Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Adaptive Destruction Processes for Diffusion Samplers

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that treating diffusion samplers as discrete-time policies and training both the generation and destruction processes, with the generation variance decoupled and learned, improves sampling quality and…

desk verdict Promising few-step diffusion sampler recipe undercut by an invalid destruction-kernel definition and an overclaimed headline result. read the letter →

arxiv 2506.01541 v1 pith:R3JXSZWK submitted 2025-06-02 cs.LG stat.ML

Timofei Gritsaev , Nikita Morozov , Kirill Tamogashev , Daniil Tiapkin , Sergey Samsonov , Alexey Naumov , Dmitry Vetrov , Nikolay Malkin This is my paper
classification cs.LGstat.ML
keywords diffusionsamplerslearnabledestructionprocessgenerationvariancediscrete-timepoliciestrajectorybalancenormalisingconstantestimationfew-stepsamplingoff-policytraining
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

Diffusion samplers are usually viewed as discretisations of continuous-time diffusions, which forces the generation and destruction processes to share the same diffusion coefficient. The paper argues that a discrete-time diffusion sampler is better treated as a pair of trainable Markovian policies, and it proposes learning both the generation and destruction transition kernels as unconstrained Gaussian densities with decoupled variances. Its central claim is that when the number of steps is small, training both processes improves sampling quality and normalising-constant estimation, on top of the already substantial gain from making the generation variance learnable. A sympathetic reader would care because few-step samplers are the practical regime where the continuous-time theory is least reliable, and the paper identifies stabilisation techniques that make joint training work.

What carries the argument

The load-bearing object is the pair of decoupled Gaussian transition kernels in Eqs. (11)--(12). The generation step is $\vec p_\theta(X_{t+\Delta t}\mid X_t)=\mathcal N(X_t+f_\theta(X_t,t)\Delta t,\ \operatorname{diag}(\gamma_\theta)\sigma^2\Delta t I_d)$ with $\gamma_\theta=\exp\{C_1\tanh(\mathrm{NN}_\theta^{(1)})\}$, and the destruction step is $\overleftarrow p_\varphi(X_t\mid X_{t+\Delta t})=\mathcal N(\operatorname{diag}(\alpha_\varphi)\tfrac{t}{t+\Delta t}X_{t+\Delta t},\ \operatorname{diag}(\beta_\varphi)\tfrac{t}{t+\Delta t}\sigma^2\Delta t I_d)$ with $\alpha_\varphi,\beta_\varphi=1\pm C_2\tanh(\mathrm{NN}_\varphi^{(\cdot)})$. These kernels stay Gaussian with diagonal covariance but no longer correspond to an SDE with matching diffusion coefficients, which is what makes joint training of both processes possible. The training signal is the time-reversal equality enforced through a second-moment divergence called trajectory balance, whose losses give the ELBO and EUBO bounds on the log-normalising constant.

What would settle it

Train the joint sampler on a target whose exact few-step reverse transition is strongly non-Gaussian or has correlated coordinates, for example a narrow curved ridge or a mixture of elongated Gaussians rotated off the coordinate axes, and compare ELBO and the learned destruction kernel against the exact reverse kernel; if a mixture or low-rank Gaussian kernel clearly outperforms the diagonal Gaussian parametrisation, the paper's central claim would need qualification.

Watch

Extended reading notes

Core claim

The paper establishes that, in discrete time, the time-reversal constraint $p_0\vec p_\theta = p_{\mathrm{target}}\overleftarrow p_\varphi$ can be optimised over both $\theta$ and $\varphi$ when the transitions are Gaussians with arbitrary means and learned variances, as long as the variances are bounded multiplicative corrections of a reference discretisation. Because continuous-time path measures with different diffusion coefficients are mutually singular, the decoupling of generation and destruction variances is only possible away from the continuous-time limit, and the paper exploits exactly that freedom. Its experiments show three things: learnable generation variance is the largest single improvement, adding a learned destruction process yields a further gain on all the tasks it tests, and the gain is most pronounced at small numbers of steps and on targets with narrow modes, such as the Hard Funnel. The paper also finds that training the destruction process by the trajectory-balance loss is more stable than training it by maximum-likelihood-style trajectory likelihood maximisation, which frequently diverges at larger step counts.

Load-bearing premise

The central claim rests on the assumption that the true few-step reverse transitions of the target are well approximated by Gaussian kernels with diagonal covariance and bounded multiplicative corrections; if the transitions are strongly non-Gaussian or correlated across coordinates, the learned destruction process cannot represent them and joint training may converge to a biased solution.

Editorial extensions

If this is right

  • On Gaussian-mixture targets, jointly learned samplers with as few as 5 steps can match or beat fixed-variance samplers with 20 steps.
  • Learning the generation variance alone gives most of the improvement; the learned destruction process adds a consistent further gain, largest at small $T$ or when modes are narrow.
  • The trajectory-balance objective is preferable to trajectory likelihood maximisation for the destruction process at larger step counts because its gradient vanishes at the global optimum and does not diverge.
  • Off-policy training with an exploratory behaviour policy is on par with or better than backpropagating through the sampling trajectory, and it avoids storing the full computation graph.
  • In the high-dimensional text-conditional GAN-latent experiment with $T=5$, joint training improves the reported bound on the normalising constant and the reward score on five of seven prompts while keeping diversity unchanged.

