Causal Foundation Models (CFMs)

• Causal Foundation Models (CFMs) are pretrained neural models that estimate causal effects or causal structures on unseen datasets through in-context learning without modifying model parameters.
• These models are used in interactive causality examination, causal model synthesis, and prediction management for synthetic datasets generated through training from SCMs (Structural Causal Models).
• CausalFM (Causal Foundation Models) and CC-PFN (Continuous treatment CCAuteous Processing Foundation Network) are specific examples, which aid in causal Directed Acyclic Graph (DAG) estimation, causal structural representations, and Bayesian model reasoning.

Causal Foundation Models (CFMs) are pretrained neural models designed to estimate causal quantities or causal structures on previously unseen datasets through in-context learning, typically without updating model parameters or fitting a bespoke estimator. Their training data are usually generated from structural causal models (SCMs), allowing factual observational contexts to be paired with interventional, counterfactual, or structural targets. The term encompasses several related paradigms: amortized treatment-effect estimation, causal discovery, graph-conditioned causal inference, continuous-treatment response modeling, time-series and continuous-time causal modeling, and causally structured representation learning. Unlike ordinary foundation models, which primarily model observational regularities, CFMs are intended to represent intervention-related quantities such as $P(Y\mid do(X=x))$, potential outcomes, causal graphs, or identified sets under partial identification. Their validity remains conditional on the causal prior, identification assumptions, simulator coverage, and the quality of available domain knowledge.

1. Conceptual foundations and scope

1.1 Causality beyond association

Causal claims concern the data-generating process rather than statistical association alone. A correlation between $X$ and $Y$ may arise because $X$ causes $Y$, $Y$ causes $X$, or a third variable $Z$ causes both:

$$X\rightarrow Y,\qquad Y\rightarrow X,\qquad X\leftarrow Z\rightarrow Y.$$

In an SCM, endogenous variables are generated by structural assignments such as

$X_i := f_i(\operatorname{Pa}(X_i),U_i),$

where $X$0 denotes the parents of $X$1 and $X$2 represents exogenous influences. Interventions replace a structural assignment, yielding quantities such as

$X$3

which generally differs from the observational conditional distribution $X$4. Counterfactuals concern alternative outcomes for the same unit, such as $X$5.

This distinction motivates the CFM agenda. A model that produces plausible causal language or predicts familiar causal associations need not possess a structural causal model, identify causal effects, discover graphs from data, or maintain counterfactual consistency. LLMs may encode “correlations on top of causation”: statistical regularities in text that describes causal relations (Willig et al., 2022). Such information can act as a prior or “head start,” but it is not equivalent to causal grounding.

1.2 The CFM paradigm

A CFM replaces repeated, dataset-specific causal estimation with amortized inference. During pretraining, the model encounters many synthetic causal datasets generated from a prior over SCMs. At deployment, an observational dataset is provided as context together with a causal query. The model then produces an estimate or predictive distribution in a forward pass.

The defining properties of a CFM in the narrowest formulation are:

1. Cross-dataset pretraining: learning from heterogeneous causal datasets or SCMs.
2. Causal objectives: predicting interventions, counterfactuals, effects, or structure rather than only observed outcomes.
3. Task-level generalization: operating on an unseen dataset, mechanism, graph, or covariate distribution.
4. Minimal adaptation: avoiding parameter updates or per-dataset optimization at inference.

A practical definition describes a CFM as a prior-data fitted network pretrained on a variety of causal tasks to estimate causal quantities from new datasets using in-context learning without updating its weights (Stith et al., 2 Sep 2026). This definition includes both PFN-style estimators of treatment effects and pretrained models for causal discovery.

The term is not used uniformly. CInA defines a first-step CFM as a model pretrained across datasets to infer covariate-balancing weights for treatment-effect estimation (Zhang et al., 2023). CausalFM defines a CFM through a causal SCM prior and posterior-predictive inference for identifiable causal settings (Ma et al., 12 Jun 2025). Other work applies the term to pretrained causal-discovery models, graph-conditioned causal predictors, structured time-series models, and causally motivated representation learners.

