---
title: Causal Abstraction Network (CAN) Framework
url: https://www.emergentmind.com/topics/causal-abstraction-network-can
type: topic
---

# Causal Abstraction Network (CAN) Framework

to=arxiv.search 򐂕json
{"query":"id:2406.00394 OR title:\"Learning Causal Abstractions of Linear Structural Causal Models\"","max_results":5,"sort_by":"submittedDate","sort_order":"descending"}
to=arxiv.search ՞նչjson
{"query":"all:\"Causal Abstraction Network\"","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}
Causal Abstraction Network (CAN) denotes a formal framework for relating a more detailed low-level model to a higher-level model while preserving causal and interventional structure under an explicit alignment. In one widely used formulation, a CAN is the triple $(M,M',\alpha)$, or equivalently the pair $\alpha=(\tau,\omega)$, witnessing that a high-level Structural Causal Model (SCM) faithfully abstracts a low-level SCM. In a more specialized later usage, CAN refers to a network-sheaf construction whose node-stalks are Gaussian SCMs and whose restriction maps are transposes of constructive linear causal abstractions. This suggests that the term names both a general abstraction framework and a particular sheaf-theoretic instantiation of that framework [2301.04709, 2509.25236].

## 1. Core definition and terminological scope

A CAN, in the general causal-abstraction literature, connects a **low-level model** and a **high-level causal model** through an **alignment**. The low-level model may be an SCM or a neural network viewed as a causal model; the high-level model is an interpretable SCM or algorithm; and the alignment specifies how low-level variables or representations implement high-level variables. In one formulation, the low-level SCM is written as $M=(V,U,F)$ and the high-level SCM as $M'=(V',U',F')$, with the CAN given by $(M,M',\alpha)$, where $\alpha$ maps low-level variables and interventions to high-level ones [2301.04709]. In another formulation, the same basic object is presented as a pair $\alpha=(\tau,\omega)$, where $\tau$ is a partial surjection on joint settings and $\omega$ is a partial surjection on interventions [2508.11214].

The alignment is commonly decomposed into a partition of low-level variables into cells and a family of surjective maps. For each high-level variable $X'$, one specifies a cell $\Pi_{X'}$ of low-level variables and a surjection
\[
\tau_{X'}:\mathrm{Val}(\Pi_{X'}) \to \mathrm{Val}(X').
\]
Collecting these induces a partial map from low-level joint assignments to high-level joint assignments, and likewise a map on partial settings or interventions [2301.04709].

A concise way to organize the main usages is the following.

| Usage | Core object | Representative papers |
|---|---|---|
| General causal-abstraction usage | Low-level model, high-level model, and an alignment such as $(M,M',\alpha)$ or $\alpha=(\tau,\omega)$ | [2301.04709], [2508.11214] |
| Mechanistic-interpretability usage | A low-level network, a high-level causal hypothesis or algorithm, and an alignment tested by interchange interventions | [2106.02997], [2303.02536] |
| Sheaf-theoretic Gaussian usage | A network sheaf whose node-stalks are Gaussian SCMs and whose restriction maps are transposes of constructive linear causal abstractions | [2509.25236], [2602.02623] |

A common source of confusion is that these usages are related but not identical. One paper explicitly states that a CAN is “not a new data structure but rather the pair $\alpha=(\tau,\omega)$ witnessing” that a complex micro-causal model implements a simpler macro-causal model [2508.11214], whereas another introduces CAN as a specific instance of network sheaves of causal knowledge [2509.25236].

## 2. Interventional semantics and constructive abstraction

The central requirement in causal abstraction is that interventions commute with the abstraction mapping. In deterministic settings with unique solutions, the constructive-abstraction condition is expressed by the commuting-diagram equation
\[
\tau(\mathrm{Solve}(M_i)) = \mathrm{Solve}(M'_{\tau(i)})
\]
for every low-level intervention $i$ in the domain of $\tau$ [2301.04709]. In probabilistic form, the requirement becomes preservation of interventional distributions:
\[
\forall i:\ \tau_* P^{M_i} = P^{M'_{\tau(i)}}.
\]
This is the sense in which the high-level model is a faithful simplification of the low-level one.

Earlier formulations describe the same idea as **$T$-abstraction**. There, a surjective map
\[
T:R(V_L)\to R(V_H)
\]
induces a mapping on interventions $W_T$, and the abstraction condition is that for every low-level intervention $i_L$,
\[
T\circ i_L(M_L)=W_T(i_L)(M_H)
\]
as functions from $R(V_L)$ to $R(V_H)$ [2106.02997]. A **constructive abstraction** is the special case in which $T$ factorizes along disjoint blocks of low-level variables, one block per high-level variable.

