Pith. sign in

REVIEW 3 major objections 5 minor 61 references

GELD: A Unified Neural Model for Efficiently Solving Traveling Salesman Problems Across Different Scales

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One pre-trained network solves TSPs from 100 to 744,710 nodes.

desk verdict A genuine architectural contribution with solid middle-scale evidence; the 744k headline is the weakest sentence, but the rest withstands review. read the letter →

arxiv 2506.06634 v1 pith:FY75XL6Y submitted 2025-06-07 cs.AI

Yubin Xiao , Di Wang , Rui Cao , Xuan Wu , Boyang Li , You Zhou This is my paper
classification cs.AI
keywords TravelingSalesmanProblemneuralcombinatorialoptimizationlarge-scalegeneralizationlinearattentionself-improvementlearningre-constructionpost-processingEuclideanTSPunifiedmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that one neural network, trained once, can solve Euclidean Traveling Salesman Problems at many sizes at once—100, 500, 1,000, 5,000, and 10,000 nodes—and scale far beyond that, without the usual divide-and-conquer decomposition. The model, GELD, pairs a lightweight global encoder with a local decoder that only considers the nearest neighbors of the current city, which keeps each decision fast. On uniformly random instances, the same pre-trained checkpoint stays within 0.06 to 2.8 percent of strong reference solutions when beam search and reconstruction are added, and as a post-processor it cuts the gaps of other neural solvers by 35 to 97 percent. Combined with a simple insertion heuristic, it solves instances up to 744,710 nodes, which the paper reports as a first for a non-divide-and-conquer neural model. If true, this means a single checkpoint could replace a zoo of size-specific neural TSP solvers and make neural TSP solving practical for real logistics and routing.

What carries the argument

The load-bearing object is the broad-global-assessment/refined-local-selection loop. The encoder's Region-Average Linear Attention (RALA) divides the normalized node coordinates into $m = m_r m_c = 9$ rectangular regions, forms one proxy embedding per region by averaging the query embeddings of its nodes, and then exchanges information between nodes and proxies through two softmax weight matrices; this costs $\mathcal{O}(n m h)$ time and $\mathcal{O}(n h)$ space with no extra learnable parameters. The decoder, in turn, restricts every step to the available $k$-nearest neighbors of the previously selected node, $k = \min\{k_m, n_t\}$ with $k_m = 100$, and scores those candidates with attention that folds in the distance matrix and uses RMSNorm. The third mechanism is reconstruction: randomly selected sub-tours are fed back through the model after re-normalizing coordinates, which reassigns nodes to different RALA regions and diversifies inputs so that beam search and parallel reconstruction can improve solutions. A two-stage training schedule—supervised on small instances, then self-improvement on larger ones with curriculum scaling—is what lets one checkpoint cover all scales.

What would settle it

Construct a Euclidean instance with two well-separated clusters arranged so that the optimal tour leaves one cluster through a city that is not among the $k$-nearest neighbors of the city it departs from; if GELD's decoder is restricted to those neighbors, it cannot output that tour, and its gap on such instances will stay bounded away from zero regardless of beam-search width or reconstruction iterations.

Watch

Extended reading notes

Core claim

GELD's central claim is that global assessment and local selection can be separated without losing solution quality: a single lightweight attention layer gives the model a coarse view of the whole instance, and a heavier decoder then chooses among only the $k$-nearest neighbors of the last visited city. To make the global view cheap, the encoder uses Region-Average Linear Attention (RALA), which partitions the plane into a fixed $3 \times 3$ grid, averages query embeddings into regional proxies, and performs attention between nodes and proxies in $\mathcal{O}(n)$ time and space instead of quadratic cost. To make the local view effective, the decoder is trained first by supervised learning on up to 100-node instances and then by self-improvement on larger instances up to 1,000 nodes. With these ingredients plus beam search and parallel re-construction, the paper reports average gaps of 0.06 to 2.8 percent on uniform instances from 100 to 10,000 nodes, and shows that the same model improves other solvers' routes by 35 to 97 percent with tens of seconds of extra compute.

Load-bearing premise

The model can only pick the next city from the $k$-nearest neighbors of the current city, so if the optimal tour needs a long edge to a city outside that local window, the model can never produce that edge and its achievable gap is bounded below by that locality error.

