acceptodds
Under review as a conference paper at ICLR 2027

DVRPTW-Bench: A Benchmark for Dynamic Vehicle Routing

Abstract

Dynamic vehicle routing with time windows (DVRPTW) requires planning routes as requests are revealed during operation. Solvers need to coordinate planning timing, dispatch, and routing over diverse, evolving request snapshots while maintaining computational efficiency. Existing benchmarks typically either use a small number of solvers to study problem difficulty or compare solver performance under fixed demand settings, therefore overlooking differences in the advantage of various methods across different scenarios. DVRPTW-Bench addresses this gap by systematically evaluating 17 approaches across 135 demand configurations with 10 instances each. The evaluated approaches cover three planning components: decision timing, dispatch, and routing. And demand is modeled through four factors: problem scale, degree of dynamism (DoD), temporal arrival patterns, and spatial demand evolution. Experiments show higher DoD increases learned dispatch's advantage in distance per served customer (DPC) relative to rolling hybrid genetic search (HGS). Temporal concentration reduces DPC more for HGS than for neural constructors, while dispersed arrivals give learned dispatch a larger advantage. Strong bursts can also reverse the benefits of urgency triggers for some neural constructors on large instances. Spatial hotspot shifts yield larger DPC reductions for HGS than for neural constructors on large instances. The benchmark suite and code are available at https://anonymous.4open.science/r/iclr-dvrptw0925.

Primary area:
datasets and benchmarks
Keywords:
Dynamic Vehicle Routing Problem, Combinatorial Optimization, Learning-Based Optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cL4cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DVRPTW-Bench, a benchmark for the dynamic vehicle routing problem with time windows, where requests are revealed during operation and solvers must decide when to plan, which requests to dispatch, and how to route them. The authors argue that existing benchmarks either study problem difficulty with a few solvers or compare methods under fixed demand settings, and therefore do not show how the relative strengths of methods change across demand conditions. To address this, they build a controllable instance generator that separately varies problem scale, degree of dynamism, temporal arrival patterns, and spatial demand evolution, together with a unified online evaluation protocol. Under this protocol, they evaluate a broad set of approaches covering planning triggers, search-based and neural routing solvers, and learned and anticipatory dispatch policies. The experiments characterize how demand conditions shape the relative performance of these methods, for example how dynamism, bursty arrivals, and shifting hotspots affect search-based versus neural solvers, and show that adaptive decision timing can improve various routing methods without retraining.

Critical strengths
  • The experiments are extensive and detailed, covering a large instance grid and many method configurations, with comprehensive per-slice results in the appendix.
  • The paper is clearly written and well organized, making the benchmark design and evaluation protocol easy to follow.
  • The authors make a clear effort to make the evaluation as complete as possible, including training-distribution variants, trigger sweeps, and search-budget studies.
  • The controlled generator, which varies each demand factor independently, enables clean attribution of performance changes to specific demand conditions.
Critical weaknesses

1. Limited advancement over prior work. Most components of the benchmark come from existing work. The problem setting closely follows the EURO–NeurIPS dynamic VRPTW, each demand factor has been studied in prior benchmarks (as Table 1 shows), and all evaluated methods and the offline reference are existing implementations. The main contribution is therefore combining these elements into a larger factorial grid under one protocol. Such a contribution can be valuable, but only if it yields insights that prior work could not provide. However, most findings are consistent with established intuitions, and the mechanisms proposed to explain them are not directly tested. Closely related resources such as dvrpsim and SVRPBench are also not included in Table 1, so the claimed gap is not fully established. I would encourage the authors to clarify which conclusions could not have been obtained from existing benchmarks.

2. Lack of real-world data and calibration. All instances are synthetic, and the generator’s settings, such as four narrow burst peaks, a fixed hotspot strength, clustered Gaussian customer locations with a central depot, and the ranges of lead times and window widths, are not calibrated against real data. A real-world delivery dataset is cited only to justify placing time windows after request creation. As a result, it is unclear whether the chosen configurations span realistic demand regimes, or whether conclusions drawn from them would hold in practice. This is a missed opportunity: calibrating the generator against real operational data, or including real-data-derived instances such as those used in the EURO–NeurIPS competition, would let the benchmark show how closely its scenarios reflect reality, which would be a valuable contribution in its own right.

