pith. sign in

arxiv: 2510.19544 · v2 · submitted 2025-10-22 · ❄️ cond-mat.dis-nn · cond-mat.stat-mech· cs.AI· cs.LG· physics.comp-ph

Demonstrating Real Advantage of Machine-Learning-Enhanced Monte Carlo for Combinatorial Optimization

Luca Maria Del Bono , Federico Ricci-Tersenghi , Francesco Zamponi This is my paper

Pith reviewed 2026-05-18 04:49 UTC · model grok-4.3

classification ❄️ cond-mat.dis-nn cond-mat.stat-mechcs.AIcs.LGphysics.comp-ph
keywords combinatorial optimizationmachine learningMonte Carlo methodsIsing spin glassesannealing algorithmsQUBO
0
0 comments X

The pith

A machine-learning Monte Carlo method outperforms Simulated Annealing and is more robust than Population Annealing on Ising spin glass problems.

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

The paper establishes that Global Annealing, which combines local Monte Carlo moves with machine learning global moves, finds lower energy configurations in three-dimensional Ising spin glasses than Simulated Annealing does. It also demonstrates greater robustness to changes in problem hardness and system size than Population Annealing, all without any hyperparameter tuning. Local moves are shown to be important for the best performance. This provides evidence that machine learning can enhance classical optimization methods to achieve real advantages in combinatorial problems.

Core claim

Global Annealing Monte Carlo surpasses Simulated Annealing in performance on QUBO problems for 3D Ising spin glasses and is more robust than Population Annealing across hardness and size without hyperparameter tuning, with local moves playing a crucial role.

What carries the argument

The Global Annealing algorithm that augments standard local moves with global moves proposed by a machine learning model.

If this is right

  • Global Annealing maintains performance across different problem hardness levels and system sizes.
  • No hyperparameter tuning is required for different instances.
  • Local moves are essential for optimal results when combined with ML global moves.
  • Machine learning-assisted methods can exceed classical techniques in combinatorial optimization.

Where Pith is reading between the lines

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

  • The success might indicate that ML models can learn general features of the energy landscape useful for optimization.
  • Similar approaches could be applied to other hard optimization problems like scheduling or graph partitioning.
  • It raises the question of how the ML model generalizes to much larger systems.

Load-bearing premise

The trained machine learning model proposes globally useful moves on new instances of varying size and hardness without needing retraining or tuning.

What would settle it

A demonstration that Global Annealing performs worse than Simulated Annealing on a new set of larger spin glass instances or requires tuning to maintain advantage would falsify the main claim.

Figures

Figures reproduced from arXiv: 2510.19544 by Federico Ricci-Tersenghi, Francesco Zamponi, Luca Maria Del Bono.

Figure 1
Figure 1. Figure 1: FIG. 1. ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Success probability as a function of the mean running time for SA, PA and GA (with and without local moves), [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Median success probability (solid lines) over 200 in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Same plot as in Fig. 3, here obtained with 10 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Probability density of the overlap as obtained by SA, PA, GA (green, red and blue, respectively) compared to the one [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

Combinatorial optimization problems are central to both practical applications and the development of optimization methods. While classical and quantum algorithms have been refined over decades, machine learning--assisted approaches are comparatively recent and have not yet consistently outperformed simple, state-of-the-art classical methods. Here, we focus on a class of Quadratic Unconstrained Binary Optimization (QUBO) problems, specifically the challenge of finding minimum energy configurations in three-dimensional Ising spin glasses. We use a Global Annealing Monte Carlo algorithm that integrates standard local moves with global moves proposed via machine learning. We show that local moves play a crucial role in achieving optimal performance. Benchmarking against Simulated Annealing and Population Annealing, we demonstrate that Global Annealing not only surpasses the performance of Simulated Annealing but also exhibits greater robustness than Population Annealing, maintaining effectiveness across problem hardness and system size without hyperparameter tuning. These results provide clear and robust evidence that a machine learning--assisted optimization method can exceed the capabilities of classical state-of-the-art techniques in a combinatorial optimization setting.

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

2 major / 2 minor

Summary. The paper introduces a Global Annealing Monte Carlo algorithm for 3D Ising spin-glass QUBO instances that augments conventional local Metropolis moves with global configuration proposals generated by a trained machine-learning model. It reports that this hybrid method outperforms Simulated Annealing in solution quality and exhibits greater robustness than Population Annealing across varying system sizes and problem hardness, all without per-instance or per-size hyperparameter retuning.

