Particle GFlowNets: Rethinking Generative Marginalization Models

Tiago da Silva, Diego Mesquita, Salem Lahlou
Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337:6366-6383, 2026.

Abstract

Generative Marginalization Models (MaMs) have been recently introduced as efficient neural sampling models for any-order autoregressive modelling of discrete distributions. By learning both the marginal and conditional probabilities of a persistent-block {Gibbs} sampler, MaMs enable fast posterior evaluation with a single neural network forward pass. While prior work has considered MaMs to be distinct from Generative Flow Networks ({GFlowNets}), a well-established paradigm for inference in discrete stochastic models, we show that they are equivalent. Then, we also extend MaMs’ sampling strategy to non-autoregressive generative processes. In particular, we describe an automatic criterion for full-state rejuvenation of the {Gibbs} sampler, derived from the Gelman-Rubin statistic, which plays a key role in speeding up learning convergence. Our experiments show that our method, called Particle {GFlowNets}, markedly accelerates training in large combinatorial spaces.

Cite this Paper

BibTeX
@InProceedings{pmlr-v337-silva26a, title = {Particle {GFlowNets}: Rethinking Generative Marginalization Models}, author = {da Silva, Tiago and Mesquita, Diego and Lahlou, Salem}, booktitle = {Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence}, pages = {6366--6383}, year = {2026}, editor = {Perković, Emilija and Malinsky, Daniel}, volume = {337}, series = {Proceedings of Machine Learning Research}, month = {17--21 Aug}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v337/main/assets/silva26a/silva26a.pdf}, url = {https://proceedings.mlr.press/v337/silva26a.html}, abstract = {Generative Marginalization Models (MaMs) have been recently introduced as efficient neural sampling models for any-order autoregressive modelling of discrete distributions. By learning both the marginal and conditional probabilities of a persistent-block {Gibbs} sampler, MaMs enable fast posterior evaluation with a single neural network forward pass. While prior work has considered MaMs to be distinct from Generative Flow Networks ({GFlowNets}), a well-established paradigm for inference in discrete stochastic models, we show that they are equivalent. Then, we also extend MaMs’ sampling strategy to non-autoregressive generative processes. In particular, we describe an automatic criterion for full-state rejuvenation of the {Gibbs} sampler, derived from the Gelman-Rubin statistic, which plays a key role in speeding up learning convergence. Our experiments show that our method, called Particle {GFlowNets}, markedly accelerates training in large combinatorial spaces.} }
Endnote
%0 Conference Paper %T Particle GFlowNets: Rethinking Generative Marginalization Models %A Tiago da Silva %A Diego Mesquita %A Salem Lahlou %B Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Emilija Perković %E Daniel Malinsky %F pmlr-v337-silva26a %I PMLR %P 6366--6383 %U https://proceedings.mlr.press/v337/silva26a.html %V 337 %X Generative Marginalization Models (MaMs) have been recently introduced as efficient neural sampling models for any-order autoregressive modelling of discrete distributions. By learning both the marginal and conditional probabilities of a persistent-block {Gibbs} sampler, MaMs enable fast posterior evaluation with a single neural network forward pass. While prior work has considered MaMs to be distinct from Generative Flow Networks ({GFlowNets}), a well-established paradigm for inference in discrete stochastic models, we show that they are equivalent. Then, we also extend MaMs’ sampling strategy to non-autoregressive generative processes. In particular, we describe an automatic criterion for full-state rejuvenation of the {Gibbs} sampler, derived from the Gelman-Rubin statistic, which plays a key role in speeding up learning convergence. Our experiments show that our method, called Particle {GFlowNets}, markedly accelerates training in large combinatorial spaces.
APA
da Silva, T., Mesquita, D. & Lahlou, S.. (2026). Particle GFlowNets: Rethinking Generative Marginalization Models. Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, in Proceedings of Machine Learning Research 337:6366-6383 Available from https://proceedings.mlr.press/v337/silva26a.html.

Related Material