arxiv: 2605.10727 · v1 · submitted 2026-05-11 · 💻 cs.LG · math.DG

Recognition: 2 theorem links

· Lean Theorem

Kernel-Gradient Drifting Models

Maria Esteban-Casadevall , Jorge Carrasco-Pollo , Max Welling , Jan-Willem van de Meent , Erik J. Bekkers , Floor Eijkelboom

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:58 UTC · model grok-4.3

classification 💻 cs.LG math.DG

keywords kernel-gradient driftingone-step generative modelingscore matchingcharacteristic kernelsRiemannian manifoldsFisher-Rao geometrysmoothed KL descentgenerative models

0 comments

The pith

Kernel gradients turn the drift in one-step generative models into the score difference between smoothed data and model distributions.

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

The paper introduces kernel-gradient drifting to replace fixed Euclidean displacement directions with directions induced by the kernel in one-step generative models. This reformulation shows that the resulting drift equals the score difference between a kernel-smoothed version of the data distribution and the model distribution. A sympathetic reader would care because the change yields identifiability when the kernel is characteristic and supplies a smoothed-KL descent view of the dynamics. The same construction works directly on Riemannian manifolds and on discrete data through the Fisher-Rao geometry of the probability simplex, enabling one-step sampling on spherical, sequence, and molecular data without distillation from larger models.

Core claim

By defining the drift direction through the gradient of the kernel with respect to the model points, the generative dynamics become the difference between the score of the kernel-smoothed empirical distribution and the score of the model distribution. For characteristic kernels this difference is identifiable, so the model distribution is uniquely recovered. Because the kernel gradient lies in the intrinsic tangent space, the same drift applies without modification on Riemannian manifolds and on the probability simplex under the Fisher-Rao metric, producing state-of-the-art one-step samples for geospatial, DNA, and molecular data.

What carries the argument

the kernel-gradient drift, obtained as the gradient of the kernel evaluated at model points, which equals the score difference between kernel-smoothed data and model distributions and supplies an intrinsic tangent vector on the target space.

If this is right

The drifting dynamics are identifiable for any characteristic kernel.
The process performs descent on a smoothed version of the KL divergence.
The construction extends unchanged to Riemannian manifolds and to the probability simplex.
One-step generation reaches state-of-the-art quality on spherical geospatial data, promoter DNA, and molecules without distillation.
Kernel gradients keep the drift inside the geometry of the data space.

Where Pith is reading between the lines

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

Different families of kernels could be chosen to match the geometry or topology of a given data domain and potentially accelerate convergence.
The score-difference view may allow direct transfer of theoretical tools from kernel methods into the analysis of drifting models.
One could test whether the same kernel-gradient construction improves sample quality when the target data live on other discrete structures such as graphs or trees.
If the smoothed-KL interpretation holds, the framework might be combined with existing variational bounds to obtain new training objectives for one-step models.

Load-bearing premise

Kernel gradients must supply well-defined intrinsic tangent vectors on the data space and the score difference must stay identifiable and stable for the chosen kernels and distributions without further regularization.

What would settle it

Train the drifting model with a non-characteristic kernel on a low-dimensional distribution that admits multiple distinct densities; if optimization yields multiple different model distributions that produce identical drifts, identifiability fails.

Figures

Figures reproduced from arXiv: 2605.10727 by Erik J. Bekkers, Floor Eijkelboom, Jan-Willem van de Meent, Jorge Carrasco-Pollo, Maria Esteban-Casadevall, Max Welling.