Significance. If the empirical claims are supported by properly aggregated statistics and a size-agnostic training protocol, the work would constitute concrete evidence that ML-enhanced Monte Carlo can deliver a measurable advantage over established classical baselines in combinatorial optimization. The explicit demonstration that local moves remain essential would also be a useful methodological contribution.

major comments (2)
  1. The central robustness claim (no hyperparameter tuning across system size and hardness) is load-bearing yet insufficiently documented. The manuscript must specify the ML architecture (e.g., whether input dimensionality is fixed or padded), the exact training distribution (sizes and hardness levels used), and whether a single set of weights is applied to all test instances or whether separate models are trained per L. Without this information the statement that the same model “maintains effectiveness … without hyperparameter tuning” cannot be evaluated.
  2. Quantitative support for the performance claims is missing from the abstract and appears only partially detailed in the results. The manuscript should report, for each method and each (L, hardness) combination: mean residual energy, success probability, number of independent runs, and error bars or standard deviations. Aggregation procedure (median over instances? best-of-N?) must also be stated explicitly.
minor comments (2)
  1. Figure captions and legends should explicitly label the ML model variant, the training set size, and whether local moves are enabled or disabled in each curve.
  2. A brief description of the loss function and training hyperparameters used for the move-proposal network would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive comments. We address each of the major comments below and indicate the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: The central robustness claim (no hyperparameter tuning across system size and hardness) is load-bearing yet insufficiently documented. The manuscript must specify the ML architecture (e.g., whether input dimensionality is fixed or padded), the exact training distribution (sizes and hardness levels used), and whether a single set of weights is applied to all test instances or whether separate models are trained per L. Without this information the statement that the same model “maintains effectiveness … without hyperparameter tuning” cannot be evaluated.

    Authors: We appreciate the referee pointing out the need for greater clarity on this central aspect of our work. In the revised manuscript, we will expand the Methods section to fully specify the ML architecture, including details on input dimensionality handling (we use padding to accommodate varying system sizes while maintaining a fixed model input size). We will also detail the training distribution, which includes a range of system sizes from L=4 to L=10 and various hardness levels generated via standard spin-glass instance creation methods. Crucially, a single trained model with one set of weights is applied to all test instances across different sizes and hardness levels, without any per-instance or per-size retraining or hyperparameter adjustment. This protocol underpins our robustness claim, and we will make this explicit to allow proper evaluation. revision: yes

  2. Referee: Quantitative support for the performance claims is missing from the abstract and appears only partially detailed in the results. The manuscript should report, for each method and each (L, hardness) combination: mean residual energy, success probability, number of independent runs, and error bars or standard deviations. Aggregation procedure (median over instances? best-of-N?) must also be stated explicitly.

    Authors: We agree that providing more detailed quantitative metrics will improve the transparency and reproducibility of our results. In the revised manuscript, we will update the Results section to include, for each method and each (L, hardness) combination, the mean residual energy, success probability, number of independent runs, and associated error bars or standard deviations. We will also explicitly describe the aggregation procedure, which involves averaging over multiple independent instances and runs. Additionally, we will consider incorporating key quantitative highlights into the abstract to better support the performance claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical benchmarks against external baselines remain independent

full rationale

The paper reports direct numerical comparisons of Global Annealing (local moves plus ML-proposed global moves) versus Simulated Annealing and Population Annealing on standard 3D Ising spin-glass QUBO instances. Performance is quantified by energy or success probability on held-out problem instances whose size and hardness are varied explicitly; these quantities are not defined in terms of any fitted parameter that is subsequently relabeled as a prediction. No self-definitional loop, fitted-input-called-prediction, or load-bearing self-citation chain appears in the derivation of the robustness claim. The ML model is trained once and then applied; whether that model truly generalizes is an empirical question that can be falsified by the reported curves, not a definitional identity. The study is therefore self-contained against external classical baselines.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the trained ML model for global proposals and on the standard assumptions of Monte Carlo sampling; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption A machine-learning model trained on spin-glass configurations can propose globally useful moves that remain effective on unseen instances of different size and hardness.
    This premise is required for the method to generalize without per-instance retraining.

pith-pipeline@v0.9.0 · 5731 in / 1271 out tokens · 36512 ms · 2026-05-18T04:49:54.424912+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

113 extracted references · 113 canonical work pages · 1 internal anchor

  1. [1]

    according to Eq

    General description of the procedure In the GA procedure, one uses a generative model to generate configurationsσ ′ of the system, approximately at equilibrium, i.e. according to Eq. (4) at a temperature β. These configurations are then used as global proposal moves for the MC procedure atβ′ =β+∆β > β, instead of the standard single-spin-flip moves of the...

