StAD: Stein Amortized Divergence for Fast Likelihoods with Diffusion and Flow

arxiv: 2605.16486 · v1 · pith:N5PXGTBXnew · submitted 2026-05-15 · 📊 stat.ML · astro-ph.IM· cs.LG

StAD: Stein Amortized Divergence for Fast Likelihoods with Diffusion and Flow

Gurjeet Jagwani , Stephen Thorp , Sinan Deger , Hiranya Peiris This is my paper

Pith reviewed 2026-05-19 21:29 UTC · model grok-4.3

classification 📊 stat.ML astro-ph.IMcs.LG

keywords Stein operatoramortized divergenceprobability flow ODEdiffusion modelsflow modelslikelihood estimationgenerative modelsdensity estimation

0 comments p. Extension

pith:N5PXGTBX Add to your LaTeX paper

What is a Pith Number?

\usepackage{pith}
\pithnumber{N5PXGTBX}

Prints a linked pith:N5PXGTBX badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

StAD uses the Langevin-Stein operator to learn PF-ODE divergences without ever computing Jacobians.

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

The paper introduces StAD, a distillation method that predicts the divergence of the probability flow ordinary differential equation in diffusion and flow models by applying the Langevin-Stein operator. This sidesteps the costly trace of the Jacobian, which is O(D squared) exactly or O(D) with noise using standard estimators. The approach matches or exceeds the Hutchinson and Hutch++ estimators in speed and variance on density estimation tasks including CIFAR-10 and ImageNet. Under regularity conditions the learned vector fields are shown to satisfy the Stein class, supporting generalization across a range of generative models.

Core claim

StAD distills the divergence of the PF-ODE using the Langevin-Stein operator so that likelihoods can be obtained without computing the Jacobian of the learned vector field; the resulting amortized estimates are competitive with Hutchinson-style methods on image benchmarks and, when regularity conditions hold, produce vector fields that belong to the Stein class and therefore generalize to varied generative models.

What carries the argument

The Langevin-Stein operator applied as a distillation target to amortize and predict the divergence of the PF-ODE vector field.

If this is right

Likelihood evaluation becomes faster and exhibits lower variance than the Hutchinson estimator on CIFAR-10 and ImageNet.
The same learned divergence predictor applies across multiple diffusion and flow architectures without retraining the Jacobian estimator.
Density estimation workflows that require repeated likelihood calls become more practical for Bayesian analysis.
Under the stated regularity conditions the method extends to a broader class of generative models beyond the training distribution.

Where Pith is reading between the lines

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

High-dimensional Bayesian workflows could evaluate model evidence at lower cost by swapping in the amortized divergence.
The Stein-class property might allow the same distilled predictor to serve as a drop-in module for related ODE-based density estimators.
Training the distillation network once could amortize divergence costs across many downstream sampling or inference runs.

Load-bearing premise

Regularity conditions exist that allow the learned vector fields to satisfy the Stein class and thereby generalize.

What would settle it

Exact likelihood values computed via the analytical Jacobian on a low-dimensional Gaussian model differ systematically from the StAD estimates by more than the reported variance reduction.

Figures

Figures reproduced from arXiv: 2605.16486 by Gurjeet Jagwani, Hiranya Peiris, Sinan Deger, Stephen Thorp.

**Figure 1.** Figure 1: Divergence (blue/red contours) visualisation for a vector field (arrows) defined by the PF-ODE at a fixed time step. 1.2. Stein’s method and Stein operators Stein’s method (Stein, 1972) was introduced to bound the distance between a sum of random variables and the Gaussian distribution. Since then, it has proved to be a powerful tool for quantifying distances between probability measures and bounding conv… view at source ↗

**Figure 2.** Figure 2: Divergence estimation for a VP-SDE diffusion model, trained to predict astrophysical fluxes and flux errors (26 dimensional conditional density). Histograms show log likelihood residuals (i.e. exact − estimated log likelihood) for Hutchinson with 1 and 8 probes, rank-4 Hutch++ with 4 probes, and StAD. Dotted line corresponds to zero residual. 4. Experiments and results We test StAD on a variety of densit… view at source ↗

**Figure 3.** Figure 3: Pareto optimal divergence estimation with PF-ODEs. These are estimates for a VPSDE diffusion model predicting a 26 dimensional conditional density over astrophysical fluxes and flux errors. The figures plotted here are the same as in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Sample-wise correlation test for NLL estimates from StAD vs. 2-probe Hutchinson, for 2048 images taken from ImageNet32x32. Each blue point corresponds to one image. plexity. For CIFAR-10, we observe that the difference in the mean NLL estimates is ∼ 0.02 bpd. Similarly, for ImageNet-32x32, we observe a difference in the mean NLL of ∼ 0.07 bpd. The NLL estimates from StAD show less variance when averaged ac… view at source ↗

**Figure 5.** Figure 5: Sample-wise NLL estimates for CIFAR-10. Left: Using the Hutchinson algorithm with 2 probes. Centre: Using StAD. Right: Image-by-image difference between StAD and Hutchinson with 2 probes: NLL[StAD] − NLL[H(2)]. Acknowledgements We thank Anik Halder and Justin Alsing for useful comments regarding the project and some of the source code. This work has been supported by funding from the European Research Co… view at source ↗

