abstract and §5.2 cannot be taken at face value. The correct verdict remains CONDITIONAL: the framework is credible and the algorithmic contribution is plausible, but the specific ridership and network-design conclusions require either empirical estimation of β_d or demonstrated robustness across a wider calibration grid. I do not see an internal inconsistency that would justify REJECT, and the authors' own sensitivity discussion is a strength rather than a reason to dismiss the work.","tokens_in":29630,"tokens_out":11718,"duration_ms":140391,"concrete_test":"Re-run the §5.2 experiments at Rsub=40 DKK over β_d ∈ {0.02, 0.05, 0.1} and floors ∈ {0.01, 0.05, 0.1}, recalibrating α_d and b̂w_d each time; if any cell gives PT demand share at/below baseline or the 12.4–32.3% interval shifts by >5 pp, the headline ridership claim is not identified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that optimized networks 'substantially increase' PT ridership (Abstract) and capture 12.4–32.3% of demand (§5.2)—is not an empirical result; it is an output of the demand model calibrated in §5.1.3. There, total demand per OD is inferred as b̂w_d = q^obs_d / p^obs_d, where p^obs_d is computed from the logit constraint (11) at the current line concept with α_d=0 and then floored at 5%. OD-specific α_d are then recalibrated so the same logit model reproduces the observed baseline share. Hence the 8% baseline is fixed by construction, and the entire ridership response to an optimized network is governed by the hand-set β_d=0.05. §5.3/Table 5 shows the stakes: at β_d=0.02 the optimized solution captures about 1,460 PT passengers (below the observed baseline of 4,885), while at β_d=0.10 it captures about 30,675. Reported network designs also swing from roughly 5 to 28 lines and 35 to 472 vehicles. The 5% share floor is equally unestimated and directly scales the total-demand denominator of every reported share. The paper is transparent about this sensitivity, but transparency does not make the headline robust: within the plausible range of an unestimated parameter, the 'substantial ridership increase' can reverse. This is the load-bearing weakness. Secondary, reinforcing issues are the same-author preprint used as the evaluator and the absence of code/data, but the calibration issue alone is decisive for the quantitative claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Adaptive Large Neighborhood Search (ALNS) matheuristic for public transport line planning with endogenous demand. The model is an MINLP that jointly chooses lines and frequencies, routes passengers on a Change & Go network, and determines the PT mode share with a logit function. The algorithm combines destroy/repair operators, a frequency-setting local search, and an embedded fixed-point column-generation heuristic for passenger assignment and demand estimation. It is evaluated on the Odense, Denmark, bus network with 1,826 OD pairs. The authors report that the optimized networks reduce average headways from about 41 minutes to 6.8–13 minutes and increase PT share from about 8% to 12–32%, while emphasizing that the resulting designs are sensitive to passenger-behavior assumptions.","tokens_in":30016,"tokens_out":10777,"duration_ms":117427,"significance":"If the algorithmic claims hold, this is a useful step toward practical line planning with demand responsiveness, dynamic line generation, and frequency optimization at a scale beyond exact methods. The paper is strong in its presentation: the MINLP is clearly written, the ALNS and construction heuristic are given in pseudocode, and the experiments include ablated operators, replicated runs, convergence plots, and alternative routing and demand assumptions. The authors are also transparent about the sensitivity of their results. However, the headline ridership figures are outputs of a demand model calibrated to the baseline by construction, with a hand-set elasticity and an unestimated share floor; moreover, the formal capacity constraint appears dimensionally inconsistent. These issues affect the quantitative conclusions and must be resolved before the paper can be accepted.","major_comments":[{"comment":"The demand model is calibrated so that the baseline PT share is reproduced by construction: total demand is set as b_wd = q_obs_d / p_obs_d, where p_obs_d is obtained from Eq. (11) with α_d=0 and a 5% floor, and α_d is then fitted so the logit model reproduces the observed baseline share. The baseline ~8% share is therefore not a model prediction, and all out-of-sample ridership gains are governed mainly by the hand-set β_d=0.05 and the 5% floor. Table 5 shows the stakes: at β_d=0.02 the optimized network serves 1,460 PT passengers (below the observed baseline of 4,885), while at β_d=0.10 it serves 30,675; line counts vary from 5 to 28 and