Pith. sign in

REVIEW 4 major objections 4 minor 29 references

Exploring Generative Networks for Manifolds with Non-Trivial Topology

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A diffusion-style sampler that chooses between a forward and a backward proposal at each step can visit all sectors of a disconnected target manifold, where normalizing flows and ordinary diffusion models collapse to a single sector.

desk verdict Nice proof-of-concept with a genuine new architecture, but the central balance condition is unproven and the numerics are qualitative. read the letter →

arxiv 2502.02127 v1 pith:CX4WNZSF submitted 2025-02-04 hep-lat

Shiyang Chen , Gert Aarts , Biagio Lucini This is my paper · ORCID
classification hep-lat
keywords GFlowNetsgenerativemodelslatticefieldtheorytopologicalfreezingdiffusionnormalizingflowsmodecollapsephi^4model
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

Normalizing flows and unsupervised diffusion models fail on target distributions whose support is disconnected---the triple-ring toy model and the two magnetisation sectors of a two-dimensional $phi^{4}$ lattice theory---because the learned map keeps the connectivity of a unimodal prior. The paper introduces a GFlowNet-inspired diffusion model in which every diffusion step proposes two states, one from the forward Langevin process and one from the backward process, and a small network fed with the action chooses which pathway to accept. Trained directly on the target density through the KL divergence, this network generates all three rings in the toy model and the expected two-peaked magnetisation on a $64\times32$ lattice. If the construction is correct, it offers a generative route around topological freezing: sampling that is ergodic across topological sectors without needing a pre-existing ensemble.

What carries the argument

The load-bearing object is the stochastic pathway selector: at each fictitious time $t$ the forward and backward Langevin networks each produce a candidate state, $z_{f,t}$ and $z_{b,t}$, both conditioned on the action $S(z_t)$, and a dedicated dense network outputs the probability of taking the forward pathway, capped below $0.5$. The identity that carries the argument is Eq. (6), which expresses $P(z_t)/P(z_{t-1})$ as a weighted combination of two delta-function terms, one accepting the forward proposal with its backward posterior and one accepting the backward proposal with its forward posterior; substituting this ratio into the KL objective of Eq. (4) makes the whole construction trainable directly on the target density. This is the paper's GFlowNet-inspired mechanism: a flow-balance condition over paths rather than an invertible map between densities. The action input lets the network see the sector structure---it can be discontinuous where the topology changes---and the below-$0.5$ cap prevents the chain from collapsing onto a single pathway.

What would settle it

Train the pathway-choosing sampler on a triple-ring target whose three ring weights are known analytically, run a long chain, and compare the empirical fraction of samples in each ring to the exact weights with statistical errors; a deviation larger than the expected fluctuations would show that Eq. (6) is not a valid balance condition. The same test can be run on the $\phi^4$ magnetisation by comparing the two-peak histogram against a long reference MCMC run at the same parameters.

Watch

Extended reading notes

Core claim

The central claim is that a diffusion generative model can sample ergodically from a manifold with non-trivial topology if, at each diffusion step, the transition is a learned stochastic choice between the forward-proposed state $z_{f,t}$ and the backward-proposed state $z_{b,t}$, with the action $S(z_t)\propto -\log P(z_t)$ supplied as input. The paper writes the ratio $P(z_t)/P(z_{t-1})$ in Eq. (6) as the sum of a forward-accepted and a backward-accepted branch, each weighted by the output of an acceptance network whose probability is capped below $0.5$ to avoid pathway collapse, and plugs this ratio into the KL divergence (Eq. (4)) to train all networks without data. On the triple-ring model the trained network generates all three rings, and on a $64\times32$ lattice $\phi^4$ model at $\lambda=0.022$, $\kappa=0.3$ it produces the two-peaked magnetisation histogram required by the $\mathbb{Z}_2$ symmetry; normalizing flows and unsupervised diffusion models trained on the same problems collapse to one sector.

Load-bearing premise

The method's guarantee of unbiased sampling depends on the ratio in Eq. (6) being exactly the right update rule for the chain; the paper states this ratio and trains the network with it, but does not prove that the trained network actually satisfies it, so if the ratio is wrong or the below-0.5 cap on pathway probability biases the walk, the three-ring and two-peak histograms are not evidence of ergodic sampling.

