pith. machine review for the scientific record. sign in

arxiv: 2605.09301 · v1 · submitted 2026-05-10 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Neural Cluster First, Route Second: One-Shot Capacitated Vehicle Routing via Differentiable Optimal Transport

Maximilian Schiffer, Samuel J. K. Chin

Authors on Pith no claims yet

Pith reviewed 2026-05-12 03:28 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords Capacitated Vehicle RoutingOptimal TransportNeural Combinatorial OptimizationNon-autoregressiveCluster-First Route-SecondSymmetry Invariance
0
0 comments X

The pith

A non-autoregressive neural solver for capacitated vehicle routing assigns customers to vehicles first via optimal transport then routes within clusters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper revives the classical cluster-first route-second strategy for the capacitated vehicle routing problem and adapts it to neural networks by replacing sequential construction with a single global assignment step. This assignment is made differentiable through an entropic optimal transport layer that respects total fleet capacity while producing a sparse plan that an exact solver can turn into feasible routes. The resulting architecture is shown to be invariant to spatial rotations and reflections as well as to the ordering of routes and the ordering inside each route. Because it avoids autoregressive decoding it scales to instances an order of magnitude larger than those used in training while keeping optimality gaps small. A reader would care because last-mile routing remains a dominant cost driver in logistics and any method that removes sequential bottlenecks can directly improve solution speed and quality at industrial scale.

Core claim

Neural CFRS is the first purely non-autoregressive one-shot neural framework for CVRP that enforces global fleet-capacity constraints end-to-end via a differentiable entropic optimal transport layer, produces a continuous transport plan that is sparsified by an exact capacitated assignment solver, and carries formal guarantees that it abstracts away E(2) spatial symmetries, inter-route permutation symmetries, and intra-route traversal symmetries; equipped with a pre-trained spatial vocabulary it achieves a 2.73 percent optimality gap on size-100 benchmarks and scales to out-of-distribution instances of size 1000 with a gap below 4 percent.

What carries the argument

The differentiable entropic optimal transport layer that produces a continuous customer-to-vehicle transport plan respecting total capacity before an exact solver sparsifies it into a discrete assignment.

If this is right

  • The method produces solutions for 1000-customer problems with an optimality gap below 4 percent even when the model is trained only on smaller instances.
  • A single-layer version of the architecture retains an approximate 5 percent gap at the 1000-customer scale.
  • The architecture is invariant to E(2) spatial transformations, to the ordering of routes, and to the ordering of stops inside routes.
  • Deployment on standard benchmarks yields a 2.73 percent optimality gap on 100-customer instances without task-specific fine-tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The non-autoregressive design may allow parallel inference across multiple problem instances on the same hardware, reducing wall-clock time in high-volume routing services.
  • The separation of assignment and routing stages could be reused as a template for other vehicle-routing variants that admit a natural clustering decomposition.
  • Because the spatial vocabulary is pre-trained, the same backbone might transfer to related assignment problems such as multi-depot routing or pickup-and-delivery without retraining the entire network.

Load-bearing premise

The approach assumes real-world spatial customer distributions and a fixed vehicle capacity together with a pre-trained spatial vocabulary that enables zero-shot scaling.

What would settle it

Running the method on 1000-customer instances drawn from a qualitatively different spatial distribution, such as highly clustered or grid-like layouts, and measuring whether the optimality gap remains under 4 percent would directly test the scaling claim.

Figures

Figures reproduced from arXiv: 2605.09301 by Maximilian Schiffer, Samuel J. K. Chin.

