Pith. sign in

REVIEW 3 major objections 5 minor 57 references

A General Neural Backbone for Mixed-Integer Linear Optimization via Dual Attention

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read An attention-driven, element-centric neural backbone for mixed-integer linear programming consistently outperforms the standard graph-neural-network backbone across instance-level, element-level, and solving-state-level tasks.

desk verdict A solid, usable attention backbone for MILP that mostly beats BGNN on prediction and branching, but the PaS solving gains are partly confounded by per-method search hyperparameters, so the strongest headline needs a tighter control. read the letter →

arxiv 2601.04509 v2 pith:NZYRFEVG submitted 2026-01-08 cs.AI

Peixin Huang , Yaoxin Wu , Yining Ma , Cathy Wu , Wei Zhang , Wen Song This is my paper · ORCID
classification cs.AI MSC 90C1168T0790C27
keywords mixed-integerlinearprogrammingneuralcombinatorialoptimizationdualattentionelement-centricrepresentationgraphnetworksbranch-and-boundpredict-and-searchlearning
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 show that the dominant way of representing mixed-integer linear programs (MILPs) for machine learning — as bipartite graphs processed by graph neural networks — is not the best possible foundation. The authors propose an element-centric view in which variables and constraints are two sets of equal-status elements, and a dual-attention mechanism that performs self-attention within each set and cross-attention between the sets, running the two streams in parallel. They claim this backbone consistently outperforms graph-based architectures on three levels of MILP learning: predicting instance properties, predicting optimal binary variable values, and learning branching policies. If right, attention-based encoders would become the stronger default choice for learning-enhanced combinatorial optimization, enabling deeper networks, better long-range reasoning, and improved solver integration.

What carries the argument

The key object is the dual-attention layer, which combines (1) intra-type self-attention over variables and over constraints separately, linearized with a kernel trick to keep time and memory linear in the number of elements, and (2) inter-type cross-attention between variables and constraints, made sparse by restricting attention to the nonzero entries of the constraint matrix. The two attention streams run in parallel and are then fused with residual connections and normalization. This mechanism is what removes the locality bottleneck: instead of passing messages only between immediate neighbors, every element can directly attend to any other element in a single layer, while the sparse cro

What would settle it

Re-run the predict-and-search experiments with identical fix-and-correct budgets for the attention-based and graph-based prediction models on the same 100 test instances; if the primal-gap and primal-integral differences shrink to zero, the claimed architectural advantage is an artifact of the tuning procedure rather than the backbone.

Watch

Extended reading notes

Core claim

The central claim is that an attention-driven, element-centric backbone produces more expressive variable and constraint embeddings than the conventional graph-message-passing backbone, and that this advantage transfers to downstream MILP tasks. Across synthetic and real-world benchmarks, the dual-attention model improves prediction accuracy — lower instance-level error, higher binary-variable classification metrics — and yields more confident, separated probability estimates. Mechanism analyses show that attention maintains non-negligible influence from nodes seven or more hops away, where graph influence decays to zero, and that the attention model can be stacked to four or more layers wit

Load-bearing premise

The comparison assumes the per-method tuning of downstream search hyperparameters (how many variables to fix and how aggressively to correct) is not what drives the reported improvements; if a graph-based model with equally-tuned search parameters matched the attention model's solving performance, the backbone would not be the cause.

Editorial extensions

If this is right

  • Attention-based backbones can serve as drop-in replacements for graph encoders across MILP learning tasks, with consistent gains in prediction quality and downstream solving performance.
  • Deeper encoders (four or more layers) become usable without the over-smoothing and gradient-decay that makes deep graph networks unstable, so representation quality improves with depth.
  • Better binary-variable predictions translate into better primal heuristics: predict-and-search frameworks find better solutions faster, and on several benchmarks surpass the solver's default behavior.
  • Learning-to-branch policies guided by attention embeddings produce smaller branch-and-bound trees, with the advantage growing on larger, out-of-distribution instances.
  • The architecture generalizes out-of-distribution to heterogeneous real-world MILP instances, suggesting robustness beyond the training distribution.

