Causal Flow: Directional Mechanisms in Causal Models

• Causal flow is defined as the directed propagation of influence in structured systems, enabling precise interventions and counterfactual predictions via invertible models.
• It leverages flow-based structural causal models and latent-space transformations to perform abduction, intervention, and counterfactual inference across various domains.
• Applications include medical-data harmonization, causal representation learning, and directed information analysis in systems like traffic networks and turbulent flows.

Searching arXiv for papers on “causal flow” and closely related formulations to ground the article in current research. Causal flow is a polysemous research term whose meaning depends on disciplinary context, but across its main usages it denotes a directed propagation of influence or information through a structured system. In causal generative modeling, it most commonly refers to a structural causal model (SCM) whose structural assignments are parameterized by invertible flow models, enabling abduction, intervention, and counterfactual prediction through exact or tractable inversion (Wang et al., 2021, Le et al., 2024). In causal representation learning, the term also encompasses flow-based priors or latent mechanisms that preserve causal dependencies among disentangled factors (Fan et al., 2023, Jin et al., 29 Jan 2026). In other domains, “causal flow” denotes directed information transfer in collective behavior, traffic networks, turbulent flows, or autoregressive token-processing architectures, where the emphasis shifts from interventional SCM semantics to operational notions of influence, predictive asymmetry, or sequential dependence (Lord et al., 2016, Molavipour et al., 2020, Martínez-Sánchez et al., 2022, Osawa et al., 2024, Wei et al., 28 Jan 2026, Niu et al., 6 Mar 2026). The unifying idea is that causal flow is not mere association: it is a structured, directional mechanism by which variation, perturbation, or information at one locus affects downstream variables, states, or representations.

1. Flow-based structural causal models

In the strict SCM sense, causal flow denotes a model in which each endogenous variable is generated from exogenous noise and its parents through an invertible flow map. A representative formulation appears in medical-data harmonization, where the SCM variables are sex $s$, age $a$, imaging site $t$, and MRI-derived feature vector $x$, with structural assignments

$$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$

and independent exogenous noises $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$ (Wang et al., 2021). The joint distribution factorizes as

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$

with $P(x\mid s,a,t)$ implemented as a conditional normalizing flow (Wang et al., 2021).

The critical property is invertibility. If $x=f_\theta(\epsilon)$ with base density $p_\epsilon$, then

$a$0

Because the inverse exists, exogenous variables can be recovered exactly from observations in deterministic SCMs, so Pearl’s abduction–action–prediction procedure becomes computationally direct (Wang et al., 2021). In the medical harmonization setting, this yields subject-specific counterfactuals of the form

$a$1

which answers: what would the same subject’s features have been if acquired at site $a$2 rather than $a$3 (Wang et al., 2021).

A related but more general formulation arises in identifiable causal flow models built from ordering information rather than a full known graph. There, node-wise probability-flow ODEs

$a$4

define a triangular monotone increasing map from exogenous variables $a$5 to observables $a$6, consistent with a causal ordering $a$7 (Le et al., 2024). The terminal-time map $a$8 serves as the structural mechanism, and the triangularity induced by ordering yields identifiability up to component-wise invertible transforms of exogenous variables (Le et al., 2024). This suggests that flow parameterizations can retain causal semantics even when only a valid causal order is available.

These SCM-based uses establish the most technically precise meaning of causal flow: an invertible causal generative mechanism, usually on a DAG, in which flows implement structural equations and thereby support exact or near-exact density evaluation, intervention, and counterfactual inference (Wang et al., 2021, Le et al., 2024, Wu et al., 4 Nov 2025).

2. Abduction, intervention, and counterfactual prediction

The importance of causal flow models lies in their support for all three levels of Pearl’s causal hierarchy. In harmonization, abduction is exact because the exogenous realization is obtained by inversion: $a$9 with analogous trivial inversions for root variables (Wang et al., 2021). Intervention is implemented by replacing a structural assignment, for example $t$0, and prediction is performed by re-running the forward flow using the abducted noise: $t$1 (Wang et al., 2021).

DoFlow extends the same logic to multivariate time series on a fixed DAG. Its SCM is

$t$2

with node-specific continuous normalizing flows conditioned on recurrent summaries of past histories (Wu et al., 4 Nov 2025). A forward CNF maps factual observations to latent variables,

$t$3

and the inverse CNF decodes either observational predictions or counterfactuals under altered hidden states (Wu et al., 4 Nov 2025). Under assumptions of exogenous independence, monotonicity in $t$4, and latent independence from history, DoFlow proves a counterfactual recovery result: the encoded latent is a bijective function of the exogenous noise, so decoding the same latent under counterfactual histories recovers the true SCM counterfactual almost surely (Wu et al., 4 Nov 2025). This is a particularly explicit realization of causal flow as exogenous-noise transport across worlds.