vehicles from 35 to 472. The abstract and conclusion state that the optimized networks \"substantially increase\" ridership. This is not robust: it holds only for the middle value of an unestimated parameter. Please either estimate β_d and the floor from","section":"§5.1.3, Eq. (11), Table 5"},{"comment":"Constraint (6) bounds passenger flow on an arc of line l by δ_{m(l)} z_l, where z_l is the number of vehicles assigned to line l. But the capacity of a line over the planning period (one peak hour) depends on the frequency f, not on the fleet size: capacity is approximately δ_m f T, while z_l satisfies z_l ≥ φ_lf ≈ f τ_l, with τ_l the round-trip time. Unless τ_l equals the planning period, Eq. (6) overestimates capacity by a factor τ_l/T, which is larger for longer lines. This can make the reported line plans infeasible with respect to actual passenger loads. Please correct the capacity constraint (e.g., express capacity in terms of the selected y_lf and frequency f, or otherwise align z_l with a capacity measure) and confirm that the SAMCF evaluation in §4.7 enforces the same corrected capacity.","section":"§3.4, Eq. (6); §3.1–§3.2"},{"comment":"The paper's central claim is that the ALNS is a \"competitive and scalable\" algorithm, but the experiments provide no comparison against exact solutions, lower bounds, or another published line-planning heuristic. The ablation study and convergence plots show only improvements relative to the algorithm's own starting solutions. Since the evaluation step is outsourced to the companion-paper SAMCF heuristic (Hansen and Lusby, 2026) and since three-hour runs continue to find small improvements late in the search, independent evidence of solution quality is needed. At minimum, report gaps to lower bounds (e.g., from an LP relaxation or from solving the MINLP on small instances) or benchmark against an existing method on a shared dataset.","section":"§5.2–§5.4, §4.7"}],"minor_comments":[{"comment":"The phrase \"substantially increasing public transport ridership\" should be qualified as being conditional on the assumed demand elasticity β_d=0.05 and the 5% mode-share floor, given the sensitivity shown in Table 5.","section":"Abstract, §5.2"},{"comment":"The method is described as using dynamic line generation, but §4.1 assumes a predefined line pool for the construction heuristic and the backbone-flow repair operator, and the random-line operator uses shortest paths in the PTN. Please clarify how much of the claimed novelty depends on the precomputed 797-line pool versus genuinely generated lines.","section":"§4.1, §4.2.2"},{"comment":"The choice β_d=0.05 is stated to be a \"reasonable representation\" of demand sensitivity, but no basis or reference is given. Please provide a justification, a calibration source, or at least a stated range of plausible values used in the sensitivity analysis.","section":"§5.1.3"},{"comment":"The SAMCF evaluation subroutine is described only by reference to a companion preprint. Since it is central to the results, please include sufficient detail (pricing subproblem, convergence checks, handling of empty PT path sets) for the paper to be self-contained.","section":"§4.7"},{"comment":"If no PT passengers are assigned to an OD pair, u_PT_d is undefined in Eq. (10). Please state how the model and the heuristic handle this case.","section":"§3.3.1, Eq. (10)"},{"comment":"No code or data availability statement is provided. For a computational paper of this type, releasing the instance generator, line pool, and calibrated OD matrix would materially improve reproducibility.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The capacity-constraint concern in §3.4 may indicate a mismatch between the written model and the implemented evaluation; the authors should be asked to clarify or correct this before the numerical claims are relied upon. The dependence on the companion preprint (Hansen and Lusby, 2026) for the evaluation step, with overlapping authorship, also raises self-containedness and reproducibility questions. The demand-calibration issue is not merely a style matter: the central ridership figures change dramatically within the parameter range the authors themselves test, so the abstract and conclusions need either empirical calibration or a conditional framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2607.13263. The ALNS framework is a genuine engineering contribution: it combines dynamic line generation, logit mode choice, and a column-generation-based evaluation into a pipeline that runs on a real 1,800-OD network in Odense. That scale is beyond exact methods in this subfield, and the authors deserve credit for the pseudocode, the ablation study, and the willingness to test sensitivity to the demand elasticity. The paper is readable, honest about its assumptions, and does not oversell the method as a proved optimum.