3. Questionable evaluation of neural baselines. The neural constructors are trained on static VRPTW instances from the RL4CO generator, whose spatial and temporal distributions differ substantially from those of the benchmark. Moreover, the “scale-matched” setting matches the nominal daily number of customers, whereas the actual planning pools contain only a few to around a dozen requests per call, so the models are applied far outside their training sizes. The models are also used without the post-optimization (for example, local search) or inference-time augmentation that is standard in the neural combinatorial optimization literature, and key details such as the decoding strategy and how departure times are chosen are not reported. Consequently, the large gap to HGS cannot be attributed to the limitations of neural construction itself rather than to distribution shift or the specific usage protocol. Training on planning-pool snapshots collected from the benchmark, or adding a simple post-optimization step, would help disentangle these factors.

4. Some scenarios and combinations are not covered. This is a minor point, but several practically relevant settings are not considered: limited fleet capacity (which the authors note as a limitation), stochastic or time-dependent travel times, request cancellations, and reassignment of vehicles en route. In addition, although the introduction emphasizes the coupling between decision timing, dispatch, and routing, the trigger study pairs triggers only with HGS and neural constructors, and not with the dispatch policies. Including these settings and combinations in future versions would make the benchmark more comprehensive.


After rebuttal

I'd thank the authors for the detailed responses. They have addressed most of my concerns, and I have accordingly raised my score to 3.

Questions for authors

See the Weaknesses.

3 comments
AuthorsOpenReview ↗

Response (1/2): Response to W1/Q1

Thank you for the detailed suggestions. Below we set out three advances over existing parameterized models and the real-data evidence.

W1 We agree that demand factors are not new: van Lon et al. (2016) generate multi-peak arrivals and Zhang et al. (2023) generate time-varying hotspots. Our contribution lies in the following three advances.

Advance 1: a more realistic and flexible demand model. Our parameterized family independently controls temporal peaks and spatial hotspots, supports asymmetric demand patterns, and can be fitted directly to real orders.

Demand patternExisting modelsDVRPTW-Bench
Temporal peaksvan Lon et al. (2016): equally spaced, identically shaped sinusoidal peaks with coupled width and normalized height.Uniform background plus a (truncated) Gaussian mixture, with adjustable peak count, locations, widths, and weights; shared by B/R scenarios and real-data fitting.
Spatial hotspotsZhang et al. (2023): two default hotspots, one cooling and one intensifying.KK clusters and PP periods with adjustable intensity α\alpha; calibration also estimates period boundaries, active hotspot sets, and weights.

Advance 2: calibration to real cities, for testing and for training. Meituan resembles R with weak hotspots (α=0.11\alpha=0.11), Shanghai resembles B with stronger hotspots (α=0.42\alpha=0.42), and Jilin resembles B without significant hotspots. Our staged calibration fits release times, lead times, windows, clusters, and hotspot dynamics (W2). The calibrated generator reproduces method rankings across all 21 paired Hangzhou real/generated instances and can supply additional test and training data. Matching training to benchmark snapshots already narrows neural gaps by 2.82–7.44 percentage points.

Advance 3: method rankings that change with demand conditions. Across 120 dynamic configurations, 22 of 36 method pairs reverse order at least once. These results identify where each method performs well or fails. The open-source code supports further targeted evaluation (https://anonymous.4open.science/r/iclr-dvrptw0925).

Which conclusions could not be obtained from existing benchmarks. The following results require varying factors or interactions that existing benchmark settings leave unexamined:

FindingEvidenceRequired control
Planning timing can recover much of learned dispatch’s benefit.Urgency-10 reduces DPC by 5.78–7.43%, approximately 70% of ML-CO’s gain; count/volume triggers increase it by 14.63–32.43%.Planning trigger
Search benefits more from bursts than neural construction.US → BS: HGS OG decreases by 8.32 points, versus 0.17–2.30 for original constructors and 3.74–5.22 after fine-tuning.Arrival profile × solver
Timing benefits depend on solver and demand.Under large strong bursts, Urgency-10 increases three constructors’ DPC by 1.98–4.01%, while HGS still improves.Trigger × solver × demand
Learned dispatch’s advantage varies with scale and concentration.Official ML-CO: +19.86% at n=20n=20, −8.34% at n=500n=500; ML-CO SL’s advantage shrinks from 16.63% to 10.22% as burst intensity increases.Scale and burst intensity
Spatial dynamics expose weaknesses beyond pool size.US → UV: similar pool sizes, but neural OG increases by 3.65–6.47 points while HGS OG decreases by 4.54.Hotspot evolution
Anticipatory dispatch benefits vary with available slack.ICD improves DPC by 1.15% on the main grid, versus 28.85% for ICD-history on Hangzhou orders with a 303-minute mean lead time.Lead time and real-data calibration

We will add SVRPBench and dvrpsim to Table 1. The reason that we didn't put these two research into Table1 is that SVRPBench studies robustness under stochastic, time-varying travel; dvrpsim provides general dynamic-routing simulation. DVRPTW-Bench complements these tools through controlled demand generation and systematic evaluation of planning timing, dispatch, and routing. The focus and conclusions of these three researches are different.

