pith. sign in

arxiv: 2606.06272 · v1 · pith:F27RRWIRnew · submitted 2026-06-04 · 💻 cs.LG · cs.AI

Your GFlowNet Secretly Learns an Optimal Transport Plan

Ian Maksimov , Nikita Morozov , Denis Belomestny , Sergey Samsonov This is my paper

Pith reviewed 2026-06-28 02:32 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords GFlowNetoptimal transportKantorovich problemminimum flowshortest path costsgenerative modelstransport plansgraph flows
0
0 comments X

The pith

Fixing the initial flow distribution in a minimum-flow GFlowNet reduces its objective to a Kantorovich optimal transport problem whose solution is recovered by sampling the trained policy.

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

The paper establishes that non-acyclic GFlowNets trained under the minimum-flow objective with a fixed initial distribution solve a Kantorovich optimal transport problem whose costs are the shortest-path distances induced by the underlying graph. This equivalence means the converged GFlowNet policy directly encodes an optimal transport plan, and trajectories sampled from it recover the corresponding optimal coupling between source and target distributions. A sympathetic reader would care because the result supplies a new way to compute transport plans on large graphs by training neural-parameterized flows rather than solving the transport problem directly.

Core claim

Fixing the initial flow distribution in a minimum-flow GFlowNet reduces its objective to a Kantorovich OT problem with graph-induced shortest path costs. At the optimum, the learned GFlowNet policy therefore encodes an optimal transport plan from the source distribution to the target distribution: sampling trajectories from the minimum-flow GFlowNet recovers the corresponding optimal coupling.

What carries the argument

The reduction of the minimum-flow GFlowNet objective (with fixed initial distribution) to the Kantorovich formulation of optimal transport, where edge costs are shortest-path distances on the graph.

If this is right

  • GFlowNet training can be used to solve optimal transport problems on large graphs by learning edge flows with neural networks.
  • The optimal transport plan is recovered simply by sampling trajectories from the trained GFlowNet.
  • The equivalence holds only for the minimum-flow objective and fixed initial distribution on non-acyclic graphs.

Where Pith is reading between the lines

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

  • Standard OT solvers could be used to warm-start or regularize GFlowNet training on moderate-sized graphs.
  • The link suggests that other flow-based generative models might admit similar reductions to transport problems under appropriate objectives.
  • GFlowNets could scale optimal transport computations to graphs that are too large for exact linear programming solvers.

Load-bearing premise

The GFlowNet must be non-acyclic and trained under the minimum-flow objective with a fixed initial flow distribution.

What would settle it

Compute the exact optimal coupling via a standard OT solver on the same graph using shortest-path costs, then compare it to the empirical distribution of trajectories sampled from a converged minimum-flow GFlowNet; any systematic mismatch falsifies the claimed equivalence.

Figures

Figures reproduced from arXiv: 2606.06272 by Denis Belomestny, Ian Maksimov, Nikita Morozov, Sergey Samsonov.

Figure 1
Figure 1. Figure 1: Visualization of the solution to the GFlowNet LP problem and Kantorovich OT optimal plan on a hypergrid environment. OT induces optimal coupling that connects source and target states and how much mass is transfered between them. GFlowNet on the other hand builds a policy that samples target states starting from source states, thus inducing a coupling with respect to the shortest path metric. Both formulat… view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the distributions used in the experiments: moon-shaped (left), corner-shaped (center), and ball-shaped (right). Where: z = s M − 1 , c =  1 2 , . . . , 1 2  , e1 = (1, 0, . . . , 0), b = c + rout 2 e1. and rin, rout are inner and outer balls radius. The ball-shaped distribution defined as follows: L(s) ≜ I [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
read the original abstract

Generative Flow Networks (GFlowNets) are a framework for sampling structured objects via stochastic trajectories in a directed graph. In this work, we establish a theoretical connection between non-acyclic GFlowNets and optimal transport (OT). We show that fixing the initial flow distribution in a minimum-flow GFlowNet reduces its objective to a Kantorovich OT problem with graph-induced shortest path costs. At the optimum, the learned GFlowNet policy therefore encodes an optimal transport plan from the source distribution to the target distribution: we show that sampling trajectories from the minimum-flow GFlowNet recovers the corresponding optimal coupling. Our formulation enables applying the GFlowNet learning framework to OT problems on large graphs via edge flows and neural parameterization. Experiments confirm agreement with exact OT solvers and demonstrate that GFlowNets can learn high-quality transport plans.