The framework was later generalized from mechanism replacement to **mechanism transformation**. Instead of requiring that a high-level mechanism arise only by replacing a low-level submodel with a single node, one defines a mechanism-transformation functional
\[
T:\{\text{low-level mechanisms on }\Pi_{X'}\}\to \mathcal F'
\]
so that
\[
f'_{X'} = T(\{f_Y:Y\in \Pi_{X'}\};\tau_{X'}).
\]
This accommodates marginalization, variable-merge, value-merge, and arbitrary user-defined transforms [2301.04709].

A further generalization recasts abstraction as natural transformations between compositional models. That account distinguishes **downward abstractions**, which map high-level queries to low-level queries, from **upward abstractions**, which map low-level concrete queries such as Do-interventions to high-level queries. It also introduces a stronger **component-level** notion in which individual mechanisms, not only whole query behaviors, are matched by a natural transformation [2602.16612]. This suggests that constructive abstraction, $Q$-$\tau$ consistency, interchange-based abstractions, and distributed alignments can be understood within a single compositional formalism.

## 3. Linear SCM CANs and causal discovery

For linear SCMs, the abstraction problem becomes algebraically explicit. In the low-level model,
\[
X = B_{\text{low}}X + E,
\]
where $B_{\text{low}}$ is strictly upper-triangular and the exogenous noise terms are mutually independent and non-Gaussian. In the high-level model,
\[
Z = B_{\text{high}}Z + U,
\]
with $B_{\text{high}}$ upper-triangular and $U$ independent and non-Gaussian. A linear abstraction is specified by a surjective linear map
\[
f(x)=Ax,\qquad A\in \mathbb R^{b\times d},\quad \mathrm{rank}(A)=b
\]
[2406.00394].

Interventional consistency implies the matrix condition
\[
A(I-B_{\text{low}})^{-1}=(I-B_{\text{high}})^{-1}G,
\]
for some linear exogenous map $G$, and the paper derives the key algebraic condition
\[
B_{\text{high}} = A B_{\text{low}} A^+,
\]
up to pseudo-inverses of $A$ or $G$ [2406.00394]. The same work proves two structural results. First, **Abstract Connectivity** states that an edge $Z_i\to Z_j$ exists in the high-level graph if and only if every low-level variable participating in $Z_i$ has a directed path to at least one low-level variable participating in $Z_j$, where the path may pass only through irrelevant low-level nodes. Second, **Block Ordering** states that the low-level ordering is a refinement of the high-level ordering: each abstract variable defines a disjoint block of low-level nodes, and these blocks can be topologically sorted in the same order as the high-level DAG [2406.00394].

These results yield a necessary-and-sufficient graphical test: a high-level model is an abstraction of a low-level model under $A$ if and only if the induced block structure is disjoint and respects the topological order, and the per-edge algebraic constraint $B_{\text{high}}=AB_{\text{low}}A^+$ holds [2406.00394].

The same paper introduces **Abs-LiNGAM**, a learning procedure that estimates the abstraction map by least squares on paired $(X,Z)$ data, constructs an artificial high-level dataset by mapping low-level samples through the learned abstraction, estimates the high-level DAG with DirectLiNGAM, infers forbidden ancestral relations between blocks, and then runs DirectLiNGAM on the low-level data subject to those constraints. The paper states that Abs-LiNGAM is “the first algorithm to leverage these constraints in practice,” and reports that with as few as $|D_J|\simeq d$ paired samples it attains the same ROC-AUC $(\approx 0.97)$ as vanilla LiNGAM on $25$–$50$-node graphs, while run times drop from $\approx 60\ \mathrm{s}$ to $\approx 40\ \mathrm{s}$ on average in the small-graph setting [2406.00394]. A plausible implication is that causal abstraction can function not only as an explanatory relation but also as a scalability device for causal discovery.

## 4. Mechanistic interpretability and distributed alignments

In mechanistic interpretability, a CAN is typically instantiated by a low-level neural network, a high-level causal or algorithmic model, and an alignment between internal representations and interpretable variables. The standard empirical test is the **interchange intervention**: one runs the model on a base input, patches in the activation or subspace value taken from a source input, and compares the resulting output with the counterfactual output of the high-level model under the corresponding intervention [2106.02997].