Editorial extensions

If this is right

  • A single pre-trained GELD checkpoint can replace separate small-scale and large-scale neural TSP solvers, because it covers 100- to 10,000-node instances with the same parameters.
  • GELD as a post-processor makes weak, fast solvers practical: applying it to their outputs yields 35 to 97 percent gap reductions at the cost of tens of seconds, as demonstrated on real-world benchmark instances.
  • The linear-complexity encoder lets GELD run with large batch sizes on a 24 GB GPU, so instances that make quadratic-attention baselines run out of memory can still be solved.
  • The two-stage training recipe—supervised labels on small instances followed by self-improvement on larger ones—produces cross-scale generalization without expensive ground-truth tours for large TSPs.
  • When combined with a simple random-insertion heuristic, GELD solves extremely large instances up to 744,710 nodes, a scale the paper reports no non-divide-and-conquer neural solver has reached before.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test is to push beyond uniform distributions: because RALA's $3\times3$ grid is fixed in normalized coordinates, adaptive or data-dependent region partitioning could reduce the near-doubling of gaps observed on clustered instances.
  • The $k$-nearest restriction means GELD's ceiling is the locality of the optimal tour; pairing the model with a rare global edge-insertion or 2-opt step could close the remaining gap without much latency.
  • The same broad-assessment/refined-selection pattern could transfer to other routing problems, and the paper itself points to capacitated vehicle routing as the next target.
  • Because GELD's improvement comes from diversifying inputs during reconstruction rather than from model size, the lesson may be that input-space diversification during inference is a cheaper route to better solutions for learned combinatorial solvers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces GELD, a neural TSP solver combining a Global-view Encoder (GE) with a Region-Average Linear Attention (RALA) mechanism and a Local-view Decoder (LD) that restricts the next-node candidate set to the k-nearest neighbors of the current node. A two-stage training scheme first supervises on small instances (n ≤ 100) and then self-improves on larger instances using beam search and parallel reconstruction (PRC). The authors claim that a single pre-trained checkpoint solves Euclidean TSPs from 100 to 10,000 nodes with gaps of 0.06–2.8% (with BS and PRC), outperforms seven state-of-the-art models in solution quality and inference speed, and, when combined with a random-insertion heuristic, produces tours for instances up to 744,710 nodes without divide-and-conquer. The empirical study covers synthetic uniform/clustered/explosion/implosion instances, TSPLIB, National TSPs, and four VLSI instances, with ablations on RALA, the second training stage, the global view, the local view, and the hyperparameters m and k_m.

Significance. If the claims hold, GELD is a meaningful advance: the O(n) encoder complexity, the cross-scale generalization of a single checkpoint, and the demonstrated post-processing gains are all valuable for practical neural TSP solving. The paper ships source code and evaluates against external references (LKH3, Concorde, best-known values), so the headline gaps are not circular. The ablation suite is unusually thorough, covering the major architectural choices and the sensitivity of m and k_m. However, the significance is currently tempered by two issues that are load-bearing for the central claims: the decoder's k-nearest-neighbor mask is a hard representational ceiling that is never quantified, and the claimed superiority over seven SOTA models is established only for the computationally expensive BS+PRC variant, not for the greedy variant that the text sometimes invokes. These issues are fixable with additional analysis and careful qualification, but they must be addressed before the claims can be accepted as stated.