Editorial extensions

If this is right

  • On the triple-ring target, the method generates all three rings, while normalizing flows and unsupervised diffusion models generate only the inner ring.
  • On a $64\times32$ two-dimensional $\phi^4$ lattice at $\lambda=0.022$, $\kappa=0.3$, the magnetisation histogram shows the two peaks demanded by $\mathbb{Z}_2$ symmetry; the comparison methods show one peak.
  • The action is fed into the proposal networks as an additional input, so the sampler can react to the sector structure through a log-density that may be discontinuous across sectors.
  • Because the model is trained by minimising the KL divergence to the target density, it needs no pre-existing ensemble of configurations.
  • If the balance condition holds at larger volumes and for gauge theories, the architecture is a candidate for generative sampling that bypasses topological freezing in lattice field theory.

Reading between the lines

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

  • The paper does not prove that the optimized chain obeys Eq. (6) as an exact balance condition; a direct test would be to compare the empirical occupation frequencies of the three rings against analytically known weights, with error bars, and to vary the $0.5$ cap to see whether the cap itself biases the stationary distribution.
  • Because the acceptance probability is capped below $0.5$, the sampler never commits fully to the forward route; in problems with more than two sectors an asymmetric or multiway choice may be a natural extension.
  • The $\phi^4$ demonstration is a single small volume; whether the mechanism remains ergodic across the larger free-energy barriers of lattice gauge theories, where topological freezing is severe, is the next stress test.
  • Adding a Metropolis-Hastings accept/reject step on top of the learned pathway choice would convert the sampler into an exactly unbiased MCMC chain and would test whether the learned balance is already sufficient.
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

4 major / 4 minor

Summary. This paper proposes a stochastic-pathway generative model, inspired by GFlowNets, intended to sample distributions with multiple disconnected topological sectors. The model augments a diffusion process with a learned choice between forward and backward proposals, includes the action as an input, and is trained by minimizing a KL divergence between a path-probability ratio and the target. The authors apply the method to a triple-ring toy model and to the magnetization distribution of the 2D phi^4 model, reporting that normalizing flows and unsupervised diffusion models collapse to a single sector while the proposed network reproduces the full distribution. The paper is a proceedings-style contribution in which the central positive claim rests on an analytical ratio, Eq. (6), and on visual histograms in Figs. 6 and 7.

Significance. The problem addressed, bypassing topological freezing and mode collapse in lattice field theory, is important and timely. The paper gives a useful negative demonstration that normalizing flows and unsupervised diffusion models collapse to a single sector on the triple-ring model and on the phi^4 magnetization distribution. However, the central positive claim is not established: Eq. (6), the balance condition that defines the training objective, is asserted without derivation and, as written, is not a valid marginal density ratio; the numerical evidence is exclusively visual, with no error bars or quantitative comparison; and no measurements of transitions between sectors are shown. Because the method's correctness rests on Eq. (6), the contribution in its current form does not provide a reliable sampler. If the derivation and validation were repaired, the underlying idea could be a valuable step toward generative sampling across topological sectors.

