FoundCause Discovers Causal Graphs with Latent Confounders.

Discovering causal relationships from observational data is a complex challenge, particularly due to the need to identify directed structures and account for unobserved, or latent, confounding variables without the benefit of interventions. Traditional methods often struggle with these aspects. Researchers have developed FoundCause, an amortized causal discovery model designed to address these limitations. FoundCause is trained entirely on large collections of synthetic structural causal models, allowing it to learn transferable statistical patterns and map datasets directly to causal graphs in a single forward pass. This approach explicitly models latent confounding using learnable latent tokens. The model incorporates several key inductive biases, including a permutation-invariant transformer encoder with alternating attention over samples and variables, and the injection of pairwise statistical features. A factorized decoder separates edge existence from direction, and a triangular refinement module reasons over higher-order causal motifs. FoundCause also handles missing data. It significantly outperforms 11 classical and 4 other amortized causal discovery methods on 15 real-world datasets, achieving substantial improvements in F1 score, AUROC, and structural Hamming distance, while offering rapid inference.