Reading between the lines

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

  • A direct test outside the paper: apply the same dual-attention backbone to other solver-side decisions such as cutting-plane selection or node selection; the paper only demonstrates branching, so its claim of a 'general backbone' for solver decisions is plausible but untested.
  • The element-centric abstraction is not specific to MILP; variable-constraint systems also appear in satisfiability and constraint programming, so the same dual-attention design could be evaluated there, though the paper only reports MILP results.
  • The linearized self-attention trades exact attention for efficiency; on extremely dense or small instances the approximation could deviate from full attention, and the paper does not quantify this approximation gap.
  • If the reported gains hold under strictly matched search budgets, the implication is that an encoder's global receptive field — not the downstream heuristic's tuning — is the decisive factor for learning-based MILP performance.
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 proposes an attention-driven neural backbone for mixed-integer linear programming (MILP). Instead of the standard bipartite-graph message passing of BGNNs, the authors treat variables and constraints as two element types and apply parallel intra-type self-attention and inter-type cross-attention. The self-attention is linearized with a sigmoid kernel and the cross-attention is sparsified using the nonzero coefficient matrix A, yielding O((n+m+e)d^2) time and O((n+m+e)d) memory per layer. The backbone is evaluated on three task groups: instance-level feasibility and objective prediction, element-level binary-variable prediction integrated into the PaS and Apollo predict-and-search frameworks, and solving-state-level branching-policy imitation learning. Additional experiments on MIPLIB test out-of-distribution generalization. The paper reports consistent improvements over BGNN-based baselines across these tasks along with mechanism analyses and ablations.

Significance. If the empirical claims are substantiated, the paper provides a useful new architectural foundation for learning-enhanced MILP, with favorable scaling behavior compared with dense attention. The strengths are the breadth of downstream tasks, the inclusion of MIPLIB generalization tests, public code, and architectural ablations that isolate the contributions of self- and cross-attention. The main limitation is that several load-bearing comparisons are confounded or lack uncertainty quantification, so the central 'consistently outperforms' claim is not yet fully supported.

major comments (3)
  1. [§4.2, Appendix D.2 (Table D5)] The PaS hyperparameters were tuned separately for BGNN and for the proposed model. For MIS, PaS-BGNN uses (k0,k1,Delta)=(300,300,40) while PaS-Ours uses (400,400,40); for CA, (40,0,40) vs (50,0,40). In predict-and-search, k0/k1 fix more variables, which can reduce search tree size and improve PG/PI independently of embedding quality. The reported 50% PI improvement on MIS (Table 4) could therefore be a search-configuration artifact rather than a backbone effect. The Apollo arm uses identical parameters, but the Apollo gains are small (e.g., WA PI 2.75 vs 2.73, PG 0.020% vs 0.019%) and are reported without variance. Please add a controlled PaS comparison with identical hyperparameters for both models, and/or a sensitivity analysis over k0,k1,Delta for the BGNN baseline.
  2. [§4.2, Tables 3–4; §4.3, Tables 5–6] Most performance tables report point estimates without repeated trials, confidence intervals, or significance tests. At the solving-state level, acc@1 gaps are small (e.g., SC 70.35 vs 71.51; CA 70.07 vs 70.20), and the tree-size comparisons in Table 6 have largely overlapping ± ranges (SC: 65.4±11% vs 64.9±10%; CFL: 350.4±43% vs 323.1±39%). Without multiple seeds or statistical tests, the claim of consistent superiority is not established for these tasks. Please report standard deviations across at least 5 independent training runs and appropriate significance tests, or temper the claim.
  3. [§4.5, Table 9] The real-world generalization claim is weakened by a metric in the same table: Ours-CA has a higher Error-Num/Error-Rate than BGNN-CA (405,113/10.29% vs 316,100/8.03%), and Ours-WA has a higher MSE (0.5174 vs 0.2412). The text explains the CA error discrepancy by label imbalance, but the WA MSE discrepancy is not addressed. Since the paper elsewhere emphasizes all-metric improvements, please reconcile these entries or justify why MCC/Bal-Acc/F1 should be prioritized over MSE in this table.
minor comments (5)
  1. [Abstract] Grammar: 'MILP ... are central' should be 'MILP ... is central'; also standardize capitalization and commas in the opening sentence.
  2. [Fig. 3] The legend uses 'GCN+PaS' and 'GCN+Apollo', but the text and Tables 3–4 refer to BGNN. Please make the notation consistent.
  3. [§4.4.1] The text says 'plot the average Macro-F1 score with standard deviation in Fig. 4d and 4f', but panels (d) and (h) are the Macro-F1 plots; panel (f) is the t-SNE visualization. Please correct the cross-reference.
  4. [Appendix D.2] Please describe how the PaS hyperparameters were tuned (e.g., grid search, validation metric, number of trials) and state explicitly in the main text that Apollo uses identical parameters for both models.
  5. [§3.1, Eqs. (5)–(6)] The sigmoid-based linear attention is asserted to approximate standard softmax self-attention, but no validation or error analysis is provided. A short discussion, or a small-scale comparison with standard softmax, would help readers assess the approximation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical architecture comparison against external baselines; no fitted constant is relabeled as a prediction.