Figure 1
Figure 1. Figure 1: The end-to-end Neural CFRS architecture. A SMAE (optional if using (x, y) inputs) is first pre-trained over the fixed spatial support X to extract E(2)-invariant representations Ψ. Phase 1 utilizes ST to construct node embeddings H and a capacity-aware greedy decoding step to dynamically extract a discrete set of K geographic anchor seeds C. In Phase 2, type embeddings Etype are first added to differentiat… view at source ↗
Figure 2
Figure 2. Figure 2: Yˆ precision and recall when compared against Y¯ exact. At τ = 0+, recall and precision is above 95% and 80% respectively, indicating that ≈ 5% of optimal edges are left out. At τ = 1−, recall and precision is above 70% and 97% respectively, indicating that ≈ 3% of nodes are assigned to suboptimal vehicles [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pareto frontier of the solver’s computational efficiency, mapping the MIP gap limit against [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Histogram detailing the distribution of optimality gaps for Neural CFRS (exact) on [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of the baseline LKH-3 solutions and the Neural CFRS (exact) predictions for the two N = 100 instances that yielded the most significant performance improvements. In these specific cases, Neural CFRS outperforms the baseline by 1.88% and 1.65%, respectively. The corresponding transport plans show a generally high level of confidence in the assignments, with many of the probabilities close … view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of optimality gaps across different problem sizes. Observe how [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A selection from the ten solutions exhibiting the largest optimality gaps for [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: A selection from the ten solutions exhibiting the largest optimality gaps for [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A selection from the ten solutions exhibiting the largest optimality gaps for [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
read the original abstract

The Capacitated Vehicle Routing Problem (CVRP) underpins modern last-mile logistics. Current Neural Combinatorial Optimization (NCO) methods construct CVRP solutions autoregressively, inheriting sequential decoding bottlenecks, sensitivity to spatial symmetries, and brittle out-of-distribution behavior. We revisit the classical Cluster-First-Route-Second (CFRS) paradigm -- long known to be asymptotically optimal but largely overlooked by NCO -- and argue that it is structurally aligned with the core strengths of deep learning: similarity and assignment over global context, rather than the construction of long sequential tours. We introduce Neural CFRS, the first purely non-autoregressive one-shot neural CFRS framework for the CVRP. It enforces global fleet-capacity constraints end-to-end via a differentiable entropic Optimal Transport layer, producing a continuous transport plan to sparsify an exact capacitated assignment solver. We provide formal theoretical guarantees that our architecture intrinsically abstracts away $E(2)$ spatial, inter-route permutation, and intra-route traversal symmetries. By equipping the framework with a pre-trained spatial vocabulary, we unlock extreme parameter efficiency and zero-shot scaling. Designed primarily for real-world spatial distributions under a constant capacity setting, Neural CFRS scales robustly to out-of-distribution $N=1000$ instances with a < 4% gap -- retaining an approximate 5% gap at this scale even as an ultra-lightweight, single-layer architecture. Furthermore, when deployed out-of-the-box on standard benchmarks, we achieve a highly competitive 2.73% optimality gap on size-100 problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper introduces Neural CFRS, the first purely non-autoregressive one-shot neural framework for the Capacitated Vehicle Routing Problem (CVRP) that revives the classical Cluster-First-Route-Second paradigm. It uses a differentiable entropic Optimal Transport layer to enforce global fleet-capacity constraints end-to-end, produces a continuous transport plan for an exact capacitated assignment solver, and claims formal theoretical guarantees that the architecture abstracts away E(2) spatial symmetries, inter-route permutation symmetries, and intra-route traversal symmetries. Equipped with a pre-trained spatial vocabulary, the method achieves parameter efficiency and zero-shot scaling, reporting a 2.73% optimality gap on size-100 problems and robust scaling to out-of-distribution N=1000 instances with a <4% gap (approximately 5% even for a single-layer variant).

