Bayesian Generative Modeling (BGM)

• Bayesian Generative Modeling (BGM) is a probabilistic framework that leverages explicit priors, latent variables, and posterior updates to manage uncertainty in model inference.
• It is applied in diverse domains such as inverse problems, causal inference, and structure learning, enhancing model reliability through structured latent spaces.
• BGM emphasizes generating distributions over hypotheses rather than single deterministic outputs, thereby supporting robust statistical estimation and prediction.

Bayesian Generative Modeling (BGM) denotes a class of probabilistic modeling strategies in which generation, inference, and uncertainty quantification are organized around explicit priors, latent variables, and posterior updating. Across the cited literature, BGM appears in several closely related forms: latent-variable priors for inverse problems, posterior samplers over structured latent objects such as directed acyclic graphs (DAGs), Bayesianized deep generators with uncertainty over weights, belief-state generators based on Bayesian updates, and simulator-based models whose parameters are inferred from observed summaries (Marschall et al., 2022, Deleu et al., 2022, Graves et al., 2023, Lienen et al., 11 Feb 2025, Mor et al., 2018). A common thread is the use of a probabilistic generative mechanism—sometimes over observations, sometimes over latent hypotheses themselves—together with posterior inference that preserves uncertainty rather than collapsing to a single deterministic estimate.

1. Conceptual scope

In the surveyed work, BGM is not a single model family but a modeling stance. One recurring formulation uses a latent variable $z$ with prior $\pi(z)$ and a conditional model $x \mid z$, so that the induced prior on the ambient variable is

$\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$

This is explicit in probabilistic generative priors for inverse problems, where $x\mid z \sim \mathcal N(g(z),\Gamma(z))$ defines a proper prior on the original high-dimensional space rather than constraining inference to the exact image of a deterministic decoder (Marschall et al., 2022). A second formulation treats the unknown object itself as structured and discrete—for example a Bayesian-network DAG $G$—and defines a posterior

$$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$

with a generative model trained to sample graphs from that posterior approximation (Deleu et al., 2022). A third formulation makes the evolving state of the generator a posterior belief rather than a noisy sample, as in Bayesian Flow Networks (BFNs) and Bayesian Sample Inference (BSI), where Bayesian updates are the core state transition (Graves et al., 2023, Lienen et al., 11 Feb 2025).

The literature also distinguishes strict Bayesian modeling from looser probabilistic or Bayesian-network-inspired usage. "Generative Model for Heterogeneous Inference" targets Bayesian-network-style conditional inference, but the work is explicit that it is not Bayesian in the modern sense of placing priors over parameters and doing posterior parameter inference; its Bayesian content lies primarily in probabilistic dependency modeling over multivariate random variables (Zhou et al., 2018). This distinction matters because BGM, in the stronger sense used by the other papers, requires an explicit probabilistic generative specification together with posterior or posterior-like inference.

2. Probabilistic foundations

A unifying Bayesian template is

$p(\theta\mid y)\propto p(\theta)\,p(y\mid \theta),$

but the surveyed papers instantiate $\theta$ very differently. In inverse problems, $\theta$ may be the ambient unknown $\pi(z)$0, yielding

$\pi(z)$1

with $\pi(z)$2 induced by a probabilistic generator and sometimes approximated by a Gaussian Laplace prior in the original space (Marschall et al., 2022). In Bayesian image reconstruction, the unknown is a latent StyleGAN2 code $\pi(z)$3, and the posterior takes the form

$\pi(z)$4

where the likelihood combines pixel-space and perceptual-space Gaussian terms after a known corruption operator $\pi(z)$5 is applied to the generated image $\pi(z)$6 (Marinescu et al., 2020).

In structure learning, the posterior target is a law over latent graph hypotheses rather than over parameters of a continuous decoder. DAG-GFlowNet chooses reward

$\pi(z)$7

so that terminal-state probabilities proportional to $\pi(z)$8 approximate $\pi(z)$9. Under modularity assumptions on the parameter prior and structure prior,

$x \mid z$0

which makes posterior inference computationally tractable at the level of local edge additions (Deleu et al., 2022).

