Task-Driven Partial Causal Learning

• The paper demonstrates that restricting causal models to task-specific objectives improves predictions and control in complex systems.
• Task-driven partial causal learning focuses on low-dimensional causal objects, such as binary indicators and ancestral matrices, rather than full DAGs.
• Predictive reframing mitigates confounding by learning statistical properties over observed subsets, enhancing generalization for downstream tasks.

Task-driven partial causal learning denotes a family of methods in which the learned causal object is restricted to what a downstream task actually needs, rather than to a complete reconstruction of the underlying causal structure. Across recent work, the target of learning may be a backdoor variable $z_t$ sufficient for planning under policy interventions in reinforcement learning, a predictor of statistical properties of variable subsets that were never jointly observed, an explicit DAG over task-relevant macro-variables for goal-directed control from visual data, binary descendant indicators $Z_{a,i}$ for selective conformal calibration under interventions, the conditional mean of task-specific causal demand parameters in confounded multi-task pricing, a partial ancestral matrix $\tilde{\mathbf T}$ for conditioning causal foundation models, or the minimal reward-ancestor abstraction needed to preserve optimal control (Rezende et al., 2020, Janzing et al., 2023, Nair et al., 2019, Asiaee et al., 2 Mar 2026, Gupta et al., 10 Feb 2026, Reuter et al., 16 Feb 2026, Wang et al., 2024). This suggests that “partial” does not denote causal incompleteness in a pejorative sense; rather, it denotes restriction of the estimand, representation, or invariance target to the part of the causal structure that is operationally required.

1. Predictive reframing of causal learning

A central formulation recasts causal learning as a supervised prediction problem over partially observed variable sets. In this view, the full variable universe is $S=\{X_1,\dots,X_n\}$, data arrive as datasets $D_1,\dots,D_\ell$ over subsets $S_j\subseteq S$, the observed subsets are the training points, and previously unobserved subsets are the test points. The prediction target is not necessarily a full joint distribution, but a statistical property $Q$ of a subset, with

$Q:\Delta_S \to \mathcal{R}.$

Examples given for $Q$ include conditional independence, partial uncorrelatedness, sign of correlation, covariance matrix, and existence of a linear additive noise model. A causal model $M$ is then judged by whether its induced predictor $Z_{a,i}$0 matches the target property on unseen subsets, rather than by whether $Z_{a,i}$1 is the uniquely true causal graph (Janzing et al., 2023).

This perspective changes the evaluation criterion. The empirical training error is defined over observed subset-statistics,

$Z_{a,i}$2

and the desired generalization target is the expected error on new subsets. The resulting notion of usefulness is pragmatic: a causal hypothesis is useful whenever it correctly predicts the relevant statistical property of an unobserved joint distribution. The paper makes the explicit point that a sparse causal graph that omits weak influences may be more useful than a dense one because it yields stronger predictive closure over missing subset-statistics (Janzing et al., 2023).

The same predictive reframing appears in interventional settings. Selective conformal inference under interventions does not require full graph recovery; it requires identification of which interventions leave a target unaffected, because exchangeability may hold only within that subset. Likewise, multi-task demand learning with endogenous prices does not identify each latent task parameter vector exactly; instead, it identifies a task-conditioned causal conditional mean. In all of these cases, the learned causal object is defined by the downstream prediction or decision problem, not by exhaustive structural reconstruction (Asiaee et al., 2 Mar 2026, Gupta et al., 10 Feb 2026).

2. Task-relevant causal objects

The learned object in task-driven partial causal learning is typically a low-dimensional causal functional, indicator, or structured side-information variable.

For selective conformal inference, the central label is the binary descendant indicator

$Z_{a,i}$3

which records whether intervention $Z_{a,i}$4 affects target $Z_{a,i}$5. This is exactly the task-relevant causal label because selective calibration only needs to know whether a calibration example is exchangeable with the test example. The paper explicitly reformulates the problem as estimating $Z_{a,i}$6 rather than learning the full DAG, and it shows that the induced contamination fraction depends asymmetrically on false positives and false negatives. False positives directly contaminate the calibration set, while false negatives only reduce calibration size (Asiaee et al., 2 Mar 2026).

For confounded multi-task demand learning, the target is not the full latent task-specific parameter $Z_{a,i}$7, since $Z_{a,i}$8 contains an unobserved deviation $Z_{a,i}$9 that is not perfectly recoverable from finite observational data. The estimand is instead

$\tilde{\mathbf T}$0

the conditional mean of the task-specific causal parameters given a carefully designed masked information set. The paper treats this as the learnable causal object that remains identifiable under endogeneity and limited within-task price variation (Gupta et al., 10 Feb 2026).

For causal foundation models, the task-relevant object is a partial ancestral matrix

$\tilde{\mathbf T}$1

whose entries encode known ancestor, known non-ancestor, and unknown relations. This representation covers no knowledge, partial knowledge, and complete ancestral knowledge within one format. The paper argues that ancestral information is often more practical than a full adjacency matrix because domain experts may know that one variable is an ancestor of another without knowing whether the effect is direct (Reuter et al., 16 Feb 2026).

