Pith. sign in

REVIEW 4 major objections 6 minor 52 references

A survey of probabilistic generative frameworks for molecular simulations

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read No single generative model wins across molecular benchmarks; the best choice depends on data dimensionality, complexity, and mode asymmetry.

desk verdict A useful benchmark survey whose pairwise rankings need repeated-seed error bars and binning sensitivity before they can be taken as statistically grounded. read the letter →

arxiv 2411.09388 v1 pith:Y44DSP6V submitted 2024-11-14 cs.LG cond-mat.dis-nncond-mat.softcond-mat.stat-mech

classification cs.LGcond-mat.dis-nncond-mat.softcond-mat.stat-mech
keywords generativemodelsnormalizingflowsdiffusionconditionalflowmatchingmolecularsimulationdihedralanglesfree-energyestimationbenchmarking
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 sets out to answer a practical question for molecular simulations: when should you use a flow-based generative model versus a diffusion model? It compares three representative architectures — Neural Spline Flows (NS), Conditional Flow Matching (CFM), and Denoising Diffusion Probabilistic Models (DDPM) — on a Gaussian mixture with tunable dimensionality and on the dihedral torsion angles of the Aib9 peptide from explicit-water molecular dynamics. The central finding is that there is no universal winner: NS is most faithful for free-energy differences between asymmetric modes, CFM is most accurate and fastest at high dimensionality with simple structure, and DDPM best reproduces complex, multimodal low-dimensional distributions. This matters because generative models are already used widely for molecular tasks, and the choice of architecture has lacked systematic comparison-based guidance.

What carries the argument

The unifying object is the change-of-measure identity $p(x)=q(x')/|J_M(x')|$, which lets any invertible map $M$ define a candidate probability density and turns generative modeling into the task of optimizing $M$. NS parameterizes $M$ with invertible neural spline layers; CFM learns a drift field along an optimal-transport bridge; DDPM learns the score $\nabla\log p(x,t)$ of a diffusion process and integrates the probability-flow ODE. The comparison itself is carried by two tunable benchmark datasets — a Gaussian mixture whose dimensionality, training set size, and mode asymmetry can be varied, and the $\{\Phi,\Psi\}$ dihedral angles of Aib9 whose residue index tunes distributional complexity — with accuracy measured by projecting generated samples onto the two principal components of the training data and computing KL divergence between binned histograms.

What would settle it

Re-run the Gaussian-mixture dimensionality sweep (dimensionality 10 to 100) with at least five bin counts or widths and ten random seeds per model, computing the rank order of NS, CFM, and DDPM by KL divergence; if the ordering is not stable across binning choices, the paper's regime-specific conclusions fail on their own metric.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a mapping between data geometry and model family. For the 4-modal Gaussian mixture, CFM achieves the lowest KL divergence once dimensionality reaches about 40 and remains stable beyond that, while NS error rises sharply at high dimensionality. When the mixture is made asymmetric in a 50-dimensional bimodal case, NS reproduces the training free-energy difference most faithfully, with CFM second and DDPM last. On the Aib9 peptide, where exterior residues sample many metastable conformations, DDPM gives the lowest per-residue KL divergence and handles small training sets as well as or better than NS, while CFM performs least well. The paper reads these results as evidence that no framework dominates and that dimensionality, complexity, and modal asymmetry jointly determine the right choice.

Load-bearing premise

Every accuracy comparison rests on projecting generated samples onto two principal components of the training data, binning them, and scoring with KL divergence using an unspecified bin count and width, and the free-energy test additionally depends on an arbitrary 0.0374 kJ/mol cutoff for excluding high-energy bins; change these choices and the reported rankings could shift.

Editorial extensions

If this is right

  • CFM should be the default first try for high-dimensional molecular data with relatively simple basins, where it also offers the fastest sample generation.
  • DDPM is the better choice for low-dimensional but strongly multimodal distributions, such as peptide torsion-angle distributions, especially when training data are scarce.
  • NS is the model to reach for when the quantity of interest is a free-energy difference between asymmetric states, despite its poor scaling to high dimensions.
  • The two datasets provide reusable benchmarks with tunable complexity, so newly introduced flow and diffusion variants can be checked against the same regime structure.