Reading between the lines

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

  • The paper does not test non-Gaussian or correlated transition kernels, so an open extension is whether replacing the diagonal Gaussian family with low-rank or mixture kernels widens the benefit to targets whose true reverse transitions are strongly correlated; the paper itself notes the transitions 'could be more general than Gaussian.'
  • Because the gains shrink as the number of steps grows, the method's practical value is concentrated in low-latency regimes; one could combine it with step distillation or schedule the number of steps during training to amplify the effect.
  • The sensitivity to the destruction-process learning rate suggests that an automated schedule, such as warm-up or adaptive tuning of $\mathrm{lr}_\varphi/\mathrm{lr}_\theta$, could make joint training less hand-tuned; the paper reports optimal ratios ranging from equal to $10^{-5}$.
  • Viewed through the paper's soft-RL reformulation, the learned destruction process is a learned exploration schedule, so the same stability techniques should transfer to other amortised samplers that currently fix their forward process.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes training the destruction (reverse) process jointly with the generation process in discrete-time diffusion samplers, using Gaussian transition kernels with learnable mean and variance corrections. The authors argue that when the number of diffusion steps is small, decoupling generation and destruction variances and learning both processes improves sampling quality and normalizing-constant estimation. They study several training objectives (trajectory balance, reverse KL/PIS, TLM, VarGrad), propose stability techniques (shared backbone, separate optimizers, target networks, prioritized replay), and report experiments on synthetic benchmarks and on text-conditional sampling in a StyleGAN3 latent space.

Significance. If the central claim is correct, the paper gives a practically useful recipe: for low-step discrete-time diffusion samplers, users should train both the generation and destruction policies rather than fixing the destruction process. The paper has clear strengths: a broad benchmark suite, multiple objectives and metrics (ELBO, EUBO, 2-Wasserstein), three seeds, full supplementary tables, a dedicated ablation study, and a nontrivial high-dimensional GAN latent-space experiment. The paper does not ship code or machine-checked proofs, but it builds on the published implementation of [70]. However, the validity of the destruction-kernel parameterization and the force of the empirical claims need to be settled before the contribution can be accepted.

major comments (3)
  1. [§4.1, Eq. (12)] The destruction kernel is written as ←p_φ(X_t | X_{t+Δt}) = N(X_t; diag(α_φ(X_t,t)) · t/(t+Δt) X_{t+Δt}, diag(β_φ(X_t,t)) · t/(t+Δt) σ²Δt I), with α_φ and β_φ functions of X_t, the variable being conditioned on. This is not a normalized Gaussian density in X_t: the normalizing constant of a Gaussian depends on the covariance, and here the covariance (and mean) vary with X_t. Consequently ∫ ←p_φ(X_t | X_{t+Δt}) dX_t is not 1 in general, and the log-density used in Eqs. (13), (16), (18), and (19) is not the log-density of any Markov kernel. As written, the training objective does not minimize a valid divergence between two Markov chains, and the reported gains of TB_{θ,φ} over fixed-destruction baselines could be artifacts of an improper density. If this is a notation slip and the networks actually take X_{t+Δt}, the text must state this and Eq. (12) must be corrected; otherwise a valid parameterization, for example a normalizing flow with a Jacobian correction, is required. This issue is load-bearing for the theory and for the interpretation of every destruction-learning experiment.
  2. [Figure 1 caption; §5.2] The caption of Figure 1 states 'Mean and std over 3 seeds, collapsed runs excluded.' The paper does not report how many runs collapsed, on which configurations, or what the excluded values were. With only three seeds, excluding failures can materially change the averaged metrics and can reverse a comparison. This is especially consequential because Table 7 reports divergent runs for TBθ+TLMφ and because Table 2 shows large seed-to-seed variation (for example, EUBO values of 119.58±26.11 and 6.55±3.48 across settings). The main figures and the §5.2 claim that learning the destruction process improves 'on all tasks' should be accompanied by a full run-level report, or at least by explicit counts of collapsed runs and a sensitivity analysis.
  3. [§5.2, Tables 4–8] The sentence 'Learning the destruction process yields an improvement over models with fixed destruction process and learned generation variance on all tasks' is not uniformly supported by the tables. For example, Table 4 (125GMM, T=2) shows TB_{θ,φ} ELBO −5.34±1.09 versus −4.92±0.16 for TBθ(learned var.); Table 8 (Manywell, T=10) shows ELBO gap −0.83±0.02 versus −0.78±0.02; and Table 7 (Hard Funnel, T=5) shows EUBO 2800.63±1351.9 for TB_{θ,φ} versus 155.51±19.6 for TBθ(learned var.), despite the ELBO being better for TB_{θ,φ}. The qualitative claim should be restated with a precise scope, and the paper should discuss these exceptions rather than presenting the improvement as universal.
minor comments (5)
  1. [Appendix B, Eq. (20)] The unsuccessful parametrization in Eq. (20) has the same conditioning problem as Eq. (12): the destruction variance is written as a function of X_t. If Eq. (12) is corrected, Eq. (20) must be corrected or annotated consistently.
  2. [§C.1, Funnel definition] The text says the first variable is drawn from N(0,1) for Easy Funnel and N(0,9) for Hard Funnel, but the displayed formula states x0∼N(0,1) in both cases; the formula for Hard Funnel should be N(0,9).
  3. [Eq. (17)] The target-network update is written as θ=(1−τ)θ+τθ, which is tautological; the target weights need a distinct symbol, for example θ_target ← (1−τ)θ_target + τθ.
  4. [Table 2] The 'Separate backbones' row reports a 2-Wasserstein value of 28.29±0.49, identical to the 'TBθ, learned gen var.' row above it; this looks like a copying error and should be verified.
  5. [§5.3] The phrase 'a 10^3 times smaller learning rate' is ambiguous; it should read 'a factor of 10^−3 smaller' or '10^3 times smaller' as appropriate, and the same applies to the 10^−4/10^−5 ratios in §C.3.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation of the main claim; only minor non-load-bearing self-citations.

