Causal Abstractions Overview

• Causal abstractions are formal relationships between causal models that map low-level details to high-level descriptions while preserving intervention outcomes.
• They employ frameworks like (τ,ω) mappings, abstraction tuples, and categorical natural transformations to ensure interventional consistency.
• Applications range from mechanistic interpretability in neural networks to data-efficient causal modeling in scientific and industrial settings.

Causal abstractions are formal relations between causal models that describe the same system at different levels of granularity. In the literature, the recurring requirement is not merely observational agreement, but agreement under interventions: after mapping low-level states, variables, or interventions into a higher-level description, the interventional behavior should commute with abstraction, exactly or approximately. This idea appears in functional frameworks based on $(\tau,\omega)$ or $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$, in graphical frameworks such as Cluster DAGs, in linear $T$-abstractions, and in categorical formulations based on natural transformations between Markov functors (Zennaro et al., 2023, Englberger et al., 6 Oct 2025).

1. Formal core: SCMs, state maps, and interventional consistency

The common substrate is the structural causal model. In the discrete setting used by Zennaro et al., an SCM $M$ consists of a DAG together with exogenous variables $\mathcal U$, endogenous variables $\mathcal X$, structural functions $\mathcal F$, and a distribution $P(\mathcal U)$, with mechanisms of the form

$x_i = f_i(pa(X_i)).$

Interventions replace structural functions by constants and induce intervened SCMs $M_\iota$ (Zennaro et al., 2023).

Two closely related functional idioms recur. One is the $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$0 framework, where $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$1 maps low-level states to high-level states and $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$2 maps low-level interventions to high-level interventions. In the formulation adopted by COTA, exactness requires

$\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$3

so intervening on the base model and then pushing forward through $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$4 yields the same distribution as intervening directly on the abstract model (Felekis et al., 2023). The other idiom is the abstraction tuple

$\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$5

where $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$6 is the set of relevant low-level variables, $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$7 is a surjective variable map, and $\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$8 is a family of surjective state maps from low-level configurations to high-level values (Zennaro et al., 2023).

For interventional consistency, Zennaro et al. encode interventional distributions as stochastic matrices and require the commutation condition

$\boldsymbol{\alpha}=\langle R,a,\alpha\rangle$9

where $T$0 and $T$1. This expresses the idea that abstracting the outcome of a low-level intervention and intervening directly at the high level should agree on the resulting distribution over coarse variables (Zennaro et al., 2023).

The same general idea can be restricted to particular layers of Pearl’s causal hierarchy. In "Neural Causal Abstractions," Xia and Bareinboim define $T$2-$T$3 consistency for specific causal queries and $T$4-$T$5 consistency for entire levels $T$6. Their construction makes explicit that an abstraction can be valid for observational questions while failing for interventional ones, or valid for interventional questions while failing for counterfactuals (Xia et al., 2024).

2. Major formalisms: graphical, linear, soft-interventional, and categorical

One major axis in the literature is the relation between graphical and functional notions of abstraction. "Aligning Graphical and Functional Causal Abstractions" proves an equivalence between the class of Cluster DAGs, consistent $T$7-abstractions with the range of abstracted variables mapped bijectively, and constructive $T$8-abstractions. The same paper extends this picture with Partial Cluster DAGs, which relax full clustering by allowing a remainder set of variables while preserving mediated causal and confounding effects (Schooltink et al., 2024).

A second axis is the linear setting. In "Learning Causal Abstractions of Linear Structural Causal Models," low-level and high-level SCMs are linear, and the abstraction is a linear map $T$9 with

$M$0

Under linearity, strong abstraction with all hard interventions allowed, and no cancelling paths, constructive abstraction is forced: the relevant-variable sets $M$1 are non-empty and pairwise disjoint. The paper further characterizes abstract edges through $M$2-direct paths, proves a block ordering theorem linking high-level causal order to low-level blocks, and derives the parametric condition

$M$3

as the core block abstraction constraint (Massidda et al., 2024).

A third axis concerns intervention type. "Causal Abstraction with Soft Interventions" extends the Beckers–Halpern framework from hard interventions to soft interventions that replace mechanisms by possibly non-constant functions without adding new causal connections. A central result is that low soft abstraction is ambiguous when soft interventions are allowed at the high level, and that the stronger definition of soft abstraction removes the ambiguity by requiring consistency for all endogenous–exogenous pairs, thereby guaranteeing a unique intervention map $M$4 with a specific and necessary explicit form (Massidda et al., 2022).