**Figure 6.** Figure 6: shows the results of our experiment. For the PSD case, shown in the top panel, we see that Hutch++ and XTrace outperform Hutchinson for all m ≥ 4 irrespective 2 1 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 2 10 number of matrix-vector multiplications, m 10 7 10 5 10 3 10 1 10 1 10 3 MAE D = 4 D = 4 D = 4 D = 16 D = 16 D = 16 D = 64 D = 64 D = 64 D = 256 D = 256 D = 256 trace estimation, PSD matrix Hutchinson Hutch++ … view at source ↗

**Figure 7.** Figure 7: Image-wise NLL comparison (StAD vs. Hutchinson with 2 probes) for 2048 images from CIFAR-10. H. CIFAR-10 and ImageNet samples In [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

**Figure 8.** Figure 8: NLL histograms for 2048 images from ImageNet-32x32. Left: Using the Hutchinson algorithm with 2 probes to estimate the NLL. Centre: Using StAD. Right: Difference between StAD and Hutchinson: NLL[StAD] − NLL[Hutchinson]. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

**Figure 9.** Figure 9: Unconditional CIFAR-10 samples from a VP-SDE diffusion model trained with a U-Net architecture. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

**Figure 10.** Figure 10: ImageNet-32x32 samples from a VP-SDE diffusion model trained with a U-Net architecture. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗

read the original abstract

Diffusion and flow-based models are ubiquitously used for generative modelling and density estimation. They admit a deterministic probability flow ordinary differential equation (PF-ODE), analogous to continuous normalizing flows (CNFs), which describes the transport of the probability mass. Obtaining the likelihood from these models is of interest to many workflows, especially Bayesian analysis, and requires solving the trace of the Jacobian to compute the divergence of the learned PF-ODE, which is either $\mathcal{O}(D^2)$ to compute exactly or $\mathcal{O}(D)$ with a noisy estimate. We introduce StAD, a new distillation method to predict and learn the divergence of the PF-ODE using the Langevin-Stein operator without ever computing the Jacobian. We show that our method is competitive with the Hutchinson and Hutch++ on CIFAR-10, ImageNet and other density estimation tasks, consistently improving the variance and speed of the likelihood predictions compared to the Hutchinson. We additionally show our method will generalize to a varied class of generative models, and show that under some regularity conditions these learned vector fields can be made to satisfy the Stein class.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

StAD distills divergence via the Langevin-Stein operator for faster PF-ODE likelihoods, but the generalization claim hinges on unverified regularity conditions.

read the letter

The main point is that this paper gives a distillation trick to learn the divergence of the probability-flow ODE using the Langevin-Stein operator instead of ever touching the Jacobian. That sidesteps the usual O(D) or O(D^2) cost for likelihoods in diffusion and flow models, which matters for Bayesian work on image data or similar tasks. They report lower variance and better speed than plain Hutchinson on CIFAR-10 and ImageNet, and they claim the approach carries over to other generative models under some regularity conditions that put the learned fields into the Stein class. If the numbers hold, it is a practical step for anyone who needs repeated likelihood evaluations without the trace overhead. The framing is straightforward and the motivation is clear: exact divergence is a known bottleneck, and amortizing it this way is a reasonable direction that builds directly on the Stein identity without circular fitting. The experimental comparisons are at least directionally useful even if the abstract does not show the raw tables. The soft spot is the generalization step. The paper invokes regularity conditions to guarantee Stein-class membership but does not appear to supply empirical residuals on held-out models or any bounds that would confirm those conditions survive on typical PF-ODE vector fields for ImageNet-scale data. Without that check, the theoretical license for the variance reduction is thinner than it looks, and the speed gains might not transfer cleanly. Minor issues include the lack of visible ablations or error bars in the summary material, which makes it harder to judge how robust the improvement really is. This paper is aimed at people who already work with diffusion or flow models and need faster likelihoods for downstream inference. A reader focused on efficient estimators for continuous normalizing flows or score-based models would find the method description worth their time. It is coherent enough on its own terms to deserve a serious referee, mainly because the core idea targets a real computational pain point and the Stein-operator angle is distinct from prior Hutchinson-style work. I would send it to review with a request for explicit verification of the Stein-class conditions and fuller experimental reporting.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces StAD, a distillation method to approximate the divergence of the PF-ODE in diffusion and flow models via the Langevin-Stein operator, avoiding explicit Jacobian trace computation. It reports competitive likelihood estimation performance with reduced variance and improved speed relative to the Hutchinson estimator on CIFAR-10 and ImageNet, and asserts generalization across generative models provided the learned vector fields satisfy the Stein class under stated regularity conditions.