major comments (4)
  1. [Section 3, Eq. (6)] Eq. (6) is the central balance condition used to define the KL objective in Eq. (4), but it is not derived and, as written, is not a ratio of marginal densities. The left-hand side P(z_t)/P(z_{t-1}) depends only on z_t and z_{t-1}, while the right-hand side contains the proposed states z_{f,t} and z_{b,t} and the selection network evaluated on those proposals. In the algorithm described, these proposals are random variables, and the transition kernel from z_{t-1} to z_t is a mixture density of the form K(z_t|z_{t-1}) = integral over the rejected proposal z_r of [P_f(z_t|z_{t-1})P_b(z_r|z_{t-1})alpha(z_{t-1},z_t,z_r) + P_b(z_t|z_{t-1})P_f(z_r|z_{t-1})(1-alpha(z_{t-1},z_r,z_t))]. No such integration over the rejected branch appears in Eq. (6), and the Dirac deltas at the proposed states do not define a genuine density with respect to z_t. Consequently, minimizing the KL divergence in Eq. (4) using Eq. (6) does not demonstrably match the generated marginal P(z_t) to the target Q(z_t).
  2. [Section 3, Fig. 4] The restriction that the selection-network output is "limited and less than 0.5 to prevent pathway collapse" is not derived from any balance condition. In a valid two-branch proposal scheme, the probability of choosing the forward proposal must satisfy a detailed-balance-like relation with the proposal densities; imposing an arbitrary upper bound below 1/2 in general breaks that relation and biases the chain. The paper gives no argument that the cap is compatible with the target distribution. Since this selection network is the mechanism claimed to enable exploration of different topological sectors, such a bias would directly affect the central claim.
  3. [Section 3, Fig. 6; Section 4, Fig. 7] The empirical demonstration is exclusively visual. Figs. 6 and 7 show scatter plots and histograms with no error bars, no number of samples, no autocorrelation or effective sample size, and no quantitative comparison to the target such as a histogram distance, a two-sample test, or a KL divergence estimate. The text itself states that "the samples are noisy" in the paragraph after Fig. 6, and the phi^4 histograms in Fig. 7 are not accompanied by an exact or reference distribution. Visual agreement with a multimodal target can be produced by a biased mixture over sectors or by initialization in all sectors; without quantitative or dynamical evidence the claim of ergodic unbiased sampling is not supported.
  4. [Section 4, Fig. 7; Section 5] No evidence of transitions between topological sectors is presented. The triple-ring target has three disjoint support components, and the magnetization distribution in the phi^4 case has two sectors. Neither Fig. 6 nor Fig. 7 reports the sequence of sector indices, the number of transitions between sectors, or the autocorrelation time of the sector variable. A model that assigns mass to all three rings from initialization but never transits between them would reproduce the static histograms while remaining topologically frozen. The word "ergodically" in the abstract is therefore not evidenced by the current data.
minor comments (4)
  1. [Section 2, Eq. (4)] Equation (4) has unbalanced parentheses: D_KL(P(z_t || Q(z_t)) should be D_KL(P(z_t) || Q(z_t)).
  2. [General notation] The symbol P is overloaded: it denotes the learnt distribution, the forward and backward proposal densities, and also the selection network, e.g., P(z_t|z_{f,t-1}, z_t, z_{b,t-1}) in Eq. (6) and Fig. 4. Using a distinct symbol such as alpha for the selection probability would make the equations substantially easier to parse.
  3. [Section 2] The statement that "the connectivity of the manifolds K(z_t) and D(z_t) will determinate the topology" is informal. The image of a connected set under a continuous map is connected, but this does not by itself imply that the diffusion sampler cannot cross low-density regions or that the support of the stationary distribution is connected; the claim should be rephrased more carefully.
  4. [Research Data and Code Access] The manuscript states that code and data are available upon request, but for reproducibility and for the quantitative checks requested above, a public repository with the exact network architecture, hyperparameters, and training details would be preferable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the target distribution enters as the training objective, and the questioned balance condition is an unsupported technical claim, not a circular reduction.

full rationale

The paper optimizes the learned path measure P(z_t) against the target Q(z_t) via the KL divergence in Eq. (4), so agreement with the target in Figs. 6 and 7 is the optimization objective, not an independent prediction. This is the standard mode of operation for generative samplers and does not constitute circular reasoning: the target action/log-density is an input, not a reconstructed output. The load-bearing balance condition in Eq. (6) is asserted without derivation and may be mathematically suspect, but an unproven or incorrect equation is a correctness risk, not a circularity, because it does not reduce the claimed result to its inputs by construction or by self-citation. The paper's references to GFlowNets [27,28] and to diffusion models [9] are external or non-load-bearing; it does not invoke a self-citation uniqueness theorem, does not fit a parameter to a subset and then predict the same quantity, and does not smuggle the ansatz in via citation. The self-citations involving co-author G. Aarts (e.g., Refs. [9,12]) are contextual and do not carry the central claim. No circular step can be exhibited, so the appropriate score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on three unproved or weakly justified premises: the topology-preservation argument for normalizing flows (Section 2), the connectivity-to-topology mapping for diffusion models (Section 2), and, most importantly, the balance condition in Eq. (6) (Section 3). The acceptance probability bound less than 0.5 is a hand-chosen parameter. There are no invented physical entities.