AuthorsOpenReview ↗

Response (2/2): Response to W2-4/Q2-4

W2 We evaluate Meituan food-delivery orders from the 2025 INFORMS TSL Data Challenge (Wang et al., 2025) and Cainiao’s LaDe-P pickup data (Wu et al., 2023), fitting on earlier dates and testing on held-out days.

(1) Real demand supports controllable hotspots. Permutation tests with 500 spatial-label shuffles, preserving order times, detect significant adjacent-period spatial drift on all seven held-out days in Shanghai and Hangzhou, one of two Meituan days, and none in Jilin. These differences support both V and S scenarios. Fitted mean hotspot intensity ranges from 0.11 to 0.42 across six datasets; R was fitted to Meituan, while B and α=0.8\alpha=0.8 provide stress settings.

(2) The same equations fit temporal and spatial demand. Staged fitting isolates release times (G1), lead times and windows (G2), static clusters (G3-S), and evolving hotspots (G3), retaining defaults for unfitted components. Across six datasets, mean held-out Wasserstein errors are 8.74/10.08/0.52 minutes for release times/lead times/window widths. Selected results are:

DatasetSignificant hotspot daysMean αW1W_1 (min): release / lead / windowSpatial-drift error: G3-S → G3
Meituan1/20.116.36 / 0.77 / 0.650.00097 → 0.00081
Shanghai7/70.4217.92 / 10.01 / 0.490.00658 → 0.00170
Hangzhou7/70.207.04 / 16.10 / 0.520.00561 → 0.00132
Jilin0/70.188.77 / 12.45 / 0.260.00606 → 0.00545

Hotspot modeling reduces spatial-drift error by 74.1%/76.5% in Shanghai/Hangzhou, versus 10.2% in Jilin, consistent with the significance tests.

(3) Calibrated data reproduce real-stream method rankings. We evaluate 21 paired real/generated instances from Hangzhou region 80: three N=100N=100 replicas per day over June 24–30. All five methods achieve 100% service, with matching rankings across all pairs. ML-CO reduces DPC relative to HGS by 13.88%/17.46% on real/generated streams, and ICD-history by 28.85%/31.16%. Details appear in Appendix J.

W3 We separately evaluate snapshot fine-tuning for distribution and scale mismatch, followed by PyVRP local search. DPC gaps to rolling HGS across 1,350 instances are:

ModelOriginalFine-tunedFine-tuned + local search
PolyNet+15.10%+7.66%+1.64%
SymNCO+15.32%+12.50%+1.73%
POMO+16.84%+12.64%+1.70%

Fine-tuning removes 2.82–7.44 percentage points, but pure constructors still trail HGS, OR-Tools, ALNS, and both dispatch methods. Local search reduces gaps to 1.64–1.73%, outperforming OR-Tools and ALNS for an additional 0.075–0.153 s/instance.

HGS still benefits more from concentrated demand: after fine-tuning and local search, constructor OG decreases by 7.20–7.31 points under US → BS and 3.82–3.89 under US → UV, versus 8.32 and 4.54 for HGS. Under hotspot shifts, fine-tuning alone benefits PolyNet; POMO and SymNCO improve mainly through local search.

Models use one greedy decode without augmentation or multistart, followed by exact capacity/time-window feasibility projection; infeasible visits return to the pending pool. Routes depart at the latest feasible time at or after the current time. Local search accepts feasible improvements under a 0.05 s soft budget per call. These rules will be documented in Appendix C.1.

W4

The benchmark assumes ample vehicles, deterministic travel, and routes fixed after departure. Undispatched cancellations require only pending-pool updates; dispatched-order changes, diversion, and limited fleets require extended execution states and resource constraints and will be acknowledged as limitations.