Categorical work recasts these constructions at a more structural level. "Causal Abstractions, Categorically Unified" treats causal models as Markov functors $M$5 and causal abstractions as deterministic natural transformations $M$6, where $M$7 embeds the restricted free category of the high-level graph into that of the low-level graph. This yields an equivalence, in $M$8, between the categorical definition and a standard interventional consistency criterion. "Causal and Compositional Abstraction" then distinguishes downward abstractions, mapping high-level queries to low-level queries, from upward abstractions, mapping concrete low-level queries to high-level ones, and introduces a stronger component-level or mechanism-level abstraction (Englberger et al., 6 Oct 2025, Lorenz et al., 18 Feb 2026).

3. Learning causal abstractions from data

A central recent development is the transition from declarative abstraction criteria to learning procedures. Zennaro et al. consider the setting where the two SCMs $M$9 and $\mathcal U$0 and the structural part $\mathcal U$1 are given, but the state-level maps $\mathcal U$2 are unknown. They formulate abstraction learning as minimizing overall abstraction error over a combinatorial space of surjective maps, relax the discrete search to stochastic matrices parameterized by unconstrained weights $\mathcal U$3, and optimize a loss combining a Jensen–Shannon-based commutativity term with a surjectivity penalty. Their main methodological claim is that the subproblems induced by multiple interventions should be solved jointly rather than independently or sequentially (Zennaro et al., 2023).

In the linear non-Gaussian regime, "Learning Causal Abstractions of Linear Structural Causal Models" introduces Abs-LiNGAM. The method first estimates the linear abstraction $\mathcal U$4 from a small paired dataset, then learns a high-level model on synthesized abstract data, uses the learned abstract graph to derive forbidden low-level paths, and finally runs DirectLiNGAM with these constraints on the low-level data. The paper reports that, once the number of paired samples is sufficient, Abs-LiNGAM matches the structure-recovery quality of standard DirectLiNGAM while significantly reducing runtime and improving scalability (Massidda et al., 2024).

Xia and Bareinboim pursue a different route in "Neural Causal Abstractions." They define abstraction constructively from intervariable clusterings and intravariable clusterings, characterize the existence of full $\mathcal U$5-consistent abstractions through the Abstract Invariance Condition (AIC), and then connect abstract identifiability to ordinary identifiability through the Dual Abstract ID theorem: $\mathcal U$6 Their NeuralAbstractID procedure trains a $\mathcal U$7-constrained Neural Causal Model over abstract variables, while Representational NCMs learn the abstraction map itself through an encoder–decoder objective in high-dimensional settings (Xia et al., 2024).

COTA removes the assumption that SCMs are fully specified. Instead, it assumes only the DAGs, a finite intervention set, an intervention map $\mathcal U$8, and observational and interventional samples from both levels. The method formulates causal abstraction learning as a multi-marginal optimal transport problem with a causally informed cost $\mathcal U$9 and chain-wise penalties derived from truncated factorization and do-calculus constraints. The resulting objective is jointly convex in the transport plans and strictly convex under entropy regularization (Felekis et al., 2023).

4. Neural networks, mechanistic interpretability, and computational explanation

In neural-network analysis, causal abstraction provides a precise sense in which an internal representation realizes an interpretable variable. "Causal Abstractions of Neural Networks" aligns hidden representations with variables in a natural-logic causal model and uses interchange interventions to test whether replacing a neural representation has the same causal effect as replacing the aligned high-level variable. In the MQNLI case study, a BERT-based model with state-of-the-art performance successfully realizes parts of the natural logic model’s causal structure, whereas a simpler baseline model fails to show any such structure (Geiger et al., 2021).

"Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability" pushes this agenda further. It generalizes the theory from mechanism replacement to arbitrary mechanism transformation, defines typed and cyclic variants, decomposes constructive abstraction into marginalization, variable merge, and value merge, and uses this language to unify activation and path patching, causal mediation analysis, causal scrubbing, causal tracing, circuit analysis, concept erasure, sparse autoencoders, differential binary masking, distributed alignment search, and steering. It also formalizes graded faithfulness through $\mathcal X$0-on-average abstraction and connects this directly to interchange intervention accuracy (Geiger et al., 2023).

The paper "Combining Causal Models for More Accurate Abstractions of Neural Networks" addresses the empirical fact that a single algorithmic hypothesis is often only partially faithful. Its proposal is to combine several simple high-level models and let the active causal model depend on the input region $\mathcal X$1, thereby treating the network as occupying different computational states for different inputs. The paper reports a trade-off between the strength of an interpretability hypothesis, defined as the number of inputs explained by the high-level models, and its faithfulness, defined as interchange intervention accuracy (Pîslar et al., 14 Mar 2025).