These constructions share a common design pattern: the causal target is minimized to the portion necessary for valid prediction, calibration, or control. A plausible implication is that the complexity bottleneck in causal learning can often be shifted from whole-graph recovery to task-defined causal sufficiency.

3. Reinforcement learning and control

In reinforcement learning, task-driven partial causal learning emerges from the tension between planning efficiency and causal validity. A standard partial model predicts only part of the observation, but if that model does not condition on the information the policy used to choose actions, then the missing information becomes a confounder. The resulting model learns the observational conditional

$\tilde{\mathbf T}$2

instead of the interventional quantity needed for planning,

$\tilde{\mathbf T}$3

The paper’s FuzzyBear example makes this concrete: a reward model trained on logged experience can become behavior-policy dependent and therefore wrong under a new policy. The proposed remedy is to choose the partial observation not arbitrarily, but as a backdoor variable $\tilde{\mathbf T}$4 in the agent’s computational graph, yielding a Causal Partial Model (CPM) with rollout equations based on backdoor adjustment. The appendix proves that the autoregressive model is causally correct, the non-causal partial model is causally incorrect, and the CPM is causally correct (Rezende et al., 2020).

A closely related but distinct line of work learns minimal task-specific causal abstractions in factored MDPs. Causal Bisimulation Modeling (CBM) uses a shared causal dynamics model and a task-specific causal reward model to identify the reward parents $\tilde{\mathbf T}$5 and their ancestors $\tilde{\mathbf T}$6. The theoretical bridge is the statement that if $\tilde{\mathbf T}$7 denotes the ancestors of the reward-relevant variables in the causal transition graph, then the abstraction

$\tilde{\mathbf T}$8

is a bisimulation abstraction for reward $\tilde{\mathbf T}$9. The paper argues that this yields the minimal abstraction needed for control: task-specific enough to remove irrelevant variables, but causal enough to retain upstream variables required to control reward-relevant ones. Empirically, it reports that learned implicit dynamics models identify causal relationships and abstractions more accurately than explicit ones, and that the derived abstractions allow a task learner to achieve near-oracle levels of sample efficiency and outperform baselines on all tasks (Wang et al., 2024).

Task-driven partial causal learning also appears in visually grounded goal-directed control. In a goal-conditioned MDP family with latent causal structures, the agent receives a short probing trajectory of visual observations and actions, induces a DAG $S=\{X_1,\dots,X_n\}$0 over switches and lights, and then conditions a goal-conditioned policy on that graph. The causal graph is not learned as a latent embedding inside the policy; it is explicitly constructed and then consumed through an attention bottleneck that selects task-relevant edges. The method handles One-to-One, Many-to-One, One-to-Many, and Masterswitch families, and the paper reports improved F1 score for graph recovery and higher success rate on unseen environments, with the graph attention bottleneck improving success by about $S=\{X_1,\dots,X_n\}$1 in One-to-One and Masterswitch and $S=\{X_1,\dots,X_n\}$2 in One-to-Many and Many-to-One (Nair et al., 2019).

4. Partial structure for interventional prediction and selective inference

In interventional prediction, the operative question is often which calibration points or structural assumptions are valid for the current query. Selective conformal inference under interventions formalizes this by defining the safe calibration set $S=\{X_1,\dots,X_n\}$3 as the interventions with $S=\{X_1,\dots,X_n\}$4. Under the selective exchangeability assumption, calibration scores from $S=\{X_1,\dots,X_n\}$5 are exchangeable with the test score, but this set is unknown and must be learned. The paper’s main coverage theorem quantifies the effect of contamination when the selected set $S=\{X_1,\dots,X_n\}$6 includes affected interventions: $S=\{X_1,\dots,X_n\}$7 It also proposes a conservative correction $S=\{X_1,\dots,X_n\}$8. On synthetic linear SEMs, under controlled contamination up to $S=\{X_1,\dots,X_n\}$9, the corrected procedure maintains $D_1,\dots,D_\ell$0 coverage while uncorrected selective conformal prediction degrades to $D_1,\dots,D_\ell$1. A proof-of-concept on Replogle K562 CRISPRi perturbation data further demonstrates applicability to real genomic screens (Asiaee et al., 2 Mar 2026).

The same theme appears in causal foundation models. A CFM in the PFN/neural-process paradigm predicts the conditional interventional distribution

$D_1,\dots,D_\ell$2

but the paper argues that inference should condition on whatever causal information is available at test time: full graph, partial graph, ancestral information, or none. Its central conditioning mechanism adds a learnable structure-derived bias to transformer attention logits, with known ancestors encouraged and known non-ancestors discouraged. The reported empirical result is that learnable attention biases are the most effective way to exploit full and partial causal information, that all graph-conditioned models outperform the no-graph baseline, and that ancestor matrices perform about as well as adjacency matrices, sometimes slightly better (Reuter et al., 16 Feb 2026).

