CausalFlow-T: Flow-Based Causal Inference

• CausalFlow-T is a family of flow-based causal frameworks that integrate autoregressive flows, CNFs on causal DAGs, and stream-transformer techniques.
• It employs invertible transformations that ensure identifiability and support interventional and counterfactual inference through structured causal orderings.
• The umbrella term covers distinct implementations such as CAREFL, DoFlow, and stream-transformer models, each offering unique methodological strengths.

The name CausalFlow-T is not explicitly used as a canonical model designation in the cited literature. In the provided arXiv record, it most plausibly denotes a family of flow-based causal frameworks in which causal structure is embedded into an invertible transformation—typically an autoregressive flow, a continuous normalizing flow defined over a causal DAG, or, in a distinct stream-theoretic line, a causal stream transformer equipped with ultrametric fixed-point semantics. This suggests an umbrella usage rather than a single standardized architecture (Khemakhem et al., 2020, Wu et al., 4 Nov 2025, Ruess, 2023).

1. Terminological scope and principal interpretations

Within the supplied sources, three technically distinct constructions are the main candidates for the label.

Interpretation in the literature	Defining structure	Main use
CAREFL / causal autoregressive flows	Fixed-order affine or additive autoregressive flows	Causal discovery, interventions, counterfactuals
DoFlow	CNFs over a causal DAG for multivariate time series	Observational, interventional, and counterfactual forecasting
Causal stream transformer framework	Causal transformer $T$ on $\omega$-streams	Fixed-point existence, induction, anytime prefix approximation

The first interpretation is Causal Autoregressive Flows, which establishes an intrinsic correspondence between a simple family of autoregressive normalizing flows and identifiable causal models. The second is DoFlow, where the query label is explicitly mapped to a time-series framework whose paper name is different. The third is a stream-semantics construction in which the name is not used directly, but the provided summary identifies the causality–contraction correspondence plus multivalued fixpoint calculus for stream transformers as the closest match to such a label (Khemakhem et al., 2020, Wu et al., 4 Nov 2025, Ruess, 2023).

A common source of ambiguity is that the term causal flow is also used in unrelated senses. One paper uses CausalFlow to denote a visual analytics system for event sequences rather than a generative causal model, and another uses causal flow as a component inside a supervised VAE for disentangled causal representation learning. These are terminologically adjacent but methodologically distinct (Xie et al., 2020, Fan et al., 2023).

2. Autoregressive flows as causal structural equation models

The central modeling idea in the autoregressive-flow line is that an autoregressive normalizing flow already imposes an ordering over variables: $x_j = T_j(z_j, x_{<T(j)}),$ which is structurally analogous to a structural equation model

$x_j = f_j(x_{<T(j)}, n_j),$

with independent noise $n_j$. The key claim is therefore that autoregressive flow orderings correspond to causal orderings, and that an appropriately restricted flow can be interpreted as a causal SEM (Khemakhem et al., 2020, Monti et al., 2020).

The restriction is crucial. All stacked sub-flows must use the same permutation $T$, and each transformer must be affine,

$T_j(u,v)=e^{s_j(v)}u+t_j(v),$

with additive flows as the special case $s_j(v)=0$. In the bivariate setting this induces

$$x_j = e^{s_j(x_{<T(j)})} z_j + t_j(x_{<T(j)}), \qquad j=1,2.$$

Here $z_1,z_2$ are independent latent noises, $\omega$0 models the causal effect of parents on the child, and $\omega$1 allows the noise scale to depend on the cause. This is a strict generalization of additive noise models, since the additive case is recovered when $\omega$2 (Khemakhem et al., 2020).

The identifiability theorem is one of the defining results of this formulation. If $\omega$3 are Gaussian and

$\omega$4

with $\omega$5 nonlinear and invertible, then the model is identifiable. If $\omega$6, identifiability becomes the additive noise case and holds for any factorial noise distribution. The paper also emphasizes why this restriction is needed: general universal autoregressive flows are too expressive, since they can represent arbitrary densities in either ordering, destroying identifiability (Khemakhem et al., 2020).

Architecturally, CAREFL uses stacked autoregressive flows with fixed orderings. Each flow consists of several affine autoregressive sub-flows, and for each sub-flow $\omega$7 and $\omega$8 are parameterized by neural networks, typically MLPs. Training is maximum likelihood,

$\omega$9

and tractability follows from the triangular Jacobian. The experiments use an isotropic Laplace base distribution for the latent variables (Khemakhem et al., 2020).