**Figure 1.** Figure 1: Riemannian kernel-gradient drifting on S 2 . Samples are transported along tangent vectors and mapped back to the sphere with the exponential map, allowing drifting to respect the geometry of the data support. 2 Drifting and the Gaussian limitation Drifting models. Drifting models Deng et al. [2026] train one-step generators by moving the iterative refinement usually performed at sampling time into the tra… view at source ↗

**Figure 2.** Figure 2: Attraction drift direction induced by the Laplacian kernel. Under the original formulation, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Generated distributions for the swissroll dataset on spherical and hyperboloid manifolds. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Classifier Two-Sample Test (C2ST) accuracy on the Euclidean Swiss-roll generation task for different Matérn smoothness values ν. Lower is better, with chance performance indicated at 0.5. To assess the effect of kernel choice, we evaluate Euclidean swissroll generation with Matérn kernels of varying smoothness ν. Results are shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Final generated distributions for the checkerboard and swissroll targets on the sphere [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗

**Figure 6.** Figure 6: Evolution of the generated distribution during training for the checkerboard target on the [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗

**Figure 7.** Figure 7: Evolution of the generated distribution during training for the checkerboard target on the [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

**Figure 8.** Figure 8: Evolution of the generated distribution during training for the swissroll target on the [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

**Figure 9.** Figure 9: Evolution of the generated distribution during training for the swissroll target on the [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

**Figure 10.** Figure 10: Evolution of the ’cat’ distribution on the probabiblity simplex during training. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗

read the original abstract

We propose kernel-gradient drifting, a one-step generative modeling framework that replaces the fixed Euclidean displacement direction in drifting models with directions induced by the kernel itself. Standard drifting is attractive because it enables fast, high-quality generation without distilling a large pretrained diffusion model, but its theory is currently understood mainly for Gaussian kernels, where the drift coincides with smoothed score matching and is identifiable. Our gradient-based reformulation exposes this score-based structure for general kernels: the resulting drift is the score difference between kernel-smoothed data and model distributions, yielding identifiability for characteristic kernels and a smoothed-KL descent interpretation of the drifting dynamics. Since kernel gradients are intrinsic tangent vectors, the same construction extends naturally to Riemannian manifolds and to discrete data via the Fisher-Rao geometry of the probability simplex. Across spherical geospatial data, promoter DNA and molecule generation, kernel-gradient drifting enables state-of-the-art one-step generation beyond the Euclidean setting without distillation.

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

This generalizes drifting models to arbitrary kernels and non-Euclidean spaces by making the drift direction come from the kernel gradient itself, which recovers a score difference and enables one-step generation on spheres and simplices.

read the letter

The one thing to take away is that this paper generalizes drifting models to arbitrary kernels and non-Euclidean domains by replacing the usual displacement with directions from the kernel gradient itself. That move exposes a score difference structure and opens the door to one-step generation on spheres and probability simplices. What is new here is the explicit reformulation that recovers the score-matching property for general kernels, not just Gaussians. The drift becomes the score of the smoothed data minus the score of the model, which gives identifiability for characteristic kernels and frames the process as descent on a smoothed KL. They then carry this over to Riemannian manifolds using the kernel's intrinsic gradients and to discrete data on the simplex via Fisher-Rao geometry. That is a clean step beyond earlier drifting work. The paper does well on the application side. Experiments cover spherical geospatial data, promoter DNA, and molecule generation, where it reaches state-of-the-art one-step performance without distilling from a big diffusion model. That practical angle matters for domains where Euclidean assumptions break. The soft spots are around the theoretical claims. The abstract states the identifiability and smoothed-KL descent but does not include the derivations or error analysis. If the full paper supplies those without extra unstated assumptions on kernel positive-definiteness or data support, the argument holds; otherwise the soundness drops. The experiments would benefit from more controls on kernel selection and comparisons to other manifold-aware methods to confirm the gains come from the gradient reformulation. This work is for people building generative models on structured data such as molecules, sequences, or geospatial points. Readers who follow score-based generative modeling or manifold learning will find the connection useful. It has a clear idea with supporting experiments, so it deserves a serious referee who can check the derivations and push on the empirical setup. I would send it for peer review.

Referee Report

0 major / 3 minor

Summary. The paper introduces kernel-gradient drifting, a one-step generative modeling method that replaces Euclidean displacement in drifting models with directions induced by general kernels. The central reformulation shows that the resulting drift equals the score difference between kernel-smoothed data and model distributions, which yields identifiability when the kernel is characteristic and admits a smoothed-KL descent interpretation of the dynamics. The construction uses intrinsic tangent vectors from kernel gradients, allowing direct extension to Riemannian manifolds and to discrete data on the probability simplex via Fisher-Rao geometry. Experiments on spherical geospatial data, promoter DNA sequences, and molecular generation report state-of-the-art one-step performance without distillation.