Significance. If the empirical speed/variance gains are reproducible with proper controls and the regularity conditions can be verified or bounded, the approach would offer a practical route to faster likelihoods in high-dimensional generative models without sacrificing the theoretical grounding of the Stein identity.

major comments (1)

Abstract and generalization section: the claim that StAD generalizes to a varied class of generative models rests on the assertion that the distilled vector fields satisfy the Stein class under the invoked regularity conditions. The manuscript supplies neither empirical verification (e.g., Stein identity residuals on held-out models or datasets) nor bounds confirming that typical PF-ODE vector fields on CIFAR-10/ImageNet obey these conditions; this is load-bearing for the theoretical justification of the reported variance and speed advantages outside the distillation distribution.

minor comments (1)

The abstract states competitive results on CIFAR-10 and ImageNet yet the manuscript should include explicit tables with error bars, run-time measurements, and ablation controls so that the variance and speed claims can be quantitatively assessed.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the concern regarding the generalization claim and the supporting evidence for the regularity conditions below, and will revise the paper to strengthen this aspect.

read point-by-point responses

Referee: Abstract and generalization section: the claim that StAD generalizes to a varied class of generative models rests on the assertion that the distilled vector fields satisfy the Stein class under the invoked regularity conditions. The manuscript supplies neither empirical verification (e.g., Stein identity residuals on held-out models or datasets) nor bounds confirming that typical PF-ODE vector fields on CIFAR-10/ImageNet obey these conditions; this is load-bearing for the theoretical justification of the reported variance and speed advantages outside the distillation distribution.

Authors: We agree that explicit empirical verification and bounds would strengthen the generalization claim. The current manuscript invokes standard regularity conditions (e.g., sufficient smoothness and decay at infinity for the vector field to belong to the Stein class) under which the identity holds, and demonstrates competitive performance on the evaluated models. In the revision we will add (i) empirical measurements of Stein identity residuals on held-out data and additional generative models, and (ii) a short discussion with references showing that the Lipschitz continuity and bounded-gradient properties typical of trained PF-ODE vector fields on CIFAR-10/ImageNet are sufficient to satisfy the invoked conditions. These additions will directly support the theoretical justification for variance and speed advantages beyond the distillation distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external Stein operator and empirical benchmarks

full rationale

The paper presents StAD as a distillation approach that replaces Jacobian trace computation with the Langevin-Stein operator for PF-ODE divergence. Competitiveness is demonstrated via direct empirical comparisons to Hutchinson and Hutch++ estimators on CIFAR-10, ImageNet, and other tasks, with reported improvements in variance and speed. Generalization to varied generative models is stated under explicit regularity conditions that enable Stein class membership; these conditions are invoked as assumptions rather than derived from the method itself. No load-bearing step reduces by construction to a fitted parameter, self-citation chain, or renamed input. The central claims remain independently verifiable against external benchmarks and the established Stein identity, satisfying the criteria for a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard domain assumptions about PF-ODEs in generative models plus an ad-hoc regularity condition for the Stein class; no free parameters or new entities are introduced in the abstract.

axioms (2)

domain assumption Diffusion and flow-based models admit a deterministic probability flow ordinary differential equation (PF-ODE).
Foundational premise stated at the start of the abstract for all subsequent divergence computations.
ad hoc to paper Under some regularity conditions these learned vector fields can be made to satisfy the Stein class.
Invoked to claim generalization to a varied class of generative models.

pith-pipeline@v0.9.0 · 5740 in / 1294 out tokens · 43489 ms · 2026-05-19T21:29:40.354359+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We introduce StAD, a new distillation method to predict and learn the divergence of the PF-ODE using the Langevin-Stein operator without ever computing the Jacobian... under some regularity conditions these learned vector fields can be made to satisfy the Stein class.
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

lim R→∞ ∮∂BR p(x)δ(x)v(x),n̂ dS = 0 ... We regularise δ such that this occurs by construction... κR ... cosine transition

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 4 internal anchors

[1]

Alsing, J., Peiris, H., Mortlock, D., Leja, J., and Leistedt, B

URL https://doi.org/10.1093/mnras/ stz1960. Alsing, J., Peiris, H., Mortlock, D., Leja, J., and Leistedt, B. Forward modeling of galaxy populations for cosmo- logical redshift distribution inference.The Astrophysical Journal Supplement Series, 264(2):29, February 2023. URL https://doi.org/10.3847/1538-4365/ ac9583. Alsing, J., Thorp, S., Deger, S., Peiris...

work page doi:10.1093/mnras/ 2023
[2]

cc/paper_files/paper/2019/file/ ba7609ee5789cc4dff171045a693a65f- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ ba7609ee5789cc4dff171045a693a65f- Paper.pdf. B´enard, C., Staber, B., and Da Veiga, S. Kernel Stein discrepancy thinning: a theoretical perspective of pathologies and a practical fix with regularization. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Adva...

work page 2019
[3]

cc/paper_files/paper/2023/file/ 9a8eb202c060b7d81f5889631cbcd47e- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ 9a8eb202c060b7d81f5889631cbcd47e- Paper-Conference.pdf. Biloˇs, M. and G¨unnemann, S. Scalable normalizing flows for permutation invariant densities. In Meila, M. and Zhang, T. (eds.),Proceedings of the 38th International Conference on Machine Learning, volume 139 ofPro- ceedings of Machine ...

