Amortized Factor Inference Networks

• AFINs are a design principle where reusable neural inference objects are learned from factorized probabilistic models to produce proposals, variational posteriors, or evidence ratios.
• They employ inverse factorization and typed-factor interfaces to transform generative models into tractable forward passes for Sequential Monte Carlo and simulation-based inference.
• Empirical results demonstrate faster convergence, improved effective sample sizes, and robustness across diverse Bayesian tasks, while noting challenges in model misspecification and high-dimensional settings.

Amortized Factor Inference Networks (AFINs) denote a class of amortized neural inference constructions in which reusable inference objects are learned from a factorized probabilistic specification rather than derived anew for each dataset. The literature suggests that the term is not restricted to a single canonical architecture. In "Inference Networks for Sequential Monte Carlo in Graphical Models" (Paige et al., 2016), AFINs are structured neural recognition models that implement a stochastic inverse of a directed graphical model and serve as proposal distributions for Sequential Monte Carlo (SMC). In "Amortized Factor Inference Networks for Posterior Inference" (Ko et al., 26 May 2026), the name is used for encode–merge–decode networks that accept a typed-factor model specification and observations and return parameters of a variational posterior. Related work explicitly places Bayes-factor estimation with Evidence Networks and local blanket-level inference with $\Delta$-AI within an AFIN-style pattern of amortizing factor-level probabilistic computations (Jeffrey et al., 2023, Falet et al., 2023).

1. Scope and terminology

The principal usages of the term organize around what is being amortized and which probabilistic object is produced. In all cases, the amortized artifact is intended for reuse across many inference instances, but the target ranges from proposal distributions to posterior parameterizations to scalar evidence ratios.

Paper	Amortized object	Primary target
(Paige et al., 2016)	Structured inverse recognition network	$q(x \mid y)$ for SMC proposals
(Ko et al., 26 May 2026)	Encode–merge–decode posterior network	$q_\phi(z)$ from typed factors
(Jeffrey et al., 2023)	Evidence Network as an AFIN instance	Bayes factor or log Bayes factor
(Falet et al., 2023)	$\Delta$-AI as an AFIN-style method	Local conditional or blanket-consistent sampler

In the 2016 formulation, the defining feature is inverse factorization of a known directed graphical model. The forward model

$$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$

is converted into an inverse model whose conditionals are oriented from observations toward latents, and a structured neural density estimator approximates the inverse conditional needed for posterior sampling (Paige et al., 2016).

In the 2026 formulation, the defining feature is instead a typed-factor interface. A model instance is written as

$$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$

and a single trained network maps the specification $(t,\theta)$ together with observations to the parameters of a variational posterior, while remaining invariant to varying numbers of observations and equivariant in latent coordinates (Ko et al., 26 May 2026).

This multiplicity of usage makes one misconception worth correcting. AFINs are not, in the literature considered here, a single standardized neural module. Rather, the term is applied to a broader design principle: amortize inference objects that respect factorized probabilistic structure, and train them so that test-time inference becomes a forward pass or a low-cost correction step.

2. Inverse-factor AFINs in directed graphical models

The original AFIN construction is defined for directed graphical models with latent variables $x=(x_1,\dots,x_N)$ and observed variables $y=(y_1,\dots,y_M)$. The posterior target is

$$\pi(x) \equiv p(x \mid y) = \frac{p(x,y)}{p(y)}, \qquad \gamma(x) \equiv p(x,y),$$

and the goal is to construct a proposal $q(x \mid y)$0 that closely approximates the intractable posterior while remaining tractable to sample and evaluate (Paige et al., 2016).

The key structural step is the construction of a stochastic inverse via inverse factorization. One first chooses a valid topological order of the original DAG and defines the Markov blanket $q(x \mid y)$1 of a node $q(x \mid y)$2 as the union of its parents, children, and the parents of its children. The order is then reversed, and inverse parents are defined by intersecting the Markov blanket with later variables in the reversed ordering:

$q(x \mid y)$3

with the observed-node analogue defined similarly. This yields an inverse conditional

$q(x \mid y)$4

and the recognition model mirrors this structure as

$q(x \mid y)$5

A central proposition in the construction is preservation of local conditional independence: if two inverse variables are conditionally independent given a third in the inverse graph, then the corresponding original variables were also conditionally independent in the original graph. This ensures that the inverse factorization does not introduce spurious independencies. The construction is explicitly designed to preserve dependencies created by explaining-away in head-to-head structures, and when several latents become jointly dependent after conditioning, they are grouped into a joint factor and modeled autoregressively:

$q(x \mid y)$6

Architecturally, the 2016 AFIN instantiates one conditional neural density estimator per inverse factor. The paper extends MADE into a conditional estimator by injecting inverse parents as conditioning inputs and constructing masks so that conditioning inputs propagate to all outputs, including the first autoregressive dimension. Continuous latents are modeled with mixture-of-Gaussians heads,

$q(x \mid y)$7

using softplus for $q(x \mid y)$8 and softmax for $q(x \mid y)$9, while discrete latents use Bernoulli or categorical heads. Repeated structures, such as per-time-step factors in HMMs or per-group factors in hierarchical models, reuse the same sub-network parameters (Paige et al., 2016).