    work page

  2. [2]

    Details on the architecture In this work we have used a shallow MADE (Masked Autoencoder for Distribution Estimation, [103]) autore- gressive architecture, which is modeled as: P(σ i|σ<i) = exp Pi−1 j=1 Wijσiσj 2 cosh Pi−1 j=1 Wijσj ,(7) whereσ <i is the set of spinsσ<i ={σ 1, . . . , σi−1}. We note that the autoregressive approach requires to choose an o...

    work page

  3. [3]

    National Centre for HPC, Big Data and Quantum Com- puting

    Details on the training procedure Training is performed by minimizing the binary cross- entropy loss (i.e., minimizing the Kullback-Leibler diver- gence betweenρ NN andρ GB). The initial training runs for40epochs using the Adam optimizer with learning rateη 0 = 10−3. We employ an exponential learning–rate schedulethathalvestherateevery10epochs. Earlystop-...

    work page 2024

  4. [4]

    Maxsat, hard and soft constraints

    Chu Min Li and Felip Manyà. Maxsat, hard and soft constraints. In Armin Biere, Hans van Maaren, and Toby Walsh, editors,Handbook of Satisfiability, volume 185 ofFrontiers in Artificial Intelligence and Applica- tions, pages 613–631. IOS Press, 2009

    work page 2009

  5. [5]

    Jensen and Bjarne Toft.Graph Coloring Problems

    Tommy R. Jensen and Bjarne Toft.Graph Coloring Problems. Wiley-Interscience Series in Discrete Mathe- matics and Optimization. Wiley, New York, 1995

    work page 1995

  6. [6]

    Goemans and David P

    Michel X. Goemans and David P. Williamson. Improved approximation algorithms for maximum cut and satisfi- ability problems using semidefinite programming.Jour- nal of the ACM, 42(6):1115–1145, 1995

    work page 1995

  7. [7]

    Garey and David S

    Michael R. Garey and David S. Johnson.Comput- ers and Intractability: A Guide to the Theory of NP- Completeness. W. H. Freeman, New York, 1979

    work page 1979

  8. [8]

    Springer, 5 edition, 2016

    MichaelL.Pinedo.Scheduling: Theory, Algorithms, and Systems. Springer, 5 edition, 2016

    work page 2016

  9. [9]

    David S. Johnson. Approximation algorithms for com- binatorial problems.Journal of Computer and System Sciences, 9(3):256–278, 1974

    work page 1974

  10. [10]

    A greedy heuristic for the set- covering problem.Mathematics of Operations Research, 4(3):233–235, 1979

    Václav Chvátal. A greedy heuristic for the set- covering problem.Mathematics of Operations Research, 4(3):233–235, 1979

    work page 1979

  11. [11]

    Lawler, Jan Karel Lenstra, A

    Eugene L. Lawler, Jan Karel Lenstra, A. H. G. Rin- nooy Kan, and David B. Shmoys, editors.The Traveling Salesman Problem: A Guided Tour of Combinatorial Optimization. Wiley, Chichester, 1985

    work page 1985

  12. [12]

    Solving max-cut to optimality by intersecting semidef- inite and polyhedral relaxations.Optimization Online, 2007

    Franz Rendl, Giovanni Rinaldi, and Angelika Wiegele. Solving max-cut to optimality by intersecting semidef- inite and polyhedral relaxations.Optimization Online, 2007

    work page 2007

  13. [13]

    Biqcrunch: A semidefinite branch-and-bound method for solving binary quadratic problems.ACM Transac- tions on Mathematical Software, 43(4):32:1–32:23, 2017

    Nathan Krislock, Jérôme Malick, and Frédéric Roupin. Biqcrunch: A semidefinite branch-and-bound method for solving binary quadratic problems.ACM Transac- tions on Mathematical Software, 43(4):32:1–32:23, 2017

    work page 2017

  14. [14]

    Biqbin: A parallel branch-and-bound solver for binary quadraticproblemswithlinearconstraints.ACM Trans- actions on Mathematical Software, 48(3):24:1–24:29, 2022

    Nicolò Gusmeroli, Timotej Hrga, Borut Lužar, Janez Povh, Melanie Siebenhofer, and Angelika Wiegele. Biqbin: A parallel branch-and-bound solver for binary quadraticproblemswithlinearconstraints.ACM Trans- actions on Mathematical Software, 48(3):24:1–24:29, 2022

    work page 2022

  15. [15]