A different use of counterfactual intervention appears in LLM-agent debugging. There, CausalFlow models an execution trace $t$5 as a sequential chain of dependent steps and defines a step-level Causal Responsibility Score

$t$6

where the trace is re-executed after replacing step $t$7 with a candidate repair $t$8 (Bonagiri et al., 25 May 2026). Although this framework does not use normalizing flows, it uses interventional counterfactual semantics over structured traces, so the phrase “CausalFlow” designates causal attribution plus counterfactual repair rather than density modeling (Bonagiri et al., 25 May 2026). This usage broadens the term from invertible SCMs to operational intervention over sequential systems.

3. Causal priors and disentangled latent spaces

In causal representation learning, causal flow denotes flow-based latent models that preserve or parameterize causal relations among learned concepts. DCVAE introduces a causal flow inside the inference model of a supervised VAE. Its transformed latent variables $t$9 are produced by an affine autoregressive flow

$x$0

where $x$1 is a lower-triangular adjacency matrix encoding a causal graph over latent coordinates (Fan et al., 2023). Because the Jacobian is triangular, the log-determinant is tractable,

$x$2

and the transformed latent posterior remains explicit (Fan et al., 2023). In this formulation, the flow serves as a latent SEM: each transformed coordinate is generated from parent latents and noise, with $x$3 enforcing causal dependencies (Fan et al., 2023).

FlexCausal shifts the flow component from the causal mechanism to the exogenous prior. The latent SCM uses additive noise,

$x$4

while each exogenous noise block $x$5 has an independent flow-based prior $x$6, implemented with Masked Autoregressive Flows (Jin et al., 29 Jan 2026). The total prior factorizes as

$x$7

and, since the map $x$8 is volume-preserving, the latent SCM prior becomes

$x$9

(Jin et al., 29 Jan 2026). This design explicitly decouples mechanism $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$0 from noise statistics $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$1, which is valuable when exogenous factors are non-Gaussian or multimodal (Jin et al., 29 Jan 2026).

Both DCVAE and FlexCausal treat causal flow as a latent-space construct rather than an observable-space SCM. Inference, disentanglement, and intervention are performed in a representation space whose geometry is organized by a DAG, with flows either implementing the causal mechanism or the exogenous prior (Fan et al., 2023, Jin et al., 29 Jan 2026). A plausible implication is that the phrase “causal flow” in representation learning increasingly denotes structured invertible transport between latent exogenous variables and causally organized concepts, rather than only between exogenous variables and observed data.

4. Causal effect estimation and distributional potential outcomes

Another line of work uses flow dynamics to model distributions of potential outcomes rather than structural mechanisms over observed variables. RepFlow formulates causal effect estimation as a joint problem of balanced representation learning and Conditional Flow Matching (Xie et al., 7 May 2026). The data are observational triples $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$2, with potential outcomes $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$3, and the target estimand is typically the CATE

$$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$4

under consistency, unconfoundedness, and overlap (Xie et al., 7 May 2026).

The model first learns a balanced representation $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$5 by minimizing the entropically regularized Wasserstein distance between treated and control representations, with an $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$6 normalization constraint

$$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$7

to stabilize Sinkhorn optimization (Xie et al., 7 May 2026). It then trains a conditional continuous normalizing flow with velocity field

$$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$8

where $$s := \epsilon_S,\qquad a := f_A(\epsilon_A),\qquad t := \epsilon_T,\qquad x := f_X(\epsilon_X; s,a,t),$$9 interpolates between outcome data $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$0 and Gaussian noise $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$1, and the loss is the conditional flow-matching objective

$\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$2

(Xie et al., 7 May 2026).

This architecture is called a causal flow because, for each representation-treatment pair $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$3, the ODE transports noise into the full conditional potential-outcome distribution $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$4, not just its mean (Xie et al., 7 May 2026). Sampling from the reverse-time ODE yields Monte Carlo estimates of $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$5, $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$6, and even the distribution of individual treatment effects $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$7 (Xie et al., 7 May 2026). In this literature, causal flow therefore means a conditional probability flow over potential outcomes, coupled to a causal identification strategy via representation balancing rather than an explicit SCM over observables.