On the 1,500 dynamic instances, replacing Fix-30 with Urgency-10 reduces ML-CO's paired DPC by 1.89% and approximately halves both its planning calls and solver time. This reduction is accompanied by a decrease in service rate from 99.25% to 95.29%, indicating a computation--quality--service trade-off rather than an unqualified improvement. One possible explanation is that the evaluated ML-CO policy was trained from online states generated under Fix-30; changing the trigger alters the timing and distribution of the states encountered at inference. Identifying an effective trigger for ML-CO may therefore require jointly designing the trigger and dispatch policy, or retraining the policy on states generated by the target trigger.

We will incorporate the additional experiments and key findings into the revised manuscript. We hope these clarifications and additional results address your concerns, and we would appreciate your reconsideration in light of this evidence. Thank you again for your constructive feedback.

Thank you for the detailed responses. They have addressed most of my concerns, and I have accordingly raised my score to 3.

Reviewer_eAm4OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a benchmark for dynamic vehicle routing with time windows. It compares routing solvers across 1,350 synthetic instances to examine how demand conditions and planning decisions affect performance.

Critical strengths

The trigger experiments offer a useful practical takeaway. Changing when to plan improves several neural solvers without retraining, showing that decision timing can matter alongside the routing model itself.

Critical weaknesses

W1. The neural solvers are trained on static VRPTW instances at predefined nominal sizes, but tested on dynamic snapshots with varying, often much smaller sizes. For example, mixed-scale training uses 20–200 customers, while the average post-initial planning pools at DoD = 0.95 contain only 1.63–13.63 requests. This leaves it unclear whether the reported gaps reflect limitations in handling dynamic demand or simply a mismatch between training and evaluation workloads.

W2. The benchmark does not fully separate the factors it claims to control. Increasing problem size also increases the number of spatial clusters, so the reported scale effects may partly reflect changes in geometry. The main trends are also not checked across different burst widths or hotspot strengths, leaving it unclear how much the conclusions depend on the chosen generator settings.

Questions for authors

Q1. Do the performance gaps persist when the neural solvers are trained on instances whose sizes match the actual dynamic planning-pool distribution?

Q2. Are the findings robust to fixed spatial structure and different generator settings?

Q3. The setup is close to existing dispatch-wave benchmarks such as EURO Meets NeurIPS. What new challenges or insights does DVRPTW-Bench add beyond testing more demand settings and solvers?

Q4. What unique value does this benchmark offer? What concrete guidance do these findings provide for future model design?

3 comments
AuthorsOpenReview ↗

Response (1/2): Response to W1-2/Q1-2

Thank you for your questions. Briefly, for Q1, training matched to actual planning pools substantially narrows the gaps, while preserving the original findings that pure constructors trail search and dispatch methods and that the gaps grow with scale. For Q2, the original findings on scale and demand remain valid under fixed cluster counts and parameter sweeps.

W1/Q1 The gap narrows but persists after training is matched to actual planning pools. We fine-tuned the three constructors on actual planning-pool snapshots collected under our protocol and re-evaluated their DPC gaps relative to rolling HGS on the paper's 1,350 instances:

ModelOriginal checkpointAfter snapshot fine-tuningDPC reduction from fine-tuning relative to the original model
PolyNet+15.10%+7.66%6.25%
SymNCO+15.32%+12.50%2.23%
POMO+16.84%+12.64%3.43%

(1) The original ordering of methods is retained. After fine-tuning, the three constructors still trail HGS, OR-Tools (+3.72%), ALNS (+5.77%), and the two dispatch methods. They move ahead of LK, Tabu, and ACO, with which they previously shared a similar performance tier (+15.57–17.82%).

(2) The original finding that the gap grows with problem scale remains valid. After fine-tuning, PolyNet's DPC gap relative to rolling HGS increases from 3.15% to 10.69%.

(3) The ordering among models is retained and becomes clearer. Fine-tuned PolyNet's gap is approximately 5 percentage points smaller than those of SymNCO and POMO, exceeding the differences between their original checkpoints (0.22–1.74 percentage points).

(4) The magnitude of the gaps requires revision. Distribution mismatch accounts for 2.82–7.44 percentage points of the original gaps. Fine-tuning improves DPC on 70–95% of instances.

Details are provided in Appendix H.1.