BFNs and BSI shift the object of inference once more. In BFNs, the generative state is the parameter vector $x \mid z$1 of a factorized input distribution

$x \mid z$2

updated by Bayes’ rule in light of noisy messages, while a neural network predicts an interdependent output distribution $x \mid z$3 (Graves et al., 2023). In BSI, the sample $x \mid z$4 itself is treated as the unknown random variable, with Gaussian belief states

$x \mid z$5

and posterior updates

$x \mid z$6

where the “measurements” $x \mid z$7 are generated around a learned prediction of the hidden sample (Lienen et al., 11 Feb 2025).

A central conceptual distinction in this literature is between direct data generation and generation over posterior hypotheses. DAG-GFlowNet is explicit that the GFlowNet does not directly model $x \mid z$8 as a latent-variable generator would; instead, data enter through the reward $x \mid z$9, so the model is generative over posterior graph hypotheses, not over raw observations (Deleu et al., 2022). This suggests that BGM includes both conventional generative models over data and generative samplers over posterior objects.

3. Latent structure and prior construction

BGM places unusual emphasis on the structure of latent space. In causal inference with high-dimensional covariates, CausalBGM uses a partition

$\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$0

where $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$1 affects both treatment and outcome, $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$2 affects only outcome, $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$3 affects only treatment, and $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$4 affects neither treatment nor outcome but helps explain covariates (Liu et al., 1 Jan 2025). BGM-IV adopts the same decomposition for nonlinear instrumental-variable regression, interpreting $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$5 as shared confounding structure, $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$6 as outcome-specific variation, $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$7 as treatment-specific variation, and $\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$8 as covariate-only nuisance information (Luo et al., 7 May 2026). In both cases, the latent space is organized by causal role rather than by purely compressive criteria.

In physically structured settings, prior construction is equally explicit. The CO$\pi(x)=\int \pi(x\mid z)\pi(z)\,dz.$9 monitoring framework first generates prior plume realizations by geostatistical geology plus multiphase flow simulation, then trains a VAE so that low-dimensional latent codes $x\mid z \sim \mathcal N(g(z),\Gamma(z))$0 are approximately Gaussian while decoded samples retain plume geometry. The inferred unknown is therefore a latent code whose pushforward through the decoder yields a saturation field, which is then mapped to seismic velocity through Brie fluid mixing, Gassmann substitution, and acoustic wave propagation (Li et al., 13 Dec 2025). Sparse Bayesian channel modeling uses a latent coefficient vector $x\mid z \sim \mathcal N(g(z),\Gamma(z))$1 in a steering-vector dictionary $x\mid z \sim \mathcal N(g(z),\Gamma(z))$2, with

$x\mid z \sim \mathcal N(g(z),\Gamma(z))$3

so sparsity and physical interpretability arise simultaneously from ARD-style diagonal covariance structure (Böck et al., 25 Feb 2025).

Learned priors also appear in more conventional deep generative settings. BRGM uses a pretrained StyleGAN2 generator as an image prior, with a structured prior over the layerwise latent vectors $x\mid z \sim \mathcal N(g(z),\Gamma(z))$4 and an additional von Mises colinearity term between layers (Marinescu et al., 2020). The inverse-problem paper on Laplace approximation argues that using a deterministic generator as a hard manifold prior yields a push-forward posterior with no Lebesgue density in ambient space, whereas the probabilistic model

$x\mid z \sim \mathcal N(g(z),\Gamma(z))$5

induces a proper ambient-space density and supports consistency results (Marschall et al., 2022). This is one of the clearest statements in the surveyed literature that a learned generator should often be treated as a probabilistic prior rather than an exact support restriction.

4. Inference and training mechanisms

The inference mechanisms used in BGM are diverse because the posterior geometries and model objects differ substantially.

Before detailing them, the following taxonomy is useful.

