Active Causal Learning Approaches

• Active causal learning is a paradigm where adaptive data collection is used to design experiments for unveiling causal structures and mechanisms.
• It employs sequential interventions based on current uncertainty to improve model identifiability and directly target causal estimands.
• Algorithmic strategies, ranging from Bayesian design to reinforcement learning, enhance sample efficiency across diverse application domains.

Active causal learning denotes a family of sequential design problems in which a learner uses current uncertainty about a causal target to choose the next experiment, intervention, or outcome query, rather than relying on a fixed dataset. Across the literature, the target may be a causal graph, a set of structural mechanisms, an interventional query, or a broader causal quantity such as the ATE, ATTE, ITE, CATE, ATT, or distribution-shifted effects. The shared principle is adaptive data acquisition: information gathered at one step changes which action is most informative at the next step (Kügelgen et al., 2019, Toth et al., 2022, Song et al., 2023, Gao et al., 29 Sep 2025).

1. Conceptual scope

A canonical active causal learning loop begins with initial data, maintains a belief state over causal hypotheses, evaluates candidate actions by an explicit utility, acquires new data, and updates the belief state before repeating. In graph-learning settings, the latent object is typically a DAG or a Markov equivalence class; in Bayesian formulations it may be a full structural causal model $M$, a graph $G$, graph-specific functions $f_i$, or a query $Y=q(M)$. Actions may be interventions of the form $do(X_j=x)$, single-vertex or multi-vertex intervention targets, treatment assignments, unit-selection queries, or factual-outcome queries in an observational pool (Kügelgen et al., 2019, Toth et al., 2022, Gao et al., 29 Sep 2025).

The topic is therefore broader than active causal discovery in the narrow sense. Some papers study discovery of graph structure from interventional data; others assume the graph is already known and actively learn the parent-to-child mechanisms; others directly target downstream estimands without first requiring a fully specified graph. In ActiveCQ, for example, the learner does not intervene to generate new treatments, but instead chooses which already-existing factual outcomes in a pool to observe; in the mechanism-learning setting, the graph is fixed and the interventions are selected to reduce posterior uncertainty about the unknown functions $f_n$ (Rubenstein et al., 2017, Gao et al., 29 Sep 2025).

This breadth matters because the object of interest changes the notion of informativeness. A graph-oriented utility rewards experiments that orient edges or shrink posterior mass over structures; a query-oriented utility rewards experiments that reduce uncertainty about a causal quantity; a mechanism-oriented utility rewards experiments that reduce functional risk. A plausible implication is that “active causal learning” is best understood as a design principle—choose data adaptively with respect to a causal target—rather than as a single algorithmic template.

2. Structure learning, identifiability, and causal representations

In structure-learning formulations, the central obstacle is identifiability. From observational data alone, a DAG is generally identifiable only up to Markov equivalence: different DAGs can share the same skeleton and v-structures, and hence the same conditional independences. Interventions improve identifiability because they delete incoming edges into the intervened variables and thereby refine the observational CPDAG into an $\mathcal I$-essential graph. This logic underlies both classical essential-graph methods and sequential policies defined on partially oriented graphs (Hauser et al., 2012, Amirinezhad et al., 2020).

One line of work optimizes identifiability directly on the current essential graph. A greedy single-vertex strategy chooses the next vertex $v$ to minimize the worst-case number of undirected edges remaining after the intervention, while an arbitrary-size target strategy minimizes the worst-case clique number of the remaining undirected chain graph. The latter yields the sharp bound

$$\min_{I' \subset [p]} \max_{D \in \mathbf D(G)} \omega\!\left(\mathcal E_{\mathcal I\cup\{I'\}}(D)\right) = \left\lceil \frac{\omega(G)}{2} \right\rceil,$$

from which it follows that

$k=\left\lceil \log_2(\omega(G))\right\rceil$

intervention targets are sufficient and, in the worst case, necessary for full identifiability. The constructive algorithm uses proper coloring of chordal chain components and runs in linear time $G$0 (Hauser et al., 2012).

A different line addresses continuous random variables with nonlinear mechanisms. There the environment is modeled as a causally sufficient nonlinear additive noise SCM,

$G$1

with GP priors over the mechanisms,

$G$2

The intervention $G$3 is chosen to maximize expected information gain about the graph, but the intervention value $G$4 lies in a continuous domain $G$5, making the action space uncountable. The resulting acquisition is approximated by Monte Carlo over interventional outcomes and optimized by Bayesian optimization, specifically GP-UCB, over $G$6 for each fixed target $G$7 (Kügelgen et al., 2019).

The representation need not be a DAG. In causal probability trees, the causal process is a root-to-leaf generative tree, and different branches can encode different causal orders. The paper on active learning of causal probability trees states that trees are “strictly more general” than standard causal graphs because they allow context-dependent causal dependencies. Interventions $G$8 are scored by expected information gain over a finite hypothesis class of trees, and in the two-variable case the expected gain reduces to a Jeffrey divergence between $G$9 and $f_i$0 (Herlau, 2022).