full rationale

The paper's central claim is empirical: a dual-attention backbone outperforms BGNN on instance-, element-, and solving-state-level MILP tasks. There is no derivation chain whose output equals its own input. Instance-level results use synthetic instances and SCIP labels; element-level prediction accuracy (Table 2) is a direct comparison of predicted binary-variable values against optimal solutions with identical training data and features, before any predict-and-search hyperparameters enter. The PaS/Apollo downstream results (Tables 3–4) do tune search hyperparameters per method (Appendix D.2, Tables D5–D6), which is a benchmark-validity concern about whether the search configuration rather than the embeddings drives the primal-gap improvements, but it is not a circular reduction: k0, k1, and Δ are search parameters, not fitted predictions renamed as results, and the backbone superiority is independently supported by Table 2, branching Tables 5–6, ablations Tables 7–8, and the MIPLIB generalization Table 9. Self-citations (e.g., [23], [31]) appear only in survey-style reference ranges and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. The long-range attribution and gradient analyses are post-hoc explanations, not part of a derivation that presupposes the conclusion. Therefore, no specific circular step can be quoted per the hard rules, and the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on model hyperparameters chosen by hand, downstream search hyperparameters tuned per method, the validity of a linear-attention approximation, and the assumption that sparse coefficient-masked cross-attention is sufficient. No new physical or mathematical entities are postulated.

free parameters (5)
  • Number of dual-attention layers T = 4
    Chosen for a balance between representation power and complexity (Section 4); not derived.
  • Attention heads H = 2
    Chosen hyperparameter; not derived or justified by a study.
  • Embedding dimension d = 64
    Set to match the BGNN baseline; chosen, not derived.
  • PaS/Apollo hyperparameters k0, k1, Delta = Varies per dataset and method; Tables D5-D6
    Tuned separately for BGNN and the attention model; downstream PG/PI improvements mix backbone effect with search configuration.
  • Epsilon in attention denominator = 1e-8
    Small constant to avoid zero denominators; not central to the claim.
assumptions (4)
  • ad hoc to paper Sigmoid-based linear attention approximates standard softmax self-attention well enough for MILP representation learning.
    Invoked in Eqs. (5)-(6) without an error bound or ablation against full attention; all experiments use this approximation.
  • domain assumption Cross-type attention restricted to the support of A preserves sufficient structural information.
    Eqs. (11)-(12) mask cross-attention by the sparse matrix A, so inter-type interactions are not actually global despite the prose claiming global information exchange.
  • domain assumption A two-layer BGNN is the correct baseline; deeper BGNN is unusable.
    Section 4.4.1 shows BGNN degrades beyond two layers, so the comparison uses 2-layer BGNN versus 4-layer attention, leaving a capacity mismatch.
  • domain assumption Gurobi/SCIP within their time limits provide correct supervised labels and best-known values.
    Used throughout Section 4 for label collection, BKV computation, and evaluation; time-limited solver outputs are treated as ground truth.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A General Neural Backbone for Mixed-Integer Linear Optimization via Dual Attention." pith.science (2026). https://pith.science/paper/NZYRFEVG

@misc{pith2026260104509,
  author       = {Pith},
  title        = {Pith review of: A General Neural Backbone for Mixed-Integer Linear Optimization via Dual Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZYRFEVG}},
  note         = {Machine review of arXiv:2601.04509}
}
read the original abstract

Mixed-integer linear programming (MILP) is a foundational framework for combinatorial optimization across science and engineering, but remains hard to solve at scale due to NP-hardness. Recent learning-based methods typically model MILP instances as variable-constraint bipartite graphs and use Graph Neural Networks (GNNs) for representation learning, yet their locality limits representation power. We propose an attention-driven neural backbone that adopts an element-centric view of variables and constraints, with dual attention performing parallel intra-type self-attention and inter-type cross-attention. Across three representative tasks at the instance, element, and solving-state levels, our model consistently outperforms conventional GNN-based architectures, highlighting attention-based, element-centric modeling as a powerful foundation for learning-enhanced combinatorial optimization.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 1 linked inside Pith