A broader philosophical synthesis appears in "How Causal Abstraction Underpins Computational Explanation." There, implementation is treated as abstraction-under-translation: a high-level computational model $\mathcal X$2 is implemented by a low-level system $\mathcal X$3 only if $\mathcal X$4 is a constructive abstraction of a translation of $\mathcal X$5. This links representational vehicles to low-level clusters singled out by the abstraction and ties explanatory relevance to prediction and generalization rather than to arbitrary state mappings (Geiger et al., 15 Aug 2025).

5. Scientific, industrial, and decision-theoretic applications

Causal abstraction has already been used for multi-resolution scientific modeling. Zennaro et al. study lithium-ion battery manufacturing, where a fine-grained WMG coating model with multiple mass-loading outputs is related to a coarser LRCS model with a single mass-loading variable. After learning an abstraction, transported WMG data improve downstream prediction of LRCS mass loading: the reported mean-squared error is approximately $\mathcal X$6 for LRCS-only training, $\mathcal X$7 when transported WMG data with support are added, and $\mathcal X$8 when transported WMG data are added without direct support at the target comma-gap value (Zennaro et al., 2023).

COTA studies the same electric battery manufacturing problem under weaker assumptions, using only DAGs and samples rather than fully specified SCMs. In that setting, the learned abstraction again serves as a data-augmentation mechanism across laboratories. The paper reports that in one setup training on LRCS data with $\mathcal X$9 and testing on LRCS data with $\mathcal F$0, the earlier abstraction-learning framework yields $\mathcal F$1 MSE, whereas COTA yields $\mathcal F$2 (Felekis et al., 2023).

The abstraction idea also supports formal transfer of causal reasoning across graphs. In the categorical ADMG setting, applying Pearl’s do-calculus rules on a high-level graphical abstraction of an ADMG yields valid equalities for the low-level model, extending earlier statements from cluster-based abstractions to a more general graphical setting (Englberger et al., 6 Oct 2025). In decision making, AT-UCB exploits an approximate $\mathcal F$3-abstraction between two causal multi-armed bandits, explores within a cheap-to-simulate and coarse-grained CMAB, and then runs UCB on a restricted base action set. Its analysis uses two Wasserstein-based discrepancy measures—Interventional Consistency error and Reward Discrepancy error—to justify elimination of arms whose abstract images are safely suboptimal (Dyer et al., 4 Sep 2025).

6. Limitations, debates, and open problems

A persistent limitation is that abstraction validity depends on what is required to be preserved. Xia and Bareinboim’s Abstract CHT states that even if $\mathcal F$4 is $\mathcal F$5-$\mathcal F$6 consistent with $\mathcal F$7, it will almost never be $\mathcal F$8-$\mathcal F$9 consistent for $P(\mathcal U)$0. Their drug-study example makes the same point concretely: a model may be $P(\mathcal U)$1-$P(\mathcal U)$2 consistent but not $P(\mathcal U)$3-$P(\mathcal U)$4 consistent. A common misconception, therefore, is that matching observational distributions is sufficient for causal abstraction; the literature treats this as too weak for preserving causal semantics (Xia et al., 2024).

Another debate concerns expressivity versus tractability. Soft abstraction shows that permitting high-level soft interventions without strengthening the definition makes the intervention map $P(\mathcal U)$5 non-unique; the stronger endogenous–exogenous consistency criterion fixes this, but at a substantial formal cost (Massidda et al., 2022). Likewise, the equivalence between graphical and functional abstractions currently relies on bijective range maps in the alignment results for CDAGs and PCDAGs, so genuine value coarse-graining falls outside that equivalence theorem (Schooltink et al., 2024).

Assumptions also remain strong in the main learnable regimes. The linear theory relies on linearity, acyclicity, independent non-Gaussian noise, causal sufficiency at the abstract level, and a given abstract dimension $P(\mathcal U)$6 (Massidda et al., 2024). Neural Causal Abstractions require a chosen C-DAG and either manually specified or learned clusterings (Xia et al., 2024). The survey literature accordingly notes that current theory is still centered on equivalence rather than true abstraction in the broader interpretability sense, and that scalable discovery of abstractions, robust approximate abstractions, and frameworks tailored specifically to modern large models remain open (Zhang, 2024).

The categorical literature enlarges the design space rather than closing it. It shows that constructive and strong $P(\mathcal U)$7-abstractions correspond to different monoidal assumptions, that strong $P(\mathcal U)$8-abstractions need not be constructive, and that component-level abstraction yields a strengthened mechanism-level variant of constructive abstraction (Englberger et al., 6 Oct 2025, Lorenz et al., 18 Feb 2026). Taken together, these results suggest that future work will have to resolve several tensions at once: exactness versus approximation, learnability versus formal completeness, simple graphical structure versus distributed representation, and faithful abstraction versus explanatory usefulness.