Causal Disentangled Representation Learning

Updated 7 July 2026

CDRL is a framework where low-dimensional latent representations capture semantically meaningful factors with an explicit causal organization.
It replaces independent priors with structured, causal priors in latent variable models, enhancing counterfactual generation and robustness.
Causal disentanglement improves identifiability and performance through tailored supervision, innovative architectures, and causal evaluation metrics.

Searching arXiv for recent and foundational papers on causal disentangled representation learning. arxiv_search(query="causal disentangled representation learning", max_results=10, sort_by="relevance") Causal disentanglement representation learning (CDRL) studies latent-variable models in which a low-dimensional representation is required to recover semantically meaningful factors together with their causal organization. In this literature, the target is not merely a statistically factorized code, but a representation whose coordinates or subspaces correspond to ground-truth factors, causal mechanisms, or interventionally stable components, so that changing one factor has the prescribed effect on descendants while leaving non-descendants invariant. Relative to classical disentanglement with independent priors, CDRL explicitly addresses causally related factors, confounding, interventions, counterfactual generation, and identifiability up to permutations and admissible reparameterizations (Jin et al., 29 Jan 2026, Shen et al., 2020, Suter et al., 2018).

1. Causal semantics of disentanglement

The causal formulation departs from the mean-field view in which latent factors are assumed independent. Several papers argue that this assumption is mismatched to many real data-generating processes, where high-level factors are causally related, spuriously correlated through confounders, or both. DEAR states that previous methods with independent priors fail to disentangle causally related factors even under supervision, because enforcing independence in the prior conflicts with fitting causally dependent factors (Shen et al., 2020). ReI makes the same point from a collider perspective, where conditioning on the observed datum induces dependencies among otherwise unrelated generating variables (Castorena, 2023).

A foundational causal account models observations $X$ as effects of generative factors $G_1,\dots,G_n$ , possibly influenced by confounders $C$ . In the “disentangled causal process” of Suter et al., there are no directed edges among the $G_i$ , and any statistical dependence between factors is explained by confounders rather than direct causal links (Suter et al., 2018, Reddy et al., 2021). This setup separates causal independence from observational independence: factors may be marginally dependent, and they generally become dependent after conditioning on $X$ , yet still remain causally disentangled.

Several works therefore define causal disentanglement interventionally. C-Disentanglement writes the criterion as

$P\bigl(Z_i \mid \mathrm{do}(Z_{-i}=z_{-i}), X\bigr)=P(Z_i\mid X),$

emphasizing that intervening on other latent coordinates should not alter $Z_i$ when the factors are causally separate (Liu et al., 2023). “Desiderata for Representation Learning: A Causal Perspective” gives an equivalent unsupervised criterion $P[Z_j\mid do(Z_{-j}=z_{-j})]=P(Z_j)$ , and links it to factorized support conditions (Wang et al., 2021). ICM-VAE strengthens the notion further: its definition of causal disentanglement is mechanism-equivalence, meaning alignment not only of coordinates but of the conditional mechanisms of a latent SCM, up to permutation and scaling or elementwise reparameterization (Komanduri et al., 2023).

A common misconception in this literature is that disentanglement is synonymous with statistical independence. The causal literature treats that identification as insufficient. “On Causally Disentangled Representations” explicitly argues that metrics designed for ordinary disentanglement may miss confounding, motivating new causal metrics and datasets (Reddy et al., 2021). This suggests that CDRL is best understood as a refinement of disentanglement in which intervention semantics, rather than mere factorization, define success.

2. Architectural and probabilistic formulations

Most CDRL models are instantiated as VAEs, bidirectional generative models, or related latent-variable architectures augmented with an SCM, a learned DAG, or a causal prior. A central pattern is to replace an isotropic or factorized prior with a structured prior that respects causal dependencies among latent factors.