  1. [1]

    Nature645, 354–361 (2025)

    Kalinin, K.P., Gladrow, J., Chu, J., Clegg, J.H., Cletheroe, D., Kelly, D.J., Rah- mani, B., Brennan, G., Canakci, B., Falck, F.,et al.: Analog optical computer for ai inference and combinatorial optimization. Nature645, 354–361 (2025)

    2025

  2. [2]

    Nature625(7995), 468–475 (2024)

    Romera-Paredes, B., Barekatain, M., Novikov, A., Balog, M., Kumar, M.P., Dupont, E., Ruiz, F.J., Ellenberg, J.S., Wang, P., Fawzi, O.,et al.: Mathematical discoveries from program search with large language models. Nature625(7995), 468–475 (2024)

    2024

  3. [3]

    Nature Machine Intelligence6(6), 589–604 (2024)

    Du, Y., Jamasb, A.R., Guo, J., Fu, T., Harris, C., Wang, Y., Duan, C., Li` o, P., Schwaller, P., Blundell, T.L.: Machine learning-aided generative molecular design. Nature Machine Intelligence6(6), 589–604 (2024)

    2024

  4. [4]

    Nature Communications11(1), 2446 (2020)

    Naseri, G., Koffas, M.A.: Application of combinatorial optimization strategies in synthetic biology. Nature Communications11(1), 2446 (2020)

    2020

  5. [5]

    Nature594(7862), 207–212 (2021) 29

    Mirhoseini, A., Goldie, A., Yazgan, M., Jiang, J.W., Songhori, E., Wang, S., Lee, Y.-J., Johnson, E., Pathak, O., Nova, A.,et al.: A graph placement methodology for fast chip design. Nature594(7862), 207–212 (2021) 29

    2021

  6. [6]

    European Journal of Operational Research 290(2), 405–421 (2021)

    Bengio, Y., Lodi, A., Prouvost, A.: Machine learning for combinatorial optimiza- tion: a methodological tour d’horizon. European Journal of Operational Research 290(2), 405–421 (2021)

    2021

  7. [7]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Li, S., Kulkarni, J., Menache, I., Wu, C., Li, B.: Towards foundation models for mixed integer linear programming. In: The Thirteenth International Conference on Learning Representations (2025)

    2025

  8. [8]

    constraint programming for shop scheduling problems: new results and outlook

    Naderi, B., Ruiz, R., Roshanaei, V.: Mixed-integer programming vs. constraint programming for shop scheduling problems: new results and outlook. INFORMS Journal on Computing35(4), 817–843 (2023)

    2023

Show all 57 references
  1. [9]

    Transportation research part B: Methodological168, 124–150 (2023)

    Zhou, H., Qin, H., Cheng, C., Rousseau, L.-M.: An exact algorithm for the two- echelon vehicle routing problem with drones. Transportation research part B: Methodological168, 124–150 (2023)

    2023

  2. [10]

    Operations Research71(6), 2130–2145 (2023)

    Ararat, C ¸ ., Meimanjan, N.: Computation of systemic risk measures: a mixed- integer programming approach. Operations Research71(6), 2130–2145 (2023)

    2023

  3. [11]

    Nucleic Acids Research 44(10), 93–93 (2016)

    Poos, A.M., Maicher, A., Dieckmann, A.K., Oswald, M., Eils, R., Kupiec, M., Luke, B., K¨ onig, R.: Mixed integer linear programming based machine learn- ing approach identifies regulators of telomerase in yeast. Nucleic Acids Research 44(10), 93–93 (2016)

    2016

  4. [12]

    Operations Research73(2), 1011–1028 (2025)

    Maragno, D., Wiberg, H., Bertsimas, D., Birbil, S ¸.˙I., Hertog, D., Fajemisin, A.O.: Mixed-integer optimization with constraint learning. Operations Research73(2), 1011–1028 (2025)

    2025

  5. [13]

    Computers & Chemical Engineering185, 108660 (2024)