major comments (3)
  1. [Section 4.1, Eq. (9), and Table 3] The decoder restricts the candidate set to the k nearest available nodes of the current node, with k_m=100. This is a hard action mask: any optimal tour edge from a to b where b is not among the k nearest neighbors of a is unreachable by the model, regardless of training, beam search, or reconstruction. The manuscript never quantifies this representational ceiling. The clustered-distribution results in Table 3 are consistent with the mask blocking inter-cluster edges (e.g., GELD+S* at TSP-5000 is 4.27% on clustered versus 2.77% on uniform; GELD+G at TSP-5000 is 11.62% versus 7.62%). The authors should either (a) report, on the test distributions, the fraction of optimal (or LKH3) tour edges that fall inside the k-nearest-neighbor set, (b) ablate with a global candidate set that always includes the destination or a sampled set of far nodes, or (c) explicitly restrict the paper's claims to distributions where optimal tours are locally realizable. Without this, the headline '0.06 to 2.8 percent' gaps are partly an artifact of the action-space restriction rather than a measure of learned policy quality.
  2. [Abstract and Section 5.2, Tables 3 and 4] The abstract and the concluding text of Section 5.2 state that GELD outperforms seven state-of-the-art models considering both solution quality and inference speed, without specifying the inference strategy. In Table 3, the greedy variant GELD+G has an average gap of 4.68% on uniform instances, worse than both UDC (4.07%) and INViT-3V (4.42%); on clustered, explosion, and implosion distributions GELD+G is also worse than INViT-3V. The real-world results in Table 4 are starker: GELD+G has an average gap of 14.39% on National TSPs versus INViT-3V's 8.75%. The claim is only supportable for GELD+S*, which uses beam search with width 16 and 1,000 PRC iterations, giving it substantially more compute than the baselines' default greedy or simple-rollout strategies. The paper must either qualify the claim to the specific inference budget used, or provide a controlled comparison where all methods are given a comparable post-processing/search budget.
  3. [Section 5.3, Table 7, and Abstract] The abstract's statement that 'GELD is shown as capable of solving TSPs with up to 744,710 nodes' should be qualified. The results in Table 7 are for the pipeline 'RI + GELD', where a random-insertion heuristic constructs the initial tour and the neural model repeatedly reconstructs random sub-tours. The neural model alone is not evaluated on these instances, and the text itself describes this as integration with a conventional heuristic. In addition, the claim of being 'first-of-its-kind to solve this large size TSP without relying on divide-and-conquer strategies' is debatable: PRC repeatedly extracts and re-solves sub-solutions, which is structurally similar to a decomposition strategy even if the paper defines D&C more narrowly. The authors should state precisely what GELD alone does and what the hybrid pipeline does, and should temper the 'first-of-its-kind' wording accordingly.
minor comments (5)
  1. [Algorithm 1, line 7] The expression 'epoch ⋅ (nmax − km) ∣ ne2' uses an unclear division symbol; this should be written as an explicit integer division or a floor operation.
  2. [Section 5.1.3] The description of baseline inference settings is incomplete: for Omni-TSP, ELG, and UDC the number of rollouts is not reported, and for INViT-3V the number of data-augmentation repetitions is not specified. This makes the timing comparisons hard to reproduce.
  3. [Table 4] The average gap in Table 4 is computed only over instances that each model successfully solves, and the number of solved instances is not always displayed. This can make a model with frequent memory failures look better than a model that solves all instances; the authors should report the full per-instance coverage in the main table or at least state the count explicitly in every cell.
  4. [Section 4.1, Eq. (10)] The masking operation uses -inf inside the softmax but then declares the output to be in R^k; the notation should clarify that the probability mass is renormalized over the candidate set Kset.
  5. [Section 5.4, Table 6] The ablation 'w/o local view' extends the decoder to a global candidate set and reports OOM at TSP-10000 for S*. This result is interesting, but the discussion could more explicitly interpret it as evidence that the k-restriction is what enables large-scale inference, rather than as evidence about solution quality.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: evaluation is against external references, and the k-NN decoding mask is a representational limitation rather than a self-referential prediction.

full rationale

The paper's reported gaps are computed against external references (LKH3, Concorde, and best-known TSPLIB/World TSP values, Eq. 12, Section 5.1.4), so the headline quality numbers are not defined by the model's own outputs. The k-nearest-neighbor decoding mask (Eq. 9-10) imposes a representational ceiling: tours using edges to nodes outside Kset are unreachable, which explains the larger gaps on clustered instances, but this is a structural limitation, not circularity, because the tour-length objective remains externally defined and the gaps are actually measured. The second-stage SIL training (Algorithm 1) trains on pseudo-labels produced by the model's own beam search and reconstruction; however, these labels are improved (shorter) tours discovered by search, and the final evaluation is against LKH3-based ground truth, so the self-referential training loop does not define the reported metrics. Self-citations (e.g., Xiao et al. 2024a, 2024b, 2025; Wang et al. 2025; Wu et al. 2024) are contextual and non-load-bearing; the core RALA mechanism and decoder design are new and assessed against independent baselines. No equation equates a fitted parameter with a predicted quantity by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The model's headline results depend on hand-chosen hyperparameters (m, k_m, n_max, B, PRC budget) and on three unproven modeling premises: locality of optimal TSP edges (used in the decoder), a fixed spatial grid for global pooling (used in RALA), and the reliability of self-generated pseudo-labels (used in stage-2 training). One additional stated justification, the optimal-sub-structure property of COP solutions, is false as stated but is not needed for reconstruction to work as local search. No new physical or external entities are postulated; the regional proxies of RALA are internal computational constructs without an independent measurable counterpart.