1.3 Levels of causal capability

CFM capabilities can be separated into increasingly demanding levels:

• Causal vocabulary and fact retrieval: producing causal statements or recalling familiar relations.
• Causal graph prediction: estimating directed, ancestral, or confounding relations.
• Causal discovery: inferring structure from observational or interventional data.
• Interventional prediction: estimating outcomes under $X$6-operations.
• Causal-effect estimation: estimating ATE, CATE, treatment-response curves, or policies.
• Counterfactual reasoning: predicting alternative outcomes for the same unit while preserving shared latent factors.
• Partial identification: representing a set or interval of causal quantities when the data and assumptions do not identify a unique value.

These capabilities should not be conflated. “Can Foundation Models Talk Causality?” found that LLMs answered some commonsense causal questions and reconstructed familiar graph relations, but did not demonstrate causal discovery from data, intervention generalization, invariant structural understanding, or formal counterfactual reasoning (Willig et al., 2022).

2. Priors, identification, and Bayesian amortization

2.1 SCM priors

Most CFM designs rely on a prior over causal data-generating processes. A prior may sample:

• graph topology and edge density;
• observed and latent variables;
• structural functions;
• treatment-assignment mechanisms;
• outcome mechanisms;
• noise distributions;
• sample sizes;
• interventions and intervention targets;
• observation schedules or missingness patterns.

The pretrained model approximates a posterior or posterior predictive distribution induced by this prior. In a generic formulation,

$X$7

where $X$8 is an observational context and $X$9 is a causal query.

CausalFM formalizes a causal setting as a paired observational/interventional distribution with an identifiable query. Its $Y$0-SCM prior assigns probability only to SCMs compatible with the relevant assumptions, such as back-door, front-door, or instrumental-variable conditions (Ma et al., 12 Jun 2025). The key distinction is between identification, which depends on causal assumptions, and estimation, which can be amortized by the pretrained model.

2.2 Why prior validity matters

A model cannot infer information that is not identified by the data and assumptions. If a prior contains mutually incompatible or unidentified causal worlds, observational data may fail to concentrate the posterior on the true causal quantity. Conversely, a prior that excludes observationally compatible SCMs can produce an invalidly narrow posterior.

“Foundation Models for Partial Causal Identification” addresses this issue through a canonical prior with full support over SCMs with finite discrete observables. The prior assigns positive probability to every finite directed and bidirected graph pattern and uses Dirichlet distributions over canonical exogenous parameters (Bellot et al., 21 Aug 2026). Its purpose is to ensure that the support of the posterior predictive distribution contains all causal-query values compatible with the observed distribution and structural assumptions.

For a scalar query $Y$1, the identified set can be written as

$Y$2

Under full support, a CFM’s asymptotic posterior support can represent this set rather than collapsing to an arbitrary point estimate. This separates sampling uncertainty, which may vanish with increasing sample size, from irreducible identification uncertainty, which remains whenever multiple SCMs imply different causal values.

2.3 In-context Bayesian inference

PFN-based CFMs train a transformer to approximate posterior-predictive inference through synthetic episodes. A typical episode consists of:

1. sampling an SCM from the causal prior;
2. generating an observational context dataset;
3. generating interventional or counterfactual targets from the same SCM;
4. training the transformer to predict the target distribution from the context and query.

The deployment procedure is therefore

$Y$3

with no parameter update. Predictive distributions are commonly represented through histogram or discretized outcome heads, enabling negative log-likelihood or cross-entropy training and, in principle, uncertainty quantification.

This approach differs from conventional causal inference, in which propensity scores, outcome models, balancing weights, matching procedures, or doubly robust estimators are fitted separately to each dataset. It also differs from ordinary PFNs because the context is observational while the target is generated under an intervention or counterfactual query.