Significance. If the derivations hold, the work provides a principled generalization of drifting models beyond Gaussian kernels and Euclidean space. The explicit score-difference and smoothed-KL connections, together with the intrinsic-geometry extension, could enable efficient, theoretically grounded one-step generation on structured domains such as molecules and sequences. The reported empirical results on three non-Euclidean tasks supply falsifiable evidence that the predicted behavior is realized in practice.

minor comments (3)

[§3.2] §3.2: the statement that kernel gradients are 'intrinsic tangent vectors' on manifolds and the simplex is asserted without an explicit local-coordinate verification or reference to the relevant differential-geometry lemma; a short paragraph or appendix derivation would remove ambiguity.
[Table 2, Figure 4] Table 2 and Figure 4: the one-step generation metrics are compared against several baselines, but the table does not report standard deviations over the multiple random seeds mentioned in the experimental protocol; adding error bars would strengthen the state-of-the-art claim.
[Abstract, §1] Abstract and §1: the phrase 'smoothed-KL descent interpretation' is used before the quantity is defined; a one-sentence gloss in the abstract would improve readability for readers outside the immediate subfield.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of our work on kernel-gradient drifting models, including the recognition of the score-difference reformulation, identifiability results for characteristic kernels, smoothed-KL descent interpretation, and extensions to Riemannian and discrete domains. We appreciate the recommendation for minor revision.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents a gradient-based reformulation of drifting models that derives the drift term as the score difference between kernel-smoothed data and model distributions for general kernels. This follows directly from the kernel gradient construction and characteristic kernel properties, without reducing to a fitted parameter renamed as a prediction or a self-citation chain. The identifiability and smoothed-KL descent claims are presented as consequences of the reformulation rather than inputs, and the extension to Riemannian manifolds and the probability simplex is justified geometrically via intrinsic tangent vectors. No load-bearing equation or premise in the provided text reduces by construction to its own inputs or prior self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard kernel theory (characteristic kernels induce identifiability) and the assumption that kernel gradients act as intrinsic vectors on the data manifold or simplex; no free parameters or new invented entities are introduced in the abstract.

axioms (2)

domain assumption Characteristic kernels ensure identifiability of the drift as score difference
Stated directly in the abstract as yielding identifiability for characteristic kernels.
domain assumption Kernel gradients are intrinsic tangent vectors on Riemannian manifolds and the probability simplex
Invoked to justify extension beyond Euclidean space.

pith-pipeline@v0.9.0 · 5476 in / 1360 out tokens · 51753 ms · 2026-05-12T04:58:46.281788+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 (J uniquely satisfies the functional equation) unclear

?

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

the resulting drift is the score difference between kernel-smoothed data and model distributions... V^∇_{p,qθ}(x) = ∇_x log ˆp_k(x)/ˆq_k(x)
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking (D=3 forcing) unclear

?

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

Riemannian kernel-gradient drift... Fisher-Rao geometry of the probability simplex

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

71 extracted references · 71 canonical work pages

[1]

2026 , journal =

A Unified View of Drifting and Score-Based Models , author=. 2026 , journal =

work page 2026
[2]

Generative drifting is secretly score matching: a spectral and variational perspective.arXiv preprint arXiv:2603.09936, 2026

Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective , author=. arXiv preprint arXiv:2603.09936 , year=

work page arXiv
[3]

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

Generative Moment Matching Networks , author =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , editor =

work page 2015
[4]

2026 , journal =

Generative Modeling via Drifting , author=. 2026 , journal =

work page 2026
[5]

2015 , eprint=

Deep Unsupervised Learning using Nonequilibrium Thermodynamics , author=. 2015 , eprint=

work page 2015
[6]

The Eleventh International Conference on Learning Representations , year=

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. The Eleventh International Conference on Learning Representations , year=

work page
[7]