free parameters (6)
  • number of RALA regions m = 9 (mr = mc = 3)
    Chosen in Section 5.1.1 to balance quality and speed; Table 8 shows m=9 is a robust middle point, not the best in every cell.
  • max local selection range k_m = 100
    Chosen in Section 5.1.1; Table 9 shows k_m=150 slightly improves gaps at the cost of speed, so 100 is a compromise.
  • max training size n_max = 1,000
    Chosen in Section 5.1.1 as a training-cost tradeoff; larger values are expected to improve generalization but were not tested.
  • beam width B = 16
    Used in the S* evaluation and in SIL pseudo-label generation (Algorithm 1); affects both quality and inference time.
  • PRC iterations = 1,000
    Fixed for all post-processing results in Tables 2, 3, 5, and 7; the reported gains depend on this budget.
  • second-stage stopping thresholds = t_max=5, epsilon=1e-3, t_imp=3
    Algorithm 1 termination conditions chosen by hand; no analysis justifies these values.
assumptions (4)
  • domain assumption In Euclidean TSP, the next stop in the optimal tour is almost always within the k nearest neighbors of the previous stop.
    Used in Section 4.1 to define the decoder candidate set Kset around Eq. (9). It is plausible empirically but not guaranteed, and clustered-instance gaps in Table 3 are consistent with occasional long optimal edges.
  • ad hoc to paper The optimal solution of a combinatorial optimization problem is composed of optimal sub-solutions.
    Stated in Section 3.3 to justify Re-Construction. In general the optimal TSP tour does not contain the optimal tour for an arbitrary node subset, so the stated justification is stronger than the method requires.
  • domain assumption Pseudo-labels produced by GELD's own BS and PRC search are reliable enough to serve as training targets.
    Stage-2 SIL training (Algorithm 1) relies on these labels; the 'w/o second stage' ablation shows empirical benefit, but no per-label check against LKH3 is reported.
  • ad hoc to paper A fixed 3-by-3 grid on normalized coordinates captures the global structure of any 2D TSP instance.
    RALA in Section 4.1 uses fixed regional proxies; Table 3 shows moderate robustness across distribution patterns, but the grid is not adapted to instance shape.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GELD: A Unified Neural Model for Efficiently Solving Traveling Salesman Problems Across Different Scales." pith.science (2026). https://pith.science/paper/FY75XL6Y

@misc{pith2026250606634,
  author       = {Pith},
  title        = {Pith review of: GELD: A Unified Neural Model for Efficiently Solving Traveling Salesman Problems Across Different Scales},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FY75XL6Y}},
  note         = {Machine review of arXiv:2506.06634}
}
read the original abstract

The Traveling Salesman Problem (TSP) is a well-known combinatorial optimization problem with broad real-world applications. Recent advancements in neural network-based TSP solvers have shown promising results. Nonetheless, these models often struggle to efficiently solve both small- and large-scale TSPs using the same set of pre-trained model parameters, limiting their practical utility. To address this issue, we introduce a novel neural TSP solver named GELD, built upon our proposed broad global assessment and refined local selection framework. Specifically, GELD integrates a lightweight Global-view Encoder (GE) with a heavyweight Local-view Decoder (LD) to enrich embedding representation while accelerating the decision-making process. Moreover, GE incorporates a novel low-complexity attention mechanism, allowing GELD to achieve low inference latency and scalability to larger-scale TSPs. Additionally, we propose a two-stage training strategy that utilizes training instances of different sizes to bolster GELD's generalization ability. Extensive experiments conducted on both synthetic and real-world datasets demonstrate that GELD outperforms seven state-of-the-art models considering both solution quality and inference speed. Furthermore, GELD can be employed as a post-processing method to significantly elevate the quality of the solutions derived by existing neural TSP solvers via spending affordable additional computing time. Notably, GELD is shown as capable of solving TSPs with up to 744,710 nodes, first-of-its-kind to solve this large size TSP without relying on divide-and-conquer strategies to the best of our knowledge.

Figures

Figures reproduced from arXiv: 2506.06634 by the authors.

Figure 1
Figure 1. Framework of our proposed GELD, which incorporates a low complexity architecture with a GE and an LD. Furthermore, the effectiveness of RC is improved by considering the need for diversified inputs in the model architectural design. the current ones. Importantly, RC is fundamentally distinct from the D&C strategy which decomposes a large problem into multiple smaller sub-problems—a process that may lead to suboptima… view at source ↗
Figure 2
Figure 2. Computation process of RALA. it requires meeting the following three criteria: 1) Compre￾hensive coverage of all node information to enable interac￾tion among nodes and facilitate global information transfer; 2) Low computational complexity to ensure scalability to larger-scale TSPs; and 3) The ability to obtain global infor￾mation in a vague manner, allowing for effective diversifying model inputs during the RC pro… view at source ↗
Figure 3
Figure 3. Visualization of solutions on TZ6117 TSP instance with 6117 nodes. Yubin Xiao et al.: Preprint submitted to Elsevier Page 14 of 17 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of solutions on FI10639 TSP instance with 10639 nodes. Yubin Xiao et al.: Preprint submitted to Elsevier Page 15 of 17 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 55 canonical work pages