3. Architectures and computational strategies

3.1 Attention as causal balancing

CInA treats observational units as transformer tokens. Self-attention operates across units rather than words, and attention coefficients are related theoretically to optimal covariate-balancing weights. The method minimizes an RKHS balancing objective and trains an attention-based classifier to distinguish treated and control units. Under linear-independence and global-optimum conditions, the value-normalization combination

$Y$4

corresponds to the optimal balancing coefficient (Zhang et al., 2023).

The method amortizes balancing across multiple datasets and performs zero-shot treatment-effect estimation through a forward pass. It addresses observed-covariate imbalance but does not solve unmeasured confounding, positivity failure, interference, or model misspecification. Its main task is binary-treatment SATE or ATE estimation.

3.2 Direct and decomposed causal discovery

Several CFMs target graph discovery rather than effect estimation.

Causal Pretraining maps multivariate time series directly to lagged causal graphs. It uses supervised synthetic time series and graph labels, with MLP, GRU, ConvMixer, and Transformer architectures. Performance improves when training data cover heterogeneous dynamics and when model capacity increases, but transfer depends on shared dynamics, causal sufficiency, absence of instantaneous effects, and time-invariant mechanisms (Stein et al., 2024).

Sea—Sample, Estimate, Aggregate separates local statistical estimation from global graph reconstruction. It samples subsets of variables, runs classical estimators such as FCI or GIES on those subsets, and aggregates their noisy partial outputs with global statistics such as inverse covariance. Its aggregator is trained on synthetic causal datasets and can correct systematic estimator errors. The architecture is supported by a result showing that marginal estimates over sufficiently small subsets can recover a graph pattern under stated assumptions (Wu et al., 2024).

DCD-PFN amortizes local Markov-boundary discovery rather than directly predicting the full adjacency matrix. A decoupling branch learns sample-wise weights intended to make non-target variables mutually independent, while an MB branch predicts membership in the target’s Markov boundary. Local boundaries are then combined through spouse removal, collider identification, Meek rules, and CPDAG reconstruction (Guan et al., 19 Jun 2026).

DAG-FM decomposes graph recovery into autoregressive leaf-node and parent-node prediction. It iteratively selects a leaf, predicts its parents, removes the leaf, and repeats. The procedure produces a causal ordering and structurally valid DAG. Its Mixture-of-Leaf-Experts routes observations among LiNGAM, ANM, HNM, PNL, general-mechanism, and dummy experts, allowing heterogeneous local mechanism families (Chen et al., 13 Jul 2026).

CDFM treats unknown causal mechanisms as latent variables and uses a variational decomposition involving mechanism inference, structure-aware reconstruction, and graph decoding. Its pretraining corpus includes multiple graph families, nonlinearities, discrete mechanisms, measurement distortions, missingness, latent confounding, and time-lagged mechanisms (Qiao et al., 13 Jul 2026).

3.3 Unified structure and outcome models

TabPFN-CFM jointly predicts causal structure and outcomes. It represents observed systems as acyclic directed mixed graphs with directed adjacency, ancestral reachability, and bidirected confounding matrices. A shared transformer supports separate structural and outcome decoders.

The model is trained on observational, interventional, and counterfactual queries. It supports:

• observational prediction;
• interventional prediction under $Y$5;
• same-unit counterfactual prediction;
• directed-edge prediction;
• ancestral-reachability prediction;
• latent-confounding prediction.

On synthetic and semi-synthetic tasks, supplied graph information improved interventional and counterfactual prediction while leaving observational prediction nearly unchanged. The model’s scope remains bounded by binary treatments, acyclicity, the synthetic prior, and the validity of supplied graph information (Zhu et al., 25 Jun 2026).

3.4 Graph-conditioned attention

General-purpose CFMs may benefit from domain knowledge supplied at test time. “Use What You Know” introduces partial ancestral matrices (PAMs), where $Y$6 denotes a known ancestor relation, $Y$7 a known non-ancestor relation, and $Y$8 an unknown relation (Reuter et al., 16 Feb 2026).