Significance. If the theoretical symmetry guarantees and empirical scaling results hold with proper verification, this work could meaningfully advance neural combinatorial optimization by shifting from autoregressive construction to a differentiable clustering-assignment approach that aligns with deep learning strengths in similarity and assignment. The parameter-efficient zero-shot scaling and explicit handling of symmetries would represent a notable strength over existing NCO methods for routing problems.

major comments (1)
  1. [Abstract] Abstract: the central scaling claim of robust zero-shot performance to N=1000 instances with a <4% optimality gap (and ~5% for the single-layer variant) is load-bearing for the headline contribution, but true global optima are intractable at this scale; the manuscript must explicitly specify in the experimental evaluation how the gap is computed (e.g., against which heuristic solver such as LKH, lower bound, or best-known solution) and confirm methodological consistency with the size-100 benchmark to substantiate the generalization result.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and constructive feedback on our manuscript. We address the single major comment below and will incorporate the requested clarifications in the revised version.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central scaling claim of robust zero-shot performance to N=1000 instances with a <4% optimality gap (and ~5% for the single-layer variant) is load-bearing for the headline contribution, but true global optima are intractable at this scale; the manuscript must explicitly specify in the experimental evaluation how the gap is computed (e.g., against which heuristic solver such as LKH, lower bound, or best-known solution) and confirm methodological consistency with the size-100 benchmark to substantiate the generalization result.

    Authors: We agree that explicit specification of the reference used for optimality-gap computation at large scales is necessary for transparency and to substantiate the zero-shot scaling claims. The revised manuscript will add a clear statement in the experimental evaluation section (and a brief parenthetical note in the abstract) detailing the exact baseline employed for N=1000 instances. We will also explicitly confirm that the identical reference is used for the size-100 benchmark, thereby ensuring methodological consistency across all reported results. This clarification strengthens the presentation without altering any numerical claims or experimental outcomes. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's core claims rest on a non-autoregressive CFRS architecture using differentiable entropic OT for capacity constraints, symmetry abstractions via architecture design, and empirical scaling results on benchmarks. These are presented as building on established OT and clustering methods without any quoted reduction of a central result (e.g., a 'prediction' or 'guarantee') to a fitted parameter or self-citation by construction. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described framework. The <4% gap on N=1000 is an empirical performance metric against heuristics (as noted by the skeptic), not a derived quantity forced by the model's own inputs. The derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract provides limited detail on implementation; the pre-trained spatial vocabulary is the main unexamined component, and symmetry-abstraction claims are stated without proof sketches. No invented physical entities are introduced.

free parameters (1)
  • pre-trained spatial vocabulary
    Equips the model for extreme parameter efficiency and zero-shot scaling; parameters are fitted upstream and not detailed in the abstract.
axioms (1)
  • domain assumption The architecture intrinsically abstracts away E(2) spatial, inter-route permutation, and intra-route traversal symmetries
    Stated as formal theoretical guarantees provided by the framework.

pith-pipeline@v0.9.0 · 5586 in / 1296 out tokens · 80624 ms · 2026-05-12T03:28:07.736120+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages

  1. [1]

    Athiwaratkun, M

    B. Athiwaratkun, M. Finzi, P. Izmailov, and A. G. Wilson. There are many consistent ex- planations of unlabeled data: Why you should average.arXiv preprint arXiv:1806.05594, 2018

    work page arXiv 2018

  2. [2]

    Bello, H

    I. Bello, H. Pham, Q. V . Le, M. Norouzi, and S. Bengio. Neural Combinatorial Optimization with Reinforcement Learning. 11 2016

    work page 2016

  3. [3]

    Bienstock, J

    D. Bienstock, J. Bramel, and D. Simchi-Levi. A Probabilistic Analysis of Tour Partitioning Heuristics for the Capacitated Vehicle Routing Problem with Unsplit Demands.Mathematics of Operations Research, 18(4):786–802, 11 1993. ISSN 0364-765X. doi: 10.1287/moor.18.4.786

    work page doi:10.1287/moor.18.4.786 1993

  4. [4]

    Bramel and D

    J. Bramel and D. Simchi-Levi. A location based heuristic for general routing problems.Opera- tions research, 43(4):649–660, 1995

    work page 1995

  5. [5]