full rationale

The central derivation is self-contained: the training objective minimizes a time-reversal divergence between the discrete-time forward chain p0 * p_theta and the backward chain p_target * p_phi, and the reported metrics (ELBO, EUBO, 2-Wasserstein) are evaluated against known targets or ground-truth samples, not used as the training loss. The statement that a trainable destruction process is at least as good as a fixed one is explicitly a feasible-set argument ('the latter is a special case of the former'), not an empirical prediction, and the claimed improvements are supported by the paper's own ablations. Self-citations to [22] (TLM) and [70] (architecture, exploration) are used for naming and stability heuristics that the paper also tests in its own experiments, so they are not load-bearing. No uniqueness theorem or prior-work ansatz is imported to force the choice of the destruction kernel. Two caveats are correctness risks rather than circularity: Eq. (12) literally writes the destruction mean and variance as functions of X_t, the variable being integrated, so the displayed density is not a valid normalized Gaussian kernel in X_t; and the conclusion acknowledges that transitions 'could be more general than Gaussian' without testing this case. Neither makes the reported improvement equivalent to the training objective by construction. Overall, the derivation chain is not circular; the minor self-citations do not carry the central claim.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The method relies on modeling choices (Gaussian kernels, bounded variance corrections) and a set of tuned hyperparameters; the central claim is empirical and depends on these choices being adequate.

free parameters (6)
  • Generation variance bound C1 = 4.0 (synthetic), 1.0 (GAN latent)
    C1 in Eq. (11) bounds the multiplicative deviation of the generation variance from the baseline sigma^2*Delta*t. Chosen by hand; affects the flexibility of the generation policy.
  • Destruction correction bound C2 = 0.9 (synthetic), 0.1 (GAN latent)
    C2 in Eq. (12) bounds the multiplicative corrections to the destruction mean and variance. Chosen by hand; a smaller value was needed for stability in the GAN experiment.
  • Destruction learning rate ratio lr_phi/lr_theta = 1.0, 0.1, 0.001, 0.00001 depending on energy
    The relative learning rate of the destruction policy is tuned per environment (Sec. C.3). The paper finds optimal lr_phi is always <= lr_theta; this tuning affects the reported improvements.
  • Exploration factor = 0.3 (GMM), 0.2 (Funnel), 0.1 (Manywell, GAN)
    Variance-increase factor for off-policy trajectory sampling, taken from prior work; a hyperparameter.
  • Replay ratio = 2 (synthetic), 5 (GAN latent)
    Number of replay trajectories per new trajectory in prioritized experience replay; tuned per task.
  • Diffusion rate sigma^2 = 5 (GMM), 1 (other)
    Baseline diffusion coefficient of the destruction process; chosen per energy.
assumptions (4)
  • domain assumption The target distribution is an absolutely continuous density exp(-E)/Z on R^d with finite normalizing constant
    Used throughout Sec. 2 and in the definition of the problem (Eq. 1).
  • domain assumption Transition kernels are restricted to Gaussian densities with diagonal covariance
    The method parametrizes both generation and destruction kernels as Gaussians (Eqs. 11, 12); the paper states transitions could be more general but does not analyze non-Gaussian cases.
  • domain assumption The discrete-time Markov chain with T steps is the object of interest and a sufficient approximation of the continuous process
    The paper deliberately abandons continuous-time equivalence and treats the discrete-time chain as the object of interest; the improvement claims are for a fixed small T.
  • domain assumption The global optimum of the joint objective is attainable by gradient-based optimization
    The paper argues the solution with a trainable destruction process is at least as good as fixed (Sec. 4.1) based on feasibility, but this presumes the optimization finds the optimum.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Destruction Processes for Diffusion Samplers." pith.science (2026). https://pith.science/paper/R3JXSZWK

@misc{pith2026250601541,
  author       = {Pith},
  title        = {Pith review of: Adaptive Destruction Processes for Diffusion Samplers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3JXSZWK}},
  note         = {Machine review of arXiv:2506.01541}
}
read the original abstract

This paper explores the challenges and benefits of a trainable destruction process in diffusion samplers -- diffusion-based generative models trained to sample an unnormalised density without access to data samples. Contrary to the majority of work that views diffusion samplers as approximations to an underlying continuous-time model, we view diffusion models as discrete-time policies trained to produce samples in very few generation steps. We propose to trade some of the elegance of the underlying theory for flexibility in the definition of the generative and destruction policies. In particular, we decouple the generation and destruction variances, enabling both transition kernels to be learned as unconstrained Gaussian densities. We show that, when the number of steps is limited, training both generation and destruction processes results in faster convergence and improved sampling quality on various benchmarks. Through a robust ablation study, we investigate the design choices necessary to facilitate stable training. Finally, we show the scalability of our approach through experiments on GAN latent space sampling for conditional image generation.

Figures

Figures reproduced from arXiv: 2506.01541 by the authors.