  1. [1]

    , author Bixby, R.E

    author Applegate, D.L. , author Bixby, R.E. , author Chvátal, V. , author Cook, W.J. , year 2007 . title The Traveling Salesman Problem: A Computational Study . publisher Princeton University Press

    work page 2007

  2. [2]

    , author Ítalo Santana , author Sörensen, K

    author Arnold, F. , author Ítalo Santana , author Sörensen, K. , author Vidal, T. , year 2021 . title PILS : Exploring high-order neighborhoods by pattern mining and injection . journal Pattern Recognition volume 116 , pages 107957

    work page 2021

  3. [3]

    , year 1994

    author Azar, Y. , year 1994 . title Lower bounds for insertion methods for tsp . journal Combinatorics, Probability and Computing volume 3 , pages 285--292

    work page 1994

  4. [4]

    , author Lodi, A

    author Bengio, Y. , author Lodi, A. , author Prouvost, A. , year 2021 . title Machine learning for combinatorial optimization: A methodological tour d'horizon . journal European Journal of Operational Research volume 290 , pages 405--421

    work page 2021

  5. [5]

    , year 2005

    author Blum, C. , year 2005 . title Ant colony optimization: Introduction and recent trends . journal Physics of Life Reviews volume 2 , pages 353--373

    work page 2005

  6. [6]

    , author Kerschke, P

    author Bossek, J. , author Kerschke, P. , author Neumann, A. , author Wagner, M. , author Neumann, F. , author Trautmann, H. , year 2019 . title Evolving diverse TSP instances by means of novel and creative mutation operators , in: booktitle Proceedings of ACM/SIGEVO Conference on Foundations of Genetic Algorithms , pp. pages 58--71

    work page 2019

  7. [7]

    , author Surana, S

    author Chalumeau, F. , author Surana, S. , author Bonnet, C. , author Grinsztajn, N. , author Pretorius, A. , author Laterre, A. , author Barrett, T. , year 2023 . title Combinatorial optimization with policy adaptation using latent space search , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 7947--7959

    work page 2023

  8. [8]

    , author Zheng, H

    author Cheng, H. , author Zheng, H. , author Cong, Y. , author Jiang, W. , author Pu, S. , year 2023 . title Select and optimize: Learning to aolve large-scale tsp instances , in: booktitle Proceedings of International Conference on Artificial Intelligence and Statistics , pp. pages 1219--1231

    work page 2023

Show all 61 references
  1. [9]

    , author Kwon, Y.D

    author Choo, J. , author Kwon, Y.D. , author Kim, J. , author Jae, J. , author Hottung, A. , author Tierney, K. , author Gwon, Y. , year 2022 . title Simulation-guided beam search for neural combinatorial optimization , in: booktitle Proceedings of Advances in Neural Informati...

    2022

  2. [10]

    , author Michel, S

    author Drakulic, D. , author Michel, S. , author Mai, F. , author Sors, A. , author Andreoli, J.M. , year 2023 . title BQ-NCO : Bisimulation quotienting for efficient neural combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Sys...

    2023

  3. [11]

    , author Song, Z

    author Fang, H. , author Song, Z. , author Weng, P. , author Ban, Y. , year 2024 . title INV i T : A generalizable routing problem solver with invariant nested view transformer , in: booktitle Proceedings of International Conference on Machine Learning , pp. pages 12973--12992

    2024

  4. [12]

    , author Qiu, K

    author Fu, Z. , author Qiu, K. , author Zha, H. , year 2021 . title Generalize a small pre-trained model to arbitrarily large TSP instances , in: booktitle Proceedings of the AAAI Conference on Artificial Intelligence , pp. pages 7474--7482

    2021

  5. [13]

    , author Shang, H

    author Gao, C. , author Shang, H. , author Xue, K. , author Li, D. , author Qian, C. , year 2024 . title Towards generalizable neural solvers for vehicle routing problems via ensemble with transferrable local policy , in: booktitle Proceedings of International Joint Conference...

    2024

  6. [14]

    , author Sommer, J

    author Geisler, S. , author Sommer, J. , author Schuchardt, J. , author Bojchevski, A. , author G \"u nnemann, S. , year 2022 . title Generalization of neural combinatorial solvers through the lens of adversarial robustness , in: booktitle Proceedings of International Conferen...

    2022

  7. [15]

    , author Cao, Z

    author Goh, Y.L. , author Cao, Z. , author Ma, Y. , author Dong, Y. , author Dupty, M.H. , author Lee, W.S. , year 2024 . title Hierarchical neural constructive solver for real-world tsp scenarios , in: booktitle Proceedings of ACM SIGKDD Conference on Knowledge Discovery and ...