Reading between the lines

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

  • One testable extension is to replace the fixed two-component PCA projection and unspecified binning with a metric that is insensitive to bin width and repeated seeds; if the rankings survive, the regime rules are robust, and if not, the reported ordering is partly an artifact of the metric.
  • Implicit in the results is a possible selection rule for molecular pipelines: estimate the effective dimensionality and modal complexity of the target distribution first, then pick the family, rather than reusing the same architecture across tasks.
  • The same benchmark logic could be applied to newer methods such as rectified flows or latent diffusion models, which the paper names as future candidates, to see whether the regime boundaries shift.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript surveys three probabilistic generative frameworks—Neural Spline Flows, Conditional Flow Matching, and Denoising Diffusion Probabilistic Models—and benchmarks them on two molecular-relevant datasets: a Gaussian mixture model with tunable dimensionality, training-set size, and mode asymmetry, and the Aib9 peptide dihedral-angle distribution from an explicit-water MD simulation. Accuracy is measured by KL divergence between binned two-dimensional PCA projections of held-out test samples and generated samples; the paper also reports generation speed and parameter counts. The central empirical claim is that no single framework is uniformly best: NS is most accurate at estimating probability density differences, CFM is most accurate at high dimensionality with low complexity, and DDPM is most accurate for the complex, multimodal Aib9 distribution. Code and data are made publicly available.

Significance. If the rankings are statistically robust, this is a useful contribution: it provides a reusable benchmark with tunable dimensionality, complexity, and asymmetry, and it offers concrete model-selection guidance in a field where systematic comparisons are scarce. The paper's strengths are the clear taxonomy, the public code and datasets, and the direct measurement of accuracy on held-out test data. The main weakness is that the central ranking claims currently rest on a metric that is incompletely specified and on single-run measurements without error bars, so the quantitative conclusions are not yet statistically grounded.

major comments (4)
  1. [Section IV] The DKL accuracy metric is not fully specified: the text states that generated samples are projected onto the two principal components of the training data and binned, but it never gives the number of bins, the bin width, or the binning scheme. Because Figures 1a, 1b, 4a, and 4b all use this metric, the reported orderings, such as CFM versus DDPM at dimensions 40 to 100 in Figure 1a and DDPM versus NS in Figure 4a, could change if the binning were changed. Please specify the binning protocol and provide a sensitivity analysis over bin counts.
  2. [Figures 1-4] No figure reports repeated-seed variation or error bars. All DKL values appear to be single runs, so the margins that separate the models in Figures 1, 2, and 4 may be within run-to-run noise. Please run multiple random seeds, report the mean and variance or confidence intervals for every measured quantity, and state a significance criterion for the ranking claims.
  3. [Section IV A, Figure 2] The free-energy comparison depends on an arbitrarily chosen cutoff of 0.0374 kJ/mol applied to exclude high-free-energy bins. No justification is given for this value, and no sensitivity analysis is reported; because the r2 values and the claimed NS superiority are computed from the resulting partition-function estimates, the ranking may not be robust to this choice. In addition, the units are inconsistent: Eq. (1) uses dimensionless βU, while the cutoff is quoted in kJ/mol with β = 1. Please justify the cutoff, report its sensitivity, and correct the units.
  4. [Appendix B; Section IV] Hyperparameters (NS 'layers', CFM/DDPM 'model dimension') appear to be selected using the same benchmark distributions on which the models are later scored, with no separate validation split described. This can bias the comparison in favor of the model whose tuning surface was explored more thoroughly. Please state the validation procedure or re-tune on a held-out portion of the training data.
minor comments (6)
  1. [Eq. (9)] Equation (9) appears incorrect as written: the continuity equation for the probability flow should involve a divergence of the vector field, not exp(−tr J), and the right-hand side has mismatched dimensions. Please correct or clarify.
  2. [Eqs. (12)-(13)] Equation (12) defines h(x,t) = x − ∇log p(x,t), which seems inconsistent with the standard linear-drift SDE in Eq. (13); please check the sign and convention so that the forward and reverse processes are consistently defined.
  3. [Section IV B] The text says 'The {Φ, Ψ}4 plots at residue 1 are shown in Figure 4c-f'; this appears to be a typo for Figure 4c-f, and the sentence should be reworded to avoid confusion with residue 4.
  4. [References] The reference list contains duplicates: Rombach et al. appears as refs 1 and 53, and Lipman et al. appears as refs 18 and 49. Please consolidate.
  5. [Figure 3a] Figure 3a reports sample generation time without stating the hardware or the number of repeated timing measurements; please specify these details so the timing comparison is reproducible.
  6. [Section II] The statement that the KL divergence provides an upper bound on ΔFpq is made without a citation or derivation; please add a reference or a brief justification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model rankings are empirical measurements against held-out data, not derivations that reduce to fitted inputs or self-citations.