    McDonald, T., Tsay, C., Schweidtmann, A.M., Yorke-Smith, N.: Mixed-integer optimisation of graph neural networks for computer-aided molecular design. Computers & Chemical Engineering185, 108660 (2024)

    2024

  6. [14]

    European Journal of Operational Research324(3), 707–731 (2025)

    Clautiaux, F., Ljubic, I.: Last fifty years of integer linear programming: A focus on recent practical advances. European Journal of Operational Research324(3), 707–731 (2025)

    2025

  7. [15]

    Mathematical Programming, 1–44 (2024)

    Scavuzzo, L., Aardal, K., Lodi, A., Yorke-Smith, N.: Machine learning aug- mented branch and bound for mixed integer linear programming. Mathematical Programming, 1–44 (2024)

    2024

  8. [16]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Ding, J.-Y., Zhang, C., Shen, L., Li, S., Wang, B., Xu, Y., Song, L.: Accelerating primal solution findings for mixed integer programs based on solution prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 1452–1459 (2020)

    2020

  9. [17]

    arXiv preprint arXiv:2012.13349 (2020)

    Nair, V., Bartunov, S., Gimeno, F., Von Glehn, I., Lichocki, P., Lobov, I., O’Donoghue, B., Sonnerat, N., Tjandraatmadja, C., Wang, P., et al.: Solving 30 mixed integer programs using neural networks. arXiv preprint arXiv:2012.13349 (2020)

    2012 arXiv

  10. [18]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Khalil, E.B., Morris, C., Lodi, A.: Mip-gnn: A data-driven framework for guid- ing combinatorial solvers. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pp. 10219–10227 (2022)

    2022

  11. [19]

    In: The Eleventh International Conference on Learning Representations (2023)

    Han, Q., Yang, L., Chen, Q., Zhou, X., Zhang, D., Wang, A., Sun, R., Luo, X.: A gnn-guided predict-and-search framework for mixed-integer linear programming. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  12. [20]

    In: International Conference on Machine Learning, pp

    Huang, T., Ferber, A.M., Zharmagambetov, A., Tian, Y., Dilkina, B.: Contrastive predict-and-search for mixed integer linear programs. In: International Conference on Machine Learning, pp. 19757–19771 (2024). PMLR

    2024

  13. [21]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Liu, H., Wang, J., Geng, Z., Li, X., Zong, Y., Zhu, F., Hao, J., Wu, F.: Apollo-milp: An alternating prediction-correction neural solving framework for mixed-integer linear programming. In: The Thirteenth International Conference on Learning Representations (2025)

    2025

  14. [22]

    Advances in Neural Information Processing Systems33, 20012–20023 (2020)

    Song, J., Yue, Y., Dilkina, B.,et al.: A general large neighborhood search framework for solving integer linear programs. Advances in Neural Information Processing Systems33, 20012–20023 (2020)

    2020

  15. [23]

    Advances in Neural Information Processing Systems 34, 30075–30087 (2021)

    Wu, Y., Song, W., Cao, Z., Zhang, J.: Learning large neighborhood search policy for integer programming. Advances in Neural Information Processing Systems 34, 30075–30087 (2021)

    2021

  16. [24]

    In: International Conference on Machine Learning, pp

    Huang, T., Ferber, A.M., Tian, Y., Dilkina, B., Steiner, B.: Searching large neigh- borhoods for integer linear programs with contrastive learning. In: International Conference on Machine Learning, pp. 13869–13890 (2023). PMLR

    2023

  17. [25]

    Advances in neural information processing systems32, 15580–15592 (2019)

    Gasse, M., Ch´ etelat, D., Ferroni, N., Charlin, L., Lodi, A.: Exact combinato- rial optimization with graph convolutional neural networks. Advances in neural information processing systems32, 15580–15592 (2019)

    2019

  18. [26]

    Advances in Neural Information Processing Systems33, 18087–18097 (2020)

    Gupta, P., Gasse, M., Khalil, E., Mudigonda, P., Lodi, A., Bengio, Y.: Hybrid models for learning to branch. Advances in Neural Information Processing Systems33, 18087–18097 (2020)

    2020

  19. [27]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Zarpellon, G., Jo, J., Lodi, A., Bengio, Y.: Parameterizing branch-and-bound search trees to learn branching policies. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 3931–3939 (2021)

    2021

  20. [28]