    2024

  8. [16]

    , author Deville, Y

    author Ha, Q.M. , author Deville, Y. , author Pham, Q.D. , author Hà, M.H. , year 2018 . title On the min-cost traveling salesman problem with drone . journal Transportation Research Part C: Emerging Technologies volume 86 , pages 597--621

    2018

  9. [17]

    , author Gao, C

    author Han, Z. , author Gao, C. , author Liu, J. , author Zhang, J. , author Zhang, S.Q. , year 2024 . title Parameter-efficient fine-tuning for large models: A comprehensive survey . arXiv:2403.14608 http://arxiv.org/abs/2403.14608. note arXiv:2403.14608

    2024 arXiv

  10. [18]

    , year 2017

    author Helsgaun, K. , year 2017 . title An extension of the L in-- K ernighan-- H elsgaun TSP solver for constrained traveling salesman and vehicle routing problems . journal Roskilde: Roskilde University , pages 24--50

    2017

  11. [19]

    , author Kwon, Y.D

    author Hottung, A. , author Kwon, Y.D. , author Tierney, K. , year 2022 . title Efficient active search for combinatorial optimization problems , in: booktitle Proceedings of International Conference on Learning Representations

    2022

  12. [20]

    , author Mahajan, M

    author Hottung, A. , author Mahajan, M. , author Tierney, K. , year 2024 . title Polynet: Learning diverse solution strategies for neural combinatorial optimization . arXiv:2402.14048 http://arxiv.org/abs/2402.14048. note arXiv:2402.14048

    2024

  13. [21]

    , author Yang, J

    author Hou, Q. , author Yang, J. , author Su, Y. , author Wang, X. , author Deng, Y. , year 2023 . title Generalize learned heuristics to solve large-scale vehicle routing problems in real-time , in: booktitle Proceedings of International Conference on Learning Representations

    2023

  14. [22]

    , author Wang, K

    author Huang, Z. , author Wang, K. , author Liu, F. , author Zhen, H.L. , author Zhang, W. , author Yuan, M. , author Hao, J. , author Yu, Y. , author Wang, J. , year 2022 . title Learning to select cuts for efficient mixed-integer programming . journal Pattern Recognition vol...

    2022

  15. [23]

    , author Li, Q

    author Hudson, B. , author Li, Q. , author Malencia, M. , author Prorok, A. , year 2022 . title Graph neural network guided local search for the traveling salesperson problem , in: booktitle Proceedings of International Conference on Learning Representations

    2022

  16. [24]

    Garmendia, A

    author I. Garmendia, A. , author Cappart, Q. , author Ceberio, J. , author Mendiburu, A. , year 2024 . title MARCO : A memory-augmented reinforcement framework for combinatorial optimization , in: booktitle Proceedings of International Joint Conference on Artificial Intelligen...

    2024

  17. [25]

    , author Ding, Y

    author Jin, Y. , author Ding, Y. , author Pan, X. , author He, K. , author Zhao, L. , author Qin, T. , author Song, L. , author Bian, J. , year 2023 . title Pointerformer: Deep reinforced multi-pointer transformer for the traveling salesman problem , in: booktitle Proceedings ...

    2023

  18. [26]

    , author Cappart, Q

    author Joshi, C.K. , author Cappart, Q. , author Rousseau, L.M. , author Laurent, T. , year 2022 . title Learning the travelling salesperson problem requires rethinking generalization . journal Constraints volume 27 , pages 70--98

    2022

  19. [27]

    , author Choi, S

    author Kim, M. , author Choi, S. , author Kim, H. , author Son, J. , author Park, J. , author Bengio, Y. , year 2025 . title Ant colony sampling with GF lownets for combinatorial optimization , in: booktitle Proceedings of International Conference on Artificial Intelligence an...

    2025

  20. [28]

    , author Park, J

    author Kim, M. , author Park, J. , author Park, J. , year 2022 . title Sym- NCO : Leveraging symmetricity for neural combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 1936--1949

    2022

  21. [29]

    , author van Hoof, H

    author Kool, W. , author van Hoof, H. , author Welling, M. , year 2019 . title Attention, learn to solve routing problems! , in: booktitle Proceedings of International Conference on Learning Representations

    2019

  22. [30]

    , author Choo, J

    author Kwon, Y.D. , author Choo, J. , author Kim, B. , author Yoon, I. , author Gwon, Y. , author Min, S. , year 2020 . title POMO : Policy optimization with multiple optima for reinforcement learning , in: booktitle Proceedings of Advances in Neural Information Processing Sys...