DEAR replaces the usual factorized latent prior with a learnable SCM prior in a bidirectional generative model, using exogenous Gaussian noise, an invertible transform, and a weighted adjacency matrix $A$ of a DAG over latent variables (Shen et al., 2020). DCVAE instead keeps the generative model in VAE form but inserts a “causal autoregressive flow” into the posterior, parameterized by a lower-triangular adjacency mask $A$ , so that the posterior transport itself follows an assumed or learned causal order (Fan et al., 2023). ICM-VAE uses a flow-based SCM in latent space together with a causal disentanglement prior conditioned on auxiliary labels, with one conditional mechanism per node in the latent causal graph $G_1,\dots,G_n$ 0 (Komanduri et al., 2023). FlexCausal modifies the encoder side by replacing the standard diagonal posterior covariance with a block-diagonal covariance VAE, preserving intra-concept dependencies while keeping distinct concept blocks uncorrelated, and combines this with a factorized flow-based prior over exogenous noise (Jin et al., 29 Jan 2026).

Other model classes adapt the same principles to nonstandard latent spaces or data modalities. CT-VAE treats vector-quantized code slots as causal variables, links them in a causal graph, and learns action-conditioned interventions that produce atomic transitions affecting a unique factor of variation (Gendron et al., 2023). causalPIMA introduces a differentiable DAG parametrization and a Gaussian-mixture prior whose mixture components are identified with outcomes of categorical DAG nodes, enabling fully unsupervised graph learning in multimodal VAEs (Walker et al., 2023). “Object-centric architectures enable efficient causal representation learning” argues that flat Euclidean latents are inadequate for multi-object scenes because the renderer is permutation-invariant over objects and hence non-injective; it therefore combines Slot Attention with sparse perturbation supervision to learn set-structured causal representations (Mansouri et al., 2023). For graph data, CCVGAE introduces a causal disentanglement layer inside a Variational Graph Auto-Encoder, while CC-Meta-Graph transfers the same concept-free causal machinery to meta-learning on graphs (Feng et al., 2023).

Method	Core causal mechanism	Supervision regime
DEAR	SCM prior in a bidirectional generative model	Weak supervision with labels and a super-graph or ordering
DCVAE	Causal autoregressive flow in the posterior	Supervised factor labels
ICM-VAE	Structural causal flow plus causal disentanglement prior	Supervised by causally related observed labels
causalPIMA	Differentiable DAG plus Gaussian-mixture causal prior	Fully unsupervised with multimodal data and known physics
CT-VAE	Causal graph over vector-quantized latent codes	Ordered image pairs with action labels
CCVGAE	Causal disentanglement layer in VGAE	Unsupervised, concept-free

These formulations differ in parameterization, but they share three recurrent design commitments: latent causal structure is made explicit; interventions are part of the model semantics rather than only a downstream analysis tool; and standard diagonal, independent latent assumptions are relaxed whenever they obstruct causal alignment.

3. Identifiability and supervision regimes

A major strand of CDRL concerns identifiability: under what assumptions can latent factors or mechanisms be recovered up to permutation, scaling, diffeomorphism, or related equivalence classes. The strongest results typically rely on auxiliary labels, interventional variation, sparse perturbations, or structural assumptions on graphs.

DEAR proves that, under infinite-capacity encoder and generator, inclusion of the true factor distribution in the SCM family, and identifiability of the DAG structure within that family, any minimizer recovers a component-wise disentangled representation of the causal factors up to one-to-one transformations (Shen et al., 2020). DCVAE proves that, with infinite capacity and supervised alignment of each latent coordinate to the corresponding true factor, any global minimizer satisfies $G_1,\dots,G_n$ 1, again up to one-to-one scalar transforms (Fan et al., 2023). ICM-VAE establishes identifiability of causal factors and mechanisms up to permutation and elementwise reparameterization under smoothness and sufficient-variability conditions (Komanduri et al., 2023).

Confounding creates a distinct identifiability problem. C-Disentanglement introduces confounder labels $G_1,\dots,G_n$ 2 as domain-expert inductive bias and proves an identifiability theorem for mixture-of-Gaussians posteriors whose component covariances are diagonal within each confounder stratum (Liu et al., 2023). ReI attacks a related problem through causal identification in a collider SCM, regularizing the posterior toward the causally identified effect of a factor rather than a purely observational conditional (Castorena, 2023). These works are explicit that purely observational statistical independence is too strong in some regimes and too weak in others.