    McSparse: Exact solutions of sparse max- imum cut and sparse unconstrained binary quadratic optimization problems

    Jonas Charfreitag, Michael Jünger, Sven Mallach, and Petra Mutzel. McSparse: Exact solutions of sparse max- imum cut and sparse unconstrained binary quadratic optimization problems. In Cynthia A. Phillips and Bettina Speckmann, editors,2022 Proceedings of the Symposium on Algorithm Engineering and Experiments (ALENEX), pages 54–66, 2022

    work page 2022

  16. [16]

    Gurobi Optimizer Refer- ence Manual, 2024

    Gurobi Optimization, LLC. Gurobi Optimizer Refer- ence Manual, 2024

    work page 2024

  17. [17]

    Gta-an atsp method: Shifting the bottleneck from algorithm to ram.arXiv preprint arXiv:2509.13327, 2025

    Wissam Nakhle. Gta-an atsp method: Shifting the bottleneck from algorithm to ram.arXiv preprint arXiv:2509.13327, 2025

    work page arXiv 2025

  18. [18]

    Kirkpatrick, C

    S. Kirkpatrick, C. D. Gelatt, and M. P. Vec- chi. Optimization by simulated annealing.Science, 220(4598):671–680, 1983

    work page 1983

  19. [19]

    Effects of changing the boundary conditions on the ground state of ising spin glasses.Physical Review B, 62(17):11677, 2000

    Enzo Marinari and Giorgio Parisi. Effects of changing the boundary conditions on the ground state of ising spin glasses.Physical Review B, 62(17):11677, 2000

    work page 2000

  20. [20]

    Extremal optimization.New opti- mization algorithms in physics, pages 227–251, 2004

    Stefan Boettcher. Extremal optimization.New opti- mization algorithms in physics, pages 227–251, 2004

    work page 2004

  21. [21]

    Circumspect descent prevails in solving random constraint satisfaction problems.Proceedings of the Na- tional Academy of Sciences, 105(40):15253–15257, 2008

    Mikko Alava, John Ardelius, Erik Aurell, Petteri Kaski, Supriya Krishnamurthy, Pekka Orponen, and Sakari Seitz. Circumspect descent prevails in solving random constraint satisfaction problems.Proceedings of the Na- tional Academy of Sciences, 105(40):15253–15257, 2008. 11

    work page 2008

  22. [22]

    Breakout local search for the max-cutproblem.Engineering Applications of Artificial Intelligence, 26(3):1162–1173, 2013

    Una Benlic and Jin-Kao Hao. Breakout local search for the max-cutproblem.Engineering Applications of Artificial Intelligence, 26(3):1162–1173, 2013

    work page 2013

  23. [23]

    Monte carlo algorithms are very effective in finding the largest independent set in sparse random graphs.Phys- ical Review E, 100(1):013302, 2019

    Maria Chiara Angelini and Federico Ricci-Tersenghi. Monte carlo algorithms are very effective in finding the largest independent set in sparse random graphs.Phys- ical Review E, 100(1):013302, 2019

    work page 2019

  24. [24]

    How we are lead- ing a 3-xorsat challenge: from the energy landscape to the algorithm and its efficient implementation on gpus (a).Europhysics Letters, 133(6):60005, 2021

    Massimo Bernaschi, Mauro Bisson, Massimiliano Fat- ica, Enzo Marinari, Vıctor Martin-Mayor, Giorgio Parisi, and Federico Ricci-Tersenghi. How we are lead- ing a 3-xorsat challenge: from the energy landscape to the algorithm and its efficient implementation on gpus (a).Europhysics Letters, 133(6):60005, 2021

    work page 2021

  25. [25]

    Quantum annealing in the transverse ising model.Physical Review E, 58(5):5355–5363, 1998

    Tadashi Kadowaki and Hidetoshi Nishimori. Quantum annealing in the transverse ising model.Physical Review E, 58(5):5355–5363, 1998

    work page 1998

  26. [26]

    A quantum adiabatic evolution algorithm applied to random instances of an np-complete problem.Sci- ence, 292(5516):472–475, 2001

    Edward Farhi, Jeffrey Goldstone, Sam Gutmann, Joshua Lapan, Andrew Lundgren, and Daniel Preda. A quantum adiabatic evolution algorithm applied to random instances of an np-complete problem.Sci- ence, 292(5516):472–475, 2001. See also arXiv:quant- ph/0104129

    work page arXiv 2001

  27. [27]

    Theory of quantum annealing of an ising spin glass.Science, 295(5564):2427–2430, 2002