The most effective mechanism is a learnable soft attention bias:

$Y$9

where known ancestors receive positive bias, known non-ancestors receive negative bias, and unknown relations receive no structural bias. Soft biasing is generally more robust than hard attention masking because it incorporates expertise without making every constraint deterministic. Graph convolution and adaptive layer normalization provide a complementary global conditioning mechanism.

3.5 Continuous and structured temporal models

For continuous treatments, CCPFN reconstructs individual treatment-response curves from observational data. Its prior contains randomized covariate, treatment, and outcome MLPs, with heterogeneous confounding and continuous treatment assignment. The model predicts the conditional expected potential outcome

$X$0

over a grid of treatment values. It uses a tri-encoder for treatment, covariates, and outcomes, and performs zero-shot inference on unseen tasks without fine-tuning (Stith et al., 14 May 2026).

Continuous-time CFMs require more than writing a discrete update in SDE notation. A genuine continuous-time prior should have a trajectory law invariant to the observation schedule. “Towards Continuous-time Causal Foundation Models” distinguishes discrete models, naive observation-grid SDE integration, and fine-grid integration with decoupled observation. In the latter, an SDE is integrated on a fine computational grid and sampled at the desired irregular observation times (Thumm et al., 26 May 2026).

3.6 Causally structured representation learning

Not all models called CFMs estimate causal effects or graphs. “Causal Foundation Models: Disentangling Physics from Instrument Properties” uses a dual-encoder architecture for structured astronomical time series. One latent space represents intrinsic stellar properties and another represents instrument properties. Observational triplets involving the same star under different instruments and different stars under the same instrument provide weak metadata-based supervision (Audenaert et al., 7 Jul 2025).

The model combines structured contrastive losses with reconstruction. Its physics representation is more useful for downstream stellar-parameter prediction than raw light curves or a single-latent baseline, particularly when labeled data are scarce. The work illustrates a broader CFM principle: representations can be organized around hypothesized stable factors and measurement processes, although contrastive invariance does not by itself establish causal identification.

4. Training data, objectives, and applications

4.1 Synthetic pretraining distributions

Synthetic SCMs are central because large collections of real datasets with verified causal ground truth are scarce. Pretraining priors vary substantially.

Causal discovery priors use Erdős–Rényi, scale-free, small-world, stochastic-block, geometric, or other graph families. Mechanisms include linear, polynomial, sigmoid, neural-network, additive-noise, heteroscedastic, post-nonlinear, physics-inspired, discrete, ordinal, measurement-error, and time-lagged processes. Noise distributions may include Gaussian, Laplace, Uniform, Student-$X$1, Cauchy, Beta, Gamma, Exponential, Gumbel, log-normal, and mixtures.

Treatment-effect priors sample covariates, treatment assignment, outcomes, confounding strength, noise, and functional forms. CausalFM uses Bayesian neural-network SCM priors for back-door, front-door, and instrumental-variable settings (Ma et al., 12 Jun 2025). CCPFN randomizes continuous-treatment response functions, confounding structures, treatment noise, and tabular transformations (Stith et al., 14 May 2026).

Some priors include latent variables and induce bidirected confounding. Others assume causal sufficiency. Some support interventions and counterfactuals explicitly, whereas others use only observational samples and structural labels. The prior therefore determines what “causal generalization” means for a given CFM.

4.2 Self-supervised, supervised, and hybrid objectives

CFM objectives include:

• posterior-predictive negative log-likelihood;
• cross-entropy over discretized causal targets;
• graph-edge binary cross-entropy;
• Markov-boundary membership loss;
• RKHS balancing objectives;
• intervention and counterfactual prediction loss;
• masked-value reconstruction;
• quantile imputation;
• contrastive invariance;
• graph consistency and acyclicity penalties.

A distinction is necessary between self-supervised causal learning and ordinary causal supervision. In CInA, treatment assignments provide the self-supervised signal for balancing, while outcomes are not needed for the primary objective. In PFN-based models, causal targets are generated from synthetic SCMs and therefore are supervised relative to the simulator, even though no real-world causal labels are required.