Family	Latent object	Inference or training mechanism
Learned ambient-space priors	$x\mid z \sim \mathcal N(g(z),\Gamma(z))$6, $x\mid z \sim \mathcal N(g(z),\Gamma(z))$7	Laplace approximation, closed-form Gaussian posterior, HMC
Structured posterior generators	$x\mid z \sim \mathcal N(g(z),\Gamma(z))$8, $x\mid z \sim \mathcal N(g(z),\Gamma(z))$9	GFlowNet detailed balance, Bayesian updates, ELBO-like losses
Causal latent models	$G$0, network weights	VI for weights, SGD/MAP or MCMC for local latents
Simulator-based models	global parameters $G$1	ABC / SMC-ABC

In learned-prior inverse problems, the Laplace-approximation paper linearizes $G$2 around an expansion point $G$3 and freezes $G$4, producing the Gaussian prior

$G$5

which then yields a closed-form Gaussian posterior in the linear-Gaussian observation model (Marschall et al., 2022). The CO$G$6 monitoring work instead samples the latent posterior with HMC in the VAE latent space, exploiting its approximately Gaussian geometry (Li et al., 13 Dec 2025). BRGM uses both MAP optimization and variational inference in latent space, minimizing

$G$7

with reparameterized Gaussian variational samples (Marinescu et al., 2020).

For structured discrete posteriors, DAG-GFlowNet trains a forward policy $G$8 under a detailed-balance consistency constraint: $G$9 The associated loss is an expectation of squared log-ratios over transitions, estimated off-policy with a replay buffer and target-network stabilization (Deleu et al., 2022). BFNs optimize discrete- or continuous-time losses derived from communication cost and variational lower bounds, while BSI derives an ELBO whose local Gaussian KL terms reduce to precision-weighted posterior mean matching (Graves et al., 2023, Lienen et al., 11 Feb 2025).

CausalBGM uses a hybrid scheme: mean-field Gaussian variational posteriors for Bayesian neural network parameters, reparameterization with Flipout for stochastic gradients, and iterative latent-variable updates for subject-specific $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$0; at test time, local latent posteriors are sampled with random-walk Metropolis–Hastings (Liu et al., 1 Jan 2025). BGM-IV is more explicit that its latent updates are MAP-like rather than amortized variational, and that the outcome term is an IV-integrated pseudo-likelihood rather than an ordinary likelihood (Luo et al., 7 May 2026). General Bayesian quantile regression for counts uses yet another route: a fully Bayesian nonparametric generative model for $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$1, followed by loss-based general Bayesian updating of the quantile-regression functional $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$2 through repeated optimization under posterior draws of latent quantiles (Yamauchi et al., 2024).

At the simulator-based end, BGM FASt uses SMC-ABC because the likelihood is mathematically impossible or computationally prohibitive to evaluate exactly. The posterior is approximated by repeatedly simulating synthetic catalogues under parameter proposals, computing summary statistics, and retaining particles with small Poissonian discrepancy to Tycho-2 (Mor et al., 2018).

5. Representative application domains

The surveyed literature shows that BGM is not tied to one domain. In Bayesian-network structure learning, the latent object is a DAG, and BGM is used to approximate a posterior over graph structures and posterior feature probabilities such as edges, paths, and Markov blankets (Deleu et al., 2022). In inverse problems and imaging, BGM appears as a learned generative prior over ambient variables or latent codes, supporting super-resolution, in-painting, linear-Gaussian inversion, and time-lapse CO$$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$3 monitoring (Marinescu et al., 2020, Marschall et al., 2022, Li et al., 13 Dec 2025). In molecular generation, GeoBFN uses SE(3)-equivariant Bayesian Flow Networks over coordinates, discretized charges, and atom types, moving the generative process into a parameter space of distributions rather than sample space (Song et al., 2024).

Causal modeling provides another major cluster. CausalBGM uses a causally partitioned latent representation to estimate individual treatment effects and average dose-response functions under high-dimensional confounding (Liu et al., 1 Jan 2025). BGM-IV extends the same philosophy to nonlinear IV estimation, replacing the ordinary confounded outcome likelihood with an IV-integrated pseudo-likelihood

$$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$4

thereby embedding endogeneity correction into latent Bayesian generative modeling (Luo et al., 7 May 2026).