Figure 1
Figure 1. We compare performance of TB𝜃 (fixed var.), TB𝜃 (learned var.), TB𝜃 + TLM𝜑, TB𝜃,𝜑 on three GMM targets. The rightmost plot shows PIS-like generation process objectives on 40GMM. Results are compared using ELBO. Mean and std over 3 seeds, collapsed runs excluded. 125GMM Easy Funnel Hard Funnel Distorted Manywell 2 3 4 5 10 20 Number of steps 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Exponent of ELBO 125 GMM TB (fixed var.) TB (lea… view at source ↗
Figure 2
Figure 2. The same plots as Fig [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. First two dimensions of target energy and samples from diffusion samplers trained on [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Decoded latents sampled with the same random seeds from outsourced diffusion samplers trained with a StyleGAN3 prior and ImageReward with prompt ‘A person with medium length hair’. In our experiment we use StyleGAN3 [32] trained on the 256×256 FFHQ dataset [33] with 𝑑l…
Figure 5
Figure 5. Figure 5: Contour levels of 2-dimensional Gaussian mixture densities used in the experiments. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: FFHQ text-conditional latent sampling results across different prompts. The figure depicts [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels

    cs.LG 2026-06 conditional novelty 6.0 of 10

    Classifier-based adaptive stopping, trained as a non-acyclic GFlowNet, shortens MCMC trajectories while preserving or improving sample quality relative to ULA and diffusion-sampler baselines.

  2. Data-to-Energy Stochastic Dynamics

    cs.LG 2025-09 conditional novelty 6.0 of 10

    A new data-to-energy iterative proportional fitting algorithm trains Schrödinger bridges when endpoint distributions are known only through unnormalised densities.

  3. Towards Adaptive External Communication in Autonomous Vehicles: A Conceptual Design Framework

    cs.HC 2025-08 unverdicted novelty 5.0 of 10

    A three-layer framework (input, processing, output) for adaptive external human-machine interfaces in autonomous vehicles is introduced to systematize design and analysis.

Reference graph

Works this paper leans on

99 extracted references · 70 canonical work pages · cited by 3 Pith papers

  1. [70]

    Sendera, M., Kim, M., Mittal, S., Lemos, P., Scimeca, L., Rector-Brooks, J., Adam, A., Bengio, Y ., and Malkin, N. (2024). Improved off-policy training of diffusion samplers.Neural Information Processing Systems (NeurIPS)

    work page 2024

  2. [1]

    J., Mittal, S., Lemos, P., Liu, C.-H., Sendera, M., Ravanbakhsh, S., Gidel, G., Bengio, Y ., Malkin, N., and Tong, A

    Akhound-Sadegh, T., Rector-Brooks, J., Bose, A. J., Mittal, S., Lemos, P., Liu, C.-H., Sendera, M., Ravanbakhsh, S., Gidel, G., Bengio, Y ., Malkin, N., and Tong, A. (2024). Iterated denoising energy matching for sampling from Boltzmann densities.International Conference on Machine Learning (ICML)

    2024

  3. [2]

    and Bengio, E

    Atanackovic, L. and Bengio, E. (2025). Investigating generalization behaviours of generative flow networks.Transactions on Machine Learning Research

    2025

  4. [3]

    Bartosh, G., Vetrov, D., and Naesseth, C. A. (2024a). Neural diffusion models.International Conference on Machine Learning (ICML)

    2024

  5. [4]

    P., and Andersson Naesseth, C

    Bartosh, G., Vetrov, D. P., and Andersson Naesseth, C. (2024b). Neural flow diffusion models: Learnable forward process for improved diffusion modelling.Neural Information Processing Systems (NeurIPS)

    2024

  6. [5]

    Bengio, E., Jain, M., Korablyov, M., Precup, D., and Bengio, Y . (2021). Flow network based generative models for non-iterative diverse candidate generation.Neural Information Processing Systems (NeurIPS)

    2021

  7. [6]

    J., Tiwari, M., and Bengio, E

    Bengio, Y ., Lahlou, S., Deleu, T., Hu, E. J., Tiwari, M., and Bengio, E. (2023). GFlowNet foundations.Journal of Machine Learning Research, 24(210):1–55

    2023

  8. [7]

    Berner, J., Richter, L., Sendera, M., Rector-Brooks, J., and Malkin, N. (2025). From discrete-time policies to continuous-time diffusion samplers: Asymptotic equivalences and faster training.arXiv preprint arXiv:2501.06148

    arXiv 2025

Show all 99 references
  1. [8]

    Berner, J., Richter, L., and Ullrich, K. (2024). An optimal control perspective on diffusion-based generative modeling.Transactions on Machine Learning Research

    2024

  2. [9]

    Blessing, D., Berner, J., Richter, L., and Neumann, G. (2025). Underdamped diffusion bridges with applications to sampling.arXiv preprint arXiv:2503.01006

    2025 arXiv

  3. [10]

    Blessing, D., Jia, X., Esslinger, J., Vargas, F., and Neumann, G. (2024). Beyond ELBOs: A large-scale evaluation of variational methods for sampling.International Conference on Machine Learning (ICML)

    2024

  4. [11]

    Bou, A., Bettini, M., Dittert, S., Kumar, V ., Sodhani, S., Yang, X., De Fabritiis, G., and Moens, V . (2023). TorchRL: A data-driven decision-making library for PyTorch.arXiv preprint arXiv:2306.00577. 11