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

0 major / 3 minor

Summary. The paper claims that fixing the initial flow distribution in a minimum-flow GFlowNet on a non-acyclic directed graph reduces the GFlowNet objective exactly to a Kantorovich optimal transport problem whose ground cost is the shortest-path distance induced by the graph. Consequently the learned edge flows define an optimal coupling whose marginals match the prescribed source and target distributions, and sampling trajectories from the trained policy recovers that coupling. The formulation is then used to solve OT instances on large graphs via neural parameterization of the flows, with experiments reported to match exact solvers.

Significance. If the reduction is correct, the result supplies a new, scalable route to OT on graphs by importing the GFlowNet training machinery and neural function approximation. Credit is due for the explicit reduction to the standard Kantorovich formulation (no free parameters introduced) and for the experimental verification against exact OT solvers.

minor comments (3)
  1. [§3.2] §3.2, Eq. (8): the statement that the minimum-flow objective is 'identical' to the OT objective should be accompanied by an explicit side-by-side display of the two loss expressions to make the algebraic equivalence immediate.
  2. [Experiments] The experimental section reports agreement with exact solvers but does not state the precise graph sizes or the number of independent runs; adding these details would strengthen reproducibility.
  3. Notation: the symbol P is used both for the GFlowNet policy and for the OT coupling; a brief remark distinguishing the two usages would avoid confusion.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the accurate summary of our manuscript and for the positive assessment of its significance. We appreciate the recommendation for minor revision. As the report lists no specific major comments, we have no points requiring point-by-point response or manuscript changes.

Circularity Check

0 steps flagged

No significant circularity; derivation reduces to external OT formulation

full rationale

The central claim is a direct mathematical reduction: fixing the initial flow in the minimum-flow GFlowNet objective yields the Kantorovich OT problem with shortest-path costs induced by the graph. This follows from flow conservation and the objective definition without any self-referential fitting, self-citation load-bearing, or ansatz smuggling. The result is externally falsifiable via exact OT solvers, and experiments confirm agreement. No step in the provided derivation chain reduces by construction to the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the work relies on standard mathematical formulations of optimal transport and GFlowNets without introducing new free parameters or invented entities.

axioms (1)
  • standard math Kantorovich formulation of optimal transport
    The GFlowNet objective is shown to reduce to this standard problem.

pith-pipeline@v0.9.1-grok · 5678 in / 1091 out tokens · 60595 ms · 2026-06-28T02:32:13.835068+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

157 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    Optimizing Backward Policies in

    Gritsaev, Timofei and Morozov, Nikita and Samsonov, Sergey and Tiapkin, Daniil , booktitle=. Optimizing Backward Policies in

  2. [2]

    arXiv preprint arXiv:2603.01786 , year=

    Learning Shortest Paths with Generative Flow Networks , author=. arXiv preprint arXiv:2603.01786 , year=

    work page arXiv

  3. [3]

    International Journal of Computer Vision , volume=

    The Earth Mover's Distance as a Metric for Image Retrieval , author=. International Journal of Computer Vision , volume=

  4. [4]

    Proceedings of the 32nd International Conference on Machine Learning , pages=

    From Word Embeddings to Document Distances , author=. Proceedings of the 32nd International Conference on Machine Learning , pages=

  5. [5]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Optimal Transport for Domain Adaptation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  6. [6]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    Wasserstein Generative Adversarial Networks , author=. Proceedings of the 34th International Conference on Machine Learning , pages=

  7. [7]

    Advances in Neural Information Processing Systems , year=

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author=. Advances in Neural Information Processing Systems , year=

  8. [8]

    and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and. Nature Methods , year =

  9. [9]

    2008 , publisher=

    Optimal Transport: Old and New , author=. 2008 , publisher=

    2008

  10. [10]

    Econometrica , volume =

    Martin Beckmann , title =. Econometrica , volume =

  11. [11]

    , title =

    Dantzig, George B. , title =. Activity Analysis of Production and Allocation , editor =

  12. [12]

    SIAM Journal on Scientific Computing , volume=

    Quadratically Regularized Optimal Transport on Graphs , author=. SIAM Journal on Scientific Computing , volume=. 2018 , publisher=. doi:10.1137/17M1132665 , url=

    work page doi:10.1137/17m1132665 2018

  13. [13]

    arXiv preprint arXiv:2509.20408 , year=

    A Theory of Multi-Agent Generative Flow Networks , author=. arXiv preprint arXiv:2509.20408 , year=

    work page arXiv

  14. [14]