A more general interpretive link is that intervention prediction can itself be viewed as prediction of statistical properties of unobserved joint distributions involving intervention variables. One paper explicitly notes that statements such as $D_1,\dots,D_\ell$3 and $D_1,\dots,D_\ell$4 can be phrased in this way. This places selective calibration, graph-conditioned causal foundation models, and predictive causal discovery within the same broad regime of inferring task-relevant statistics under partial observability (Janzing et al., 2023).

5. Confounded multi-task transfer

In multi-task demand learning, each task has observed covariates $D_1,\dots,D_\ell$5, a small within-task history of price-demand pairs $D_1,\dots,D_\ell$6, and a heterogeneous linear demand curve

$D_1,\dots,D_\ell$7

The central complication is endogeneity: prices may depend on the full history and on the latent task-level deviation $D_1,\dots,D_\ell$8. The paper shows that pooled regression is generically inconsistent and that standard meta-learning converges to a policy-confounded estimand because the query price remains endogenous even after conditioning on observed support history (Gupta et al., 10 Feb 2026).

The proposed Decision-Conditioned Masked-Outcome Meta-Learning (DCMOML) conditions on the realized price path because decisions themselves carry information about latent fundamentals, but withholds two outcomes so that the supervision point is not recoverable from the input. Specifically, if $D_1,\dots,D_\ell$9 is the last index whose price differs from the final price, the learner masks outcomes at $S_j\subseteq S$0, randomizes the query index over those two points, and learns from the masked information set $S_j\subseteq S$1. Under the paper’s “No selection on the query shock under two-point withholding” assumption, the population risk is uniquely minimized by

$S_j\subseteq S$2

This is an explicit instance of partial causal learning: the identified object is the conditional mean of task-specific causal parameters, not their full latent realization (Gupta et al., 10 Feb 2026).

This formulation also clarifies the role of carefully designed information sets in causal transfer. Conditioning on too little information leaves confounding unaddressed; conditioning on too much can destroy identifiability because the target enters only through low-rank observational moments. The paper’s masking and randomization scheme is therefore not a heuristic regularizer but part of the identification design (Gupta et al., 10 Feb 2026).

6. Guarantees, misconceptions, and limitations

Several papers provide formal guarantees tailored to task-driven partial objectives. For predictive causal discovery over subset-statistics, VC-dimension bounds are derived for model classes used as predictors. For DAGs predicting conditional independences,

$S_j\subseteq S$3

for polytrees,

$S_j\subseteq S$4

and for collider-free paths used to predict sign of pairwise correlations,

$S_j\subseteq S$5

These results are coupled with VC-style generalization bounds, making the claim precise that partial causal models can generalize from observed subsets to unseen ones when the induced predictor class has controlled complexity (Janzing et al., 2023).

A recurring misconception is that any compressed causal representation is acceptable. The reinforcement-learning results show the opposite: arbitrary partial models can be causally incorrect because they learn observational correlations confounded by hidden information used by the behavior policy. The paper explicitly notes that the empty backdoor $S_j\subseteq S$6 is valid only when actions are independent of state; otherwise it is not causally correct (Rezende et al., 2020).

A second misconception is that the best model must be the most graphically faithful. Multiple papers reject that equation. Sparse graphs may be better predictors of missing subset-statistics than denser but more accurate graphs, and descendant indicators or conditional means may be the relevant causal targets even when full graph recovery is impossible or unnecessary (Janzing et al., 2023, Asiaee et al., 2 Mar 2026, Gupta et al., 10 Feb 2026).

A third misconception concerns evaluation. In formal-language learnability studies, the causal graph $S_j\subseteq S$7 with $S_j\subseteq S$8 is used to show that correlational estimates of task learnability are confounded because the data-generating automaton affects both dataset properties and learning difficulty. The paper introduces the binning semiring to intervene on task-specific frequency and uses decomposed Kullback-Leibler divergences to localize performance to targeted sub-tasks. Its empirical conclusion is that causal and correlational curves differ systematically and can even invert at low counts (Snæbjarnarson et al., 8 Jun 2026).

The limitations reported in the literature are domain-specific but structurally consistent. Visual causal induction from images assumes macro-variables, DAG structure, a hand-designed probing policy, and supervised graph labels during training, so it is not fully unsupervised causal discovery (Nair et al., 2019). CBM assumes a predefined factored state space and factorized transitions, and notes difficulties with nearly deterministic transitions for naive implicit CMI estimation (Wang et al., 2024). Graph-conditioned causal foundation models can answer the causal query implied by wrong graph information, which is still the correct answer relative to the assumed graph but not necessarily relative to the data-generating process (Reuter et al., 16 Feb 2026). Contamination-robust conformal correction can be conservative; if $S_j\subseteq S$9, the method returns an infinite interval (Asiaee et al., 2 Mar 2026).

Taken together, these works define task-driven partial causal learning not as a single algorithmic family but as a methodological stance. The stance is that causal learning should be aligned with the invariances, interventions, and error modes of the downstream task; that partial structure is often sufficient when it is causally chosen rather than arbitrarily compressed; and that success is measured by validity and utility on the target task, not by full structural recovery alone.