work page 2023
[4]

A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 770f8e448d07586afbf77bb59f698587- Paper.pdf. Chen, R. T. Q., Rubanova, Y ., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equa- tions. In Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Process...

work page internal anchor Pith review Pith/arXiv arXiv 2019
[5]

NICE: Non-linear Independent Components Estimation

URL https://doi.org/10.3847/1538- 3881/ad54bf. 10 StAD: Stein Amortized Divergence Dinh, L., Krueger, D., and Bengio, Y . NICE: non-linear independent components estimation. In Bengio, Y . and LeCun, Y . (eds.),3rd International Conference on Learn- ing Representations, May 2015. URL http://arxiv. org/abs/1410.8516. Dormand, J. and Prince, P. A family of ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538- 2015
[6]

cc/paper_files/paper/2025/file/ 6d13e085b79d454da5910e4ca82a3d9d- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2025/file/ 6d13e085b79d454da5910e4ca82a3d9d- Paper-Conference.pdf. Girard, D. A. A fast ‘Monte-Carlo cross-validation’ pro- cedure for large least squares problems with noisy data. Numerische Mathematik, 56:1–23, 1989. URL https: //doi.org/10.1007/BF01395775. Gittens, A. and Mahoney, M. Revisiting the ...

work page doi:10.1007/bf01395775 2025
[7]

Ho, J., Jain, A., and Abbeel, P

URL https://proceedings.mlr.press/ v119/grathwohl20a.html. Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 6840–6851. Curran Associates, Inc.,

work page
[8]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b- Paper.pdf. Huang, Y ., Transue, T., Wang, S.-H., Feldman, W. M., Zhang, H., and Wang, B. Improving flow match- ing by aligning flow divergence. In Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Ma- haraj, T., Wagstaff, K., and Zhu, J. (ed...

work page 2020
[9]

11 StAD: Stein Amortized Divergence Hutchinson, M

URL https://proceedings.mlr.press/ v267/huang25ag.html. 11 StAD: Stein Amortized Divergence Hutchinson, M. F. A stochastic estimator of the trace of the influence matrix for Laplacian smoothing splines.Com- munications in Statistics – Simulation and Computation, 18(3):1059–1076, 1989. URL https://doi.org/ 10.1080/03610918908812806. Hyv¨arinen, A. Estimati...

work page doi:10.1080/03610918908812806 1989
[10]

cc/paper_files/paper/2022/file/ a98846e9d9cc01cfb87eb694d946ce6b- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2022/file/ a98846e9d9cc01cfb87eb694d946ce6b- Paper-Conference.pdf. Karras, T., Aittala, M., Lehtinen, J., Hellsten, J., Aila, T., and Laine, S. Analyzing and improving the training dy- namics of diffusion models. In2024 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp. 24174–...

work page 2022
[11]

Shan Jia, Mingzhen Huang, Zhou Zhou, Yan Ju, Jialing Cai, and Siwei Lyu

IEEE Computer Society. URL https://doi. org/10.1109/CVPR52733.2024.02282. Kingma, D. and Gao, R. Understanding diffusion objectives as the ELBO with simple data augmen- tation. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Ad- vances in Neural Information Processing Systems, volume 36, pp. 65484–65516. Curran Associat...

work page doi:10.1109/cvpr52733.2024.02282 2024
[12]

cc/paper_files/paper/2023/file/ ce79fbf9baef726645bc2337abb0ade2- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ ce79fbf9baef726645bc2337abb0ade2- Paper-Conference.pdf. Kobyzev, I., Prince, S. J., and Brubaker, M. A. Normaliz- ing flows: An introduction and review of current meth- ods.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3964–3979, 2021. URL https: //doi.org/10.1109/TP...

work page doi:10.1109/tpami.2020.2992934 2023
[13]

cc/paper_files/paper/2016/file/ b3ba8f1bee1238a2f37603d90b58898d- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ b3ba8f1bee1238a2f37603d90b58898d- Paper.pdf. Liu, Q., Lee, J., and Jordan, M. A kernelized Stein dis- crepancy for goodness-of-fit tests. In Balcan, M. F. and Weinberger, K. Q. (eds.),Proceedings of The 33rd Inter- national Conference on Machine Learning, volume 48 of Proceedings of Machine ...

work page 2016
[14]

Liu, X., Du, H., Deng, W., and Zhang, R

URL https://openreview.net/forum? id=1k4yZbbDqX. Liu, X., Du, H., Deng, W., and Zhang, R. Optimal stochas- tic trace estimation in generative modeling. In Li, Y ., Mandt, S., Agrawal, S., and Khan, E. (eds.),Proceedings of The 28th International Conference on Artificial Intelli- gence and Statistics, volume 258 ofProceedings of Ma- chine Learning Research...

work page
[15]