    International Conference on Machine Learning , pages=

    Random Policy Evaluation Uncovers Policies of Generative Flow Networks , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  15. [15]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Gdpo: Learning to directly align language models with diversity using gflownets , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  16. [16]

    arXiv preprint arXiv:2509.15207 , year=

    Flowrl: Matching reward distributions for llm reasoning , author=. arXiv preprint arXiv:2509.15207 , year=

    work page arXiv

  17. [17]

    SIAM Journal on Mathematical Analysis , volume=

    Barycenters in the Wasserstein space , author=. SIAM Journal on Mathematical Analysis , volume=. 2011 , publisher=

    2011

  18. [18]

    Econometrica: Journal of the Econometric Society , pages=

    A continuous model of transportation , author=. Econometrica: Journal of the Econometric Society , pages=. 1952 , publisher=

    1952

  19. [19]

    International Conference on Learning Representations , year=

    Wasserstein Auto-Encoders , author=. International Conference on Learning Representations , year=

  20. [20]

    Journal of Machine Learning Research , volume=

    POT: Python Optimal Transport , author=. Journal of Machine Learning Research , volume=

  21. [21]

    Revisiting Non-Acyclic

    Morozov, Nikita and Maksimov, Ian and Tiapkin, Daniil and Samsonov, Sergey , booktitle =. Revisiting Non-Acyclic. 2025 , volume =

    2025

  22. [22]

    IEEE transactions on Systems Science and Cybernetics , volume=

    A formal basis for the heuristic determination of minimum cost paths , author=. IEEE transactions on Systems Science and Cybernetics , volume=. 1968 , publisher=

    1968

  23. [23]

    and Moulines, E

    Douc, R. and Moulines, E. and Priouret, P. and Soulier, P. , TITLE =. 2018 , PAGES =

    2018

  24. [24]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  25. [25]

    2017 , publisher=

    Markov chains and mixing times , author=. 2017 , publisher=

    2017

  26. [26]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  27. [27]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  28. [28]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  29. [29]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  30. [30]

    M. J. Kearns , title =

  31. [31]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  32. [32]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  33. [33]

    Suppressed for Anonymity , author=

  34. [34]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  35. [35]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Flow network based generative models for non-iterative diverse candidate generation , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    2024 , eprint=

    Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets , author=. 2024 , eprint=

    2024

  38. [38]

    Journal of Physics: Conference Series , volume=

    HPC resources of the higher school of economics , author=. Journal of Physics: Conference Series , volume=. 2021 , organization=

    2021

  39. [39]

    Advances in Neural Information Processing Systems , volume=

    Maximum entropy monte-carlo planning , author=. Advances in Neural Information Processing Systems , volume=

  40. [40]

    International Conference on Machine Learning , pages=

    Learning GFlowNets from partial episodes for improved convergence and stability , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    Trajectory balance: Improved credit assignment in GFlowNets , author=. Advances in Neural Information Processing Systems , volume=

  42. [42]

    International Conference on Artificial Intelligence and Statistics , pages=

    Generative flow networks as entropy-regularized rl , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

    2024

  43. [43]

    International conference on machine learning , pages=

    Reinforcement learning with deep energy-based policies , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  44. [44]

    Nature , volume=

    Mastering atari, go, chess and shogi by planning with a learned model , author=. Nature , volume=. 2020 , publisher=

    2020

  45. [45]

    Journal of Machine Learning Research , volume=

    Gflownet foundations , author=. Journal of Machine Learning Research , volume=

  46. [46]

    1974 , publisher=

    Advanced Combinatorics: The Art of Finite and Infinite Expansions , author=. 1974 , publisher=

    1974

  47. [47]

    The Twelfth International Conference on Learning Representations , year=

    Amortizing intractable inference in large language models , author=. The Twelfth International Conference on Learning Representations , year=

  48. [48]

    The Twelfth International Conference on Learning Representations , year=

    Order-Preserving GFlowNets , author=. The Twelfth International Conference on Learning Representations , year=

  49. [49]

    International Conference on Machine Learning , pages=

    Biological sequence design with gflownets , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  50. [50]

    Let the Flows Tell: Solving Graph Combinatorial Problems with GFlowNets , volume =

    Zhang, Dinghuai and Dai, Hanjun and Malkin, Nikolay and Courville, Aaron C and Bengio, Yoshua and Pan, Ling , booktitle =. Let the Flows Tell: Solving Graph Combinatorial Problems with GFlowNets , volume =

  51. [51]

    and Rupp, Matthias and von Lilienfeld, O

    Ramakrishnan, Raghunathan and Dral, Pavlo O. and Rupp, Matthias and von Lilienfeld, O. Anatole , journal =. Quantum chemistry structures and properties of 134 kilo molecules , volume =