Taken together, these formulations suggest that active causal structure learning is defined less by a single causal representation than by a recurring task: use interventions to break residual symmetries among competing causal explanations.

3. Query-specific active causal inference

A major development is the move from graph-first design to query-specific design. In Active Bayesian Causal Inference (ABCI), the unknown SCM $f_i$1 is latent, the quantity of interest is written as $f_i$2, and the learner works directly with the query posterior

$f_i$3

The next intervention is selected by maximizing the mutual information

$f_i$4

so the experiment is chosen for its value to the target query rather than for full graph recovery. The paper explicitly argues that the standard two-stage pipeline—first infer $f_i$5, then estimate the effect of interest—is often uneconomical because the query may not require a fully specified model (Toth et al., 2022).

ACE develops the same query-conditioned logic for expensive experiments in potential-outcome settings. It fits separate GP models for

$f_i$6

and defines acquisition functions by posterior variance reduction for the causal estimand of interest. For scenario 1, treatment is chosen by

$f_i$7

for scenario 2A the method chooses both unit and treatment,

$f_i$8

for scenario 2B it uses the expected variance reduction

$f_i$9

and for scenario 3 it uses an ACE-UCB rule for ITE targeting. The supported estimands include ATE, ATTE, ATO, other WATE-type quantities, and ITE (Song et al., 2023).

Two recent frameworks generalize this query-orientation further. ActiveCQ writes many causal quantities as integrals of a regression function,

$Y=q(M)$0

models $Y=q(M)$1 with a GP, and derives acquisition rules from the posterior uncertainty of the causal quantity itself, with utilities based on information gain and total variance reduction. Causal-EPIG advances the principle of “causal objective alignment,” arguing that acquisition should target unobservable causal quantities rather than indirect proxies, and instantiates this with

$Y=q(M)$2

and

$Y=q(M)$3

The former targets joint potential outcomes; the latter targets the CATE directly (Gao et al., 29 Sep 2025, Gao et al., 26 Sep 2025).

The same query-centric pattern appears under interference. Active Learning in Causal Inference with Interference models direct and spillover effects as functions of the neighborhood-treatment intensity $Y=q(M)$4, selects the most uncertain setting by

$Y=q(M)$5

and then uses a genetic algorithm to find concrete network treatment assignments that approximate the targeted $Y=q(M)$6 (Zhu et al., 2024).

This suggests a general reorientation of the field: instead of asking which experiment best reconstructs an entire latent causal world model, many recent methods ask which experiment best sharpens the specific causal quantity that will actually be used downstream.

4. Mechanism learning and adaptive model revision

A separate strand assumes that graph structure is known or sufficiently trusted and treats active causal learning as mechanism identification. In “Probabilistic Active Learning of Functions in Structural Causal Models,” the SCM is fixed, each structural mechanism has a GP prior,

$Y=q(M)$7

and learning quality is measured by a weighted integrated squared error,

$Y=q(M)$8

The posterior mean minimizes the expected risk, and the intervention value is scored by the myopic criterion

$Y=q(M)$9

The paper gives both a sampling-based approximation and a dynamic-programming decomposition for certain chain-structured intervention sets (Rubenstein et al., 2017).

Optimal intervention design in known causal models can also be framed as active learning about the intervention itself. In linear Gaussian SCMs with known DAG and shift interventions,

$do(X_j=x)$0

the infinite-sample optimal intervention for target mean $do(X_j=x)$1 is

$do(X_j=x)$2

The learner uses a DAG-respecting Bayesian prior, updates a posterior over $do(X_j=x)$3, and chooses the next intervention by minimizing a closed-form causal integrated variance (CIV),

$do(X_j=x)$4

The paper provides an information-theoretic lower bound relating CIV to mutual information and a consistency result showing that the acquisition landscape becomes well behaved near $do(X_j=x)$5 as data accumulate (Zhang et al., 2022).

Online identification of IT systems extends this mechanism-learning perspective to nonstationary operational environments. The system is modeled as

$do(X_j=x)$6

the learner maintains independent GP posteriors over the causal functions, and the overall objective minimizes discounted estimation loss plus intervention cost,

$do(X_j=x)$7

A rollout-based intervention policy chooses experiments from the current belief state, and the paper proves Bayes-optimality of the posterior mean, asymptotic consistency under standard GP conditions, and an improvement guarantee for the rollout policy relative to its base policy (Hammar et al., 2 Sep 2025).

Some work relaxes the assumption that the current causal model is structurally adequate. In active causal structure learning with latent variables for autonomous robots, the agent uses a dynamic decision network, monitors discrepancies between predicted and observed transitions via a surprise divergence,

$do(X_j=x)$8

maps negative utility surprise into evidence for a hidden variable, augments the graph with that variable, and re-estimates CPTs with Hard Weighted EM. The detour example is explicitly about revising structure when a transparent barrier causes repeated prediction failure (Riscos et al., 2024).

5. Algorithmic strategies

The algorithmic repertoire of active causal learning is unusually diverse. Some methods remain close to classical optimal design and exact graph algorithms: essential-graph search and chordal coloring produce polynomial-time intervention targets with worst-case identifiability guarantees; Monte Carlo information-gain objectives are optimized with GP-UCB over continuous intervention values; and CIV provides a closed-form acquisition in known linear SCMs, avoiding Monte Carlo over either model parameters or outcomes (Hauser et al., 2012, Kügelgen et al., 2019, Zhang et al., 2022).