    2023 arXiv

  5. [12]

    Brunswic, L., Li, Y ., Xu, Y ., Feng, Y ., Jui, S., and Ma, L. (2024). A theory of non-acyclic generative flow networks.Association for the Advancement of Artificial Intelligence (AAAI), 38(10)

    2024

  6. [13]

    F., Elvira, V ., Martino, L., Luengo, D., Miguez, J., and Djuric, P

    Bugallo, M. F., Elvira, V ., Martino, L., Luengo, D., Miguez, J., and Djuric, P. M. (2017). Adaptive importance sampling: The past, the present, and the future.IEEE Signal Processing Magazine, 34(4):60–79

    2017

  7. [14]

    Chen, T., Liu, G.-H., and Theodorou, E. A. (2022). Likelihood training of Schrödinger bridge using forward-backward SDEs theory.International Conference on Learning Representations (ICLR)

    2022

  8. [15]

    Cretu, M., Harris, C., Igashov, I., Schneuing, A., Segler, M., Correia, B., Roy, J., Bengio, E., and Lio, P. (2025). SynFlowNet: Design of diverse and novel molecules with synthesis constraints. International Conference on Learning Representations (ICLR)

    2025

  9. [16]

    De Bortoli, V ., Thornton, J., Heng, J., and Doucet, A. (2021). Diffusion Schrödinger bridge with applications to score-based generative modeling.Neural Information Processing Systems (NeurIPS)

    2021

  10. [17]

    Deleu, T., Góis, A., Emezue, C., Rankawat, M., Lacoste-Julien, S., Bauer, S., and Bengio, Y . (2022). Bayesian structure learning with generative flow networks.Uncertainty in Artificial Intelligence (UAI)

    2022

  11. [18]

    Deleu, T., Nishikawa-Toomey, M., Subramanian, J., Malkin, N., Charlin, L., and Bengio, Y . (2023). Joint Bayesian inference of graphical structure and parameters with a single generative flow network.Neural Information Processing Systems (NeurIPS)

    2023

  12. [19]

    Deleu, T., Nouri, P., Malkin, N., Precup, D., and Bengio, Y . (2024). Discrete probabilistic inference as control in multi-path environments.Proceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence, pages 997–1021

    2024

  13. [20]

    Geist, M., Scherrer, B., and Pietquin, O. (2019). A theory of regularized Markov decision processes.International Conference on Machine Learning (ICML)

    2019

  14. [21]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y . (2014). Generative adversarial nets.Neural Information Processing Systems (NeurIPS)

    2014

  15. [22]

    Gritsaev, T., Morozov, N., Samsonov, S., and Tiapkin, D. (2025). Optimizing backward policies in GFlownets via trajectory likelihood maximization.International Conference on Learning Representations (ICLR)

    2025

  16. [23]

    K., Yan, B., Domingo-Enrich, C., Sriram, A., Wood, B., Levine, D., Hu, B., Amos, B., Karrer, B., Fu, X., Liu, G.-H., and Chen, R

    Havens, A., Miller, B. K., Yan, B., Domingo-Enrich, C., Sriram, A., Wood, B., Levine, D., Hu, B., Amos, B., Karrer, B., Fu, X., Liu, G.-H., and Chen, R. T. Q. (2025). Adjoint sampling: Highly scalable diffusion samplers via adjoint matching.arXiv preprint arXiv:2504.11713

    2025 arXiv

  17. [24]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. (2016). Gaussian error linear units (gelus)

    2016

  18. [25]

    Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models.Neural Information Processing Systems (NeurIPS), 33:6840–6851

    2020

  19. [26]

    D., Gelman, A., et al

    Hoffman, M. D., Gelman, A., et al. (2014). The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo.Journal of Machine Learning Research, 15(1):1593–1623

    2014

  20. [27]

    J., Jain, M., Elmoznino, E., Kaddar, Y ., Lajoie, G., Bengio, Y ., and Malkin, N

    Hu, E. J., Jain, M., Elmoznino, E., Kaddar, Y ., Lajoie, G., Bengio, Y ., and Malkin, N. (2024). Amortizing intractable inference in large language models.International Conference on Learning Representations (ICLR)

    2024

  21. [28]

    J., Malkin, N., Jain, M., Everett, K., Graikos, A., and Bengio, Y

    Hu, E. J., Malkin, N., Jain, M., Everett, K., Graikos, A., and Bengio, Y . (2023). GFlowNet-EM for learning compositional latent variable models.International Conference on Machine Learning (ICML). 12

    2023

  22. [29]

    F., Ekbote, C

    Jain, M., Bengio, E., Hernandez-Garcia, A., Rector-Brooks, J., Dossou, B. F., Ekbote, C. A., Fu, J., Zhang, T., Kilgour, M., Zhang, D., Simine, L., Das, P., and Bengio, Y . (2022). Biological sequence design with GFlowNets.International Conference on Machine Learning (ICML)

    2022

  23. [30]

    Jang, H., Jang, Y ., Kim, M., Park, J., and Ahn, S. (2024a). Pessimistic backward policy for GFlowNets.Neural Information Processing Systems (NeurIPS)

    2024

  24. [31]

    Jang, H., Kim, M., and Ahn, S. (2024b). Learning energy decompositions for partial inference in GFlownets.International Conference on Learning Representations (ICLR)

    2024

  25. [32]

    Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., and Aila, T. (2021). Alias-free generative adversarial networks.Neural Information Processing Systems (NeurIPS)

    2021

  26. [33]

    Karras, T., Laine, S., and Aila, T. (2019). A style-based generator architecture for generative adversarial networks.Computer Vision and Pattern Recognition (CVPR)

    2019

  27. [34]

    Kim, M., Choi, S., Kim, H., Son, J., Park, J., and Bengio, Y . (2025a). Ant colony sampling with GFlowNets for combinatorial optimization.Artificial Intelligence and Statistics (AISTATS)

    2025

  28. [35]

    Kim, M., Choi, S., Yun, T., Bengio, E., Feng, L., Rector-Brooks, J., Ahn, S., Park, J., Malkin, N., and Bengio, Y . (2025b). Adaptive teachers for amortized samplers.International Conference on Learning Representations (ICLR)

    2025

  29. [36]

    Kingma, D. P. and Ba, J. (2015). Adam: A method for stochastic optimization.International Conference on Learning Representations (ICLR)

    2015

  30. [37]

    A., and Zhang, C

    Kong, L., Cui, J., Sun, H., Zhuang, Y ., Prakash, B. A., and Zhang, C. (2023). Autoregressive diffusion model for graph generation.International Conference on Machine Learning (ICML)

    2023

  31. [38]

    Kostenetskiy, P., Chulkevich, R., and Kozyrev, V . (2021). HPC resources of the Higher School of Economics.Journal of Physics: Conference Series, 1740:012050

    2021

  32. [39]

    N., Bengio, Y ., and Malkin, N

    Lahlou, S., Deleu, T., Lemos, P., Zhang, D., V olokhova, A., Hernández-Garcıa, A., Ezzine, L. N., Bengio, Y ., and Malkin, N. (2023). A theory of continuous generative flow networks. International Conference on Machine Learning (ICML)

    2023

  33. [40]

    J., Kawaguchi, K., Gidel, G., Bengio, Y ., Malkin, N., and Jain, M

    Lee, S., Kim, M., Cherif, L., Dobre, D., Lee, J., Hwang, S. J., Kawaguchi, K., Gidel, G., Bengio, Y ., Malkin, N., and Jain, M. (2025). Learning diverse attacks on large language models for robust red-teaming and safety tuning.International Conference on Learning Representatio...

    2025

  34. [41]

    Madan, K., Lamb, A., Bengio, E., Berseth, G., and Bengio, Y . (2025). Towards improv- ing exploration through sibling augmented GFlowNets.International Conference on Learning Representations (ICLR)

    2025

  35. [42]

    C., Bosc, T., Bengio, Y ., and Malkin, N

    Madan, K., Rector-Brooks, J., Korablyov, M., Bengio, E., Jain, M., Nica, A. C., Bosc, T., Bengio, Y ., and Malkin, N. (2023). Learning GFlowNets from partial episodes for improved convergence and stability.International Conference on Machine Learning (ICML)

    2023

  36. [43]

    Malkin, N., Jain, M., Bengio, E., Sun, C., and Bengio, Y . (2022). Trajectory balance: Improved credit assignment in GFlowNets.Neural Information Processing Systems (NeurIPS)

    2022

  37. [44]

    J., Everett, K

    Malkin, N., Lahlou, S., Deleu, T., Ji, X., Hu, E. J., Everett, K. E., Zhang, D., and Bengio, Y . (2023). GFlowNets and variational inference.International Conference on Learning Representa- tions (ICLR)

    2023

  38. [45]

    and Fleuret, F

    Máté, B. and Fleuret, F. (2023). Learning interpolations between Boltzmann densities.Transac- tions on Machine Learning Research (TMLR)

    2023

  39. [46]

    I., Stimper, V ., Simm, G

    Midgley, L. I., Stimper, V ., Simm, G. N., Schölkopf, B., and Hernández-Lobato, J. M. (2023). Flow annealed importance sampling bootstrap.International Conference on Learning Representa- tions (ICLR)

    2023

  40. [47]

    A., Veness, J., Bellemare, M

    Mnih, V ., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., et al. (2015). Human-level control through deep reinforcement learning.nature, 518(7540):529–533. 13

    2015

  41. [48]

    Mohammadpour, S., Bengio, E., Frejinger, E., and Bacon, P.-L. (2024). Maximum entropy GFlowNets with soft Q-learning.Artificial Intelligence and Statistics (AISTATS)

    2024

  42. [49]

    Morozov, N., Maksimov, I., Tiapkin, D., and Samsonov, S. (2025). Revisiting non-acyclic GFlowNets in discrete environments.arXiv preprint arXiv:2502.07735

    2025 arXiv

  43. [50]

    Nachum, O., Norouzi, M., Xu, K., and Schuurmans, D. (2017). Bridging the gap between value and policy based reinforcement learning.Neural Information Processing Systems (NeurIPS)

    2017

  44. [51]

    Neal, R. M. (1998). Annealed importance sampling.arXiv preprint arXiv:physics/9803008

    1998 arXiv

  45. [52]

    Neal, R. M. (2003). Slice sampling.The annals of statistics, 31(3):705–767

    2003

  46. [53]

    Neal, R. M. et al. (2011). MCMC using Hamiltonian dynamics.Handbook of Markov Chain Monte Carlo, 2(11):2

    2011

  47. [54]

    Neu, G., Jonsson, A., and Gómez, V . (2017). A unified view of entropy-regularized Markov decision processes.arXiv preprint arXiv:1705.07798