This usage differs materially from SCM-based causal flow. The causal semantics derive from the potential-outcomes framework and balancing assumptions, whereas the flow itself parameterizes conditional outcome distributions. Still, the common thread is transport between latent noise and causally interpretable distributions, with interventions realized by conditioning on treatment status (Xie et al., 7 May 2026).

5. Information-theoretic causal flow

Outside generative modeling, causal flow often denotes directed information transfer rather than interventional SCM transport. In collective animal behavior, causation entropy is defined as

$\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$8

quantifying how much the past of agent $\epsilon_S,\epsilon_A,\epsilon_T,\epsilon_X$9 improves prediction of agent $P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$0’s future beyond all other agents $P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$1 (Lord et al., 2016). The optimal causation entropy principle identifies the minimal parent set for each target and is operationalized by the oCSE discovery criterion

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$2

and the removal rule

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$3

for pruning indirect causes (Lord et al., 2016). Applied to midge swarms, this yields a directed network of information channels that is dynamically reconfiguring and often nonlocal in space (Lord et al., 2016).

Vehicular traffic uses a closely related but distinct quantity: directed information. For processes $P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$4,

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$5

and, under a Markov assumption, this simplifies to

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$6

(Molavipour et al., 2020). A Directed Information Graph is then built by thresholding normalized edge weights

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$7

to represent effective connectivity in a traffic network (Molavipour et al., 2020).

A more foundational treatment argues that standard mutual and conditional mutual information are not inherently causal because conditioning introduces “ghost channels,” hypothetical mechanisms induced by probabilistic conditioning rather than actual system channels (Ay, 2020). That work constructs channel-coupled sigma-algebras and defines information flow $P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$8 using marginals adjusted to the channel’s own measurable structure, proving a causal chain rule for these flows (Ay, 2020).

These information-theoretic formulations share with SCM-based causal flow the emphasis on directionality and the rejection of mere correlation, but they generally remain observational rather than interventional. They infer directed influence from predictive asymmetry or channel structure instead of from explicit do-operations (Lord et al., 2016, Molavipour et al., 2020, Ay, 2020).

6. Causal flow in physical and computational dynamics

In fluid mechanics, causal flow denotes the directed influence of coherent structures or local perturbations on future flow evolution. In wall-mounted-square-cylinder turbulence, proper-orthogonal-decomposition modes are analyzed with conditional transfer entropy to determine which modes cause which others (Martínez-Sánchez et al., 2022). Transfer entropy is

$P_M(s,a,t,x)=P(s)P(a)P(t)P(x\mid s,a,t),$9

and conditional transfer entropy among ROM modes identifies vortex-breaker modes as the dominant causal sources, especially over higher-order modes, while vortex-generator modes show no significant causal relationships (Martínez-Sánchez et al., 2022). This use of causal flow is not about generative modeling at all; it is about extracting a directed network of influence among coherent structures (Martínez-Sánchez et al., 2022).

A more interventionist fluid-mechanics usage appears in turbulent channel flow, where localized numerical perturbations are injected into small cells and their future amplification is tracked (Osawa et al., 2024). The perturbation energy is

$P(x\mid s,a,t)$0

and relative significance is defined as

$P(x\mid s,a,t)$1

(Osawa et al., 2024). Cells are called causally significant if perturbations there undergo large relative amplification. The dominant scaling variable is ambient shear; significant cells are associated with sweeps moving perturbations into stronger near-wall shear, whereas irrelevant cells are associated with ejections moving them outward (Osawa et al., 2024). This meaning of causal flow is therefore explicitly interventional but perturbative rather than SCM-based.

In autoregressive transformers, the phrase is adapted again. Stem analyzes self-attention through “causal information flow,” emphasizing that early tokens act as recursive anchors whose information propagates through every later position and layer (Niu et al., 6 Mar 2026). The intra-layer aggregation is

$P(x\mid s,a,t)$2

and recursive dependence across layers means pruning early tokens causes global distortion (Niu et al., 6 Mar 2026). Stem’s Token Position-Decay and Output-Aware Metric are designed to preserve this asymmetric flow of information under sparse attention (Niu et al., 6 Mar 2026). DeepSeek-OCR 2 uses “Visual Causal Flow” for a semantics-driven ordering of visual tokens: causal query tokens attend to all visual tokens and only previous query tokens via a lower-triangular mask, producing a learned reading order for document understanding (Wei et al., 28 Jan 2026). In both cases, “causal” refers to autoregressive sequential dependence rather than formal causal inference (Wei et al., 28 Jan 2026, Niu et al., 6 Mar 2026).