In more exploratory settings, identifiability is either weaker or depends on structure in the data modality. causalPIMA is fully unsupervised, but it does not claim full uniqueness of the causal model; it returns a plausible DAG learned jointly with multimodal latent features (Walker et al., 2023). For multi-object scenes, the object-centric study shows that respecting set structure reduces the required number of perturbations from $G_1,\dots,G_n$ 3 for a naive $G_1,\dots,G_n$ 4 encoder to $G_1,\dots,G_n$ 5 for an object-centric encoder with shared property projection (Mansouri et al., 2023). In linear causal disentanglement, one perfect intervention on each latent variable is sufficient and, in the worst case, necessary to recover parameters under perfect interventions, using coupled tensor decomposition of higher-order cumulants (Carreno et al., 2024). For dynamical systems, a graphical criterion over the causal graph on $G_1,\dots,G_n$ 6 yields identifiability up to permutation and element-wise diffeomorphism, and local state-dependent causal structure can strictly strengthen identifiability beyond what a single global graph permits (Baumgartner et al., 15 Mar 2026).

The supervision spectrum is therefore broad. Some models require full concept labels, some require only sparse perturbations or a super-graph, some rely on confounder strata or multimodal consistency, and some are unsupervised but correspondingly weaker in identifiability guarantees. A plausible implication is that CDRL trades supervision for structural assumptions rather than eliminating assumptions altogether.

4. Interventions, counterfactuals, and causal evaluation

Interventional semantics are a defining feature of the area. In DCVAE, the learned single-step flow equations are treated as structural equations; a do-intervention replaces one equation and downstream variables are recomputed in topological order before decoding, producing counterfactual images consistent with the learned graph (Fan et al., 2023). DEAR distinguishes ordinary latent traversals from true interventional generation, where fixing one latent and ancestral sampling through the SCM changes descendants but not parents (Shen et al., 2020). ICM-VAE adopts the classical abduction–action–prediction pipeline: infer exogenous noise, intervene on a chosen latent factor, propagate through SCM flows, and decode the counterfactual (Komanduri et al., 2023). FlexCausal adds an explicit counterfactual consistency loss over intervened, descendant, and invariant blocks, together with a manifold-aware relative intervention $G_1,\dots,G_n$ 7 to keep counterfactuals on the semantic manifold (Jin et al., 29 Jan 2026).

Evaluation methodologies reflect this emphasis. The Interventional Robustness Score (IRS) measures how robust a subset of learned latents is to interventions on nuisance generative factors, and comes with an $G_1,\dots,G_n$ 8 estimation algorithm from labeled observational data via backdoor adjustment (Suter et al., 2018). “On Causally Disentangled Representations” introduces Unconfoundedness (UC) and Counterfactual Generativeness (CG), along with the CANDLE dataset, precisely because standard disentanglement metrics can fail to detect confounding (Reddy et al., 2021). “Desiderata for Representation Learning” proposes IOSS, an independence-of-support score for unsupervised causal disentanglement, and relates independent support to a causal criterion under positivity (Wang et al., 2021). CT-VAE adds an action retrieval task, with Action Accuracy and Factor Accuracy, because its discrete causal latent space is organized around transitions rather than only reconstruction fidelity (Gendron et al., 2023).

A recurrent result across papers is that metrics based only on mutual information, total correlation, or ordinary latent traversals may overstate disentanglement. The IRS paper shows that high mutual information can coexist with poor interventional robustness (Suter et al., 2018). The causal perspective papers consequently treat do-based evaluation as indispensable whenever generative factors are correlated or confounded.

5. Empirical regimes and representative findings

Empirical studies span synthetic image benchmarks, scientific multimodal data, graphs, time series, dynamical systems, and reinforcement-learning environments. The most repeated qualitative finding is that independent-prior baselines such as $G_1,\dots,G_n$ 9-VAE, $C$ 0-TCVAE, vanilla VAE, or related factorized methods tend to entangle factors when those factors are correlated or causally related (Fan et al., 2023, Shen et al., 2020).

On supervised image benchmarks, ICM-VAE reports near-perfect DCI disentanglement ( $C$ 1) and outperforms $C$ 2-VAE, iVAE, CausalVAE, and SCM-VAE on Pendulum, Water Flow, and CausalCircuit in both DCI and IRS (Komanduri et al., 2023). ReI reports that VAE+ReI remains above approximately $C$ 3 DCI on dSprites under independent, pairwise, and 1-to-all correlation regimes, while several baselines degrade sharply; on MPI3D, ReI stays around $C$ 4 under correlation where the cited baseline TC-VAE drops from approximately $C$ 5 to approximately $C$ 6 (Castorena, 2023). CCVGAE and CC-Meta-Graph report up to $C$ 7 and $C$ 8 absolute improvements over baselines in AUC, respectively (Feng et al., 2023).