    Transactions on Machine Learning Research (2022) 31

    Gupta, P., Khalil, E.B., Ch´ etelat, D., Gasse, M., Lodi, A., Bengio, Y., Kumar, M.P.: Lookback for learning to branch. Transactions on Machine Learning Research (2022) 31

    2022

  21. [29]

    Advances in Neural Information Processing Systems35, 18514–18526 (2022)

    Scavuzzo, L., Chen, F., Ch´ etelat, D., Gasse, M., Lodi, A., Yorke-Smith, N., Aardal, K.: Learning to branch with tree mdps. Advances in Neural Information Processing Systems35, 18514–18526 (2022)

    2022

  22. [30]

    In: The Twelfth International Conference on Learning Representations (2024)

    Zhang, C., Ouyang, W., Yuan, H., Gong, L., Sun, Y., Guo, Z., Dong, Z., Yan, J.: Towards imitation learning to branch for mip: A hybrid reinforcement learning based sample augmentation approach. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  23. [31]

    In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp

    Sun, Y., Wang, K., Hu, Z., Wu, R., Wu, Y., Song, W., Shen, X., Lv, T., Fan, C.: Mgmatch: Fast matchmaking with nonlinear objective and constraints via multimodal deep graph learning. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. ...

    2024

  24. [32]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Feng, S., Yang, Y.: Sorrel: Suboptimal-demonstration-guided reinforcement learn- ing for learning to branch. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 11212–11220 (2025)

    2025

  25. [33]

    In: International Conference on Machine Learning, pp

    Tang, Y., Agrawal, S., Faenza, Y.: Reinforcement learning for integer program- ming: Learning to cut. In: International Conference on Machine Learning, pp. 9367–9376 (2020). PMLR

    2020

  26. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence46(12), 9697–9713 (2024)

    Wang, J., Wang, Z., Li, X., Kuang, Y., Shi, Z., Zhu, F., Yuan, M., Zeng, J., Zhang, Y., Wu, F.: Learning to cut via hierarchical sequence/set model for effi- cient mixed-integer programming. IEEE Transactions on Pattern Analysis and Machine Intelligence46(12), 9697–9713 (2024)

    2024

  27. [35]

    In: International Conference on Machine Learning, pp

    Paulus, M.B., Zarpellon, G., Krause, A., Charlin, L., Maddison, C.: Learning to cut by looking ahead: Cutting plane selection via imitation learning. In: International Conference on Machine Learning, pp. 17584–17600 (2022). PMLR

    2022

  28. [36]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Ling, H., Wang, Z., Wang, J.: Learning to stop cut generation for efficient mixed- integer linear programming. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 20759–20767 (2024)

    2024

  29. [37]

    In: International Conference on Machine Learning, pp

    Puigdemont, P., Skoulakis, S., Chrysos, G., Cevher, V.: Learning to remove cuts in integer linear programming. In: International Conference on Machine Learning, pp. 41235–41255 (2024). PMLR

    2024

  30. [38]

    Journal of Machine Learning Research24(130), 1–61 (2023)

    Cappart, Q., Ch´ etelat, D., Khalil, E.B., Lodi, A., Morris, C., Veliˇ ckovi´ c, P.: Com- binatorial optimization and reasoning with graph neural networks. Journal of Machine Learning Research24(130), 1–61 (2023)

    2023

  31. [39]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Li, Q., Han, Z., Wu, X.-M.: Deeper insights into graph convolutional networks for semi-supervised learning. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32 (2018) 32

    2018

  32. [40]

    In: International Conference on Learning Representations (2022)

    Topping, J., Di Giovanni, F., Chamberlain, B.P., Dong, X., Bronstein, M.M.: Understanding over-squashing and bottlenecks on graphs via curvature. In: International Conference on Learning Representations (2022)

    2022

  33. [41]

    In: International Conference on Learning Representations (2023)

    Chen, Z., Liu, J., Wang, X., Lu, J., Yin, W.: On representing mixed-integer linear programs by graph neural networks. In: International Conference on Learning Representations (2023)

    2023

  34. [42]

    Advances in Neural Information Processing Systems37, 123991–124024 (2024)

    Chen, Z., Liu, J., Chen, X., Wang, W., Yin, W.: Rethinking the capacity of graph neural networks for branching strategy. Advances in Neural Information Processing Systems37, 123991–124024 (2024)

    2024

  35. [43]