full rationale

The paper's central claims are comparative empirical benchmark results. The accuracy metric is defined in Section IV as a PCA projection of generated and held-back samples onto training-data principal components, followed by binning and KL divergence computation. This is an external measurement of sample quality, not a quantity whose definition already encodes the conclusion. The free-energy comparison in Section IV A applies the same histogram-based estimator to both training and generated samples; the reported r2 therefore measures whether generated samples reproduce the training distribution, which is an empirical property of the samples rather than a construct imposed by the model equations. Hyperparameters are optimized as described in Appendix B, but the final scores are computed on data held back from training (10% test set for the GMM; held-back test set for Aib9), so no fitted parameter is renamed as a prediction. The Aib9 trajectory is generated using the authors' prior simulation package, but that package is code-reproducible and publicly available, and the trajectory is used as empirical input data, not as a load-bearing theorem or uniqueness argument. No ansatz, uniqueness theorem, or definitional equivalence is imported from prior work by the same authors; the theoretical background (Boltzmann distribution, KL divergence, flow and diffusion equations) is standard and does not by itself force any ranking. Concerns about the unrepeated KL metric or the arbitrary 0.0374 kJ/mol cutoff are statistical robustness issues, not circularity.

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

The benchmarks are self-contained empirical measurements rather than derivations, so the ledger mainly captures evaluation-metric choices: an arbitrary free-energy cutoff, a two-PC binning metric, tuned model hyperparameters, and an equal-training-time rule. There are no invented physical entities.

free parameters (5)
  • Free-energy histogram cutoff = 0.0374 kJ/mol
    Arbitrary threshold below which histogram bins are discarded when computing partition-function ratios in the free-energy estimation experiment; directly affects the rankings in Figure 2.
  • PCA projection dimension = 2
    All DKL evaluations project both training and generated samples onto the first two principal components of the training data; this choice defines the accuracy metric and may favor models that match dominant variance modes.
  • NS network depth ('layers') = tuned, exact value not stated
    Hyperparameter optimized for each dataset according to Appendix B, Figure 5a; final values are not reported, so replication requires re-tuning.
  • CFM/DDPM model dimension = tuned, exact value not stated
    Width and implicit depth hyperparameter optimized per dataset (Appendix B, Figures 5b and 5c); final values are not reported.
  • Training time budget = not stated
    Models are trained for an equal amount of wall time rather than to convergence; relative accuracy rankings depend on this budget and may change with longer or shorter training.
assumptions (5)
  • standard math Boltzmann distribution p(x)=e^{-βU}/Z and change-of-measure identity p(x)=q(x')/|J_M(x')|
    These identities underlie the theoretical framing in Sections II and III and are used to justify training generative models as free-energy-minimizing transformations.
  • domain assumption The 200 ns OpenMM simulation at 450 K with TIP3 water provides a trustworthy equilibrium sample of Aib9 dihedral angles
    Appendix Table I lists simulation parameters, but the paper provides no convergence checks or independent validation that the trajectory is equilibrated; if the MD sample is biased, the Aib9 benchmark target is wrong.
  • ad hoc to paper DKL between binned 2D PCA projections of training and generated samples is a valid accuracy metric
    Section IV specifies PCA projection and binning but does not give bin count or width, nor does it validate that two principal components preserve the model differences in the 18-dimensional dihedral space.
  • ad hoc to paper Equal training time is a fair comparison across models
    Section IV A states models are allowed to train for an equal amount of time; if convergence rates differ, this choice biases the relative accuracy rankings.
  • ad hoc to paper The free-energy cutoff of 0.0374 kJ/mol excludes only negligible high-energy bins
    Section IV A imposes this cutoff when computing Z1/Z2; no sensitivity analysis is shown, so the reported free-energy rankings could depend on the cutoff value.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A survey of probabilistic generative frameworks for molecular simulations." pith.science (2026). https://pith.science/paper/Y44DSP6V