    Christiaens and G

    J. Christiaens and G. V . Berghe. Slack induction by string removals for vehicle routing problems. Transportation Science, 54(2):417–433, 2020. ISSN 15265447. doi: 10.1287/trsc.2019.0914

    work page doi:10.1287/trsc.2019.0914 2020

  6. [6]

    M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

    work page 2013

  7. [7]

    BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In J. Burstein, C. Doran, and T. Solorio, editors, Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Paper...

    work page doi:10.18653/v1/n19-1423 2019

  8. [8]

    Drakulic, S

    D. Drakulic, S. Michel, F. Mai, A. Sors, and J.-M. Andreoli. BQ-NCO: Bisimula- tion Quotienting for Efficient Neural Combinatorial Optimization. In A. Oh, T. Nau- mann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neu- ral Information Processing Systems, volume 36, pages 77416–77429. Curran Associates, Inc., 2023. URL https://proc...

    work page 2023

  9. [9]

    J. K. Falkner and L. Schmidt-Thieme. Neural Capacitated Clustering. 2 2023

    work page 2023

  10. [10]

    Feydy, T

    J. Feydy, T. Séjourné, F.-X. Vialard, S.-i. Amari, A. Trouvé, and G. Peyré. Interpolating between optimal transport and mmd using sinkhorn divergences. InThe 22nd international conference on artificial intelligence and statistics, pages 2681–2690, 2019

    work page 2019

  11. [11]

    M. L. Fisher and R. Jaikumar. A generalized assignment heuristic for vehicle routing.Networks, 11(2):109–124, 1981

    work page 1981

  12. [12]

    B. A. Foster and D. M. Ryan. An Integer Programming Approach to the Vehicle Scheduling Problem.Operational Research Quarterly (1970-1977), 27(2):367, 1976. ISSN 00303623. doi: 10.2307/3009018

    work page doi:10.2307/3009018 1970

  13. [13]

    Fu, K.-B

    Z.-H. Fu, K.-B. Qiu, and H. Zha. Generalize a small pre-trained model to arbitrarily large tsp instances. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 7474–7482, 2021

    work page 2021

  14. [14]

    B. E. Gillett and L. R. Miller. A heuristic algorithm for the vehicle-dispatch problem.Operations research, 22(2):340–349, 1974

    work page 1974

  15. [15]

    Gurobi Optimizer Reference Manual, 2026

    Gurobi Optimization LLC. Gurobi Optimizer Reference Manual, 2026. URL https://www. gurobi.com

    work page 2026

  16. [16]

    Helsgaun

    K. Helsgaun. An extension of the Lin-Kernighan-Helsgaun TSP solver for constrained traveling salesman and vehicle routing problems.Roskilde: Roskilde University, pages 24–50, 2017

    work page 2017

  17. [17]

    Q. Hou, J. Yang, Y . Su, X. Wang, and Y . Deng. Generalize Learned Heuristics to Solve Large- scale Vehicle Routing Problems in Real-time. InThe Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=6ZajpxqTlQ. 10

    work page 2023

  18. [18]

    N. Iyer, V . Thejas, N. Kwatra, R. Ramjee, and M. Sivathanu. Wide-minima density hypothesis and the explore-exploit learning rate schedule.Journal of Machine Learning Research, 24(65): 1–37, 2023

    work page 2023

  19. [19]

    C. K. Joshi, T. Laurent, and X. Bresson. An Efficient Graph Convolutional Network Technique for the Travelling Salesman Problem. 6 2019

    work page 2019

  20. [20]

    Khosla, P

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan. Supervised contrastive learning.Advances in neural information processing systems, 33:18661–18673, 2020

    work page 2020

  21. [21]

    M. Kim, J. Park, and J. Park. Sym-NCO: Leveraging Symmetricity for Neural Combinatorial Optimization. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 1936–1949. Curran Associates, Inc., 2022. URL https://proceedings.neurips.cc/paper_files/paper/ 2022/file/0c...

    work page 1936

  22. [22]