    In: Forty-second International Conference on Machine Learning (2025)

    Chen, Z., Chen, X., Liu, J., Wang, X., Yin, W.: Expressive power of graph neural networks for (mixed-integer) quadratic programs. In: Forty-second International Conference on Machine Learning (2025)

    2025

  36. [44]

    Advances in Neural Information Processing Systems30, 6000–6010 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems30, 6000–6010 (2017)

    2017

  37. [45]

    In: The Twelfth International Conference on Learning Representations (2024)

    Deng, C., Yue, Z., Zhang, Z.: Polynormer: Polynomial-expressive graph trans- former in linear time. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  38. [46]

    Technical report, Optimization Online (February 2024)

    Bolusani, S., Besan¸ con, M., Bestuzheva, K., Chmiela, A., Dion ´ ısio, J., Donkiewicz, T., Doornmalen, J., Eifler, L., Ghannam, M., Gleixner, A., Graczyk, C., Halbig, K., Hedtke, I., Hoen, A., Hojny, C., Hulst, R., Kamp, D., Koch, T., Kofler, K., Lentz, J., Manns, J., Mexi, G...

    2024

  39. [47]

    In: NeurIPS 2021 Competitions and Demonstrations Track, pp

    Gasse, M., Bowly, S., Cappart, Q., Charfreitag, J., Charlin, L., Ch´ etelat, D., Chmiela, A., Dumouchelle, J., Gleixner, A., Kazachkov, A.M.,et al.: The machine learning for combinatorial optimization competition (ml4co): Results and insights. In: NeurIPS 2021 Competitions and...

    2021

  40. [48]

    In: Learning Meets Combinatorial Algorithms at NeurIPS2020 (2020)

    Prouvost, A., Dumouchelle, J., Scavuzzo, L., Gasse, M., Ch´ etelat, D., Lodi, A.: Ecole: A gym-like library for machine learning in combinatorial optimization solvers. In: Learning Meets Combinatorial Algorithms at NeurIPS2020 (2020)

    2020

  41. [49]

    https://www.gurobi.com

    LLC Gurobi Optimization: Gurobi Optimizer (2021). https://www.gurobi.com

    2021

  42. [50]

    Operations Research Letters41(6), 611–614 (2013) 33

    Berthold, T.: Measuring the impact of primal heuristics. Operations Research Letters41(6), 611–614 (2013) 33

    2013

  43. [51]

    In: International Conference on Learning Representations (2018)

    Ancona, M., Ceolini, E., ¨Oztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. In: International Conference on Learning Representations (2018)

    2018

  44. [52]

    Brody, S., Alon, U., Yahav, E.: How attentive are graph attention networks? In: International Conference on Learning Representations (2022)

    2022

  45. [53]

    Mathematical Programming Computation13(3), 443–490 (2021)

    Gleixner, A., Hendel, G., Gamrath, G., Achterberg, T., Bastubbe, M., Berthold, T., Christophel, P., Jarck, K., Koch, T., Linderoth, J.,et al.: Miplib 2017: data- driven compilation of the 6th mixed-integer programming library. Mathematical Programming Computation13(3), 443–490 (2021)

    2017

  46. [54]

    Bergman, D., Cire, A.A., Van Hoeve, W.-J., Hooker, J.: Decision Diagrams for Optimization vol. 1. Springer, Berlin, Heidelberg (2016)

    2016

  47. [55]

    In: Proceedings of the 2nd ACM Conference on Electronic Commerce, pp

    Leyton-Brown, K., Pearson, M., Shoham, Y.: Towards a universal test suite for combinatorial auction algorithms. In: Proceedings of the 2nd ACM Conference on Electronic Commerce, pp. 66–76 (2000)

    2000

  48. [56]

    In: Padberg, M.W

    Balas, E., Ho, A.: Set covering algorithms using cutting planes, heuristics, and subgradient optimization: a computational study. In: Padberg, M.W. (ed.) Combinatorial Optimization, pp. 37–60. Springer, Berlin, Heidelberg (2009)

    2009

  49. [57]

    European Journal of Operational Research50(3), 280–297 (1991) 34

    Cornu´ ejols, G., Sridharan, R., Thizy, J.-M.: A comparison of heuristics and relaxations for the capacitated plant location problem. European Journal of Operational Research50(3), 280–297 (1991) 34

    1991

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.