    Giuseppe E Santoro, Roman Martonák, Erio Tosatti, and Roberto Car. Theory of quantum annealing of an ising spin glass.Science, 295(5564):2427–2430, 2002

    work page 2002

  28. [28]

    Chakrabarti

    Arnab Das and Bikas K. Chakrabarti. Colloquium: Quantum annealing and analog quantum computation. Reviews of Modern Physics, 80(3):1061–1081, 2008

    work page 2008

  29. [29]

    Entanglement- assisted variational algorithm for discrete optimization problems.arXiv preprint arXiv:2501.09078, 2025

    Lorenzo Fioroni and Vincenzo Savona. Entanglement- assisted variational algorithm for discrete optimization problems.arXiv preprint arXiv:2501.09078, 2025

    work page arXiv 2025

  30. [30]

    A Quantum Approximate Optimization Algorithm

    Edward Farhi, Jeffrey Goldstone, and Sam Gut- mann. A quantum approximate optimization algorithm. arXiv:1411.4028, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  31. [31]

    Iterative quantum algorithms for maximum independent set.Physical Re- view A, 110(5):052435, 2024

    Lucas T Brady and Stuart Hadfield. Iterative quantum algorithms for maximum independent set.Physical Re- view A, 110(5):052435, 2024

    work page 2024

  32. [32]

    The quantum approximate optimization algorithm needs to seethewholegraph: Worstcaseexamples

    Edward Farhi, David Gamarnik, and Sam Gutmann. The quantum approximate optimization algorithm needs to see the whole graph: Worst case examples. arXiv:2005.08747, 2020

    work page arXiv 2005

  33. [33]

    The quantum adi- abatic algorithm applied to random optimization prob- lems: The quantum spin glass perspective.Physics Re- ports, 523(3):127–205, 2013

    Victor Bapst, Laura Foini, Florent Krzakala, Guilhem Semerjian, and Francesco Zamponi. The quantum adi- abatic algorithm applied to random optimization prob- lems: The quantum spin glass perspective.Physics Re- ports, 523(3):127–205, 2013

    work page 2013

  34. [34]

    Machine learning for combinatorial optimization: A methodologicaltourd’horizon.European Journal of Op- erational Research, 290(2):405–421, 2021

    Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: A methodologicaltourd’horizon.European Journal of Op- erational Research, 290(2):405–421, 2021

    work page 2021

  35. [35]

    Khalil, Yuyu Zhang, Bistra Dilk- ina, and Le Song

    Hanjun Dai, Elias B. Khalil, Yuyu Zhang, Bistra Dilk- ina, and Le Song. Learning combinatorial optimization algorithms over graphs. InNeurIPS, 2017

    work page 2017

  36. [36]

    Atten- tion, learn to solve routing problems! InICLR, 2019

    Wouter Kool, Herke van Hoof, and Max Welling. Atten- tion, learn to solve routing problems! InICLR, 2019

    work page 2019

  37. [37]

    Predict- ing ground state configuration of energy landscape en- semble using graph neural network.arXiv preprint arXiv:2008.08227, 2020

    Seong Ho Pahng and Michael P Brenner. Predict- ing ground state configuration of energy landscape en- semble using graph neural network.arXiv preprint arXiv:2008.08227, 2020

    work page arXiv 2008

  38. [38]

    Calculation of the ground states of spin glasses us- ing a restricted boltzmann machine.JETP Letters, 115(8):466–470, 2022

    Alena O Korol’, V Yu Kapitan, Aleksandr Vasil’evich Perzhu, Mikhail Alexandrovich Padalko, D Yu Kapi- tan, Roman Andreevich Volotovskii, Egor Vadimovich Vasil’ev, Aleksey Evgenievich Rybin, Pavel Alekseevich Ovchinnikov, Petr Dmitrievich Andriushchenko, et al. Calculation of the ground states of spin glasses us- ing a restricted boltzmann machine.JETP Let...

    work page 2022

  39. [39]

    Searching for spin glass ground states through deep reinforcement learn- ing.Nature communications, 14(1):725, 2023

    Changjun Fan, Mutian Shen, Zohar Nussinov, Zhong Liu, Yizhou Sun, and Yang-Yu Liu. Searching for spin glass ground states through deep reinforcement learn- ing.Nature communications, 14(1):725, 2023

    work page 2023

  40. [40]

    Sequential stochastic combinatorial optimization using hierarchal reinforcement learning.arXiv preprint arXiv:2502.05537, 2025