free parameters (1)
  • Acceptance probability upper bound = < 0.5
    Section 3: 'The output probability is limited and less than 0.5 to prevent pathway collapse.' This hand-chosen bound changes the transition probabilities and is not derived from a detailed balance or flow-matching condition.
assumptions (4)
  • standard math Normalizing flows are homeomorphisms between prior and target manifolds (Eq. (1)).
    Invoked in Section 2 to argue NFs cannot change topology; a standard property of diffeomorphisms, but the relevance to multimodal targets is qualitative.
  • domain assumption Connectivity of K(z_t) and D(z_t) determines the topology of the diffusion output z_{t+1}.
    Section 2, after Eq. (2); stated without proof. The Langevin noise term makes the support of z_{t+1} non-trivial, so this premise is not rigorous.
  • ad hoc to paper Eq. (6) is a valid balance equation for the forward and backward proposal process.
    Section 3; the ratio mixes delta functions and density functions, and no proof is given that optimizing the KL in Eq. (4) with this ratio yields unbiased samples from P(z_t).
  • domain assumption GFlowNet foundations (Refs. [27,28]) justify the training objective used here.
    The paper cites GFlowNets but does not show the mapping between Eq. (6) and a known GFlowNet loss; this connection is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Generative Networks for Manifolds with Non-Trivial Topology." pith.science (2026). https://pith.science/paper/CX4WNZSF

@misc{pith2026250202127,
  author       = {Pith},
  title        = {Pith review of: Exploring Generative Networks for Manifolds with Non-Trivial Topology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CX4WNZSF}},
  note         = {Machine review of arXiv:2502.02127}
}
read the original abstract

The expressive power of neural networks in modelling non-trivial distributions can in principle be exploited to bypass topological freezing and critical slowing down in simulations of lattice field theories. Some popular approaches are unable to sample correctly non-trivial topology, which may lead to some classes of configurations not being generated. In this contribution, we present a novel generative method inspired by a model previously introduced in the ML community (GFlowNets). We demonstrate its efficiency at exploring ergodically configuration manifolds with non-trivial topology through applications such as triple ring models and two-dimensional lattice scalar field theory.

Figures

Figures reproduced from arXiv: 2502.02127 by the authors.

Figure 1
Figure 1. The architecture for forward and backward processes at a fictitious time 𝑡. (a) Target distribution. (b) Prior distribution. (c) Distribution generated using normalizing flow. (d) Distribution generated using an unsupervised diffusion model [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Target and prior distributions (above) and incomplete distributions generated using normalizing flow and an unsupervised diffusion model (below). 4 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Improved architecture for the forward and backward processes at fictitious time 𝑡. One more neural net 𝑃(𝑧𝑓 ,𝑡, 𝑧𝑡 , 𝑧𝑏,𝑡) with the proposed states 𝑧𝑓 ,𝑡 and 𝑧𝑏,𝑡 from the forward and backward processes is introduced to decide which pathways will be accepted. The network flattens the augmented pair 𝑧𝑓 ,𝑡 and 𝑧𝑏,𝑡, and then computes the probability of the upper pathway via two dense layers with 50 neurons and ReLU an… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The architecture for 𝑃(𝑧𝑡 |𝑧𝑓 ,𝑡−1, 𝑧𝑡 , 𝑧𝑏,𝑡−1) and 𝑃(𝑧𝑡−1|𝑧𝑏,𝑡, 𝑧𝑡 , 𝑧𝑓 ,𝑡−1) at fictitious time 𝑡. 𝑃(𝑧𝑡) 𝑃(𝑧𝑡−1) = 𝛿(𝑧𝑡 − 𝑧𝑓 ,𝑡) 𝑃(𝑧𝑡 |𝑧𝑓 ,𝑡−1, 𝑧𝑡 , 𝑧𝑏,𝑡−1)𝑃𝑓 ,𝑡−1 (𝑧𝑓 ,𝑡 |𝑧𝑡−1, 𝑆(𝑧𝑡−1)) 𝑃(𝑧𝑡−1|𝑧𝑏,𝑡, 𝑧𝑡 , 𝑧𝑓 ,𝑡)𝑃𝑏,𝑡(𝑧𝑡−1|𝑧𝑓 ,𝑡, 𝑆(𝑧𝑓 ,𝑡)) + 𝛿(𝑧𝑡 − 𝑧𝑏,𝑡) 𝑃(𝑧𝑡 |𝑧𝑓 ,𝑡−1…
Figure 5
Figure 5. Figure 5: The structure of the novel network at diffusion time 𝑡 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Distribution generated using our novel neural net. where 𝜅 is the hopping parameter and 𝜆 the bare coupling constant. The model has a 𝑍2 symmetry. The symmetry can be observed by estimating the magnetization, 𝑀 = 1 𝑁𝑇𝑁𝐿 ∑︁ 𝑥∈Λ 𝜙𝑥 . (8) We have estimated the magnetizati…
Figure 7
Figure 7. Figure 7: Histograms of the magnetisation 𝑀 in the two-dimensional scalar field theory obtained using the novel architecture introduced in this work (above), normalising flow (middle) and unsupervised diffusion models (below). three disconnected components, both NFs and unsuperv…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 canonical work pages