3. Causal discovery, interventions, and counterfactuals in the flow formulation

In the bivariate case, causal discovery is performed by fitting two models, one for each ordering, and comparing held-out likelihoods. The causal score is the test-set log-likelihood ratio

$x_j = T_j(z_j, x_{<T(j)}),$0

If $x_j = T_j(z_j, x_{<T(j)}),$1, one concludes $x_j = T_j(z_j, x_{<T(j)}),$2 causes $x_j = T_j(z_j, x_{<T(j)}),$3; if $x_j = T_j(z_j, x_{<T(j)}),$4, one concludes $x_j = T_j(z_j, x_{<T(j)}),$5 causes $x_j = T_j(z_j, x_{<T(j)}),$6. The rationale is that flows provide exact normalized likelihoods, so causal direction can be scored by a likelihood-ratio test style method rather than by regression residual heuristics or linearity assumptions (Khemakhem et al., 2020).

The same formulation supports interventional queries. For

$x_j = T_j(z_j, x_{<T(j)}),$7

the intervention replaces the structural equation for $x_j = T_j(z_j, x_{<T(j)}),$8 by the constant $x_j = T_j(z_j, x_{<T(j)}),$9 and leaves downstream equations unchanged. In the sequential version, one samples $x_j = f_j(x_{<T(j)}, n_j),$0, fixes $x_j = f_j(x_{<T(j)}, n_j),$1, and computes the remaining $x_j = f_j(x_{<T(j)}, n_j),$2 in causal order. In the parallel affine-flow version, because the inverse map is available, the implementation uses

$x_j = f_j(x_{<T(j)}, n_j),$3

for a placeholder $x_j = f_j(x_{<T(j)}, n_j),$4, with $x_j = f_j(x_{<T(j)}, n_j),$5 in implementation, and then computes

$x_j = f_j(x_{<T(j)}, n_j),$6

This is described as marginalization over latent variables (Khemakhem et al., 2020).

Counterfactual inference exploits invertibility even more directly. The abduction step is

$x_j = f_j(x_{<T(j)}, n_j),$7

which recovers the latent noise corresponding to an observation. After the action step modifies the relevant structural equation, prediction is performed by forward propagation,

$x_j = f_j(x_{<T(j)}, n_j),$8

The resulting distinction is explicit in the paper: interventions marginalize over latent noise, whereas counterfactuals condition on the inferred latent noise (Khemakhem et al., 2020, Monti et al., 2020).

Empirically, the method is reported to be the only one consistently recovering the true causal direction across all synthetic settings considered in the final CAREFL paper. On the 108-pair Cause-Effect Pairs benchmark, CAREFL achieved about 73% correct direction, compared with 66% for Linear LR, 69% for ANM, and 69% for RECI. On synthetic four-dimensional SEMs, CAREFL achieved lower MSE on interventional expectations than ANM baselines. On es-fMRI stimulation data, the reported median absolute errors were 0.586 for CAREFL, 0.655 for ANM, and 0.643 for Linear SEM. The earlier flow-based causal discovery paper reports the same 108-pair benchmark result and emphasizes accurate interventional and counterfactual predictions on synthetic data (Khemakhem et al., 2020, Monti et al., 2020).

4. Temporal formulation: DoFlow as a CausalFlow-T time-series model

A separate time-series instantiation is given by DoFlow, where the provided summary explicitly states that the query calls it CausalFlow-T, although the paper’s model name is DoFlow. Its problem setting is multivariate forecasting under observational, interventional, and counterfactual queries in systems governed by a causal DAG (Wu et al., 4 Nov 2025).

The structural causal model is written as

$x_j = f_j(x_{<T(j)}, n_j),$9

where each node depends on its own past, the past of its DAG parents, and exogenous noise independent across nodes and time. History is summarized with an RNN per node,

$n_j$0

and the conditioning state is

$n_j$1

For each node, DoFlow learns a time-conditioned CNF with ODE

$n_j$2

Training uses Conditional Flow Matching with linear interpolant

$n_j$3

and the model retains explicit likelihoods of future trajectories (Wu et al., 4 Nov 2025).

The paper’s counterfactual recovery result is stated under three assumptions: $n_j$4 and

$n_j$5

It first proves the existence of a differentiable bijection $n_j$6 such that

$n_j$7

and then states the counterfactual recovery corollary

$n_j$8

Operationally, observational and interventional forecasting decode fresh latent samples unless $n_j$9, in which case the value is clamped. Counterfactual forecasting instead follows abduction–action–prediction by reusing the factual latent noise: $T$0 This is the temporal analogue of reusing the same exogenous randomness under an altered causal context (Wu et al., 4 Nov 2025).