URL https://proceedings.mlr.press/ v258/liu25k.html. Lu, C. and Song, Y . Simplifying, stabilizing and scaling continuous-time consistency models. In Yue, Y . (ed.), 13th International Conference on Learning Represen- tations, April 2025. URL https://openreview. net/forum?id=LyJi5ugyJx. Luhman, E. and Luhman, T. Knowledge distillation in iter- ative gener...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[16]

cc/paper_files/paper/2023/file/ f115f619b62833aadc5acb058975b0e6- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ f115f619b62833aadc5acb058975b0e6- Paper-Conference.pdf. Maoutsa, D., Reich, S., and Opper, M. Interacting particle solutions of Fokker–Planck equations through gradient- log-density estimation.Entropy, 22(8):802, July 2020. URLhttps://doi.org/10.3390/e22080802. Meyer, R. A., Musco, C., Musco...

work page doi:10.3390/e22080802 2023
[17]

cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740- Paper.pdf. Perez, E., Strub, F., de Vries, H., Dumoulin, V ., and Courville, A. FiLM: visual reasoning with a general conditioning layer. InProceedings of the Thirty-Second AAAI Conference on Artificial In- telligence, AAAI’18/IAAI’18/EAAI’18. AAAI Press,

work page 2019
[18]

v32i1.11671

URL https://doi.org/10.1609/aaai. v32i1.11671. 13 StAD: Stein Amortized Divergence Persson, D., Cortinovis, A., and Kressner, D. Improved variants of the Hutch++ algorithm for trace estimation. SIAM Journal on Matrix Analysis and Applications, 43 (3):1162–1185, 2022. URL https://doi.org/10. 1137/21M1447623. Ranganath, R., Tran, D., Altosaar, J., and Blei,...

work page doi:10.1609/aaai 2022
[19]

cc/paper_files/paper/2016/file/ d947bf06a885db0d477d707121934ff8- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ d947bf06a885db0d477d707121934ff8- Paper.pdf. Raonic, B., Mishra, S., and Lanthaler, S. Towards a cer- tificate of trust: Task-aware OOD detection for scientific AI. In V ondrick, C. (ed.),14th International Conference on Learning Representations, April 2026. URL https: //openreview.net/forum...

work page 2016
[20]

Riabiz, M., Chen, W

URL https://openreview.net/forum? id=FbssShlI4N. Riabiz, M., Chen, W. Y ., Cockayne, J., Swietach, P., Niederer, S. A., Mackey, L., and Oates, C. J. Optimal thinning of MCMC output.Journal of the Royal Sta- tistical Society Series B: Statistical Methodology, 84(4): 1059–1081, April 2022. URL https://doi.org/ 10.1111/rssb.12503. Robbins, H. E. An empirical...

work page doi:10.1111/rssb.12503 2022
[21]

Fundamentals of Enzyme Kinetics: Michaelis-Menten and Non-Michaelis- Type (Atypical) Enzyme Kinetics

URL https://doi.org/10.1007/978-1- 4612-0919-5_26. Ronneberger, O., Fischer, P., and Brox, T. U-Net: Convolu- tional networks for biomedical image segmentation. In Navab, N., Hornegger, J., Wells, W. M., and Frangi, A. F. (eds.),Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, volume 9351 ofLecture Notes in Computer Science, pp. 2...

work page doi:10.1007/978-1- 2015
[22]

cc/paper_files/paper/2025/file/ d79ac139911df25d27f14bbf008deaee- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2025/file/ d79ac139911df25d27f14bbf008deaee- Paper-Conference.pdf. Saibaba, A. K., Alexanderian, A., and Ipsen, I. C. Ran- domized matrix-free trace and log-determinant esti- mators.Numerische Mathematik, 137:353–395, Oc- tober 2017. URL https://doi.org/10.1007/ s00211-017-0880-z. Salimans, T. and Ho, ...

work page doi:10.1007/978-3-031-73016-0_6 2025
[23]

URL https://openreview.net/forum? id=WNzy9bRDvG. Song, Y . and Ermon, S. Generative modeling by estimating gradients of the data distribution. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d’Alch´e-Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 32, pp. 11895–11907. Curran Associates, Inc.,

work page
[24]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93- Paper.pdf. Song, Y . and Ermon, S. Improved techniques for training score-based generative models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 12438–12448. Cur...

work page 2019
[25]

cc/paper_files/paper/2020/file/ 92c3b916311a5517d9290576e3ea37ad- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 92c3b916311a5517d9290576e3ea37ad- Paper.pdf. Song, Y ., Durkan, C., Murray, I., and Ermon, S. Max- imum likelihood training of score-based diffusion models. In Ranzato, M., Beygelzimer, A., Dauphin, Y ., Liang, P., and Vaughan, J. W. (eds.),Advances in Neural Information Processing Systems, ...

work page 2020
[26]

A note on the evaluation of generative models

URL https://proceedings.mlr.press/ v202/song23a.html. Stein, C. A bound for the error in the normal approxima- tion to the distribution of a sum of dependent random variables. In Le Cam, L. M., Neyman, J., and Scott, E. L. (eds.),Proceedings of the sixth Berkeley sympo- sium on mathematical statistics and probability, volume 2: Probability theory, volume ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538-4357/ae0936 1972
[27]