    2017 arXiv

  48. [55]

    M., Christensen, A., Dittadi, A., and Winther, O

    Nielsen, B. M., Christensen, A., Dittadi, A., and Winther, O. (2024). DiffEnc: Variational diffusion with a learned encoder.International Conference on Learning Representations (ICLR)

    2024

  49. [56]

    Noé, F., Olsson, S., Köhler, J., and Wu, H. (2019). Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457):eaaw1147

    2019

  50. [57]

    and Richter, L

    Nüsken, N. and Richter, L. (2021). Solving high-dimensional Hamilton–Jacobi–Bellman PDEs using neural networks: perspectives from the theory of controlled diffusions and measures on path space.Partial differential equations and applications, 2(4):48

    2021

  51. [58]

    and Richter, L

    Nüsken, N. and Richter, L. (2023). Interpolating between BSDEs and PINNs: Deep learning for elliptic and parabolic boundary value problems.Journal of Machine Learning

    2023

  52. [59]

    Pan, L., Malkin, N., Zhang, D., and Bengio, Y . (2023). Better training of GFlowNets with local credit and incomplete trajectories.International Conference on Machine Learning (ICML)

    2023

  53. [60]

    and Cipcigan, F

    Phillips, D. and Cipcigan, F. (2024). MetaGFN: Exploring distant modes with adapted metady- namics for continuous GFlowNets.arXiv preprint arXiv:2408.15905

    2024 arXiv

  54. [61]

    Puterman, M. L. (2014).Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons

    2014

  55. [62]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021). Learning transferable visual models from natural language supervision.International Conference on Machine Learning (ICML)

    2021

  56. [63]

    Rector-Brooks, J., Madan, K., Jain, M., Korablyov, M., Liu, C.-H., Chandar, S., Malkin, N., and Bengio, Y . (2023). Thompson sampling for improved exploration in GFlowNets.arXiv preprint arXiv:2306.17693

    2023 arXiv

  57. [64]

    and Berner, J

    Richter, L. and Berner, J. (2024). Improved sampling via learned diffusions.International Conference on Learning Representations (ICLR)

    2024

  58. [65]

    Richter, L., Boustati, A., Nüsken, N., Ruiz, F., and Akyildiz, O. D. (2020). Vargrad: a low- variance gradient estimator for variational inference.Neural Information Processing Systems (NeurIPS)

    2020

  59. [66]

    Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation.Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 1...

    2015

  60. [67]

    M., and Kuleshov, V

    Sahoo, S., Gokaslan, A., De Sa, C. M., and Kuleshov, V . (2024). Diffusion models with learned adaptive noise.Neural Information Processing Systems (NeurIPS). 14

    2024

  61. [68]

    Samsonov, S., Lagutin, E., Gabrié, M., Durmus, A., Naumov, A., and Moulines, E. (2022). Local-global MCMC kernels: the best of both worlds.Neural Information Processing Systems (NeurIPS)

    2022

  62. [69]

    Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2016). Prioritized experience replay. International Conference on Learning Representations (ICLR)

    2016

  63. [71]

    R., Cherkasov, A., Kim, W

    Shen, T., Seo, S., Lee, G., Pandey, M., Smith, J. R., Cherkasov, A., Kim, W. Y ., and Ester, M. (2024). TacoGFN: Target-conditioned GFlowNet for structure-based drug design.Transactions on Machine Learning Research

    2024

  64. [72]

    D., Campbell, A., and Doucet, A

    Shi, Y ., Bortoli, V . D., Campbell, A., and Doucet, A. (2023). Diffusion Schrödinger bridge matching.Neural Information Processing Systems (NeurIPS)

    2023

  65. [73]

    B., de Souza da Silva, E., Souza, A

    Silva, T., Alves, R. B., de Souza da Silva, E., Souza, A. H., Garg, V ., Kaski, S., and Mesquita, D. (2025). When do GFlowNets learn the right distribution?International Conference on Learning Representations (ICLR)

    2025

  66. [74]

    A., and Mesquita, D

    Silva, T., de Souza, D. A., and Mesquita, D. (2024). Streaming Bayes GFlowNets.Neural Information Processing Systems (NeurIPS)

    2024

  67. [75]

    Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Riedmiller, M. (2014). Determinis- tic policy gradient algorithms.International Conference on Machine Learning (ICML)

    2014

  68. [76]

    A., Maheswaranathan, N., and Ganguli, S

    Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N., and Ganguli, S. (2015). Deep unsu- pervised learning using nonequilibrium thermodynamics.International Conference on Machine Learning (ICML)

    2015

  69. [77]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2021). Score- based generative modeling through stochastic differential equations.International Conference on Learning Representations (ICLR)

    2021

  70. [78]

    Song, Z., Yang, C., Wang, C., An, B., and Li, S. (2024). Latent logic tree extraction for event sequence explanation from LLMs.International Conference on Machine Learning (ICML)

    2024

  71. [79]

    Stromme, A. (2023). Sampling from a Schrödinger bridge.Artificial Intelligence and Statistics (AISTATS)

    2023

  72. [80]

    Sun, J., Berner, J., Azizzadenesheli, K., and Anandkumar, A. (2024). Physics-informed neural networks for sampling.ICLR 2024 Workshop on AI4DifferentialEquations In Science

    2024

  73. [81]