\n\nThe soft spot is the demand calibration in Section 5.1.3. Total demand is inferred by dividing observed PT trips by a logit share computed with alpha_d=0 and a 5% floor; then alpha_d is recalibrated so the baseline share matches. That makes the 8% baseline and all ridership gains outputs of the model. The hand-set beta_d=0.05 is the main driver: Table 5 shows beta=0.02 gives about 1,460 PT passengers (below the observed 4,885) and beta=0.10 gives about 30,675. Reported networks swing from 5 to 28 lines and 35 to 472 vehicles. So the abstract's 'substantially increasing ridership' claim is not robust within the plausible range the paper itself tests. This is load-bearing for the quantitative conclusions.\n\nSecondary issues: no lower bounds or comparison against prior heuristics (Canca et al. is closest), and the evaluation subroutine is a same-author preprint with no code or data released. These are addressable but currently limit how much independent confidence the empirical results can carry.\n\nStill, the method is the contribution, not the specific numbers. For people working on transit network design heuristics, this is a useful reference on how to assemble these components at scale. It deserves a serious referee, but the authors should be pressed to reframe the conclusions as conditional on calibration and, ideally, add a lower-bound or benchmark comparison. I'd read it, and I might cite the method if I needed a recent example of ALNS with logit demand. For a reading group, it would be a good case study in how calibration choices shape headline results.\n\nRegards.","headline":"Worth a serious look for the method; don't cite the ridership numbers without checking the calibration sensitivity.","tokens_in":30505,"tokens_out":3192,"would_cite":true,"duration_ms":32545,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B06","90C59"],"pacs":[],"model":"deepseek-v4-flash","headline":"Public transport line planning can be done at city scale with demand that responds to service quality; the optimal networks favor fewer, far more frequent lines, and hinge on the measured demand elasticity.","keywords":["line planning","public transport","adaptive large neighborhood search","endogenous demand","mode choice","dynamic line generation","logit model","frequency setting"],"falsifier":"Measure the actual mode-choice elasticity for the Odense area from smart-card or survey data, re-run the optimization with the estimated parameter and re-calibrated constants, and compare the resulting network. If the real elasticity is at or below about 0.02, the model itself indicates only about 5 lines would be optimal — meaning the paper's central 'fewer, higher-frequency lines' conclusion would collapse.","tokens_in":29521,"feed_emoji":"🚌","tokens_out":8413,"duration_ms":72814,"temperature":0.7,"pith_summary":"This paper sets out to show that public transport line planning can be optimized at realistic city scale even when passenger demand is allowed to respond to the quality of the service offered. The authors develop an adaptive large neighborhood search algorithm that jointly selects lines, sets frequencies, routes passengers, and estimates mode choice through a logit model, generating new lines dynamically during the search rather than relying on a fixed candidate set. Applied to the Odense, Denmark bus network with about 1,800 origin-destination pairs, the algorithm produces networks that concentrate service into fewer, more frequent lines — average headways fall from roughly 41 minutes to 6.8–13 minutes — and captured public transport demand rises from about 8% of total trips to 12–32% depending on the subsidy level. The paper further shows that the resulting network design is sensitive to the assumed demand elasticity, swinging from a 5-line network at low sensitivity to a 28-line network at high sensitivity, which is why the authors argue careful demand calibration is essential.","feed_headline":"A bus-network redesign that cuts waits from 41 to 7 minutes","feed_subtitle":"An adaptive search algorithm with built-in demand response shows how — and why the elasticity assumption matters.","key_machinery":"The key mechanism is the coupling of an ALNS metaheuristic with an embedded evaluation loop. Destroy and repair operators (removing, shortening, extending, and adding lines) modify a line concept; a local search procedure adjusts operating frequencies by estimating marginal objective improvements; and each candidate solution is evaluated by a column-generation-based fixed-point heuristic that alternates between solving a capacitated multi-commodity passenger flow and updating OD-specific demand via a logit mode-choice function. The logit function — with OD-specific constants calibrated to observed