3. Offline training and use as SMC proposals

The original AFIN training objective is an inclusive-KL criterion averaged over datasets:

$q_\phi(z)$0

Because the expectation is taken under the tractable joint $q_\phi(z)$1, stochastic gradients are available from ancestral samples of the forward model:

$q_\phi(z)$2

No gradients through the generative model are required; only the recognition network is differentiated. Training is performed offline with synthetic data, using mini-batch stochastic optimization with Adam and a held-out synthetic validation set; when validation worsens or after a fixed number of steps, synthetic training and validation data are regenerated to maintain diversity (Paige et al., 2016).

The inclusive-KL choice is not incidental. It encourages mass-covering proposals, reducing the risk that $q_\phi(z)$3 assigns negligible density where the true posterior has support. In the SMC setting this is important because support mismatch or overly narrow proposals induce weight degeneracy. The proposal learned by an AFIN is then inserted directly into an SMC scheme with sequential targets $q_\phi(z)$4 and unnormalized densities $q_\phi(z)$5:

$q_\phi(z)$6

For temporal models, the incremental weight takes the familiar form

$q_\phi(z)$7

and resampling, for example when $q_\phi(z)$8, combats degeneracy (Paige et al., 2016).

The paper demonstrates this program on three representative model classes. For non-conjugate polynomial regression with global latent weights, all weights are jointly dependent given the data and are modeled with a single joint conditional network. For hierarchical power plant pump failures, one shared inverse factor is learned for the repeated local latent variables together with a joint estimator for the global hyperparameters. For a factorial HMM with $q_\phi(z)$9 devices, a single time-shared conditional models the jointly dependent device states at each time step. Reported outcomes include accurate marginal likelihood estimates with as few as 5 particles in the pump-failure model, faster convergence than prior and naive likelihood-weighted proposals, and substantially increased effective sample size in the factorial HMM relative to proposals from the transition dynamics (Paige et al., 2016).

4. Typed-factor AFINs for zero-shot posterior inference

The 2026 AFIN formulation generalizes amortized posterior inference beyond a single fixed model by training over a family of typed-factor Bayesian models. The latent variable is $\Delta$0, observations are $\Delta$1, and the model factorization is expressed as one prior factor and $\Delta$2 likelihood factors:

$\Delta$3

Each factor carries a type label $\Delta$4 drawn from a finite catalog and factor-specific parameters $\Delta$5, so the inference problem is conditioned not only on observations but also on an explicit model specification. The amortized posterior is written

$\Delta$6

where $\Delta$7 encodes the typed-factor specification and $\Delta$8 is the inference network (Ko et al., 26 May 2026).

Training is simulation-based and uses a forward-KL objective. A fresh task is sampled at each step by drawing $\Delta$9, $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$0, a prior type, likelihood types, their parameters, and then synthetic latent and observation draws. The network parameters are learned by maximizing

$$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$1

which is equivalent up to constants to minimizing

$$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$2

As in the 2016 setting, forward-KL training is emphasized because it tends to produce mass-covering approximations that are well suited as proposals for importance-sampling correction (Ko et al., 26 May 2026).

The architecture is an encode–merge–decode pipeline inspired by conjugate Bayesian updates. Each factor is first mapped through a lightweight type-specific adapter and a shared encoder:

$$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$3

where $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$4 and $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$5 for likelihood factors. Each factor embedding has a node component $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$6 and a pair component $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$7. A sequence of BoxTransformer blocks then performs attention over the factor axis:

$$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$8

with node and pair attention scores averaged over latent-coordinate axes. After $$p(x, y) = \prod_{i=1}^N f_i\big(x_i \mid pa(x_i)\big)\,\prod_{j=1}^M g_j\big(y_j \mid pa(y_j)\big)$$9 blocks, factor embeddings are pooled by summation,

$$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$0

and a decoder maps $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$1 to the parameters of a variational family (Ko et al., 26 May 2026).

Two decoder families are used. The Gaussian decoder outputs $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$2 for a full-rank Gaussian $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$3, with the precision parameterized to be symmetric positive definite by construction. The flow decoder uses a conditional RealNVP with $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$4 masked coupling layers conditioned on projections of $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$5 and a final affine transformation $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$6. Dimension-independence is achieved through BoxMLP and BoxTransformer modules whose trainable parameter shapes do not depend on latent dimension $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$7 and which are permutation-equivariant in latent coordinates. Factor-order invariance is enforced by attention over the factor set without positional encodings for likelihood factors (Ko et al., 26 May 2026).