    Xinsong Feng, Zihan Yu, Yanhai Xiong, and Haipeng Chen. Sequential stochastic combinatorial optimization using hierarchal reinforcement learning.arXiv preprint arXiv:2502.05537, 2025

    work page arXiv 2025

  41. [41]

    Unsupervised learning with gnns for qubo-based combinatorial opti- mization.EURO Journal on Computational Optimiza- tion, page 100116, 2025

    Olga Krylova and Frank Phillipsona. Unsupervised learning with gnns for qubo-based combinatorial opti- mization.EURO Journal on Computational Optimiza- tion, page 100116, 2025

    work page 2025

  42. [42]

    Combinatorial optimization with physics- inspired graph neural networks.Nature Machine Intel- ligence, 4(4):367–377, 2022

    Martin JA Schuetz, J Kyle Brubaker, and Helmut G Katzgraber. Combinatorial optimization with physics- inspired graph neural networks.Nature Machine Intel- ligence, 4(4):367–377, 2022

    work page 2022

  43. [43]

    Maria Chiara Angelini and Federico Ricci-Tersenghi. Modern graph neural networks do worse than classical greedy algorithms in solving combinatorial optimization problems like maximum independent set.Nature Ma- chine Intelligence, 5(1):29–31, 2023

    work page 2023

  44. [44]

    Inability of a graph neural network heuristic to outperform greedy algorithms in solving combinatorial optimization problems.Nature Machine Intelligence, 5(1):24–25, 2023

    Stefan Boettcher. Inability of a graph neural network heuristic to outperform greedy algorithms in solving combinatorial optimization problems.Nature Machine Intelligence, 5(1):24–25, 2023

    work page 2023

  45. [45]

    One networktoapproximatethemall: Amortizedvariational inference of ising ground states

    Sebastian Sanokowski, Wilhelm Berghammer, Johannes Kofler, Sepp Hochreiter, and Sebastian Lehner. One networktoapproximatethemall: Amortizedvariational inference of ising ground states. InMachine Learn- ing and the Physical Sciences workshop, NeurIPS 2022, 2022

    work page 2022

  46. [46]

    Nonlocal monte carlo via reinforcement learn- ing.arXiv preprint arXiv:2508.10520, 2025

    Dmitrii Dobrynin, Masoud Mohseni, and John Paul Strachan. Nonlocal monte carlo via reinforcement learn- ing.arXiv preprint arXiv:2508.10520, 2025

    work page arXiv 2025

  47. [47]

    Scalable discrete diffusion samplers: Combinatorial optimization and statistical physics.arXiv preprint arXiv:2502.08696, 2025

    Sebastian Sanokowski, Wilhelm Berghammer, Mar- tin Ennemoser, Haoyu Peter Wang, Sepp Hochre- iter, and Sebastian Lehner. Scalable discrete diffusion samplers: Combinatorial optimization and statistical physics.arXiv preprint arXiv:2502.08696, 2025

    work page arXiv 2025

  48. [48]

    Deep reinforced learning heuristic tested on spin-glass ground states: The larger picture

    Stefan Boettcher. Deep reinforced learning heuristic tested on spin-glass ground states: The larger picture. Nature Communications, 14(1):5658, 2023

    work page 2023

  49. [49]

    Reply to: Deep re- inforced learning heuristic tested on spin-glass ground states: The larger picture.Nature communications, 14(1):5659, 2023

    Changjun Fan, Mutian Shen, Zohar Nussinov, Zhong Liu, Yizhou Sun, and Yang-Yu Liu. Reply to: Deep re- inforced learning heuristic tested on spin-glass ground states: The larger picture.Nature communications, 14(1):5659, 2023

    work page 2023

  50. [50]

    Adaptive monte carlo augmented with nor- malizing flows.Proceedings of the National Academy of Sciences, 119(10):e2109420119, 2022

    Marylou Gabrié, Grant M Rotskoff, and Eric Vanden- Eijnden. Adaptive monte carlo augmented with nor- malizing flows.Proceedings of the National Academy of Sciences, 119(10):e2109420119, 2022

    work page 2022

  51. [51]

    Performance of machine-learning- assisted monte carlo in sampling from simple statistical physics models.Phys

    Luca Maria Del Bono, Federico Ricci-Tersenghi, and Francesco Zamponi. Performance of machine-learning- assisted monte carlo in sampling from simple statistical physics models.Phys. Rev. E, 112:045307, Oct 2025

    work page 2025

  52. [52]

    Pytorch: An impera- tive style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Jun- 12 jie Bai, and Soumith Chintala. Pytorch: An impera- tive style, hig...

    work page 2019

  53. [53]