demand and a fixed sensitivity beta — is what links service quality to demand, making ridership","core_discovery":"The paper's central claim is that a matheuristic based on adaptive large neighborhood search can solve line planning with endogenous demand, dynamic line generation, and passenger assignment on a real network of roughly 1,800 origin-destination pairs — beyond what exact methods can handle. The optimized line concepts concentrate resources on fewer, higher-frequency services, cutting average headways from about 41 minutes to 6.8–13 minutes and lifting captured public transport demand from about 8% to 12–32% of total trips depending on subsidy. Network structure is highly sensitive to the logit elasticity beta: 0.02 yields about 5 lines, 0.05 about 20, and 0.10 about 28. The authors take this","pith_inferences":["The paper's headline ridership gains are outputs of a demand model calibrated to observed trips with a fixed sensitivity parameter and a 5% baseline floor; if the true elasticity lies outside the 0.02–0.10 range, the absolute numbers — and possibly the qualitative 'fewer, more frequent lines' recommendation — would change.","Because the calibration only covers OD pairs present in the observed data, the model cannot represent latent trips induced by the improved service; the reported ridership increases may be a lower bound if new trips appear.","The paper shows that a threshold-based mode-choice model produces sparse networks that look good under its own assumptions but perform poorly when re-evaluated with the logit model; this suggests that re-evaluation under a common model, as the paper does, is a useful check for any demand-responsive planning study.","The algorithm's dependence on the initial solution (baseline operations vs a constructive heuristic) suggests that running several restarts and taking the best network, or combining solutions, would make the planning recommendation more robust."],"forward_implications":["If the optimization is correct, a city can cut average headways from roughly 41 minutes to 6.8–13 minutes by operating fewer lines, giving riders much shorter waits without necessarily adding vehicles at low subsidy levels.","Because lines are generated dynamically during the search, planners no longer need to pre-specify a candidate line pool; the initial line pool only seeds the search rather than constraining the final design.","Captured public transport demand rises from about 8% to 12–32% as the per-passenger subsidy increases, producing a cost-ridership frontier that a transport authority could use to set investment levels.","The three-hour runtime on a real network with about 1,800 origin-destination pairs suggests the approach is practical for routine planning studies rather than only one-off research.","The wide variation in optimal networks across the logit elasticity parameter (5 to 28 lines) means the demand-model calibration is a central input to planning, not a detail."],"fun_headline_variants":["ALNS bus redesign cuts waits from 41 to 7 minutes","Smarter line planning trims waits to 7 minutes","Demand-aware bus lines: waits drop from 41 to 7 min","ALNS finds bus lines that slash headways to 7 minutes","From 41 to 7 minutes: how ALNS redesigns bus networks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The demand response is governed by a single fixed elasticity parameter and OD-specific constants calibrated to observed demand under current operations, plus a 5% floor on baseline mode shares; if these values misrepresent how travelers actually switch modes, the optimized networks and their ridership gains are artifacts of the calibration.","fun_headline_variants_meta":{"raw":{"variants":["ALNS bus redesign cuts waits from 41 to 7 minutes","Smarter line planning trims waits to 7 minutes","Demand-aware bus lines: waits drop from 41 to 7 min","ALNS finds bus lines that slash headways to 7 minutes","From 41 to 7 minutes: how ALNS redesigns bus networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001152,"raw_usage":{"total_tokens":4614,"prompt_tokens":748,"completion_tokens":3866,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":3773}},"tokens_in":492,"tokens_out":3866,"duration_ms":28488,"temperature":1.0,"reasoning_tokens":3773,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:41:14.516648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual mode-choice elasticity for the Odense area from smart-card or survey data, re-run the optimization with the estimated parameter and re-calibrated constants, and compare the resulting network. If the real elasticity is at or below about 0.02, the model itself indicates only about 5 lines would be optimal — meaning the paper's central 'fewer, higher-frequency lines' conclusion would collapse.","supporting_citations":[],"review_version":1}