The corresponding accuracy metric is **Interchange Intervention Accuracy** (IIA). In one formulation, IIA is the fraction of test interventions for which
\[
\arg\max_y p_{\mathcal N}(y\mid x,I_{\mathrm{DNN}})=\mathcal A(x,I_{\mathrm{alg}}),
\]
where $\mathcal N$ is the network and $\mathcal A$ is the algorithmic model [2507.08802]. In practice, exact constructive abstraction is rarely available, so IIA is used as a graded measure of approximate abstraction [2303.02536].

A major development was **Distributed Alignment Search (DAS)**, which removes two restrictions of earlier methods: brute-force search over alignments and the assumption that high-level variables align with disjoint sets of neurons. DAS parameterizes an orthogonal rotation $R^\theta$ of a representation space and performs **distributed interchange intervention** by rotating into a basis with designated subspaces, injecting source values into selected subspaces, preserving the base component in the complement, and rotating back. The rotation is learned by gradient descent on an interchange-intervention training objective [2303.02536].

Empirically, DAS was used on **Hierarchical Equality** and **Monotonicity NLI (MoNLI)**. On Hierarchical Equality, DAS finds a rotation aligning two disjoint subspaces of a hidden layer to the latent comparisons $[w=x]$ and $[y=z]$ with $100\%$ IIA, whereas brute-force localist search reaches at most approximately $.60$ and the best disjoint window approximately $.73$. On MoNLI, DAS on the CLS vector in layer $9$ yields perfect IIA for a high-level model with variables for negation and lexical entailment [2303.02536].

Earlier causal-abstraction analysis of MQNLI reported a different but related methodology: candidate layer-position alignments were evaluated by building a graph whose edges indicate successful pairwise interchange behavior, and the size of the largest clique measured the extent of perfect abstraction on a subset of the data. In that study, a BERT model displayed large cliques for several natural-logic subcomputations, while a BiLSTM baseline did not [2106.02997]. This suggests that CAN methodology is sensitive not only to whether a model predicts correctly, but to whether its internal counterfactual structure matches an interpretable causal hypothesis.

## 5. Non-linearity, vacuity, and the representation dilemma

A central controversy concerns the complexity of the alignment map. One paper proves an impossibility result for **input-restricted distributed abstraction**: under assumptions including countable input space, layer-wise injectivity, layer-wise strict surjectivity on the final arg max, a matchable partial order, and perfect task performance, there always exists a potentially highly non-linear bijective alignment $\tau$ making any algorithm an input-restricted distributed abstraction of any suitable network [2507.08802].

The paper’s empirical results sharpen the point. On the hierarchical equality toy MLP, an identity alignment fails with $\mathrm{IIA}\approx .5$, linear alignments reproduce the usual pattern of strong early-layer but weaker deeper-layer performance, and a non-linear RevNet alignment yields approximately $100\%$ IIA in every layer. On Pythia language models for indirect object identification, even randomly initialized models reach $\mathrm{IIA}>80\%$ with RevNet alignment despite zero task accuracy, and smaller models incapable of solving the task still admit perfect RevNet alignments [2507.08802].

The paper terms this the **non-linear representation dilemma**: if $\tau$ is unconstrained, causal abstraction becomes vacuous; if $\tau$ is restricted, the conclusions depend critically on a representational hypothesis such as linearity, sparsity, or a privileged basis [2507.08802]. A common misconception is therefore that causal abstraction alone is sufficient for mechanistic interpretability. The literature instead supports a narrower claim: causal abstraction provides a framework for mechanistic interpretability, but meaningful conclusions require assumptions about how information is encoded.

## 6. Sheaf-theoretic CANs and Gaussian consistency

A later line of work introduces CAN as a specific instance of **network sheaves of causal knowledge**. Here one has a graph whose nodes index zero-mean Gaussian SCMs and whose edges carry abstraction maps between them. In one formulation, a **consistent CAN** consists of a zero-mean Gaussian SCM at each node and, for each ordered pair $(\ell\to h)$ in a partial order, a linear map $P_{h\ell}$ with orthonormal rows,
\[
P_{h\ell}P_{h\ell}^\top = I_h,
\]
such that whenever $\ell\to m\to h$ are composable one has
\[
P_{h\ell}=P_{hm}P_{m\ell}.
\]
The condition $P_{h\ell}P_{h\ell}^\top=I_h$ is called the **Semantic Embedding Principle (SEP)** [2602.02623].