Another group of methods learns the acquisition policy itself. In deep reinforcement learning for causal structure learning, the state is the current CPDAG or chain graph, the action is the next variable to intervene on, and the learned policy is

$do(X_j=x)$9

The graph is embedded by a GNN, the reward is the number of edges oriented after the intervention, and the policy is trained with Q-learning and replay. In Active Intervention Targeting (AIT), the learner samples plausible DAGs from a soft adjacency belief and scores candidate targets by the discrepancy

$f_n$0

favoring interventions whose post-interventional outcomes differ substantially across plausible graphs but are stable within each graph. CAASL goes further: it learns a transformer-based amortized intervention policy in a Hidden-Parameter MDP, trains it with SAC/REDQ, and uses a reward derived from the improvement of an amortized graph posterior $f_n$1. At test time the policy is real-time, adaptive, and does not require likelihood access (Amirinezhad et al., 2020, Scherrer et al., 2021, Annadani et al., 2024).

LLM-based causal discovery introduces yet another design layer: active selection of natural-language queries about graph structure. The proposed fairness-driven framework uses BFS-style graph expansion together with a dynamic score

$f_n$2

where $f_n$3, $f_n$4, and the next pair is selected by maximizing $f_n$5. The goal is to reduce unnecessary LLM queries while maintaining graph quality for downstream fairness analysis (Zanna et al., 21 Mar 2025).

This algorithmic diversity reflects different operating constraints. Exact graph-theoretic and Bayesian-design methods supply sharper guarantees; amortized and RL-based methods emphasize runtime, likelihood-free deployment, and zero-shot or out-of-distribution generalization. The reported results follow that split: the DRL approach is especially fast on dense graphs, CAASL generalizes to higher-dimensional environments and unseen intervention types, and LLM scoring reduces the practical query burden by prioritizing high-value pairs (Amirinezhad et al., 2020, Annadani et al., 2024, Zanna et al., 21 Mar 2025).

6. Domains, empirical regularities, and misconceptions

Active causal learning has been instantiated in a wide range of empirical settings: gene regulatory networks and DREAM sub-networks, Perturb-CITE-seq intervention design, interference-aware online experiments on Tencent game data, QM9 molecular design for high dipole moment, cloud and microservice testbeds, LLM debiasing, and human active “blicket detector” experiments. This breadth is not accidental: each domain combines expensive data acquisition with a causal target that cannot be recovered efficiently from passive observation alone (Amirinezhad et al., 2020, Zhang et al., 2022, Zhu et al., 2024, Fox et al., 2024, Du et al., 2024, Jiang et al., 2022).

A recurring empirical pattern is that adaptive design usually improves sample efficiency relative to random or passive baselines, but the size of the gain depends strongly on the quality of the state estimate, the realism of the intervention model, and the difficulty of the target task. In the sample case for essential-graph learning, estimation error flattens the survival curves and can make random interventions surprisingly similar to optimal ones at small sample sizes. In DRL-based structure learning, performance is strongest in dense graphs and under small intervention budgets, while low-density graphs can be harder. In Causal-EPIG, the preferred acquisition is explicitly context-dependent: a direct CATE-targeting rule can be most sample-efficient in some regimes, while joint-potential-outcome targeting is more robust in others (Hauser et al., 2012, Amirinezhad et al., 2020, Gao et al., 26 Sep 2025).

Several misconceptions recur in the literature. First, active causal learning is not synonymous with causal discovery: some methods target mechanisms after discovery, some target downstream causal quantities directly, and some work only in observational pool-based querying rather than in experimental intervention design. Second, predictive improvement and causal improvement need not coincide. In the chemistry application, the random-forest $f_n$6 for dipole moment is about the same for active and random sampling, while active selection improves recovery of the global causal graph, illustrating that graph recovery and predictive accuracy optimize different objectives (Fox et al., 2024). Third, many methods rely on strong assumptions—causal sufficiency, faithfulness, known graph structure, linearity, discrete variables, perfect randomized interventions, or access to hidden states and confidence scores—so gains are conditional on model adequacy rather than assumption-free (Zhang et al., 2022, Du et al., 2024).

The topic also includes learning active causal strategies themselves. Human experiments with hierarchical Bayesian overhypotheses show that intervention choices can be directed not only toward the current task but also toward transferable beliefs about causal functional form. Related work on passively trained agents and LLMs argues that passive data can teach an “experiment first, then seek goals” strategy, provided active intervention remains available at test time (Jiang et al., 2022, Lampinen et al., 2023).

Across these strands, the unifying theme is not a fixed formalism but a fixed asymmetry between passive and active data collection: causal targets become easier to identify when the learner can choose what to reveal next. The main open design choice is which uncertainty to reduce—structural, mechanistic, or query-specific—and the contemporary literature increasingly treats that choice as the central modeling decision rather than as an implementation detail.