  1. [1]

    Wolff,CRITICAL SLOWING DOWN, Nucl

    U. Wolff,CRITICAL SLOWING DOWN, Nucl. Phys. B Proc. Suppl.17(1990) 93. 8 Exploring Generative Networks for Manifolds with Non-Trivial Topology Shi-Yang Chen

    work page 1990

  2. [2]

    Del Debbio, G.M

    L. Del Debbio, G.M. Manca and E. Vicari,Critical slowing down of topological modes, Phys. Lett. B594 (2004) 315 [hep-lat/0403001]

    arXiv 2004

  3. [3]

    ALPHA collaboration,Critical slowing down and error analysis in lattice QCD simulations, Nucl. Phys. B845(2011) 93 [1009.5228]

    arXiv 2011

  4. [4]

    Cranmer, G

    K. Cranmer, G. Kanwar, S. Racanière, D.J. Rezende and P.E. Shanahan,Advances in machine-learning-based sampling motivated by lattice quantum chromodynamics, Nature Rev. Phys.5 (2023) 526 [2309.01156]

    arXiv 2023

  5. [5]

    Kanwar,Flow-based sampling for lattice field theories, PoSLATTICE2023(2024) 114 [2401.01297]

    G. Kanwar,Flow-based sampling for lattice field theories, PoSLATTICE2023(2024) 114 [2401.01297]

    arXiv 2024

  6. [6]

    Aarts, K

    G. Aarts, K. Fukushima, T. Hatsuda, A. Ipp, S. Shi, L. Wang et al.,Physics-driven learning for inverse problems in quantum chromodynamics,Nature Rev. Phys.(2025) [2501.05580]

    arXiv 2025

  7. [7]

    Pawlowski and J.M

    J.M. Pawlowski and J.M. Urban,Reducing Autocorrelation Times in Lattice Simulations with Generative Adversarial Networks, Mach. Learn. Sci. Tech.1 (2020) 045011 [1811.03533]

    arXiv 2020

  8. [8]

    L. Wang, Y. Jiang, L. He and K. Zhou,Continuous-Mixture Autoregressive Networks Learning the Kosterlitz-Thouless Transition,Chin. Phys. Lett.39(2022) 120502 [2005.04857]

    work page Pith review arXiv 2022

Show all 29 references
  1. [9]

    L. Wang, G. Aarts and K. Zhou,Diffusion models as stochastic quantization in lattice field theory, JHEP05(2024) 060 [2309.17082]

    2024 arXiv

  2. [10]

    L. Wang, G. Aarts and K. Zhou,Generative Diffusion Models for Lattice Field Theory, in 37th Conference on Neural Information Processing Systems, 2023 [2311.03578]

    2023 arXiv

  3. [11]

    Q. Zhu, G. Aarts, W. Wang, K. Zhou and L. Wang,Diffusion models for lattice gauge field simulations, in38th conference on Neural Information Processing Systems, 2024 [2410.19602]

    2024 arXiv

  4. [12]

    Aarts, D.E

    G. Aarts, D.E. Habibi, L. Wang and K. Zhou,On learning higher-order cumulants in diffusion models, in38th conference on Neural Information Processing Systems, 2024 [2410.21212]

    2024 arXiv

  5. [13]

    Rezende and S

    D.J. Rezende and S. Mohamed,Variational Inference with Normalizing Flows, in International conference on machine learning, pp. 1530–1538, PMLR, 2015 [1505.05770]