  52. [52]

    2020 , eprint=

    Molecular Mechanics-Driven Graph Neural Network with Multiplex Graph for Molecular Structures , author=. 2020 , eprint=

    2020

  53. [53]

    International Conference on Machine Learning , pages=

    Multi-objective gflownets , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  54. [54]

    Advances in Neural Information Processing Systems , volume=

    DynGFN: Towards Bayesian Inference of Gene Regulatory Networks with GFlowNets , author=. Advances in Neural Information Processing Systems , volume=

  55. [55]

    International Conference on Artificial Intelligence and Statistics , pages=

    Maximum entropy GFlowNets with soft Q-learning , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

    2024

  56. [56]

    nature , volume=

    Mastering the game of Go with deep neural networks and tree search , author=. nature , volume=. 2016 , publisher=

    2016

  57. [57]

    International conference on computers and games , pages=

    Efficient selectivity and backup operators in Monte-Carlo tree search , author=. International conference on computers and games , pages=. 2006 , organization=

    2006

  58. [58]

    Science , volume =

    David Silver and Thomas Hubert and Julian Schrittwieser and Ioannis Antonoglou and Matthew Lai and Arthur Guez and Marc Lanctot and Laurent Sifre and Dharshan Kumaran and Thore Graepel and Timothy Lillicrap and Karen Simonyan and Demis Hassabis , title =. Science , volume =. 2018 , doi =

    2018

  59. [59]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

  60. [60]

    Digital Discovery , volume=

    Gflownets for ai-driven scientific discovery , author=. Digital Discovery , volume=. 2023 , publisher=

    2023

  61. [61]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  62. [62]

    A unified view of entropy-regularized Markov decision processes

    A unified view of entropy-regularized markov decision processes , author=. arXiv preprint arXiv:1705.07798 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  63. [63]

    International Conference on Machine Learning , pages=

    A theory of regularized markov decision processes , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  64. [64]

    Equivalence Between Policy Gradients and Soft Q-Learning

    Equivalence between policy gradients and soft q-learning , author=. arXiv preprint arXiv:1704.06440 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  65. [65]

    European conference on machine learning , pages=

    Bandit based monte-carlo planning , author=. European conference on machine learning , pages=. 2006 , organization=

    2006

  66. [66]

    nature , volume=

    Human-level control through deep reinforcement learning , author=. nature , volume=. 2015 , publisher=

    2015

  67. [67]

    Prioritized Experience Replay , booktitle =

    Tom Schaul and John Quan and Ioannis Antonoglou and David Silver , editor =. Prioritized Experience Replay , booktitle =. 2016 , url =

    2016

  68. [68]

    International Conference on Machine Learning , pages=

    Better training of gflownets with local credit and incomplete trajectories , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  69. [69]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Ergodic Generative Flows , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , volume =

    2025

  70. [70]

    2019 , publisher=

    Computational optimal transport: With applications to data science , author=. 2019 , publisher=

    2019

  71. [71]

    2021 , publisher=

    Topics in optimal transportation , author=. 2021 , publisher=

    2021

  72. [72]

    2009 , publisher=

    Optimal transport: old and new , author=. 2009 , publisher=

    2009

  73. [73]

    The Thirteenth International Conference on Learning Representations , year=

    Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks , author=. The Thirteenth International Conference on Learning Representations , year=

  74. [74]

    Advances in Neural Information Processing Systems , volume=

    On divergence measures for training gflownets , author=. Advances in Neural Information Processing Systems , volume=

  75. [75]

    International Conference on Machine Learning , pages=

    Generative flow networks for discrete probabilistic modeling , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  76. [76]

    arXiv preprint arXiv:2305.14594 , year=

    torchgfn: A PyTorch GFlowNet library , author=. arXiv preprint arXiv:2305.14594 , year=

    work page arXiv

  77. [77]

    Taming the Noise in Reinforcement Learning via Soft Updates , booktitle =

    Roy Fox and Ari Pakman and Naftali Tishby , editor =. Taming the Noise in Reinforcement Learning via Soft Updates , booktitle =. 2016 , url =

    2016

  78. [78]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor , author =. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , editor =

    2018

  79. [79]

    Advances in Neural Information Processing Systems , volume=

    Munchausen reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  80. [80]

    International Conference on Learning Representations , year=

    Mirror Descent Policy Optimization , author=. International Conference on Learning Representations , year=

Showing first 80 references.