    Tiapkin, D., Morozov, N., Naumov, A., and Vetrov, D. P. (2024). Generative flow networks as entropy-regularized RL.Artificial Intelligence and Statistics (AISTATS)

    2024

  74. [82]

    Tong, A., Malkin, N., Fatras, K., Atanackovic, L., Zhang, Y ., Huguet, G., Wolf, G., and Bengio, Y . (2024). Simulation-free schrödinger bridges via score and flow matching.Artificial Intelligence and Statistics (AISTATS)

    2024

  75. [83]

    Uehara, M., Zhao, Y ., Biancalani, T., and Levine, S. (2024). Understanding reinforce- ment learning-based fine-tuning of diffusion models: A tutorial and review.arXiv preprint arXiv:2407.13734

    2024 arXiv

  76. [84]

    van Krieken, E., Thanapalasingam, T., Tomczak, J., Van Harmelen, F., and Ten Teije, A. (2023). A-NeSI: A scalable approximate method for probabilistic neurosymbolic inference.Neural Information Processing Systems (NeurIPS)

    2023

  77. [85]

    Vargas, F., Grathwohl, W., and Doucet, A. (2023). Denoising diffusion samplers.International Conference on Learning Representations (ICLR)

    2023

  78. [86]

    D., and Nüsken, N

    Vargas, F., Ovsianas, A., Fernandes, D., Girolami, M., Lawrence, N. D., and Nüsken, N. (2022). Bayesian learning via neural schrödinger–föllmer flows.Statistics and Computing, 33(1):3. 15

    2022

  79. [87]

    Vargas, F., Padhy, S., Blessing, D., and Nüsken, N. (2024). Transport meets variational inference: Controlled monte carlo diffusions.International Conference on Learning Representations (ICLR)

    2024

  80. [88]

    Vargas, F., Thodoroff, P., Lamacraft, A., and Lawrence, N. (2021). Solving Schrödinger bridges via maximum likelihood.Entropy, 23(9):1134

    2021

  81. [89]

    Venkatraman, S., Hasan, M., Kim, M., Scimeca, L., Sendera, M., Bengio, Y ., Berseth, G., and Malkin, N. (2025). Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models.International Conference on Machine Learning (ICML)

    2025

  82. [90]

    V olokhova, A., Koziarski, M., Hernández-García, A., Liu, C.-H., Miret, S., Lemos, P., Thiede, L., Yan, Z., Aspuru-Guzik, A., and Bengio, Y . (2024). Towards equilibrium molecular conformation generation with GFlowNets.Digital Discovery, 3:1038–1047

    2024

  83. [91]

    Xu, J., Liu, X., Wu, Y ., Tong, Y ., Li, Q., Ding, M., Tang, J., and Dong, Y . (2023). Imagereward: Learning and evaluating human preferences for text-to-image generation.Neural Information Processing Systems (NeurIPS)

    2023

  84. [92]

    Younsi, A., Abubaker, A., Seddik, M. E. A., Hacid, H., and Lahlou, S. (2025). Accurate and diverse LLM mathematical reasoning via automated PRM-guided GFlowNets.arXiv preprint arXiv:2504.19981

    2025

  85. [93]

    Zhang, D., Chen, R. T. Q., Liu, C.-H., Courville, A., and Bengio, Y . (2024). Diffusion generative flow samplers: Improving learning signals through partial trajectory optimization.International Conference on Learning Representations (ICLR)

    2024

  86. [94]

    Zhang, D., Chen, R. T. Q., Malkin, N., and Bengio, Y . (2023a). Unifying generative models with GFlowNets and beyond.arXiv preprint arXiv:2209.02606

    2023 arXiv

  87. [95]

    Zhang, D., Dai, H., Malkin, N., Courville, A., Bengio, Y ., and Pan, L. (2023b). Let the flows tell: Solving graph combinatorial problems with GFlowNets.Neural Infromation Processing Systems (NeurIPS)

    2023

  88. [96]

    Zhang, D., Rainone, C., Peschl, M., and Bondesan, R. (2023c). Robust scheduling with GFlowNets.International Conference on Learning Representations (ICLR)

    2023

  89. [97]

    and Chen, Y

    Zhang, Q. and Chen, Y . (2022). Path integral sampler: A stochastic control approach for sampling.International Conference on Learning Representations (ICLR)

    2022

  90. [98]

    Y ., Yan, Z., Layne, E., Malkin, N., Zhang, D., Jain, M., Blanchette, M., and Bengio, Y

    Zhou, M. Y ., Yan, Z., Layne, E., Malkin, N., Zhang, D., Jain, M., Blanchette, M., and Bengio, Y . (2024). PhyloGFN: Phylogenetic inference with generative flow networks.International Conference on Learning Representations (ICLR)

    2024

  91. [99]

    log exp(−E(𝑋 1))←−𝑝𝜑(𝑋 0,...,(𝑇−1)Δ𝑡 |𝑋 1) 𝑝0(𝑋 0)− →𝑝𝜃(𝑋Δ𝑡...,1 |𝑋 0) # EUBO=E 𝑋0,...,1∼𝑝 target(𝑋1)𝑝𝜑(𝑋0,...,(𝑇−1)Δ𝑡 |𝑋 1)

    Zimmermann, H., Lindsten, F., van de Meent, J.-W., and Naesseth, C. A. (2023). A variational perspective on generative flow networks.Transactions on Machine Learning Research. 16 A More on diffusion samplers A.1 On KL divergence between processes with different variances. If t...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.