Empirically, the paper reports that a single trained AFIN achieves posterior accuracy comparable to NUTS and several variational baselines while requiring 2 to 4 orders of magnitude less test-time compute. The synthetic benchmark consists of 16 prior–likelihood combinations across three difficulty levels, with posterior quality measured by posterior mean error (M1), covariance Frobenius error (M2), and sliced Wasserstein-2 (SW2). Stress tests beyond the training range report, for example, OOD-$$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$8 settings such as $$p(z, y_{1:N} \mid t_0,\theta_0,\{t_n,\theta_n\}_{n=1}^N) = p(z \mid t_0,\theta_0)\prod_{n=1}^N p(y_n \mid z, t_n,\theta_n),$$9 and $(t,\theta)$0 where AFIN+SNIS achieves $(t,\theta)$1 in $(t,\theta)$2, compared with NUTS at $(t,\theta)$3 in $(t,\theta)$4 and full-rank VI at $(t,\theta)$5 in $(t,\theta)$6. Training uses AdamW for $(t,\theta)$7 steps with cosine decay from $(t,\theta)$8, batch size 32, gradient accumulation over 4 steps, EMA weights for evaluation, and completes in approximately 24 hours on a single NVIDIA H100 (80GB) (Ko et al., 26 May 2026).

5. Evidence factors and local blanket objectives

A related extension of the AFIN idea treats the amortized target not as a posterior density but as a scalar factor comparing models. "Evidence Networks" states that Amortized Factor Inference Networks are neural estimators that amortize the computation of “factors”—ratios or scalar functionals that compare models—across many datasets and parameter configurations, and identifies Evidence Networks as concrete instances specialized to the Bayes factor (Jeffrey et al., 2023). In that setting, training data are synthetic draws from competing models labeled by their generating model, and the network is optimized with a symmetric loss whose Bayes-optimal solution is a function of the Bayes factor

$(t,\theta)$9

For the l-POP-Exponential loss,

$x=(x_1,\dots,x_N)$0

the Bayes-optimal mapping satisfies

$x=(x_1,\dots,x_N)$1

Under equal model priors, $x=(x_1,\dots,x_N)$2. The paper emphasizes that this direct estimation of log-odds avoids numerical instability from taking ratios of independently estimated model probabilities, and reports that the method is explicitly independent of the dimensionality of the parameter space and scales mildly with posterior complexity (Jeffrey et al., 2023).

A second neighboring formulation is $x=(x_1,\dots,x_N)$3-AI, which explicitly describes itself as an amortized inference algorithm for sparse probabilistic graphical models and presents an AFIN-style interpretation in which factor- or blanket-local computations are amortized by learned predictors (Falet et al., 2023). The learned sampler is a Bayesian network

$x=(x_1,\dots,x_N)$4

and the central constraint matches local factor ratios from a Markov network to local conditional ratios in the Bayesian network. For configurations $x=(x_1,\dots,x_N)$5 and $x=(x_1,\dots,x_N)$6 differing only at variable $x=(x_1,\dots,x_N)$7,

$x=(x_1,\dots,x_N)$8

This yields the squared log-ratio loss

$x=(x_1,\dots,x_N)$9

which is partition-function-free because the unknown normalizing constant cancels. The paper’s equivalence theorem states that, under positivity and the I-map relation induced by chordalization, satisfaction of the local one-flip constraints for all variables implies equality of the joint distributions $y=(y_1,\dots,y_M)$0 (Falet et al., 2023).

Taken together, these two works broaden the conceptual envelope of AFINs. One line amortizes scalar comparison factors such as Bayes factors; the other amortizes blanket-level local conditionals while tying them into a globally consistent Bayesian-network sampler. This suggests that “factor inference” in the later literature is not confined to latent-state posteriors but can also encompass evidence ratios and local consistency relations.

6. Limitations, failure modes, and open questions

The limitations reported across the literature are structurally consistent: the quality of amortization is bounded by the validity of the model family, the expressiveness of the learned family, and the extent to which local or typed-factor structure remains manageable.

For inverse-factor AFINs in SMC, two failure modes are explicit. If the generative model is misspecified relative to the true data-generating process, the offline-trained proposal may perform poorly on real data. Support mismatch, especially when $y=(y_1,\dots,y_M)$1 is too narrow, causes weight degeneracy. In addition, large Markov blankets for discrete latents produce high-dimensional conditionals that can be difficult to approximate with finite neural capacity. The paper also notes non-uniqueness of the inverse graph: different valid inverse orderings exist, and learning to choose or ensemble orderings is identified as a possible direction (Paige et al., 2016).