Zheng, Q., Le, M., Shaul, N., Lipman, Y ., Grover, A., and Chen, R

URL https://proceedings.mlr.press/ v267/yun25a.html. Zheng, Q., Le, M., Shaul, N., Lipman, Y ., Grover, A., and Chen, R. T. Q. Guided flows for generative modeling and decision making.preprint, November 2023. URL https://arxiv.org/abs/2311.13443. Zhou, M., Zheng, H., Wang, Z., Yin, M., and Huang, H. Score identity distillation: Exponentially fast distilla...

work page arXiv 2023
[28]

Zhou, M., Gu, Y ., Zheng, H., Song, L., He, G., Zhang, Y ., Hu, W., and Yang, Y

URL https://proceedings.mlr.press/ v235/zhou24x.html. Zhou, M., Gu, Y ., Zheng, H., Song, L., He, G., Zhang, Y ., Hu, W., and Yang, Y . Score distillation of flow matching models.preprint, September 2025. URL https:// arxiv.org/abs/2509.25127. A. Stein class conditions The boundary condition for the Stein class (Liu & Wang, 2016; Gorham & Mackey, 2017; Ba...

work page arXiv 2025
[29]

Of these formulations, the TrigFlow continuous- time consistency model (Lu & Song, 2025) is a particularly natural fit forStAD

and mean flow (Geng et al., 2025), that were devel- oped for faster sampling under diffusion and flow-based models. Of these formulations, the TrigFlow continuous- time consistency model (Lu & Song, 2025) is a particularly natural fit forStAD. 18 StAD: Stein Amortized Divergence Lu & Song (2025) introduce TrigFlow, that optimises the objective L(θ) =E t E...

work page 2025

[1] [1]

Alsing, J., Peiris, H., Mortlock, D., Leja, J., and Leistedt, B

URL https://doi.org/10.1093/mnras/ stz1960. Alsing, J., Peiris, H., Mortlock, D., Leja, J., and Leistedt, B. Forward modeling of galaxy populations for cosmo- logical redshift distribution inference.The Astrophysical Journal Supplement Series, 264(2):29, February 2023. URL https://doi.org/10.3847/1538-4365/ ac9583. Alsing, J., Thorp, S., Deger, S., Peiris...

work page doi:10.1093/mnras/ 2023

[2] [2]

cc/paper_files/paper/2019/file/ ba7609ee5789cc4dff171045a693a65f- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ ba7609ee5789cc4dff171045a693a65f- Paper.pdf. B´enard, C., Staber, B., and Da Veiga, S. Kernel Stein discrepancy thinning: a theoretical perspective of pathologies and a practical fix with regularization. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Adva...

work page 2019

[3] [3]

cc/paper_files/paper/2023/file/ 9a8eb202c060b7d81f5889631cbcd47e- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ 9a8eb202c060b7d81f5889631cbcd47e- Paper-Conference.pdf. Biloˇs, M. and G¨unnemann, S. Scalable normalizing flows for permutation invariant densities. In Meila, M. and Zhang, T. (eds.),Proceedings of the 38th International Conference on Machine Learning, volume 139 ofPro- ceedings of Machine ...

work page 2023

[4] [4]

A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 770f8e448d07586afbf77bb59f698587- Paper.pdf. Chen, R. T. Q., Rubanova, Y ., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equa- tions. In Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Process...

work page internal anchor Pith review Pith/arXiv arXiv 2019

[5] [5]

NICE: Non-linear Independent Components Estimation

URL https://doi.org/10.3847/1538- 3881/ad54bf. 10 StAD: Stein Amortized Divergence Dinh, L., Krueger, D., and Bengio, Y . NICE: non-linear independent components estimation. In Bengio, Y . and LeCun, Y . (eds.),3rd International Conference on Learn- ing Representations, May 2015. URL http://arxiv. org/abs/1410.8516. Dormand, J. and Prince, P. A family of ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538- 2015

[6] [6]

cc/paper_files/paper/2025/file/ 6d13e085b79d454da5910e4ca82a3d9d- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2025/file/ 6d13e085b79d454da5910e4ca82a3d9d- Paper-Conference.pdf. Girard, D. A. A fast ‘Monte-Carlo cross-validation’ pro- cedure for large least squares problems with noisy data. Numerische Mathematik, 56:1–23, 1989. URL https: //doi.org/10.1007/BF01395775. Gittens, A. and Mahoney, M. Revisiting the ...

work page doi:10.1007/bf01395775 2025

[7] [7]

Ho, J., Jain, A., and Abbeel, P

URL https://proceedings.mlr.press/ v119/grathwohl20a.html. Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 6840–6851. Curran Associates, Inc.,

work page

[8] [8]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b- Paper.pdf. Huang, Y ., Transue, T., Wang, S.-H., Feldman, W. M., Zhang, H., and Wang, B. Improving flow match- ing by aligning flow divergence. In Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Ma- haraj, T., Wagstaff, K., and Zhu, J. (ed...

work page 2020

[9] [9]