The Eleventh International Conference on Learning Representations , year=

Building Normalizing Flows with Stochastic Interpolants , author=. The Eleventh International Conference on Learning Representations , year=

work page
[8]

The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

Fisher Flow Matching for Generative Modeling over Discrete Data , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

work page
[9]

Variational flow matching for graph generation

Eijkelboom, Floor and Bartosh, Grigory and Naesseth, Christian and van de Meent, Jan-Willem and Welling, Max. Variational flow matching for graph generation. Advances in Neural Information Processing Systems 37

work page
[10]

Amortised

S. Amortised. Proceedings of the International Conference on Learning Representations , year =

work page
[11]

2025 , eprint=

Flow map matching with stochastic interpolants: A mathematical framework for consistency models , author=. 2025 , eprint=

work page 2025
[12]

2025 , journal =

Generalised Flow Maps for Few-Step Generative Modelling on Riemannian Manifolds , author=. 2025 , journal =

work page 2025
[13]

2024 , eprint=

Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion , author=. 2024 , eprint=

work page 2024
[14]

Conference on Computer Vision and Pattern Recognition (CVPR) , year=

One-Step Diffusion with Distribution Matching Distillation , author=. Conference on Computer Vision and Pattern Recognition (CVPR) , year=

work page
[15]

2025 , journal =

Mean Flows for One-step Generative Modeling , author=. 2025 , journal =

work page 2025
[16]

2026 , journal =

Riemannian MeanFlow , author=. 2026 , journal =

work page 2026
[17]

Geodesic exponential kernels: When curvature and linearity conflict , booktitle =

Feragen, Aasa and Lauze, Francois and Hauberg, Soren , year =. Geodesic exponential kernels: When curvature and linearity conflict , booktitle =

work page
[18]

The Geometric Kernels Package: Heat and Mat

Peter Mostowsky and Vincent Dutordoir and Iskander Azangulov and No. The Geometric Kernels Package: Heat and Mat. Journal of Machine Learning Research , year =

work page
[19]

2026 , journal =

Categorical Flow Maps , author=. 2026 , journal =

work page 2026
[20]

Riemannian Continuous Normalizing Flows , year =

Mathieu, Emile and Nickel, Maximilian , booktitle =. Riemannian Continuous Normalizing Flows , year =

work page
[21]

arXiv preprint arXiv:2510.03511 , year=

Platonic Transformers: A Solid Choice for Equivariance , author=. arXiv preprint arXiv:2510.03511 , year=

work page arXiv
[22]

Advances in Neural Information Processing Systems , volume=

Equivariant Flow Matching , author=. Advances in Neural Information Processing Systems , volume=

work page
[23]

Proceedings of the 37th International Conference on Machine Learning , series=

Equivariant Flows: Exact Likelihood Generative Learning for Symmetric Densities , author=. Proceedings of the 37th International Conference on Machine Learning , series=. 2020 , publisher=

work page 2020
[24]

arXiv preprint arXiv:2506.18340 , year=

Controlled Generation with Equivariant Variational Flow Matching , author=. arXiv preprint arXiv:2506.18340 , year=

work page arXiv
[25]

Introduction to Smooth Manifolds , author =

work page
[26]

Statistics & Probability Letters , volume =

Bruno Pelletier , title =. Statistics & Probability Letters , volume =

work page
[27]

Submanifold density estimation , year =

Ozakin, Arkadas and Gray, Alexander , booktitle =. Submanifold density estimation , year =

work page
[28]

arXiv preprint arXiv:2601.16777 , year =

Kernel smoothing on manifolds , author =. arXiv preprint arXiv:2601.16777 , year =

work page arXiv
[29]

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

Which Training Methods for GANs do Actually Converge? , author =. Proceedings of the 35th International Conference on Machine Learning , series =

work page
[30]

Cleanthous, Galatia and Georgiadis, Athanasios G and Kerkyacharian, Gerard and Petrushev, Pencho and Picard, Dominique , JOURNAL =

work page
[31]