    Analysis of the relation between quadratic unconstrained binary optimization and the spin-glass ground-state problem.Phys

    Stefan Boettcher. Analysis of the relation between quadratic unconstrained binary optimization and the spin-glass ground-state problem.Phys. Rev. Res., 1:033142, Dec 2019

    work page 2019

  54. [54]

    Ising formulations of many np problems

    Andrew Lucas. Ising formulations of many np problems. Frontiers in physics, 2:5, 2014

    work page 2014

  55. [55]

    Crystal statistics

    Lars Onsager. Crystal statistics. i. a two-dimensional model with an order-disorder transition.Physical re- view, 65(3-4):117, 1944

    work page 1944

  56. [56]

    The- ory of spin glasses.Journal of Physics F: Metal Physics, 5(5):965, 1975

    Samuel Frederick Edwards and Phil W Anderson. The- ory of spin glasses.Journal of Physics F: Metal Physics, 5(5):965, 1975

    work page 1975

  57. [57]

    On the computational complexity of ising spin glass models.Journal of Physics A: Math- ematical and General, 15(10):3241, 1982

    Francisco Barahona. On the computational complexity of ising spin glass models.Journal of Physics A: Math- ematical and General, 15(10):3241, 1982

    work page 1982

  58. [58]

    Simulated temper- ing: a new monte carlo scheme.Europhysics letters, 19(6):451, 1992

    Enzo Marinari and Giorgio Parisi. Simulated temper- ing: a new monte carlo scheme.Europhysics letters, 19(6):451, 1992

    work page 1992

  59. [59]

    Exchange monte carlo method and application to spin glass simulations

    Koji Hukushima and Koji Nemoto. Exchange monte carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6):1604– 1608, 1996

    work page 1996

  60. [60]

    A cluster monte carlo algorithm for 2-dimensionalspinglasses.The European Physical Jour- nal B-Condensed Matter and Complex Systems, 22:479– 484, 2001

    Jérôme Houdayer. A cluster monte carlo algorithm for 2-dimensionalspinglasses.The European Physical Jour- nal B-Condensed Matter and Complex Systems, 22:479– 484, 2001

    work page 2001

  61. [61]

    Efficient cluster algorithm for spin glasses in any space dimension.Physical review letters, 115(7):077201, 2015

    Zheng Zhu, Andrew J Ochoa, and Helmut G Katz- graber. Efficient cluster algorithm for spin glasses in any space dimension.Physical review letters, 115(7):077201, 2015

    work page 2015

  62. [62]

    Population annealing and its application to a spin glass

    Koji Hukushima and Yukito Iba. Population annealing and its application to a spin glass. InAIP Conference Proceedings, volume 690, pages 200–206, 2003

    work page 2003

  63. [63]

    Population annealing with weighted averages: A monte carlo method for rough free-energy landscapes.Physical Review E, 82:026704, 2010

    Jonathan Machta. Population annealing with weighted averages: A monte carlo method for rough free-energy landscapes.Physical Review E, 82:026704, 2010

    work page 2010

  64. [64]

    WenlongWang, JonathanMachta, andHelmutG.Katz- graber. Comparing monte carlo methods for finding ground states of ising spin glasses: Population anneal- ing, simulated annealing, and parallel tempering.Phys- ical Review E, 92:013303, 2015

    work page 2015

  65. [65]

    High- performance combinatorial optimization based on clas- sical mechanics.Science Advances, 7(6):eabe7953, 2021

    Hayato Goto, Kotaro Endo, Masaru Suzuki, Yoshisato Sakai, Taro Kanao, Yohei Hamakawa, Ryo Hidaka, Masaya Yamasaki, and Kosuke Tatsumura. High- performance combinatorial optimization based on clas- sical mechanics.Science Advances, 7(6):eabe7953, 2021

    work page 2021

  66. [66]

    EMHEB Ekanayake and Nikhil Shukla. Different paths, same destination: Designing physics-inspired dynamical systems with engineered stability to minimize the ising hamiltonian.Physical Review Applied, 24(2):024008, 2025

    work page 2025

  67. [67]

    Spinglasspeps

    Tomasz Śmierzchalski, Anna M Dziubyna, Kon- rad Jałowiecki, Zakaria Mzaouali, Łukasz Pawela, Bartłomiej Gardas, and Marek M Rams. Spinglasspeps. jl: Tensor-network package for ising-like optimiza- tion on quasi-two-dimensional graphs.arXiv preprint arXiv:2502.02317, 2025

    work page arXiv 2025

  68. [68]