    2020

  23. [31]

    , author Choo, J

    author Kwon, Y.D. , author Choo, J. , author Yoon, I. , author Park, M. , author Park, D. , author Gwon, Y. , year 2021 . title Matrix encoding networks for neural combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp....

    2021

  24. [32]

    , author Yan, Z

    author Li, S. , author Yan, Z. , author Wu, C. , year 2021 . title Learning to delegate for large-scale vehicle routing , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 26198--26211

    2021

  25. [33]

    , author Guo, J

    author Li, Y. , author Guo, J. , author Wang, R. , author Yan, J. , year 2023 . title T2 T : From distribution learning in training to gradient search in testing for combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , p...

    2023

  26. [34]

    , author Lin, X

    author Liu, F. , author Lin, X. , author Wang, Z. , author Zhang, Q. , author Xialiang, T. , author Yuan, M. , year 2024 . title Multi-task learning for routing problem with cross-problem zero-shot generalization , in: booktitle Proceedings of ACM SIGKDD Conference on Knowledg...

    2024

  27. [35]

    , author Lin, X

    author Luo, F. , author Lin, X. , author Liu, F. , author Zhang, Q. , author Wang, Z. , year 2023 . title Neural combinatorial optimization with heavy decoder: Toward large scale generalization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , ...

    2023

  28. [36]

    , author Lin, X

    author Luo, F. , author Lin, X. , author Wu, Y. , author Wang, Z. , author Xialiang, T. , author Yuan, M. , author Zhang, Q. , year 2025 . title Boosting neural combinatorial optimization for large-scale vehicle routing problems , in: booktitle Proceedings of International Con...

    2025

  29. [37]

    , author Cao, Z

    author Ma, Y. , author Cao, Z. , author Chee, Y.M. , year 2023 . title Learning to search feasible and infeasible regions of routing problems with flexible neural k-opt , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 49555--49578

    2023

  30. [38]

    , author Bai, Y

    author Min, Y. , author Bai, Y. , author Gomes, C.P. , year 2023 . title Unsupervised learning for solving the travelling salesman problem , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 47264--47278

    2023

  31. [39]

    , author Hansen, P

    author Mladenović, N. , author Hansen, P. , year 1997 . title Variable neighborhood search . journal Computers & Operations Research volume 24 , pages 1097--1100

    1997

  32. [40]

    , author Jin, Y

    author Pan, X. , author Jin, Y. , author Ding, Y. , author Feng, M. , author Zhao, L. , author Song, L. , author Bian, J. , year 2023 . title H- TSP : Hierarchically solving the large-scale traveling salesman problem , in: booktitle Proceedings of the AAAI Conference on Artifi...

    2023

  33. [41]

    , author Grimm, D.G

    author Pirnay, J. , author Grimm, D.G. , year 2024 . title Self-improvement for neural combinatorial optimization: Sample without replacement, but improvement . journal Transactions on Machine Learning Research

    2024

  34. [42]

    , author Sun, Z

    author Qiu, R. , author Sun, Z. , author Yang, Y. , year 2022 . title Dimes: A differentiable meta solver for combinatorial optimization problems , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 25531--25546

    2022

  35. [43]

    , author Hu, Y

    author Sheng, J. , author Hu, Y. , author Zhou, W. , author Zhu, L. , author Jin, B. , author Wang, J. , author Wang, X. , year 2022 . title Learning to schedule multi-numa virtual machines via reinforcement learning . journal Pattern Recognition volume 121 , pages 108254

    2022

  36. [44]

    , author Yang, Y

    author Sun, Z. , author Yang, Y. , year 2023 . title DIFUSCO : Graph-based diffusion solvers for combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 3706--3731

    2023

  37. [45]

    , author Shazeer, N

    author Vaswani, A. , author Shazeer, N. , author Parmar, N. , author Uszkoreit, J. , author Jones, L. , author Gomez, A.N. , author Kaiser, . , author Polosukhin, I. , year 2017 . title Attention is all you need , in: booktitle Proceedings of Advances in Neural Information Pro...

    2017

  38. [46]

    , author Zhou, Y

    author Wang, M. , author Zhou, Y. , author Cao, Z. , author Xiao, Y. , author Wu, X. , author Pang, W. , author Jiang, Y. , author Yang, H. , author Zhao, P. , author Li, Y. , year 2025 . title An efficient diffusion-based non-autoregressive solver for traveling salesman probl...