W2/Q2

(1) The original scale finding (Table 9(a)) does not depend on cluster count. With cluster count fixed at K = 2, 4, or 10 (US/UV, DoD = 0.5/0.95), the methods' DPC gaps relative to HGS are as follows (ranges across the three cluster counts):

Methodn=20n=100n=500Matched main grid n=20→500
OR-Tools0.00–0.010.34–0.422.94–3.320.00→2.70
ALNS0.10–0.281.25–1.584.81–5.160.41→5.15
ML-CO−14.40–−10.44−21.63–−18.23−12.93–−12.25−10.06→−12.91
ICD−0.26–0.00−2.52–−0.76−2.46–−1.17−0.02→−1.40
Fine-tuned PolyNet0.98–1.472.46–3.738.03–10.611.60→7.95 (origin model)
Fine-tuned SymNCO4.11–6.427.47–12.7410.55–16.284.10→10.08 (origin model)
Fine-tuned POMO4.88–5.888.17–14.2810.25–16.424.82→9.80 (origin model)

With cluster count fixed, the gaps of search methods and neural constructors still grow with scale. The method ranking agrees with Table 9(a).

(2) The original demand findings (Section 4.3) remain valid in the parameter sweeps. We tested three values each for burst width, burst intensity, and hotspot intensity:

Parameter changeHGS OGML-COICDFine-tuned PolyNetFine-tuned POMO
Burst width: 60 → 15 min49.52 → 39.06−15.97 → −11.18−0.98 → −2.84+6.74 → +10.40+7.96 → +12.17
Burst intensity: 0.50 → 0.9546.07 → 37.52−16.63 → −10.22−1.68 → −1.89+6.95 → +11.51+8.95 → +13.30
Hotspot intensity α: 0.4 → 1.046.00 → 41.58−16.10 → −11.16−1.41 → −1.90+5.92 → +10.36+9.19 → +23.90

The method ranking is unchanged across all nine values and agrees with the original main grid. As demand becomes more concentrated, HGS OG decreases, ML-CO's lead narrows, and the fine-tuned constructors' gaps widen. The original setting is one point along this monotonic trend.

Full results are provided in Appendix I.

AuthorsOpenReview ↗

Response (2/2): Response to Q3-4

Q3 The EURO–NeurIPS dynamic track uses fixed one-hour epochs and uniformly samples request attributes from static instances, corresponding to fixed planning intervals, stationary arrivals, and static spatial demand in our benchmark. DVRPTW-Bench extends this setting in three ways:

(1) Planning timing as a decision (Section 4.4, Appendix E). Training-free Urgency-10 lowers DPC by 5.78–7.43% for HGS and four neural constructors, approximately 70% of ML-CO’s gain. Count- and volume-based triggers instead increase DPC by 14.63–32.43%, while Urgency-10 worsens three constructors by 1.98–4.01% under large-scale strong bursts. These results reveal interactions between timing, demand, and solver choice.

(2) Interpretable demand controls that expose distinct failure modes. Our model independently controls temporal peaks and evolving spatial hotspots:

Demand patternExisting modelsDVRPTW-Bench
Temporal peaksvan Lon et al. (2016): equally spaced sinusoidal peaks with coupled width and normalized height.Uniform background plus a (truncated) Gaussian mixture, with independently adjustable peak locations, widths, and weights; shared by B/R scenarios and real-data fitting.
Spatial hotspotsZhang et al. (2023): two default hotspots, one cooling and the other intensifying.Hotspots rotate across KK clusters and PP periods with tunable intensity α\alpha; calibration also estimates period boundaries, active sets, and weights.

From US to UV, planning-pool sizes barely change, yet neural OG increases by 3.65–6.47 percentage points while HGS OG decreases by 4.54 points. Parameter sweeps further reveal differences hidden by fixed settings: the fine-tuned POMO–PolyNet gap grows from 3.27 to 13.54 points as α\alpha increases from 0.4 to 1.0, while ML-CO’s advantage over HGS shrinks from 16.63% to 10.22% as burst intensity rises from 0.50 to 0.95.

(3) Real-city calibration and condition-dependent rankings. Across 120 dynamic configurations, 22 of 36 method pairs reverse order at least once. In Hangzhou, where mean order lead time is 303 minutes versus 30–120 minutes in the main grid, ICD-history outperforms ML-CO (−28.85% vs −13.88%), reversing their main-grid ranking. Calibration on earlier days reproduces this ranking across all 21 paired real/generated instances. This supports evaluating methods under city-specific demand conditions.

Q4 The benchmark locates where a method fails: in which decision component and under which demand structure. The code is open source (https://anonymous.4open.science/r/iclr-dvrptw0925), so later work can generate targeted data to probe a method. Our results point to two directions:

  1. Learn when to plan. Timing depends on slack, arrival patterns, and the downstream solver, motivating learned or solver-aware triggers, joint triggering–dispatch policies, and burst-aware deferral.
  2. Design constructors for spatial generalization. POMO’s starting-node diversity may become less effective under concentrated demand, while PolyNet’s complementary policies may offer greater adaptability. We will test this hypothesis by varying the number of policies. Further directions include spatially conditioned policies, hotspot-intensity curricula, and historical hotspot information.