    Batchtnmc: Efficient sampling of two- dimensional spin glasses using tensor network monte carlo.arXiv preprint arXiv:2509.19006, 2025

    Tao Chen, Jingtong Zhang, Jing Liu, Youjin Deng, and Pan Zhang. Batchtnmc: Efficient sampling of two- dimensional spin glasses using tensor network monte carlo.arXiv preprint arXiv:2509.19006, 2025

    work page arXiv 2025

  69. [69]

    Solving the quantum many-body problem with artificial neural net- works.Science, 355(6325):602–606, 2017

    Giuseppe Carleo and Matthias Troyer. Solving the quantum many-body problem with artificial neural net- works.Science, 355(6325):602–606, 2017

    work page 2017

  70. [70]

    Solving statisti- cal mechanics using variational autoregressive networks

    Dian Wu, Lei Wang, and Pan Zhang. Solving statisti- cal mechanics using variational autoregressive networks. Physical review letters, 122(8):080602, 2019

    work page 2019

  71. [71]

    Enhancing the efficiency of variational autoregressive networks through renor- malization group.Physical Review E, 112(3):035310, 2025

    Sihan Wang and Zhirong Liu. Enhancing the efficiency of variational autoregressive networks through renor- malization group.Physical Review E, 112(3):035310, 2025

    work page 2025

  72. [72]

    Isingformer: Augmenting parallel tempering with learned proposals

    Saleh Bunaiyan, Corentin Delacour, Shuvro Chowd- hury, Kyle Lee, and Kerem Y Camsari. Isingformer: Augmenting parallel tempering with learned proposals. arXiv preprint arXiv:2509.23043, 2025

    work page arXiv 2025

  73. [73]

    Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457):eaaw1147, 2019

    Frank Noé, Simon Olsson, Jonas Köhler, and Hao Wu. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457):eaaw1147, 2019

    work page 2019

  74. [74]

    Skipping the replica exchange ladder with normalizing flows.The Journal of Physical Chem- istry Letters, 13(50):11643–11649, 2022

    Michele Invernizzi, Andreas Krämer, Cecilia Clementi, and Frank Noé. Skipping the replica exchange ladder with normalizing flows.The Journal of Physical Chem- istry Letters, 13(50):11643–11649, 2022

    work page 2022

  75. [75]

    Temperature steerable flows and boltzmann gen- erators.Physical Review Research, 4(4):L042005, 2022

    Manuel Dibak, Leon Klein, Andreas Krämer, and Frank Noé. Temperature steerable flows and boltzmann gen- erators.Physical Review Research, 4(4):L042005, 2022

    work page 2022

  76. [76]

    Equivariant flows: exact likelihood generative learning for symmet- ric densities

    Jonas Köhler, Leon Klein, and Frank Noé. Equivariant flows: exact likelihood generative learning for symmet- ric densities. InInternational conference on machine learning, pages 5361–5370. PMLR, 2020

    work page 2020

  77. [77]

    Equivariant flow-based sampling for lattice gauge the- ory.Physical Review Letters, 125(12):121601, 2020

    Gurtej Kanwar, Michael S Albergo, Denis Boyda, Kyle Cranmer, Daniel C Hackett, Sébastien Racaniere, Danilo Jimenez Rezende, and Phiala E Shanahan. Equivariant flow-based sampling for lattice gauge the- ory.Physical Review Letters, 125(12):121601, 2020

    work page 2020

  78. [78]

    Flow-based generative models for markov chain monte carlo in lattice field theory.Physical Re- view D, 100(3):034515, 2019

    Michael S Albergo, Gurtej Kanwar, and Phiala E Shanahan. Flow-based generative models for markov chain monte carlo in lattice field theory.Physical Re- view D, 100(3):034515, 2019

    work page 2019

  79. [79]

    Learning lattice quantum field theories with equivariant continu- ous flows.arXiv preprint arXiv:2207.00283, 2022

    Mathis Gerdes, Pim de Haan, Corrado Rainone, Roberto Bondesan, and Miranda CN Cheng. Learning lattice quantum field theories with equivariant continu- ous flows.arXiv preprint arXiv:2207.00283, 2022

    work page arXiv 2022

  80. [80]

    Scaling up machine learning for quantum field theory with equivariant continuous flows.arXiv preprint arXiv:2110.02673, 2021

    Pim de Haan, Corrado Rainone, Miranda CN Cheng, and Roberto Bondesan. Scaling up machine learning for quantum field theory with equivariant continuous flows.arXiv preprint arXiv:2110.02673, 2021

    work page arXiv 2021

Showing first 80 references.