Object-centric CDRL demonstrates a particularly strong sample-efficiency effect. In scenes with $C$ 9 objects and $G_i$ 0 shared properties, the object-centric architecture requires only $G_i$ 1 independent offsets, compared with $G_i$ 2 for a standard vector encoder; empirically, Slot Attention reaches near-perfect MCC on four-object scenes with $G_i$ 3 paired samples, while the comparable injective CNN baseline requires $G_i$ 4 (Mansouri et al., 2023). In dynamical systems, SPARTAN with local sparse attention achieves MCC $G_i$ 5 in settings where a static-graph baseline fails, supporting the claim that local causal structure can be necessary for full identifiability (Baumgartner et al., 15 Mar 2026). In the fully unsupervised multimodal setting of causalPIMA, latent clusters and learned edges align with generative orders in synthetic circles and with physically plausible relations in 3D-printed lattice data (Walker et al., 2023).

This body of results suggests a consistent empirical pattern: when the evaluation regime includes correlation shifts, interventions, or confounding, causal structure in the latent model is often more important than raw reconstruction quality, which several papers report to be similar across competing decoders (Baumgartner et al., 15 Mar 2026, Suter et al., 2018).

6. Applications, limitations, and open problems

CDRL has been applied beyond image synthesis to fairness, control, graphs, and time series. CF-VAE uses structured latent representations to obtain counterfactual fairness, proving that downstream predictors using only the causally ordered non-sensitive representation $G_i$ 6 are counterfactually fair and reporting lower unfairness scores than prior fairness baselines on Law School and Adult (Xu et al., 2022). In multi-UAV collision avoidance, causal representation disentanglement separates causal from non-causal visual factors and improves success rate and SPL in unseen scenarios relative to SAC+RAE (Zhuang et al., 2024). In autonomous driving at unsignalized intersections, CDRL inside a VGAE is used to split graph embeddings into causal and spurious features before graph RL, yielding the highest average reward during training and lower collision rate during testing than learning-based baselines (Wang et al., 30 Jul 2025). CDRL4AD extends the paradigm to multivariate time-series anomaly detection by combining time-lagged causal discovery, graph representations, and a VAE; it reports top-tier F1 and AUC across eight real-world datasets and uses root-cause scores for diagnosis (Kim et al., 13 Oct 2025). CaDeM adapts causal disentanglement to multiplex graphs by separating common and private components through alignment, self-supervision, and backdoor adjustment, improving node and graph tasks on synthetic and real datasets (Nasiri et al., 25 Mar 2026).

Despite this breadth, several limitations recur. Many methods require a known causal graph, a super-graph, concept labels, paired interventions, or confounder annotations (Shen et al., 2020, Jin et al., 29 Jan 2026, Liu et al., 2023). Hidden confounders and selection bias can corrupt structure learning when only partial ordering is known (Shen et al., 2020). Several methods note scaling limits, including challenges beyond roughly ten causal factors, the computational overhead of flow-based priors or entropy-based mutual-information estimators, and the difficulty of large categorical state spaces or large multiplex layer sets (Fan et al., 2023, Jin et al., 29 Jan 2026, Gendron et al., 2023). causalPIMA explicitly notes that, in purely observational unsupervised settings, full identifiability of the unique causal model is not guaranteed (Walker et al., 2023). The object-centric work shows that even apparently benign injectivity assumptions fail in multi-object scenes unless the representation architecture respects permutation symmetry (Mansouri et al., 2023).

A second persistent controversy concerns evaluation. The causal literature repeatedly argues that ordinary disentanglement scores may remain high under confounding, collider bias, or intervention instability, and that causal metrics such as IRS, UC, CG, or IOSS are needed to expose those failures (Suter et al., 2018, Reddy et al., 2021, Wang et al., 2021). This suggests that future progress in CDRL will depend as much on causal evaluation protocols and benchmark design as on encoder or decoder architecture. More broadly, the recent literature indicates a shift from flat latent vectors toward graph-structured, set-structured, temporal, multimodal, and dynamical representations; this suggests that “causal disentanglement” is increasingly being treated as a property of structured latent systems rather than of isolated coordinates alone.