We will incorporate the additional experiments and key findings into the revised manuscript. We hope these clarifications and additional results address your concerns, and we would appreciate your reconsideration in light of this evidence. Thank you again for your constructive feedback.

Thank you for your clarification. I have updated the score accordingly.

Reviewer_HQUvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DVRPTW-Bench, a benchmark for dynamic vehicle routing with time windows. It evaluates routing solvers, dispatch policies, and planning triggers across 1,350 instances covering 135 demand configurations. The benchmark considers variations in problem scale, degree of dynamism, temporal arrival patterns, and spatial demand evolution. The experiments compare classical search methods, neural route constructors, learned dispatch, and anticipatory dispatch, together with analyses of computational efficiency and training distributions. The results illustrate how demand conditions influence the relative performance of different methods and how planning timing affects routing efficiency and service quality.

Critical strengths
  1. The benchmark organizes dynamic routing into three meaningful components: planning timing, dispatch, and route construction. This framework helps readers understand how different decisions contribute to overall performance.
  2. The experiments cover a broad range of methods and demand configurations. The additional analyses of computational budgets and training distributions provide useful insights beyond aggregate performance rankings.
  3. The paper reports both service rates and routing costs, supported by paired comparisons and confidence intervals. These evaluations help characterize the trade-offs between service quality, routing efficiency, and computational effort.
Critical weaknesses
  1. The abstract omits “problem” from the full name of DVRPTW. The term should be written as “dynamic vehicle routing problem with time windows” to ensure consistency with the acronym.
  2. The paper does not fully separate the effects of model design, training distribution, and evaluation settings. Consequently, some explanations for the observed performance differences remain insufficiently supported.
  3. The comparisons could more clearly establish the additional value of sophisticated dispatch methods over simpler strategies.
  4. The benchmark covers a restricted set of operational assumptions. The extent to which its findings generalize to broader dynamic routing settings remains unclear.
  5. The paper provides limited evidence that the proposed benchmark adequately captures real-world routing scenarios, leaving its practical applicability insufficiently established.
Questions for authors
  1. Please clarify the respective contributions of model capability and training–evaluation distribution differences to the observed performance.
  2. Please explain the additional benefits of the more sophisticated methods over simpler strategies under comparable computational resources.
  3. Please discuss the generalizability of the main findings beyond the current benchmark assumptions and provide supporting evidence.
  4. Please discuss the extent to which the proposed benchmark captures real-world routing scenarios and its applicability to practical operations.
2 comments
AuthorsOpenReview ↗

Response (1/2): Response to W1-3/Q1-2

Thank you for your constructive comments. We re-examined the original findings by fine-tuning the neural constructors on actual planning-pool snapshots, sweeping burst and hotspot parameters, and evaluating demand modeling with real order data. The main findings remain valid; the magnitude of the neural constructors' performance gaps requires revision.

W1: The title and abstract now consistently use "dynamic vehicle routing problem with time windows," and the associated typographical errors have been corrected.

W2/Q1: Separating the effects of model capability and training distribution. We fine-tuned POMO, PolyNet, and SymNCO on actual planning-pool snapshots collected under our protocol, and re-evaluated their DPC gaps relative to rolling HGS on the paper's 1,350 instances:

ModelOriginal checkpointAfter fine-tuningDPC reduction relative to the original model
PolyNet+15.10%+7.66%6.25%
SymNCO+15.32%+12.50%2.23%
POMO+16.84%+12.64%3.43%

(1) The original ordering of methods is retained. After fine-tuning, the three constructors still trail HGS, OR-Tools (+3.72%), ALNS (+5.77%), and the two dispatch methods. They move ahead of LK, Tabu, and ACO, with which they previously shared a similar performance tier (+15.57–17.82%).

(2) The original finding that the gap grows with problem scale remains valid. PolyNet's DPC gap relative to rolling HGS originally increased from 6.43% at n = 20 to 21.22% at n = 500; after fine-tuning, it increases from 3.15% to 10.69%.

(3) The ordering among models becomes clearer. Fine-tuned PolyNet's DPC gap relative to rolling HGS is approximately 5 percentage points smaller than those of SymNCO and POMO.