For typed-factor AFINs, the principal constraint is the finite factor catalog. New factor families require defining a new adapter and retraining, even if that retraining is lighter than per-task VI or MCMC. The node–pair representation has an $y=(y_1,\dots,y_M)$2 pair component, so very high-dimensional latent spaces may require sparse or low-rank alternatives. Performance also degrades when latent dimension exceeds the training range; SNIS correction helps, but very large $y=(y_1,\dots,y_M)$3 may still favor MCMC. The paper further notes that deterministic transformations and deep hierarchical programs are not explicitly represented in the current interface (Ko et al., 26 May 2026).

For Evidence Networks, practical stability depends on the loss transform and on calibration. The recommended l-POP-Exponential loss uses $y=(y_1,\dots,y_M)$4 by default, but the paper reports that in the Dark Energy Survey application $y=(y_1,\dots,y_M)$5 induced overfitting and $y=(y_1,\dots,y_M)$6 trained more stably. Class imbalance implements model priors and must be corrected by subtracting prior-odds terms when necessary. Out-of-distribution inputs remain a concern, and ensembles or disagreement heuristics are suggested to flag such cases (Jeffrey et al., 2023).

For $y=(y_1,\dots,y_M)$7-AI, the central bottleneck is locality after chordalization. Chordalization can enlarge neighborhoods and reduce the computational advantage of the local objective. Poor exploration can fail to cover rare blankets, and the method assumes that PGM structure is known or specified; structure learning is outside its scope. For high-treewidth or highly multimodal graphs, child subsampling or junction-tree-style parameterizations may be needed (Falet et al., 2023).

Across these variants, one broad open question recurs. The literature suggests that AFINs are most effective when structural invariances can be made explicit—through inverse graphical structure, typed-factor interfaces, or local blanket decompositions—while still retaining enough expressivity to cover multimodality, long-range dependence, or distribution shift. Current proposals include richer conditional families such as flows, hybrid amortized–adaptive schemes that add online refinement, improved factor grouping, sparsity-aware masking, and broader support for probabilistic programs with arbitrary control flow (Paige et al., 2016, Ko et al., 26 May 2026).

7. Position within amortized inference research

AFINs occupy a distinct position relative to more familiar amortized inference paradigms. In the 2016 formulation, they differ from VAEs because the generative model is assumed fixed and known, the recognition network is trained offline using synthetic data from that model, and the training objective minimizes the inclusive KL rather than the reverse-KL term that appears in the ELBO. The resulting object is a model-specific proposal designed for SMC rather than a variational posterior tied to one empirical dataset (Paige et al., 2016).

The 2026 formulation distinguishes itself from classical amortized neural posterior estimation and simulation-based inference by exposing an explicit typed-factor interface and by supporting changes in prior families, likelihood families, latent dimensionality, observation count, and heterogeneous observation types within one trained network. The paper contrasts this with Bayesian ICL, Simformer, Distribution Transformers, NPE-PFN, and Masked Language Inference, arguing that AFIN uniquely combines dimension-independent modules, typed heterogeneity, invariance to $y=(y_1,\dots,y_M)$8 and factor ordering, and NUTS-level posterior quality without per-task finetuning (Ko et al., 26 May 2026).

Evidence Networks and $y=(y_1,\dots,y_M)$9-AI clarify two additional boundaries. Evidence Networks argue against learning $$\pi(x) \equiv p(x \mid y) = \frac{p(x,y)}{p(y)}, \qquad \gamma(x) \equiv p(x,y),$$0 separately and taking ratios, reporting that in a 20-dimensional time-series comparison with Neural Spline Flows the RMSE in $$\pi(x) \equiv p(x \mid y) = \frac{p(x,y)}{p(y)}, \qquad \gamma(x) \equiv p(x,y),$$1 for the density-ratio approach was more than a factor of 10 larger than direct Evidence-Network estimation with the l-POP loss (Jeffrey et al., 2023). $$\pi(x) \equiv p(x \mid y) = \frac{p(x,y)}{p(y)}, \qquad \gamma(x) \equiv p(x,y),$$2-AI, by contrast, positions itself against global trajectory-level objectives such as those used in GFlowNets, replacing long-horizon credit assignment with exact local factor-ratio supervision in sparse PGMs (Falet et al., 2023).

The resulting picture is that AFINs identify a methodological axis rather than a single model family. Along that axis, the amortized object may be a posterior proposal, a variational posterior parameterization, a Bayes factor, or a locally consistent sampler; the unifying principle is that probabilistic factors, local blankets, or typed model components are exposed to the network so that inference can be compiled into a reusable neural artifact.