Data Fusion with Causal Foundation Models

Causal foundation models (CFMs) amortise causal inference over priors of synthetic structural causal models (SCMs), predicting the effect of an experiment on a specific variable. However, observational data alone may leave multiple causal models compatible with available evidence, while experimental data with interventions on exactly the variable of interest might be unavailable. This work studies CFMs as a method to combine finite observational and surrogate-interventional datasets in order to predict a target conditional interventional distribution (CID) more accurately than with observational data alone. We first formalise the conceptual benefits of surrogate experiments. Building on this analysis, we introduce FusionCFM, a causal foundation model that uses an intervention-aware representation and hierarchical three-axis attention to exchange information across variables, samples, and experimental regimes. We evaluate it against a wide range of baselines across a varied set of restricted and arbitrary graph families, multiple graph sizes, as well as linear--Gaussian, and nonlinear, heterogeneous mechanisms. Our results show that FusionCFM is effective at targeting the CID, and effectively uses interventional information to reduce predictive uncertainty relative to observation-only methods.