(4) The magnitude of the gaps requires revision. Distribution mismatch accounts for 2.82–7.44 percentage points of the original gaps. Fine-tuning improves DPC on 70–95% of instances.

Detailed settings are provided in Appendix H.1.

W3/Q2 Additional benefits of complex methods at comparable computational cost. Section 4.5 and Appendix F of the original pdf report the time–quality relationships for dispatch methods and neural methods, respectively. The following table adds the fine-tuned neural constructors and compares methods by computational budget per instance:

Budget regimeRuntime per instanceBest methodKey result
Low latencyApproximately 1 s or lessFine-tuned neural constructorsFine-tuned PolyNet takes 0.50 s, with a +7.66% gap to HGS; low-budget HGS (0.05 s/call) takes approximately 1.2 s, with a +9.5% gap.
Moderate (before ML-CO saturates)Approximately 2–8.6 sML-COAt 0.03 s/call (approximately 2.1 s/instance), ML-CO already matches HGS with 1 s/call (24.2 s/instance). At 0.05 s/call (approximately 2.5 s/instance), its DPC is 6.72% lower, within 0.3 percentage points of its saturated improvement (−6.94%). HGS with a comparable runtime (0.1–0.2 s/call, 2.5–4.9 s/instance) still has a DPC 2.04–6.38% higher than the 1 s/call HGS baseline (US scenarios).
Moderate or higher budget (after ML-CO saturates)From approximately 8.6 sML-COAt 8.6 s, ML-CO lowers DPC by 6.93% relative to saturated HGS (12.1 s); for n = 500 and DoD ≥ 0.8, it achieves a 12.61% reduction in 2.8 s. ICD takes over ten times as long as HGS and still underperforms ML-CO.
AuthorsOpenReview ↗

Response (2/2): Response to W4-5/Q3-4

W4/Q3 The burst and hotspot findings in Section 4.3 of the original manuscript are based on U, R, B and S, V scenarios with several fixed parameter settings. We tested three values each for burst width, burst intensity, and hotspot intensity, keeping all other settings at their original defaults.

Parameter changeHGS OGML-COICDFine-tuned PolyNetFine-tuned POMO
Burst width: 60 → 15 min49.52 → 39.06−15.97 → −11.18−0.98 → −2.84+6.74 → +10.40+7.96 → +12.17
Burst intensity: 0.50 → 0.9546.07 → 37.52−16.63 → −10.22−1.68 → −1.89+6.95 → +11.51+8.95 → +13.30
Hotspot intensity α: 0.4 → 1.046.00 → 41.58−16.10 → −11.16−1.41 → −1.90+5.92 → +10.36+9.19 → +23.90

(1) The original method hierarchy is unchanged across all nine parameter values: ML-CO > ICD > HGS > OR-Tools > ALNS > fine-tuned PolyNet > fine-tuned POMO.

(2) The original finding that as demand becomes more concentrated, HGS OG decreases, ML-CO's lead narrows, and the fine-tuned constructors' gaps widen remains valid, with monotonic changes across the tested parameter values. Increasing burst intensity from 0.50 to 0.95 and reducing burst width from 60 to 15 minutes lower HGS OG by 8.55 and 10.46 percentage points, respectively, while narrowing ML-CO's lead and widening fine-tuned PolyNet's gap.

(3) The original finding that HGS benefits from time-varying hotspots while neural constructors adapt less effectively remains valid. As α increases from 0.4 to 1.0, HGS OG falls by 4.42 percentage points, while the gaps of fine-tuned PolyNet and POMO widen to +10.36% and +23.90%, respectively.

Full results are provided in Section 4.6 and Appendix I.

W5/Q4 We evaluated the original demand model using Meituan food-delivery orders released for the 2025 INFORMS TSL Data Challenge (Wang et al., 2025) and LaDe-P, the pickup subset of Cainiao's LaDe dataset (Wu et al., 2023). All fitting uses earlier dates, with evaluation on held-out days.

(1) Treating hotspots as a controllable factor is justified. Permutation tests that preserve order times and randomly shuffle spatial cluster labels 500 times show that spatial drift between adjacent periods (JS divergence) significantly exceeds the no-hotspot baseline on all seven held-out days in both Shanghai and Hangzhou. Daily z-scores are 5.38–11.04 and 5.28–9.40, respectively. One of Meituan's two days is significant, while Jilin shows no significant evidence. Cities with and without hotspots correspond to the original V and S scenarios, respectively.