Semi-intrinsic Mean Shift on Riemannian Manifolds , volume =

Caseiro, Rui and Henriques, Joao and Martins, Pedro and Batista, Jorge , year =. Semi-intrinsic Mean Shift on Riemannian Manifolds , volume =

work page
[32]

Geometric structures arising from kernel density estimation on Riemannian manifolds , volume =

Kim, Yoon Tae and Park, Hyun Suk , year =. Geometric structures arising from kernel density estimation on Riemannian manifolds , volume =. Journal of Multivariate Analysis , doi =

work page
[33]

Strong uniform consistency with rates for kernel density estimators with general kernels on manifolds , volume =

Wu, Hau-Tieng and Wu, Nan , year =. Strong uniform consistency with rates for kernel density estimators with general kernels on manifolds , volume =

work page
[34]

Borovitskiy, Viacheslav and Terenin, Alexander and Mostowsky, Peter and Deisenroth, Marc , booktitle =. Mat\'

work page
[35]

Kernel Methods on Riemannian Manifolds with Gaussian RBF Kernels , volume =

Jayasumana, Sadeep and Hartley, Richard and Salzmann, Mathieu and li, Hongdong and Harandi, Mehrtash , year =. Kernel Methods on Riemannian Manifolds with Gaussian RBF Kernels , volume =

work page
[36]

arXiv preprint arXiv:2506.19245 , year=

Universal kernels via harmonic analysis on Riemannian symmetric spaces , author=. arXiv preprint arXiv:2506.19245 , year=

work page arXiv
[37]

Transactions on Machine Learning Research , year=

The Score-Difference Flow for Implicit Generative Modeling , author=. Transactions on Machine Learning Research , year=

work page
[38]

Advances in Neural Information Processing Systems , year=

Measuring Sample Quality with Stein's Method , author=. Advances in Neural Information Processing Systems , year=

work page
[39]

Characteristic Kernels on Groups and Semigroups , volume =

Fukumizu, Kenji and Gretton, Arthur and Sch\". Characteristic Kernels on Groups and Semigroups , volume =. Advances in Neural Information Processing Systems , publisher =

work page
[40]

2023 , publisher =

Avdeyev, Pavel and Shi, Chenlai and Tan, Yuhao and Dudnyk, Kseniia and Zhou, Jian , booktitle =. 2023 , publisher =

work page 2023
[41]

2024 , eprint=

Dirichlet Flow Matching with Applications to DNA Sequence Design , author=. 2024 , eprint=

work page 2024
[42]

2026 , journal =

Gradient Flow Drifting: Generative Modeling via Wasserstein Gradient Flows of KDE-Approximated Divergences , author=. 2026 , journal =

work page 2026
[43]

2026 , journal =

Drifting Fields are not Conservative , author=. 2026 , journal =

work page 2026
[44]

Transactions on Machine Learning Research , year=

Flow map matching with stochastic interpolants: A mathematical framework for consistency models , author=. Transactions on Machine Learning Research , year=

work page
[45]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

How to build a consistency model: Learning flow maps via self-distillation , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

work page
[46]

Parallelbench: Understanding the trade-offs of parallel decoding in diffusion llms.arXiv preprint arXiv:2510.04767, 2025

Kang, Woosuk and Galim, Kamil and Oh, Sehoon and Lee, Minjae and Zeng, Yuchen and Zhang, Shijie and Hooper, Coleman and Hu, Yibo and Koo, Hyeon Ik and Cho, Nam Ik and Lee, Kangwook , title =. arXiv preprint arXiv:2510.04767 , year =. 2510.04767 , archivePrefix =

work page arXiv
[47]

arXiv preprint arXiv:2410.08709 , year =

Hayakawa, Satoshi and Takida, Yuhta and Imaizumi, Masaaki and Wakaki, Hiromichi and Mitsufuji, Yuki , title =. arXiv preprint arXiv:2410.08709 , year =. 2410.08709 , archivePrefix =

work page arXiv
[48]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

Mean Flows for One-step Generative Modeling , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

work page
[49]

and He, Kaiming , year =

Geng, Zhengyang and Lu, Yiyang and Wu, Zongze and Shechtman, Eli and Kolter, J. and He, Kaiming , year =. Improved Mean Flows: On the Challenges of Fastforward Generative Models , journal =

work page
[50]