These usages demonstrate that causal flow has become a broader systems concept: directed, path-dependent propagation constrained by architecture, dynamics, or temporal order. The term is therefore context-sensitive; absent qualification, its meaning can range from do-calculus-compatible structural transport to predictive influence or autoregressive dependence.

7. Applications, advantages, and limitations

Across domains, causal flow methods are used where interventions, distribution shifts, or structured influence matter more than observational fit alone. In medical imaging, flow-based SCM harmonization generates site-counterfactual MRI features that preserve subject-specific variation and improves cross-domain generalization over ComBat variants and IRM in age prediction and Alzheimer’s classification (Wang et al., 2021). In causal representation learning, flow priors and flow mechanisms improve disentanglement, identifiability, and counterfactual generation quality, particularly when latent factors are non-Gaussian or multimodal (Fan et al., 2023, Jin et al., 29 Jan 2026). In treatment-effect estimation, flow matching supports distributional potential-outcome modeling and uncertainty quantification rather than only point-effect estimation (Xie et al., 7 May 2026). In time-series forecasting, DoFlow provides coherent observational, interventional, and counterfactual forecasts, while explicit likelihoods support anomaly detection in hydropower systems and treatment planning (Wu et al., 4 Nov 2025).

Several recurring advantages appear across these formulations. Invertibility permits exact or direct abduction of latent or exogenous variables (Wang et al., 2021, Le et al., 2024, Wu et al., 4 Nov 2025). Exact or tractable likelihoods permit principled model comparison and anomaly detection (Wang et al., 2021, Xie et al., 7 May 2026, Wu et al., 4 Nov 2025). Flexible non-Gaussian modeling lets flow-based mechanisms capture heavy-tailed, multimodal, or otherwise complex distributions that simple Gaussian SCMs or mean-field VAEs cannot represent (Fan et al., 2023, Jin et al., 29 Jan 2026). Counterfactual tractability arises because interventions can be implemented by modifying structural assignments while preserving inferred noise (Wang et al., 2021, Wu et al., 4 Nov 2025).

The limitations are equally consistent. Most formulations assume a known or fixed causal graph or ordering, and causal discovery remains partial or auxiliary (Wang et al., 2021, Le et al., 2024, Jin et al., 29 Jan 2026, Wu et al., 4 Nov 2025). They often require causal sufficiency, meaning no unobserved common causes between modeled variables and outcomes (Wang et al., 2021, Jin et al., 29 Jan 2026, Xie et al., 7 May 2026, Wu et al., 4 Nov 2025). Counterfactual guarantees may depend on monotonicity, additive-noise structure, or independence conditions that can fail in realistic settings (Wu et al., 4 Nov 2025). Flow models also incur computational overhead, especially CNFs with ODE solvers or autoregressive spline flows in high dimensions (Wang et al., 2021, Jin et al., 29 Jan 2026, Wu et al., 4 Nov 2025). In domains like OCR or sparse attention, the word “causal” may denote sequential constraints rather than interventionally meaningful causality, which can create conceptual ambiguity (Wei et al., 28 Jan 2026, Niu et al., 6 Mar 2026).

A common misconception is that any directional or autoregressive model implements causal flow in the strong causal-inference sense. The literature does not support that equivalence. Flow-based SCMs and counterfactual models explicitly encode interventions and exogenous-noise semantics (Wang et al., 2021, Le et al., 2024, Wu et al., 4 Nov 2025). By contrast, transfer-entropy analyses, directed-information graphs, and autoregressive token orderings quantify directed dependence or information propagation without necessarily licensing do-calculus interpretations (Lord et al., 2016, Molavipour et al., 2020, Wei et al., 28 Jan 2026, Niu et al., 6 Mar 2026). The same label therefore spans a spectrum from strict causal semantics to weaker notions of directed influence.

Taken together, the literature suggests that “causal flow” is best treated not as a single method but as a family of formalisms united by directed mechanistic transport. In its strongest sense, it is an SCM implemented with invertible flows for exact or tractable intervention and counterfactual reasoning (Wang et al., 2021, Le et al., 2024, Wu et al., 4 Nov 2025). In weaker but still technically meaningful senses, it is directed information transfer in nonlinear systems (Lord et al., 2016, Molavipour et al., 2020, Martínez-Sánchez et al., 2022), perturbation relevance under controlled interventions (Osawa et al., 2024), or autoregressive information routing in deep architectures (Wei et al., 28 Jan 2026, Niu et al., 6 Mar 2026). This suggests that future usage will likely continue to bifurcate between formally causal generative modeling and broader notions of directional flow, making contextual definition essential whenever the term is used.