The Gaussian node-stalk at node $i$ is determined by a linear SEM
\[
X^{(i)} = B^{(i)}X^{(i)}+\epsilon^{(i)},
\]
with covariance
\[
\Sigma^{(i)}=(I-B^{(i)})^{-1}\Omega^{(i)}(I-B^{(i)})^{-T}.
\]
A constructive linear causal abstraction from a lower-dimensional or more detailed level to a higher-level one is constrained by a spectral condition: if $\Sigma_\ell$ and $\Sigma_h$ have ordered eigenvalues $0<\lambda_1\le \cdots \le \lambda_\ell$ and $0<\kappa_1\le \cdots \le \kappa_h$, then a CLCA can exist only if
\[
\lambda_i \le \kappa_i \le \lambda_{i+\ell-h},\qquad i=1,\dots,h.
\]
Under this condition one can choose $P_{h\ell}$ so that
\[
P_{h\ell}^\top \Sigma_\ell P_{h\ell}=\Sigma_h
\]
[2602.02623].

Related work develops the algebraic side of this construction. CANs have adjacency, incidence, and Laplacian block operators; their $0$-cochains are node assignments and their $1$-cochains are edge assignments; and the space of global sections is characterized by the Laplacian kernel. One theorem states that if the coarsest model has dimension $h$ and all nodes are reachable from it along directed edges, then $\dim \ker L = h$, and a $0$-cochain is a global section if and only if it lies in $\ker L$ [2509.25236]. This imports cohomology, consistency, and smoothness into the analysis of abstraction networks.

Learning in this setting is formulated as a sum of Kullback–Leibler divergences over candidate edges and decomposes into edge-specific local problems. The proposed solver, **SPECTRAL**, uses iterative updates with closed-form substeps and is suitable for positive definite and semidefinite covariance matrices [2602.02623]. Experiments on synthetic data report competitive performance in the causal-abstraction learning task and successful recovery of chain, star, and tree CAN structures, with false positive rate $0$ in all reported global-structure settings and true positive rate above $95\%$ at $R=100$ restarts [2602.02623].

## 7. Approximate discovery, sparsification, and diagnostic bucketing

Recent work extends CAN methodology beyond exact abstraction and explicit graph learning. One direction reframes structured pruning as search over **approximate causal abstractions**. A trained feed-forward network is treated as a deterministic SCM; a reduced SCM is obtained by hard or soft interventions that replace units with constants or affine surrogates; and faithfulness is measured by an **Interventional Risk**
\[
\mathcal L_{\mathcal I}(M_H;M_L)
=
\mathbb E_{x\sim \mathcal D}\,
\mathbb E_{I\sim \mathcal I}\,
d\bigl(\tau(M_L^{\omega(I)}(x)),\,M_H^I(x)\bigr),
\]
which is equivalent to maximizing IIA when $d$ is $0$–$1$ [2602.24266]. A second-order Taylor surrogate yields closed-form constant and affine replacements, and under stationarity and uniform curvature the score reduces exactly to variance-based pruning:
\[
c_j^*=\bar A_j,\qquad s_j=\frac{\alpha}{2}\mathrm{Var}(A_{:,j}).
\]
This both recovers variance-based pruning as a special case and clarifies when it fails [2602.24266].

Another direction treats CAN evaluation as a diagnostic rather than a single scalar score. Instead of reporting one global IIA, **bucketing** builds an interchangeability graph over correct inputs, places edges between interchange-consistent pairs, and identifies dense quasi-cliques as well-interpreted regions. A perfectly interpreted region is a clique; in practice the paper uses $\gamma$-quasi-cliques, with $\gamma=0.98$ in the reported experiments [2605.02234]. The resulting four-step recipe—filter correct inputs, choose a candidate alignment, build the interchangeability graph and extract dense regions, then train a classifier to generalize the bucket boundary—was used to discover missing distinctions and previously unmodeled intermediate variables in toy logic, entity-binding, and RAVEL settings [2605.02234].

These approximate and diagnostic extensions preserve the central CAN idea: explanation depends on how low-level interventions translate to high-level ones. They also indicate a broader methodological shift. CANs are no longer used only to certify exact abstraction; they are used to search for sparse intervention-faithful abstractions, to localize regions of faithfulness, and to refine high-level hypotheses when a proposed abstraction fails.

Source: https://www.emergentmind.com/topics/causal-abstraction-network-can