@misc{pith2026241109388,
  author       = {Pith},
  title        = {Pith review of: A survey of probabilistic generative frameworks for molecular simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y44DSP6V}},
  note         = {Machine review of arXiv:2411.09388}
}
read the original abstract

Generative artificial intelligence is now a widely used tool in molecular science. Despite the popularity of probabilistic generative models, numerical experiments benchmarking their performance on molecular data are lacking. In this work, we introduce and explain several classes of generative models, broadly sorted into two categories: flow-based models and diffusion models. We select three representative models: Neural Spline Flows, Conditional Flow Matching, and Denoising Diffusion Probabilistic Models, and examine their accuracy, computational cost, and generation speed across datasets with tunable dimensionality, complexity, and modal asymmetry. Our findings are varied, with no one framework being the best for all purposes. In a nutshell, (i) Neural Spline Flows do best at capturing mode asymmetry present in low-dimensional data, (ii) Conditional Flow Matching outperforms other models for high-dimensional data with low complexity, and (iii) Denoising Diffusion Probabilistic Models appears the best for low-dimensional data with high complexity. Our datasets include a Gaussian mixture model and the dihedral torsion angle distribution of the Aib\textsubscript{9} peptide, generated via a molecular dynamics simulation. We hope our taxonomy of probabilistic generative frameworks and numerical results may guide model selection for a wide range of molecular tasks.

Figures

Figures reproduced from arXiv: 2411.09388 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 35 canonical work pages

  1. [1]

    Rombach , author A

    author author R. Rombach , author A. Blattmann , author D. Lorenz , author P. Esser ,\ and\ author B. Ommer ,\ in\ @noop booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition \ ( year 2022 )\ pp.\ pages 10684--10695 NoStop

  2. [2]

    Achiam , author S

    author author J. Achiam , author S. Adler , author S. Agarwal , author L. Ahmad , author I. Akkaya , author F. L. \ Aleman , author D. Almeida , author J. Altenschmidt , author S. Altman , author S. Anadkat , et al. ,\ @noop journal journal arXiv preprint arXiv:2303.08774 \ ( year 2023 ) NoStop

  3. [3]

    Ho , author T

    author author J. Ho , author T. Salimans , author A. Gritsenko , author W. Chan , author M. Norouzi ,\ and\ author D. J. \ Fleet ,\ @noop journal journal Advances in Neural Information Processing Systems \ volume 35 ,\ pages 8633 ( year 2022 ) NoStop

  4. [4]

    Generative artificial intelligence for computational chemistry: a roadmap to predicting emergent phenomena

    author author P. Tiwary , author L. Herron , author R. John , author S. Lee , author D. Sanwal ,\ and\ author R. Wang ,\ @noop journal journal arXiv preprint arXiv:2409.03118 \ ( year 2024 ) NoStop

  5. [5]

    Zheng , author J

    author author S. Zheng , author J. He , author C. Liu , author Y. Shi , author Z. Lu , author W. Feng , author F. Ju , author J. Wang , author J. Zhu , author Y. Min , et al. ,\ @noop journal journal Nature Machine Intelligence \ ,\ pages 1 ( year 2024 ) NoStop

  6. [6]

    Jumper , author R

    author author J. Jumper , author R. Evans , author A. Pritzel , author T. Green , author M. Figurnov , author O. Ronneberger , author K. Tunyasuvunakool , author R. Bates , author A. Zidek , author A. Potapenko , et al. ,\ @noop journal journal nature \ volume 596 ,\ pages 583 ( year 2021 ) NoStop

  7. [7]

    Abramson , author J

    author author J. Abramson , author J. Adler , author J. Dunger , author R. Evans , author T. Green , author A. Pritzel , author O. Ronneberger , author L. Willmore , author A. J. \ Ballard , author J. Bambrick , et al. ,\ @noop journal journal Nature \ ,\ pages 1 ( year 2024 ) NoStop

  8. [8]

    Krishna , author J

    author author R. Krishna , author J. Wang , author W. Ahern , author P. Sturmfels , author P. Venkatesh , author I. Kalvet , author G. R. \ Lee , author F. S. \ Morey-Burrows , author I. Anishchenko , author I. R. \ Humphreys , et al. ,\ @noop journal journal Science \ volume 384 ,\ pages eadl2528 ( year 2024 ) NoStop