    2015 arXiv

  6. [14]

    F. Noé, S. Olsson, J. Köhler and H. Wu,Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning, Science365 (2019) eaaw1147 [1812.01729]

    2019 arXiv

  7. [15]

    Albergo, G

    M.S. Albergo, G. Kanwar and P.E. Shanahan,Flow-based generative models for Markov chain Monte Carlo in lattice field theory, Phys. Rev. D100 (2019) 034515 [1904.12072]

    2019 arXiv

  8. [16]

    Nicoli, S

    K.A. Nicoli, S. Nakajima, N. Strodthoff, W. Samek, K.-R. Müller and P. Kessel, Asymptotically unbiased estimation of physical observables with neural samplers, Phys. Rev. E 101 (2020) 023304 [1910.13496]. 9 Exploring Generative Networks for Manifolds with Non-Trivial Topology ...

    2020 arXiv

  9. [17]

    Kanwar, M.S

    G. Kanwar, M.S. Albergo, D. Boyda, K. Cranmer, D.C. Hackett, S. Racanière et al., Equivariant Flow-Based Sampling for Lattice Gauge Theory, Phys. Rev. Lett.125 (2020) 121601 [2003.06413]

    2020 arXiv

  10. [18]

    Nicoli, C.J

    K.A. Nicoli, C.J. Anders, L. Funcke, T. Hartung, K. Jansen, P. Kessel et al.,Estimation of Thermodynamic Observables in Lattice Field Theories with Deep Generative Models,Phys. Rev. Lett.126 (2021) 032001 [2007.07115]

    2021 arXiv

  11. [19]

    Nicoli, C.J

    K.A. Nicoli, C.J. Anders, T. Hartung, K. Jansen, P. Kessel and S. Nakajima,Detecting and mitigating mode-collapse for flow-based sampling of lattice field theories, Phys. Rev. D108 (2023) 114501 [2302.14082]

    2023 arXiv

  12. [20]

    R.T. Chen, Y. Rubanova, J. Bettencourt and D.K. Duvenaud,Neural ordinary differential equations, Advances in neural information processing systems31(2018) [1806.07366]

    2018 arXiv

  13. [21]

    de Haan, C

    P. de Haan, C. Rainone, M.C.N. Cheng and R. Bondesan,Scaling Up Machine Learning For Quantum Field Theory with Equivariant Continuous Flows, 2110.02673

    arXiv

  14. [22]

    Gerdes, P

    M. Gerdes, P. de Haan, C. Rainone, R. Bondesan and M.C.N. Cheng,Learning lattice quantum field theories with equivariant continuous flows,SciPost Phys.15(2023) 238 [2207.00283]

    2023 arXiv

  15. [23]

    Caselle, E

    M. Caselle, E. Cellini and A. Nada,Sampling the lattice Nambu-Goto string using Continuous Normalizing Flows, Journal of High Energy Physics02 (2024) 048 [2307.01107]

    2024 arXiv

  16. [24]

    H. Wu, J. Köhler and F. Noé,Stochastic normalizing flows, Advances in Neural Information Processing Systems33(2020) 5933 [2002.06707]

    2020 arXiv

  17. [25]

    Caselle, E

    M. Caselle, E. Cellini, A. Nada and M. Panero,Stochastic normalizing flows as non-equilibrium transformations,JHEP07 (2022) 015 [2201.08862]

    2022 arXiv

  18. [26]

    Caselle, E

    M. Caselle, E. Cellini and A. Nada,Numerical determination of the width and shape of the effective string using Stochastic Normalizing Flows, 2409.15937

    arXiv

  19. [27]

    Bengio, M

    E. Bengio, M. Jain, M. Korablyov, D. Precup and Y. Bengio,Flow network based generative models for non-iterative diverse candidate generation, 2106.04399

    arXiv

  20. [28]

    Bengio, T

    Y. Bengio, T. Deleu, E.J. Hu, S. Lahlou, M. Tiwari and E. Bengio,Gflownet foundations, 2111.09266

  21. [29]

    Albergo, D

    M.S. Albergo, D. Boyda, D.C. Hackett, G. Kanwar, K. Cranmer, S. Racanière et al., Introduction to Normalizing Flows for Lattice Field Theory, 2101.08176. 10

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.