4.3 Treatment-effect estimation

CausalFM estimates CATE under back-door adjustment by training on observational context datasets and synthetic individual treatment effects generated by paired interventions. Its concrete implementation uses three-class CATE prediction and is evaluated with PEHE on ten synthetic datasets and the Jobs benchmark. It is competitive with S-, T-, X-, DR-, RA-learners, and TARNet without per-dataset retraining (Ma et al., 12 Jun 2025).

CInA instead learns covariate-balancing weights across datasets. It targets binary-treatment ATE estimation and reports competitive performance under mechanism shifts, random-DAG simulations, and semi-synthetic datasets, with inference approximately $X$2 times faster than repeatedly fitting competing estimators in one reported setting (Zhang et al., 2023).

CCPFN extends the paradigm to continuous treatments. It reconstructs individualized treatment-response curves, evaluated using mean integrated squared error, and also supports treatment-policy evaluation through dosage policy error. Its performance is strongest for response-curve reconstruction, although policy optimization results are mixed (Stith et al., 14 May 2026).

4.4 Causal discovery

CDFM reports strong synthetic performance across 15 mechanism families and graph sizes up to $X$3, with AUROC and $X$4 exceeding the listed baselines in aggregate. Its results include causal-chamber and Tübingen evaluations, but specialized methods can outperform it under exactly matched assumptions (Qiao et al., 13 Jul 2026).

DAG-FM obtains strong precision, recall, F1, and normalized SHD on heterogeneous synthetic mechanisms and real datasets such as Sachs and Causal Chamber. Direct-LiNGAM remains superior in a pure LiNGAM setting, illustrating the tradeoff between generality and specialization (Chen et al., 13 Jul 2026).

DCD-PFN emphasizes local structure and reports strong CPDAG performance under topology, mechanism, and noise shifts. Its output is generally conservative: precision is high but recall remains limited in some settings (Guan et al., 19 Jun 2026).

Sea demonstrates a complementary design in which classical discovery methods act as local sensors and a pretrained aggregator reconciles their outputs. Its reported experiments include synthetic graphs with hundreds of variables, K562 perturbation data, Sachs, and random-DAG simulations (Wu et al., 2024).

4.5 Molecular and scientific applications

Molecular CFMs are expected to combine molecular foundation-model representations with prior biological knowledge, perturbational data, temporal structure, and mechanistic systems-biology models. Relevant data include CRISPR screens, Perturb-seq, drug inhibition, genetic screens, transcriptomics, phosphoproteomics, and single-cell multi-omics.

A molecular CFM should distinguish observational prediction from intervention prediction, support context-dependent mechanisms, and validate outputs against held-out perturbations and biological reality. Masked gene prediction alone does not imply causal direction because genes are not naturally ordered like language tokens and masked observations need not correspond to downstream variables (Lobentanzer et al., 2024).

The channel-foundation-model literature uses “CFM” in a different, domain-specific sense: Channel Foundation Models are pretrained on heterogeneous wireless channel data for tasks such as CSI estimation, beam prediction, and positioning. CSI-CLIP pretraining on more than 700,000 DeepMIMO samples across 35 scenarios improved positioning and beam prediction in the reported experiments (Xu et al., 24 Jun 2026). These models are channel-specialized foundation models, not causal foundation models in the SCM sense.

5. Evaluation, evidence, and common misconceptions

5.1 Evaluation dimensions

CFM evaluation should distinguish:

1. observational prediction;
2. graph recovery;
3. causal-effect estimation;
4. intervention prediction;
5. counterfactual consistency;
6. partial-identification coverage;
7. transfer across mechanisms and graph topologies;
8. robustness to variable renaming and paraphrase;
9. invariance under environment and measurement shifts;
10. calibration and uncertainty;
11. computational cost and adaptation requirements.