Show all 52 references
  1. [9]

    Qiao , author W

    author author Z. Qiao , author W. Nie , author A. Vahdat , author T. F. \ Miller ,\ and\ author A. Anandkumar ,\ https://doi.org/10.1038/s42256-024-00792-z journal journal Nature Machine Intelligence \ volume 6 ,\ pages 195?208 ( year 2024 ) NoStop

  2. [10]

    Corso , author H

    author author G. Corso , author H. St\" a rk , author B. Jing , author R. Barzilay ,\ and\ author T. Jaakkola ,\ https://doi.org/10.48550/ARXIV.2210.01776 title Diffdock: Diffusion steps, twists, and turns for molecular docking ( year 2022 ) NoStop

  3. [11]

    No\' e , author S

    author author F. No\' e , author S. Olsson , author J. K\" o hler ,\ and\ author H. Wu ,\ https://doi.org/10.48550/ARXIV.1812.01729 title Boltzmann generators -- sampling equilibrium states of many-body systems with deep learning ( year 2018 ) NoStop

  4. [12]

    Wang , author L

    author author Y. Wang , author L. Herron ,\ and\ author P. Tiwary ,\ @noop journal journal Proceedings of the National Academy of Sciences \ volume 119 ,\ pages e2203656119 ( year 2022 ) NoStop

  5. [13]

    Herron , author K

    author author L. Herron , author K. Mondal , author J. S. \ Schneekloth ,\ and\ author P. Tiwary ,\ @noop journal journal arXiv preprint arXiv:2308.14885 \ ( year 2023 ) NoStop

  6. [14]

    Dinh , author J

    author author L. Dinh , author J. Sohl-Dickstein ,\ and\ author S. Bengio ,\ https://doi.org/10.48550/ARXIV.1605.08803 title Density estimation using real nvp ( year 2016 ) NoStop

  7. [15]

    Tong , author N

    author author A. Tong , author N. Malkin , author G. Huguet , author Y. Zhang , author J. Rector-Brooks , author K. Fatras , author G. Wolf ,\ and\ author Y. Bengio ,\ @noop journal journal arXiv preprint arXiv:2302.00482 \ volume 2 ( year 2023 a ) NoStop

  8. [16]

    Song , author J

    author author Y. Song , author J. Sohl-Dickstein , author D. P. \ Kingma , author A. Kumar , author S. Ermon ,\ and\ author B. Poole ,\ in\ https://openreview.net/forum?id=PxTIG12RRHS booktitle International Conference on Learning Representations \ ( year 2021 ) NoStop

  9. [17]

    Durkan , author A

    author author C. Durkan , author A. Bekasov , author I. Murray ,\ and\ author G. Papamakarios ,\ @noop journal journal Advances in neural information processing systems \ volume 32 ( year 2019 ) NoStop

  10. [18]

    Lipman , author R

    author author Y. Lipman , author R. T. Q. \ Chen , author H. Ben-Hamu , author M. Nickel ,\ and\ author M. Le ,\ https://doi.org/10.48550/ARXIV.2210.02747 title Flow matching for generative modeling ( year 2022 a ) NoStop

  11. [19]

    Ho , author A

    author author J. Ho , author A. Jain ,\ and\ author P. Abbeel ,\ @noop journal journal Advances in neural information processing systems \ volume 33 ,\ pages 6840 ( year 2020 ) NoStop

  12. [20]

    Karakonstantis , author E

    author author X. Karakonstantis , author E. Fernandez-Grande ,\ and\ author P. Gerstoft ,\ https://arxiv.org/abs/2404.06928 title Efficient sound field reconstruction with conditional invertible neural networks ( year 2024 ),\ https://arxiv.org/abs/2404.06928 arXiv:2404.06928 ...

  13. [21]

    Kim , author K

    author author S. Kim , author K. J. \ Shih , author R. Badlani , author J. F. \ Santos , author E. Bakhturina , author M. T. \ Desta , author R. Valle , author S. Yoon ,\ and\ author B. Catanzaro ,\ in\ https://openreview.net/forum?id=zNA7u7wtIN booktitle Thirty-seventh Confer...