Count-data quantile regression uses a latent-threshold generative model,

$$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$5

together with a Pitman–Yor mixture prior over the joint law of $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$6, and then general Bayesian updating for quantile-specific spline effects (Yamauchi et al., 2024). In wireless communications and sensing, sparse Bayesian generative modeling jointly estimates the channel $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$7 and sparse physical coefficients $$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$8 without extra online optimization beyond posterior mean computation (Böck et al., 25 Feb 2025). In astronomy, BGM FASt treats the Milky Way as a stochastic simulator parameterized by IMF, SFH, and density-law parameters, then infers those parameters from catalog-level summaries by ABC (Mor et al., 2018). Trajectory prediction for interacting vehicles uses coordination variables, latent noise, and Bayesian weight uncertainty to generate diversified future hypotheses in conflict scenarios (Li et al., 2019).

These examples suggest that BGM is especially natural when prior structure is rich, geometry matters, and posterior uncertainty is scientifically actionable.

6. Uncertainty, misconceptions, and limitations

A major attraction of BGM is that uncertainty is a first-class output. DAG-GFlowNet estimates posterior feature probabilities from sampled graphs rather than scoring a single structure (Deleu et al., 2022). BRGM samples multiple reconstructions from an approximate posterior over StyleGAN latents (Marinescu et al., 2020). CausalBGM reports posterior intervals for individual treatment effects and dose-response functions (Liu et al., 1 Jan 2025). CO$$P(G\mid \mathcal D)\propto P(G)P(\mathcal D\mid G),$$9 monitoring uses posterior saturation ensembles to identify poorly constrained plume regions and the effect of rock-physics bias (Li et al., 13 Dec 2025). The vehicle-interaction system argues that uncertainty over generator weights yields larger variance and diversity of future trajectories (Li et al., 2019).

Several recurring misconceptions are explicitly challenged in the literature. First, a learned generator is not automatically a proper Bayesian prior on the ambient variable. If inference is restricted to the exact manifold $p(\theta\mid y)\propto p(\theta)\,p(y\mid \theta),$0, the push-forward posterior may have no Lebesgue density in the original space and Bayes estimators may be inconsistent when the truth is off-manifold (Marschall et al., 2022). Second, not every method branded “Bayesian” is Bayesian in the same sense: heterogeneous inference with EAR is best interpreted as a scalable black-box surrogate for Bayesian-network-style conditionals rather than posterior inference over model parameters or structures (Zhou et al., 2018). Third, not every posterior-like object is the posterior of a coherent full likelihood. BGM-IV is explicit that its IV objective produces an IV quasi-posterior, and the count-quantile model is explicit that $p(\theta\mid y)\propto p(\theta)\,p(y\mid \theta),$1 is updated by general Bayesian loss-based optimization rather than a direct likelihood (Luo et al., 7 May 2026, Yamauchi et al., 2024).

Limitations are equally prominent. DAG-GFlowNet becomes harder to train as the posterior sharpens with dataset size, because local score ratios become extreme (Deleu et al., 2022). The Laplace-approximation approach is restricted to linear-Gaussian inverse problems and can be crude when the induced prior is multimodal or strongly curved (Marschall et al., 2022). Mean-field Bayesian GANs may underestimate posterior correlations, and covariance-spectrum diversity is only one notion of diversity (Valizadeh et al., 30 Oct 2025). CausalBGM and BGM-IV depend strongly on latent-structure assumptions and initialization, and neither provides a full convergence theory for its alternating optimization procedure (Liu et al., 1 Jan 2025, Luo et al., 7 May 2026). VAE priors in subsurface monitoring inherit approximation error from the decoder and from fixed rock-physics mappings (Li et al., 13 Dec 2025).

Taken together, the surveyed work suggests that BGM is best understood not as a single algorithm but as a probabilistic program for combining structured priors, generative simulation, and posterior computation. Its most distinctive contribution is to treat uncertainty over latent structure, hypotheses, or scientific state variables as an object to be modeled and generated explicitly, rather than as a residual after point prediction.