Metrics reported across the literature include MAE, PEHE, MISE, dosage policy error, MSE, NLL, AUROC, AUC, precision, recall, F1, SHD, normalized SHD, CPDAG SHD, orientation accuracy, and coverage of identified sets.

High predictive accuracy is not sufficient evidence of causal competence. AUROC can remain high while thresholded graph recovery is poor. A model may achieve good observational MSE while failing under intervention. A counterfactual score against DoWhy-generated outcomes validates consistency with that simulator, not experimentally verified individual-level counterfactuals.

5.2 Scaling and transfer

Several studies report improvements from broader priors and larger models. Causal Pretraining finds that heterogeneous training distributions improve transfer and that larger models can cross capacity thresholds for graph prediction (Stein et al., 2024). CDFM reports increasing performance with graph and sample size within its training regime (Qiao et al., 13 Jul 2026). Causal discovery models such as Sea and DAG-FM report transfer across graph families, nonlinearities, and noise distributions (Wu et al., 2024, Chen et al., 13 Jul 2026).

These results do not establish distribution-free scaling laws. Performance remains conditional on:

• coverage of the synthetic prior;
• dimensionality and sample-size ranges;
• causal sufficiency or latent-confounding assumptions;
• stationarity and temporal structure;
• positivity and overlap;
• graph and mechanism similarity;
• optimization success;
• measurement and preprocessing compatibility.

5.3 Domain knowledge and graph misspecification

Graph-conditioned CFMs generally improve when supplied with correct causal information. Partial ancestral information can approach the performance of full-graph conditioning without requiring a complete DAG (Reuter et al., 16 Feb 2026). TabPFN-CFM improves interventional and counterfactual prediction when the correct graph is supplied (Zhu et al., 25 Jun 2026).

However, incorrect graph information can induce confidently wrong predictions. Soft attention biases are more flexible than hard masks, but they do not make false assumptions harmless. A CFM should distinguish known, known-false, and unknown relations, and should report uncertainty under graph misspecification.

5.4 Common misconceptions

A CFM is not automatically causal because it uses attention. Attention can implement balancing, graph aggregation, or representation learning, but raw attention weights are not causal effects. In CInA, the balancing object is a value-normalization combination rather than the attention matrix itself (Zhang et al., 2023).

A causal answer is not necessarily causal inference. Producing “$X$5 causes $X$6” does not establish identification of $X$7 or an average treatment effect (Willig et al., 2022).

Synthetic counterfactual supervision is not experimental validation. A model can reproduce the causal semantics of a simulator while failing on real mechanisms, latent confounding, or measurement processes.

Zero-shot does not mean assumption-free. A model may avoid test-time parameter updates while still relying strongly on the assumptions encoded in its pretraining prior.

Graph prediction is not always causal discovery. Querying a LLM with variable names tests retrieval of causal facts, not graph recovery from empirical observations. Similarly, using a fitted SCM to generate augmentation data does not establish that the SCM is causally identified.

Causal generalization is not ordinary domain generalization. Stable performance across environments can be suggestive, but causal transport requires assumptions about invariant mechanisms, interventions, and the data-generating process.

6. Limitations and future directions

6.1 Fundamental causal limitations

CFMs cannot remove non-identifiability. Observationally equivalent SCMs may imply different interventions or counterfactuals. Hidden confounding, selection bias, insufficient overlap, measurement error, interference, feedback cycles, and model misspecification remain fundamental concerns.

Many current models assume:

• acyclic graphs;
• causal sufficiency or limited latent confounding;
• i.i.d. samples;
• binary treatment;
• static or time-invariant mechanisms;
• adequate positivity;
• finite and compatible schemas;
• synthetic priors that cover the target domain.

Continuous-time and molecular systems frequently violate these assumptions through feedback, delayed effects, non-Markovian dynamics, cyclic networks, and context-dependent mechanisms.