  14. [22]

    Guo , author Y

    author author X. Guo , author Y. Yang , author C. Ye , author S. Lu , author B. Peng , author H. Huang , author Y. Xiang ,\ and\ author T. Ma ,\ in\ https://doi.org/10.1109/isbi53787.2023.10230524 booktitle 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) \ ...

  15. [23]

    Mehdi , author D

    author author S. Mehdi , author D. Wang , author S. Pant ,\ and\ author P. Tiwary ,\ @noop journal journal Journal of chemical theory and computation \ volume 18 ,\ pages 3231 ( year 2022 ) NoStop

  16. [24]

    Botan , author E

    author author V. Botan , author E. H. \ Backus , author R. Pfister , author A. Moretto , author M. Crisma , author C. Toniolo , author P. H. \ Nguyen , author G. Stock ,\ and\ author P. Hamm ,\ @noop journal journal Proceedings of the National Academy of Sciences \ volume 104 ...

  17. [25]

    Jarzynski ,\ @noop journal journal Physical Review E \ volume 65 ,\ pages 046122 ( year 2002 ) NoStop

    author author C. Jarzynski ,\ @noop journal journal Physical Review E \ volume 65 ,\ pages 046122 ( year 2002 ) NoStop

  18. [26]

    author author A. M. \ Hahn \ and\ author H. Then ,\ @noop journal journal Physical Review E?Statistical, Nonlinear, and Soft Matter Physics \ volume 79 ,\ pages 011113 ( year 2009 ) NoStop

  19. [27]

    No\' e , author S

    author author F. No\' e , author S. Olsson , author J. Kohler ,\ and\ author H. Wu ,\ @noop journal journal Science \ volume 365 ,\ pages eaaw1147 ( year 2019 ) NoStop

  20. [28]

    Wirnsberger , author A

    author author P. Wirnsberger , author A. J. \ Ballard , author G. Papamakarios , author S. Abercrombie , author S. Racaniere , author A. Pritzel , author D. Jimenez Rezende ,\ and\ author C. Blundell ,\ @noop journal journal The Journal of Chemical Physics \ volume 153 ( year ...

  21. [29]

    Rizzi , author P

    author author A. Rizzi , author P. Carloni ,\ and\ author M. Parrinello ,\ @noop journal journal The journal of physical chemistry letters \ volume 12 ,\ pages 9449 ( year 2021 ) NoStop

  22. [30]

    Invernizzi , author A

    author author M. Invernizzi , author A. Kramer , author C. Clementi ,\ and\ author F. Noe ,\ @noop journal journal The Journal of Physical Chemistry Letters \ volume 13 ,\ pages 11643 ( year 2022 ) NoStop

  23. [31]

    author author D. P. \ Kingma \ and\ author P. Dhariwal ,\ @noop journal journal Advances in neural information processing systems \ volume 31 ( year 2018 ) NoStop

  24. [32]

    Papamakarios , author T

    author author G. Papamakarios , author T. Pavlakou ,\ and\ author I. Murray ,\ @noop journal journal Advances in neural information processing systems \ volume 30 ( year 2017 ) NoStop

  25. [33]

    Wu , author J

    author author H. Wu , author J. Kohler ,\ and\ author F. Noe ,\ @noop journal journal Advances in Neural Information Processing Systems \ volume 33 ,\ pages 5933 ( year 2020 ) NoStop

  26. [34]

    author author R. T. Q. \ Chen , author Y. Rubanova , author J. Bettencourt ,\ and\ author D. Duvenaud ,\ https://doi.org/10.48550/ARXIV.1806.07366 title Neural ordinary differential equations ( year 2018 ) NoStop

  27. [35]

    author author M. F. \ Hutchinson ,\ @noop journal journal Communications in Statistics-Simulation and Computation \ volume 18 ,\ pages 1059 ( year 1989 ) NoStop

  28. [36]

    Sohl-Dickstein , author E

    author author J. Sohl-Dickstein , author E. Weiss , author N. Maheswaranathan ,\ and\ author S. Ganguli ,\ in\ @noop booktitle International conference on machine learning \ ( organization PMLR ,\ year 2015 )\ pp.\ pages 2256--2265 NoStop

  29. [37]