11 StAD: Stein Amortized Divergence Hutchinson, M

URL https://proceedings.mlr.press/ v267/huang25ag.html. 11 StAD: Stein Amortized Divergence Hutchinson, M. F. A stochastic estimator of the trace of the influence matrix for Laplacian smoothing splines.Com- munications in Statistics – Simulation and Computation, 18(3):1059–1076, 1989. URL https://doi.org/ 10.1080/03610918908812806. Hyv¨arinen, A. Estimati...

work page doi:10.1080/03610918908812806 1989

[10] [10]

cc/paper_files/paper/2022/file/ a98846e9d9cc01cfb87eb694d946ce6b- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2022/file/ a98846e9d9cc01cfb87eb694d946ce6b- Paper-Conference.pdf. Karras, T., Aittala, M., Lehtinen, J., Hellsten, J., Aila, T., and Laine, S. Analyzing and improving the training dy- namics of diffusion models. In2024 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp. 24174–...

work page 2022

[11] [11]

Shan Jia, Mingzhen Huang, Zhou Zhou, Yan Ju, Jialing Cai, and Siwei Lyu

IEEE Computer Society. URL https://doi. org/10.1109/CVPR52733.2024.02282. Kingma, D. and Gao, R. Understanding diffusion objectives as the ELBO with simple data augmen- tation. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Ad- vances in Neural Information Processing Systems, volume 36, pp. 65484–65516. Curran Associat...

work page doi:10.1109/cvpr52733.2024.02282 2024

[12] [12]

cc/paper_files/paper/2023/file/ ce79fbf9baef726645bc2337abb0ade2- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ ce79fbf9baef726645bc2337abb0ade2- Paper-Conference.pdf. Kobyzev, I., Prince, S. J., and Brubaker, M. A. Normaliz- ing flows: An introduction and review of current meth- ods.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3964–3979, 2021. URL https: //doi.org/10.1109/TP...

work page doi:10.1109/tpami.2020.2992934 2023

[13] [13]

cc/paper_files/paper/2016/file/ b3ba8f1bee1238a2f37603d90b58898d- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ b3ba8f1bee1238a2f37603d90b58898d- Paper.pdf. Liu, Q., Lee, J., and Jordan, M. A kernelized Stein dis- crepancy for goodness-of-fit tests. In Balcan, M. F. and Weinberger, K. Q. (eds.),Proceedings of The 33rd Inter- national Conference on Machine Learning, volume 48 of Proceedings of Machine ...

work page 2016

[14] [14]

Liu, X., Du, H., Deng, W., and Zhang, R

URL https://openreview.net/forum? id=1k4yZbbDqX. Liu, X., Du, H., Deng, W., and Zhang, R. Optimal stochas- tic trace estimation in generative modeling. In Li, Y ., Mandt, S., Agrawal, S., and Khan, E. (eds.),Proceedings of The 28th International Conference on Artificial Intelli- gence and Statistics, volume 258 ofProceedings of Ma- chine Learning Research...

work page

[15] [15]

URL https://proceedings.mlr.press/ v258/liu25k.html. Lu, C. and Song, Y . Simplifying, stabilizing and scaling continuous-time consistency models. In Yue, Y . (ed.), 13th International Conference on Learning Represen- tations, April 2025. URL https://openreview. net/forum?id=LyJi5ugyJx. Luhman, E. and Luhman, T. Knowledge distillation in iter- ative gener...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[16] [16]

cc/paper_files/paper/2023/file/ f115f619b62833aadc5acb058975b0e6- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2023/file/ f115f619b62833aadc5acb058975b0e6- Paper-Conference.pdf. Maoutsa, D., Reich, S., and Opper, M. Interacting particle solutions of Fokker–Planck equations through gradient- log-density estimation.Entropy, 22(8):802, July 2020. URLhttps://doi.org/10.3390/e22080802. Meyer, R. A., Musco, C., Musco...

work page doi:10.3390/e22080802 2023

[17] [17]

cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740- Paper.pdf. Perez, E., Strub, F., de Vries, H., Dumoulin, V ., and Courville, A. FiLM: visual reasoning with a general conditioning layer. InProceedings of the Thirty-Second AAAI Conference on Artificial In- telligence, AAAI’18/IAAI’18/EAAI’18. AAAI Press,

work page 2019

[18] [18]

v32i1.11671

URL https://doi.org/10.1609/aaai. v32i1.11671. 13 StAD: Stein Amortized Divergence Persson, D., Cortinovis, A., and Kressner, D. Improved variants of the Hutch++ algorithm for trace estimation. SIAM Journal on Matrix Analysis and Applications, 43 (3):1162–1185, 2022. URL https://doi.org/10. 1137/21M1447623. Ranganath, R., Tran, D., Altosaar, J., and Blei,...

work page doi:10.1609/aaai 2022

[19] [19]

cc/paper_files/paper/2016/file/ d947bf06a885db0d477d707121934ff8- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ d947bf06a885db0d477d707121934ff8- Paper.pdf. Raonic, B., Mishra, S., and Lanthaler, S. Towards a cer- tificate of trust: Task-aware OOD detection for scientific AI. In V ondrick, C. (ed.),14th International Conference on Learning Representations, April 2026. URL https: //openreview.net/forum...