    W. Kool, H. van Hoof, and M. Welling. Attention, Learn to Solve Routing Problems! In International Conference on Learning Representations, 2019. URL https://openreview. net/forum?id=ByxBFsRqYm

    work page 2019

  23. [23]

    W. Kool, H. van Hoof, J. Gromicho, and M. Welling. Deep Policy Dynamic Programming for Vehicle Routing Problems. pages 190–213. 2022. doi: 10.1007/978-3-031-08011-1{\_}14

    work page doi:10.1007/978-3-031-08011-1 2022

  24. [24]

    Y .-D. Kwon, J. Choo, B. Kim, I. Yoon, Y . Gwon, and S. Min. Pomo: Policy optimization with multiple optima for reinforcement learning.Advances in Neural Information Processing Systems, 33:21188–21198, 2020

    work page 2020

  25. [25]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Fixing Weight Decay Regularization in Adam, 2018. URL https://openreview.net/forum?id=rk6qdGgCZ

    work page 2018

  26. [26]

    F. Luo, X. Lin, F. Liu, Q. Zhang, and Z. Wang. Neural combinatorial optimization with heavy decoder: Toward large scale generalization.Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  27. [27]

    Y . Min, Y . Bai, and C. P. Gomes. Unsupervised learning for solving the travelling salesman problem.Advances in neural information processing systems, 36:47264–47278, 2023

    work page 2023

  28. [28]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, and others. Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems, 32, 2019

    work page 2019

  29. [29]

    A. A. Pessoa, M. Poss, R. Sadykov, and F. Vanderbeck. Branch-Cut-and-Price for the Robust Capacitated Vehicle Routing Problem with Knapsack Uncertainty.Operations Research, 69(3): 739–754, 5 2021. ISSN 0030-364X. doi: 10.1287/opre.2020.2035

    work page doi:10.1287/opre.2020.2035 2021

  30. [30]

    R. Qiu, Z. Sun, and Y . Yang. DIMES: A Differentiable Meta Solver for Combinatorial Opti- mization Problems. In A. H. Oh, A. Agarwal, D. Belgrave, and K. Cho, editors,Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id= 9u05zr0nhx

    work page 2022

  31. [31]

    Rasku, T

    J. Rasku, T. Kärkkäinen, and N. Musliu. Meta-survey and implementations of classical capaci- tated vehicle routing heuristics with reproduced results.Toward Automatic Customization of Vehicle Routing Systems, pages 133–260, 2019

    work page 2019

  32. [32]

    Renaud, F

    J. Renaud, F. F. Boctor, and G. Laporte. An Improved Petal Heuristic for the Vehicle Routeing Problem.The Journal of the Operational Research Society, 47(2):329, 2 1996. ISSN 01605682. doi: 10.2307/2584352

    work page doi:10.2307/2584352 1996

  33. [33]

    J. M. P. Silva, E. Uchoa, and A. Subramanian. Cluster branching for vehicle routing problems. INFORMS Journal on Computing, 2025

    work page 2025

  34. [34]

    Sun and Y

    Z. Sun and Y . Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization. Advances in Neural Information Processing Systems, 36, 2024. 11

    work page 2024

  35. [35]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    work page 2017

  36. [36]

    T. Vidal. Hybrid genetic search for the CVRP: Open-source implementation and SWAP* neighborhood.Computers & Operations Research, 140:105643, 4 2022. ISSN 03050548. doi: 10.1016/j.cor.2021.105643

    work page doi:10.1016/j.cor.2021.105643 2022

  37. [37]

    Vinyals, M

    O. Vinyals, M. Fortunato, and N. Jaitly. Pointer Networks. InAdvances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL https://papers.nips. cc/paper/2015/hash/29921001f2f04bd3baee84a12e98098f-Abstract.html

    work page arXiv 2015

  38. [38]

    N. A. Wouda, L. Lan, and W. Kool. PyVRP: A High-Performance VRP Solver Package. INFORMS Journal on Computing, 36(4):943–955, 7 2024. ISSN 1091-9856. doi: 10.1287/ijoc. 2023.0055

    work page doi:10.1287/ijoc 2024

  39. [39]