    Song , author J

    author author Y. Song , author J. Sohl-Dickstein , author D. P. \ Kingma , author A. Kumar , author S. Ermon ,\ and\ author B. Poole ,\ @noop journal journal arXiv preprint arXiv:2011.13456 \ ( year 2020 ) NoStop

  30. [38]

    author author N. M. \ Boffi \ and\ author E. Vanden-Eijnden ,\ @noop journal journal Machine Learning: Science and Technology \ volume 4 ,\ pages 035012 ( year 2023 ) NoStop

  31. [39]

    Hyvarinen \ and\ author P

    author author A. Hyvarinen \ and\ author P. Dayan ,\ @noop journal journal Journal of Machine Learning Research \ volume 6 ( year 2005 ) NoStop

  32. [40]

    Schrodinger ,\ in\ @noop booktitle Annales de linstitut Henri Poincare ,\ Vol

    author author E. Schrodinger ,\ in\ @noop booktitle Annales de linstitut Henri Poincare ,\ Vol. volume 2 \ ( year 1932 )\ pp.\ pages 269--310 NoStop

  33. [42]

    De Bortoli , author J

    author author V. De Bortoli , author J. Thornton , author J. Heng ,\ and\ author A. Doucet ,\ @noop journal journal Advances in Neural Information Processing Systems \ volume 34 ,\ pages 17695 ( year 2021 ) NoStop

  34. [43]

    Chen , author T

    author author Y. Chen , author T. T. \ Georgiou ,\ and\ author M. Pavon ,\ @noop journal journal Annual Review of Control, Robotics, and Autonomous Systems \ volume 4 ,\ pages 89 ( year 2021 ) NoStop

  35. [44]

    Follmer ,\ @noop journal journal Lect

    author author H. Follmer ,\ @noop journal journal Lect. Notes Math \ volume 1362 ,\ pages 101 ( year 1988 ) NoStop

  36. [45]

    Leonard ,\ @noop journal journal arXiv preprint arXiv:1308.0215 \ ( year 2013 b ) NoStop

    author author C. Leonard ,\ @noop journal journal arXiv preprint arXiv:1308.0215 \ ( year 2013 b ) NoStop

  37. [46]

    author author M. S. \ Albergo , author N. M. \ Boffi ,\ and\ author E. Vanden-Eijnden ,\ @noop journal journal arXiv preprint arXiv:2303.08797 \ ( year 2023 a ) NoStop

  38. [47]

    author author M. S. \ Albergo \ and\ author E. Vanden-Eijnden ,\ @noop journal journal arXiv preprint arXiv:2209.15571 \ ( year 2022 ) NoStop

  39. [48]

    author author M. S. \ Albergo , author M. Goldstein , author N. M. \ Boffi , author R. Ranganath ,\ and\ author E. Vanden-Eijnden ,\ @noop journal journal arXiv preprint arXiv:2310.03725 \ ( year 2023 b ) NoStop

  40. [49]

    Lipman , author R

    author author Y. Lipman , author R. T. \ Chen , author H. Ben-Hamu , author M. Nickel ,\ and\ author M. Le ,\ @noop journal journal arXiv preprint arXiv:2210.02747 \ ( year 2022 b ) NoStop

  41. [50]

    Tong , author N

    author author A. Tong , author N. Malkin , author K. Fatras , author L. Atanackovic , author Y. Zhang , author G. Huguet , author G. Wolf ,\ and\ author Y. Bengio ,\ @noop journal journal arXiv preprint arXiv:2307.03672 \ ( year 2023 b ) NoStop

  42. [51]

    Buchenberg , author N

    author author S. Buchenberg , author N. Schaudinnus ,\ and\ author G. Stock ,\ https://doi.org/10.1021/ct501156t journal journal Journal of Chemical Theory and Computation \ volume 11 ,\ pages 1330?1336 ( year 2015 ) NoStop

  43. [52]

    Liu , author C

    author author X. Liu , author C. Gong ,\ and\ author Q. Liu ,\ @noop journal journal arXiv preprint arXiv:2209.03003 \ ( year 2022 ) NoStop

  44. [53]

    Karras , author M

    author author T. Karras , author M. Aittala , author T. Aila ,\ and\ author S. Laine ,\ @noop journal journal Advances in neural information processing systems \ volume 35 ,\ pages 26565 ( year 2022 ) NoStop

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.