(2) The generation equations can fit real demand. We isolate temporal and spatial effects through staged fitting: G1 fits release times; G2 adds lead times and time windows; G3-S adds static spatial clusters; and G3 adds time-varying hotspots. Unfitted components retain benchmark defaults, with uniform S locations before G3-S. All fitting uses training days. Across six datasets, mean held-out Wasserstein errors are 8.74, 10.08, and 0.52 minutes for release times, lead times, and window widths, respectively. Selected results are:

DatasetDays with significant hotspotsMean fitted αW1W_1 (min): release / lead time / window widthSpatial-drift error: G3-S → G3
Meituan1/20.116.36 / 0.77 / 0.650.00097 → 0.00081
LaDe-P Shanghai7/70.4217.92 / 10.01 / 0.490.00658 → 0.00170
LaDe-P Hangzhou7/70.207.04 / 16.10 / 0.520.00561 → 0.00132
LaDe-P Jilin0/70.188.77 / 12.45 / 0.260.00606 → 0.00545

Time-varying hotspots reduce spatial-drift error by 74.1% in Shanghai and 76.5% in Hangzhou, versus 10.2% in Jilin, where the permutation test is insignificant. The framework thus supports data-driven assessment of whether hotspot dynamics materially improve the fit.

33 The scenario range covers observed demand. R was fitted to Meituan, while B and α = 0.8 serve as stress settings. Without fixing hotspots in advance, fitted mean α ranges from 0.11 to 0.42 across six datasets, placing observed demand closer to S and weak V.

44 Method findings transfer to real order streams. We construct 21 paired real/fitted G3 instances from seven held-out days (June 24–30) in Hangzhou region 80, with three N=100N=100 replicas per day; training and fitting use earlier dates only. HGS, OR-Tools, ALNS, ML-CO, and ICD-history achieve 100% service, with matching rankings across all 21 pairs. Relative to HGS, ML-CO reduces DPC by 13.88%/17.46% on real/fitted streams, and ICD-history by 28.85%/31.16%.

Details are provided in Appendix J.

We will incorporate the additional experiments and key findings into the revised manuscript. We hope these clarifications and additional results address your concerns, and we would appreciate your reconsideration in light of this evidence. Thank you again for your constructive feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We thank all reviewers for their careful reading and constructive comments. The reviews raised three shared questions: whether the gaps of neural constructors reflect the methods or their training and usage, whether the demand findings depend on particular generator settings, and how well the benchmark reflects real operations. We addressed them by re-testing the original findings. The main conclusions hold, and we revised the magnitude of the neural gaps.

Neural constructors. After fine-tuning on planning-pool snapshots collected under our protocol, the constructors' gap to rolling HGS falls from 15.10–16.84% to 7.66–12.64%. They still trail the search and dispatch methods, and the gap still grows with scale. Adding local search reduces the gap to about 1.7%, while HGS still benefits more from concentrated demand. The original ordering and trends therefore hold, although training–evaluation mismatch had inflated the size of the gap.

Robustness to generator settings. Across three values each of burst width, burst intensity and hotspot intensity, the method ranking is unchanged and the effects of demand concentration change monotonically, with the original settings lying on these trends. With the number of spatial clusters held fixed, the scale effects remain.

Relation to real operations. Fitted on Meituan and LaDe-P orders and tested on held-out days, the generator reproduces the release-time, lead-time and window distributions with mean errors of about ten minutes or less. Hotspot drift is significant in some cities and absent in others, which supports keeping hotspots as a controllable factor. The original strong-burst and α = 0.8 settings act as stress cases. On paired real and generated Hangzhou instances, the online methods rank identically.

Insights beyond existing benchmarks. Beyond integrating existing demand factors and methods, DVRPTW-Bench offers three advances. First, its demand model controls temporal peaks and spatial hotspots independently and can represent asymmetric patterns that existing generators cannot. Second, the same model can be fitted to real cities, so it can generate test and training data that match a target operation. Third, evaluating timing, dispatch and routing under one protocol across these conditions shows that method rankings change with demand and that when to plan is a decision in its own right. The benchmark therefore shows where each method performs well or fails rather than producing a single ranking.

Revisions. We corrected the terminology and added dvrpsim and SVRPBench to Table 1. Fine-tuned and post-optimized neural results appear in Section 4.2.2, Table 4 and Appendix H.1. Parameter sweeps and fixed cluster counts appear in Section 4.6 and Appendix I, and real-data calibration and transfer in Appendix J. Appendix C.1 documents the decoding and departure rules, and the limitations section is expanded.

We thank the reviewers again for comments that have strengthened the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.