work page 2016

[20] [20]

Riabiz, M., Chen, W

URL https://openreview.net/forum? id=FbssShlI4N. Riabiz, M., Chen, W. Y ., Cockayne, J., Swietach, P., Niederer, S. A., Mackey, L., and Oates, C. J. Optimal thinning of MCMC output.Journal of the Royal Sta- tistical Society Series B: Statistical Methodology, 84(4): 1059–1081, April 2022. URL https://doi.org/ 10.1111/rssb.12503. Robbins, H. E. An empirical...

work page doi:10.1111/rssb.12503 2022

[21] [21]

Fundamentals of Enzyme Kinetics: Michaelis-Menten and Non-Michaelis- Type (Atypical) Enzyme Kinetics

URL https://doi.org/10.1007/978-1- 4612-0919-5_26. Ronneberger, O., Fischer, P., and Brox, T. U-Net: Convolu- tional networks for biomedical image segmentation. In Navab, N., Hornegger, J., Wells, W. M., and Frangi, A. F. (eds.),Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, volume 9351 ofLecture Notes in Computer Science, pp. 2...

work page doi:10.1007/978-1- 2015

[22] [22]

cc/paper_files/paper/2025/file/ d79ac139911df25d27f14bbf008deaee- Paper-Conference.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2025/file/ d79ac139911df25d27f14bbf008deaee- Paper-Conference.pdf. Saibaba, A. K., Alexanderian, A., and Ipsen, I. C. Ran- domized matrix-free trace and log-determinant esti- mators.Numerische Mathematik, 137:353–395, Oc- tober 2017. URL https://doi.org/10.1007/ s00211-017-0880-z. Salimans, T. and Ho, ...

work page doi:10.1007/978-3-031-73016-0_6 2025

[23] [23]

URL https://openreview.net/forum? id=WNzy9bRDvG. Song, Y . and Ermon, S. Generative modeling by estimating gradients of the data distribution. In Wallach, H. M., Larochelle, H., Beygelzimer, A., d’Alch´e-Buc, F., Fox, E. B., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 32, pp. 11895–11907. Curran Associates, Inc.,

work page

[24] [24]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93- Paper.pdf. Song, Y . and Ermon, S. Improved techniques for training score-based generative models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 12438–12448. Cur...

work page 2019

[25] [25]

cc/paper_files/paper/2020/file/ 92c3b916311a5517d9290576e3ea37ad- Paper.pdf

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 92c3b916311a5517d9290576e3ea37ad- Paper.pdf. Song, Y ., Durkan, C., Murray, I., and Ermon, S. Max- imum likelihood training of score-based diffusion models. In Ranzato, M., Beygelzimer, A., Dauphin, Y ., Liang, P., and Vaughan, J. W. (eds.),Advances in Neural Information Processing Systems, ...

work page 2020

[26] [26]

A note on the evaluation of generative models

URL https://proceedings.mlr.press/ v202/song23a.html. Stein, C. A bound for the error in the normal approxima- tion to the distribution of a sum of dependent random variables. In Le Cam, L. M., Neyman, J., and Scott, E. L. (eds.),Proceedings of the sixth Berkeley sympo- sium on mathematical statistics and probability, volume 2: Probability theory, volume ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538-4357/ae0936 1972

[27] [27]

Zheng, Q., Le, M., Shaul, N., Lipman, Y ., Grover, A., and Chen, R

URL https://proceedings.mlr.press/ v267/yun25a.html. Zheng, Q., Le, M., Shaul, N., Lipman, Y ., Grover, A., and Chen, R. T. Q. Guided flows for generative modeling and decision making.preprint, November 2023. URL https://arxiv.org/abs/2311.13443. Zhou, M., Zheng, H., Wang, Z., Yin, M., and Huang, H. Score identity distillation: Exponentially fast distilla...

work page arXiv 2023

[28] [28]

Zhou, M., Gu, Y ., Zheng, H., Song, L., He, G., Zhang, Y ., Hu, W., and Yang, Y

URL https://proceedings.mlr.press/ v235/zhou24x.html. Zhou, M., Gu, Y ., Zheng, H., Song, L., He, G., Zhang, Y ., Hu, W., and Yang, Y . Score distillation of flow matching models.preprint, September 2025. URL https:// arxiv.org/abs/2509.25127. A. Stein class conditions The boundary condition for the Stein class (Liu & Wang, 2016; Gorham & Mackey, 2017; Ba...

work page arXiv 2025

[29] [29]

Of these formulations, the TrigFlow continuous- time consistency model (Lu & Song, 2025) is a particularly natural fit forStAD

and mean flow (Geng et al., 2025), that were devel- oped for faster sampling under diffusion and flow-based models. Of these formulations, the TrigFlow continuous- time consistency model (Lu & Song, 2025) is a particularly natural fit forStAD. 18 StAD: Stein Amortized Divergence Lu & Song (2025) introduce TrigFlow, that optimises the objective L(θ) =E t E...

work page 2025