    H. Ye, J. Wang, Z. Cao, H. Liang, and Y . Li. DeepACO: neural-enhanced ant systems for combinatorial optimization.Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  40. [40]

    H. Ye, J. Wang, H. Liang, Z. Cao, Y . Li, and F. Li. Glop: Learning global partition and local construction for solving large-scale routing problems in real-time. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 20284–20292, 2024. 12 A Problem Setting In this appendix we state the CVRP setting used throughout the paper in ...

    work page 2024

  41. [41]

    Seed Generation: hseed(Ψ, d)7→C∈R K×d model, extracting K dynamic seed represen- tations

    work page

  42. [42]

    Global Assignment: hassign(Ψ, d, C)7→(∆, bY) , where the Clustering Transformer (CT) and entropic Optimal Transport (OT) layer output a latent distance matrix ∆∈R N×K and a continuous transport planbY∈[0,1] N×K

    work page

  43. [43]

    OR Decoding: hCAP(∆,bY , d, Q)7→ Y , where the MIP solver enforces uniform fleet capacityQto recover the discrete boolean assignment matrix Y∈ {0,1} N×K

    work page

  44. [44]

    TSP Recovery: hTSP(Y , Xcoord)7→ T , mapping each column of Y to an independent TSP tour over the corresponding customer subset, computed under Euclidean distances derived fromX coord. The full pipeline is the composition C=h seed(Ψ, d), (∆,bY) =h assign(Ψ, d, C), Y=h CAP(∆,bY , d, Q), T=h TSP(Y , Xcoord).(7) We pass Xcoord explicitly to hTSP to make the ...

    work page

  45. [45]

    Phase 1 and 2.Both ST and CT are standard Transformer encoders, with 8 heads and 6 layers, a hidden dimension of 128 (feedforward 512),k attn = 20, dropout= 0.1.L con has a scalar weight of0.05, and all the other losses are not weighted

    work page

  46. [46]

    Phase 2.The OT layer is configured with ϵ= 0.01 and with a maximum iteration of

    work page

  47. [47]

    The full loss used is Ltotal = Lseed + 0.05× L con + LBCE +L OT

    We use ϵ= 0.001 and with a maximum iteration of 1000 for inference but found that this did not have a significant impact on performance. The full loss used is Ltotal = Lseed + 0.05× L con + LBCE +L OT

    work page

  48. [48]

    To ensure standardized evaluation across all decoding variants, Gurobi is configured with a strict time limit of 100s and a target MIPGap of 0.001

    Phase 3.For the OR solver, we use Gurobi version 12.0.3 [15], parallelized across 16 CPU cores per batch. To ensure standardized evaluation across all decoding variants, Gurobi is configured with a strict time limit of 100s and a target MIPGap of 0.001. Furthermore, we define the confidence threshold for hard assignment decoding as τhigh = 0.99 , and the ...

    work page

  49. [49]

    19 Note on Measuring Inference TimeTo properly contextualize the computational metrics reported throughout our results, we first clarify the composition of our inference pipeline

    Weight Saving Strategy.Utilizing the built-in PyTorch [ 28] framework for Stochastic Weight Averaging [1], we initiate the weight averaging process at the midpoint of training (the50%epoch) and systematically save the model states at each subsequent epoch. 19 Note on Measuring Inference TimeTo properly contextualize the computational metrics reported thro...

    work page

  50. [50]

    Standard Benchmark Setting

    on the LEHD [26] benchmark datasets. We found that maintaining the 10% nearest-neighbor sparsity ratio during inference (e.g., dynamically setting K_SPARSE= 50 for N= 500 and 100 for N= 1000 ) yielded the strongest overall generalization performance for GLOP across all evaluated scales. Table 3 shows that GLOP has poor performance on small problem sizes. ...

    work page 2074