    2025

  39. [47]

    , author Wang, D

    author Wu, X. , author Wang, D. , author Wen, L. , author Xiao, Y. , author Wu, C. , author Wu, Y. , author Yu, C. , author Maskell, D.L. , author Zhou, Y. , year 2024 . title Neural combinatorial optimization algorithms for solving vehicle routing problems: A comprehensive su...

    2024 arXiv

  40. [48]

    , author Yang, X

    author Xia, Y. , author Yang, X. , author Liu, Z. , author Liu, Z. , author Song, L. , author Bian, J. , year 2024 . title Position: Rethinking post-hoc search-based neural approaches for solving large-scale traveling salesman problems , in: booktitle Proceedings of Internatio...

    2024

  41. [49]

    , author Wang, D

    author Xiao, Y. , author Wang, D. , author Li, B. , author Chen, H. , author Pang, W. , author Wu, X. , author Li, H. , author Xu, D. , author Liang, Y. , author Zhou, Y. , year 2024 a. title Reinforcement learning-based non-autoregressive solver for traveling salesman problem...

    2024

  42. [50]

    , author Wang, D

    author Xiao, Y. , author Wang, D. , author Li, B. , author Wang, M. , author Wu, X. , author Zhou, C. , author Zhou, Y. , year 2024 b. title Distilling autoregressive models to obtain high-performance non-autoregressive solvers for vehicle routing problems with faster inferenc...

    2024

  43. [51]

    , author Wang, D

    author Xiao, Y. , author Wang, D. , author Wu, X. , author Wu, Y. , author Li, B. , author Du, W. , author Wang, L. , author Zhou, Y. , year 2025 . title Improving generalization of neural vehicle routing problem solvers through the lens of model architecture . journal Neural ...

    2025

  44. [52]

    , author Weng, J

    author Yan, D. , author Weng, J. , author Huang, S. , author Li, C. , author Zhou, Y. , author Su, H. , author Zhu, J. , year 2022 . title Deep reinforcement learning with credit assignment for combinatorial optimization . journal Pattern Recognition volume 124 , pages 108466

    2022

  45. [53]

    , author Wang, J

    author Ye, H. , author Wang, J. , author Cao, Z. , author Liang, H. , author Li, Y. , year 2023 . title Deep ACO : Neural-enhanced ant systems for combinatorial optimization , in: booktitle Proceedings of Advances in Neural Information Processing Systems , pp. pages 43706--43728

    2023

  46. [54]

    , author Wang, J

    author Ye, H. , author Wang, J. , author Liang, H. , author Cao, Z. , author Li, Y. , author Li, F. , year 2024 . title GLOP : Learning global partition and local construction for solving large-scale routing problems in real-time , in: booktitle Proceedings of the AAAI Confere...

    2024

  47. [55]

    , author Zhao, H

    author Yu, K. , author Zhao, H. , author Huang, Y. , author Yi, R. , author Xu, K. , author Zhu, C. , year 2024 . title DISCO : Efficient diffusion solver for large-scale combinatorial optimization problems . arXiv:2406.19705 http://arxiv.org/abs/2406.19705. note arXiv:2406.19705

    2024 arXiv

  48. [56]

    , author Sennrich, R

    author Zhang, B. , author Sennrich, R. , year 2019 . title Root mean square layer normalization , in: booktitle Proceedings of Advances in Neural Information Processing Systems

    2019

  49. [57]

    , author Zhou, C

    author Zheng, Z. , author Zhou, C. , author Tong, X. , author Yuan, M. , author Wang, Z. , year 2024 . title UDC : A unified neural divide-and-conquer framework for large-scale combinatorial optimization problems , in: booktitle Proceedings of Advances in Neural Information Pr...

    2024

  50. [58]

    , author Lin, X

    author Zhou, C. , author Lin, X. , author Wang, Z. , author Tong, X. , author Yuan, M. , author Zhang, Q. , year 2024 . title Instance-conditioned adaptation for large-scale generalization of neural combinatorial optimization . arXiv:2405.01906 http://arxiv.org/abs/2405.01906....

    2024 arXiv

  51. [59]

    , author Wu, Y

    author Zhou, J. , author Wu, Y. , author Song, W. , author Cao, Z. , author Zhang, J. , year 2023 . title Towards omni-generalizable neural methods for vehicle routing problems , in: booktitle Proceedings of International Conference on Machine Learning , pp. pages 42769--42789

    2023

  52. [60]

    , author Wang, H

    author Zong, Z. , author Wang, H. , author Wang, J. , author Zheng, M. , author Li, Y. , year 2022 . title Rbg: Hierarchically solving large-scale routing problems in logistic systems via reinforcement learning , in: booktitle Proceedings of ACM SIGKDD Conference on Knowledge ...

    2022

  53. [61]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.