The Fourteenth International Conference on Learning Representations , year=

Terminal Velocity Matching , author=. The Fourteenth International Conference on Learning Representations , year=

work page
[51]

The Eleventh International Conference on Learning Representations , year=

Flow Matching for Generative Modeling , author=. The Eleventh International Conference on Learning Representations , year=

work page
[52]

Denoising Diffusion Probabilistic Models , volume =

Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , booktitle =. Denoising Diffusion Probabilistic Models , volume =

work page
[53]

International Conference on Learning Representations , year=

Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

work page
[54]

International Conference on Learning Representations , year=

Progressive Distillation for Fast Sampling of Diffusion Models , author=. International Conference on Learning Representations , year=

work page
[55]

Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models , journal =

Hu, Tianyang and Li, Zhenguo and Luo, Weijian and Sun, Jiacheng and Zhang, Zhihua and Zhang, Shifeng , year =. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models , journal =

work page
[56]

Chen, Ricky T. Q. and Lipman, Yaron , booktitle =. Flow Matching on General Geometries , year =

work page
[57]

The Fourteenth International Conference on Learning Representations , year=

Riemannian Variational Flow Matching for Material and Protein Design , author=. The Fourteenth International Conference on Learning Representations , year=

work page
[58]

and Ho, Jonathan and Tarlow, Daniel and van den Berg, Rianne , booktitle =

Austin, Jacob and Johnson, Daniel D. and Ho, Jonathan and Tarlow, Daniel and van den Berg, Rianne , booktitle =. Structured Denoising Diffusion Models in Discrete State-Spaces , year =

work page
[59]

The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

Simple and Effective Masked Diffusion Language Models , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

work page
[60]

Simplified and Generalized Masked Diffusion for Discrete Data , year =

Shi, Jiaxin and Han, Kehang and Wang, Zhe and Doucet, Arnaud and Titsias, Michalis , booktitle =. Simplified and Generalized Masked Diffusion for Discrete Data , year =

work page
[61]

A Continuous Time Framework for Discrete Denoising Models , year =

Campbell, Andrew and Benton, Joe and De Bortoli, Valentin and Rainforth, Thomas and Deligiannidis, George and Doucet, Arnaud , booktitle =. A Continuous Time Framework for Discrete Denoising Models , year =

work page
[62]

The Eleventh International Conference on Learning Representations , year=

Score-based Continuous-time Discrete Diffusion Models , author=. The Eleventh International Conference on Learning Representations , year=

work page
[63]

2022 , journal =

Continuous diffusion for categorical data , author=. 2022 , journal =

work page 2022
[64]

The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

Categorical Flow Matching on Statistical Manifolds , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

work page
[65]

International Journal of Computer Vision , year =

Subbarao, Raghav and Meer, Peter , title =. International Journal of Computer Vision , year =

work page
[66]

, title =

Lee, John M. , title =

work page
[67]

The Fourteenth International Conference on Learning Representations , year=

PairFlow: Closed-Form Source-Target Coupling for Few-Step Generation in Discrete Flow Models , author=. The Fourteenth International Conference on Learning Representations , year=

work page
[68]

Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , year =

Zang, Chengxi and Wang, Fei , title =. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , year =

work page
[69]

International Conference on Learning Representations , year=

Top-N: Equivariant Set and Graph Generation without Exchangeability , author=. International Conference on Learning Representations , year=

work page
[70]

and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S

Wu, Zhenqin and Ramsundar, Bharath and Feinberg, Evan N. and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S. and Leswing, Karl and Pande, Vijay S. , title =. Chemical Science , volume =. 2018 , publisher =

work page 2018
[71]

The Thirteenth International Conference on Learning Representations , year=

Simple Guidance Mechanisms for Discrete Diffusion Models , author=. The Thirteenth International Conference on Learning Representations , year=

work page