6.2 Prior misspecification and uncertainty

The prior is both the source of generalization and a potential source of systematic bias. A model trained on a narrow mechanism family may learn simulator-specific artifacts. A full-support prior can protect asymptotic support for discrete partial-identification problems, but its internal posterior density remains prior-dependent. In restricted priors, apparently narrow predictive distributions may reflect prior exclusion rather than empirical identification.

Future CFMs require calibrated uncertainty over:

• causal graphs;
• graph equivalence classes;
• latent confounders;
• intervention effects;
• counterfactual outcomes;
• extrapolation outside treatment support;
• mechanism and environment shifts.

6.3 Scaling and schema heterogeneity

Transformer architectures often scale quadratically in the number of observations, variables, or pairwise interactions. CInA has $X$8 attention cost per forward iteration. Sea uses axial attention but retains costs involving variable subsets and local estimates. CFM architectures must also handle differing numbers of variables, missing covariates, categorical and continuous features, temporal structures, treatment spaces, and estimands.

Potential directions include permutation-equivariant graph architectures, hierarchical representations, inducing-point methods, linear attention, sparse attention, schema adapters, and variable-length causal interfaces.

6.4 Interventions, counterfactuals, and active experimentation

Future models should train on explicit interventions rather than relying only on observational contexts. Relevant objectives include:

• intervention-distribution prediction;
• counterfactual consistency under shared exogenous noise;
• sequential and longitudinal treatment regimes;
• policy evaluation and policy learning;
• mediation and natural direct or indirect effects;
• active intervention selection;
• uncertainty-guided experimental design.

Molecular CFMs, for example, could combine observational omics with CRISPR and drug perturbations, while physical and wireless CFMs could represent interventions on environment, frequency, antenna configuration, blockage, or user position. The latter are proposed extensions rather than capabilities demonstrated by channel foundation-model work (Xu et al., 24 Jun 2026).

6.5 Toward general-purpose CFMs

A general-purpose CFM would ideally combine:

• a broad but explicitly characterized SCM prior;
• graph, ancestry, and confounding representations;
• observational, interventional, and counterfactual queries;
• continuous and discrete treatments;
• static, longitudinal, and continuous-time dynamics;
• uncertainty over graphs and mechanisms;
• domain-knowledge conditioning;
• robust handling of partial identification;
• calibrated abstention and sensitivity analysis;
• evaluation against real interventions and held-out mechanisms.

The emerging literature suggests several complementary routes. CInA amortizes covariate balancing; CausalFM and CCPFN amortize Bayesian causal estimation; Sea, DCD-PFN, CDFM, and DAG-FM amortize structure discovery; TabPFN-CFM combines structural and outcome prediction; graph-conditioned models incorporate partial expertise; and continuous-time models separate trajectory laws from observation schedules (Zhang et al., 2023, Ma et al., 12 Jun 2025, Stith et al., 14 May 2026, Wu et al., 2024, Guan et al., 19 Jun 2026, Qiao et al., 13 Jul 2026, Chen et al., 13 Jul 2026, Zhu et al., 25 Jun 2026, Reuter et al., 16 Feb 2026, Thumm et al., 26 May 2026).

CausalMixFT illustrates a more limited use of causality: fitting a dataset-specific SCM to generate structured synthetic data for tabular-foundation-model fine-tuning. Across 33 TabArena datasets and 2,310 runs, it improved median normalized ROC-AUC from approximately $X$9 for standard fine-tuning to $Y$0 and reduced the median validation–test performance discrepancy from $Y$1 to $Y$2 (Bühler et al., 7 Jan 2026). The method is best characterized as causal-data-augmented adaptation rather than a CFM, because its SCM is fitted separately for each target dataset and used primarily for observational augmentation.

The central methodological lesson is that a CFM should be evaluated not by whether it produces plausible causal prose or a single accurate benchmark number, but by whether it preserves causal structure under interventions, counterfactuals, mechanism shifts, graph uncertainty, measurement changes, and incomplete domain knowledge. Foundation-model scale can amortize causal inference and discovery, but it does not replace identification assumptions, experimental evidence, structural validation, or uncertainty about what the data cannot determine.