The reported empirical results are correspondingly temporal. On synthetic DAG-based datasets, DoFlow is described as strong on observational forecasting, effective on interventional forecasting, and the only method in the comparison that directly supports counterfactual forecasting. On hydropower data it achieves the best reported RMSE among baselines, with 1.13 for observational forecasting and 1.21 for interventional forecasting. On the cancer treatment task it substantially outperforms CRN, RMSN, and MSM in normalized RMSE; for $T$1, step 3, the values reported are 1.25% for DoFlow, 2.43% for CRN, 3.16% for RMSN, and 6.75% for MSM. The paper also reports anomaly detection from log-likelihoods, with outages detected before they occur, sometimes 10–20 minutes in advance (Wu et al., 4 Nov 2025).

5. Stream-transformer interpretation and fixed-point semantics

In a distinct formal direction, the construction that most plausibly corresponds to the label CausalFlow-T is the combination of logic-based causality for stream transformers, ultrametric prefix-distance semantics, multivalued contraction and Hausdorff-Lipschitz analysis, fixed-point iteration over strongest post and weakest pre transformers, and fixpoint induction plus anytime prefix approximation. Here the basic object is a stream inclusion

$T$2

where $T$3 may be nondeterministic and multivalued (Ruess, 2023).

Causality is defined by prefix agreement. The paper’s $T$4-causality condition is

$T$5

With the prefix ultrametric

$T$6

the key theorem is the exact equivalence

$T$7

For nonempty compact-valued transformers, weak causality corresponds to nonexpansiveness and strong causality to contraction (Ruess, 2023).

The metric space of streams is spherically complete, which enables strong multivalued fixed-point principles. For strongly causal

$T$8

the strongest post and weakest pre transformers have unique fixpoints obtained by Picard iteration, with bound

$T$9

Moreover, if $T_j(u,v)=e^{s_j(v)}u+t_j(v),$0 is strongly causal, not identically empty, and each $T_j(u,v)=e^{s_j(v)}u+t_j(v),$1 is closed, then there exists $T_j(u,v)=e^{s_j(v)}u+t_j(v),$2 such that $T_j(u,v)=e^{s_j(v)}u+t_j(v),$3. Under stronger contraction, the fixpoint is unique (Ruess, 2023).

This formulation is not a probabilistic causal model in the normalizing-flow sense. Its contribution lies in recasting causal stream inclusion solving as a metric fixed-point problem on a spherically complete ultrametric space of $T_j(u,v)=e^{s_j(v)}u+t_j(v),$4-streams. The resulting iterations serve both as proof principles and as anytime algorithms, with quantitative guarantees on the already established finite prefix of a solution (Ruess, 2023).

6. Related terminology, misconceptions, and methodological boundaries

A recurrent misconception is that all uses of causal flow refer to the same model class. The supplied literature does not support that view. CausalFlow: Visual Analytics of Causality in Event Sequences is a visualization framework that integrates automatic causal discovery methods into event-sequence analysis and presents a timeline-plus-Sankey design called causal flow; it is not an invertible generative SCM. DCVAE inserts a causal flow into a supervised VAE encoder so that the learned latent representation reflects a causal adjacency matrix $T_j(u,v)=e^{s_j(v)}u+t_j(v),$5; its main target is causal disentangled representation learning rather than causal discovery from densities or causal time-series forecasting (Xie et al., 2020, Fan et al., 2023).

The assumptions also differ sharply across formulations. The autoregressive-flow approach assumes an acyclic causal structure, independent latent disturbances, and fixed autoregressive orderings; its main intervention discussion is clearest for root-node interventions, and the paper notes that general nonlinear SEMs do not enjoy identifiability without further assumptions (Khemakhem et al., 2020, Monti et al., 2020). DoFlow’s strongest theoretical guarantee is limited to univariate node settings, monotone SCMs, and assumptions (A1)–(A3), with counterfactual validity depending on the correctness of the causal DAG and the quality of latent encoding (Wu et al., 4 Nov 2025). The stream-transformer line, by contrast, is framed around causal guardedness, contraction, and closed-valued multivalued transformers rather than probabilistic identifiability (Ruess, 2023).

There is also a conceptual boundary with work on control flow in structural causal models. “Causality & Control Flow” argues that difficult preemption cases are better handled by modeling control-flow variables explicitly inside structural equations, rather than through unrestricted contingencies. That work is adjacent in spirit because it emphasizes mechanism and execution order, but it is not a flow-based density model and does not define CausalFlow-T as such (Künnemann et al., 2019).

Taken together, the literature supports a precise but non-unitary interpretation. CausalFlow-T is best understood as a convenient label for methods that make causality operational through a structured, often invertible transformation: fixed-order autoregressive flows for causal discovery and counterfactual reasoning, CNF-based causal DAG models for time-series forecasting, or causal stream transformers analyzed through ultrametric fixed-point theory. The common thread is not a single architecture but the use of structured flow semantics to turn causal assumptions into computable inference procedures (Khemakhem et al., 2020, Wu et al., 4 Nov 2025, Ruess, 2023).