Kolmogorov-Arnold Causal Generative Models

Causal generative models provide a principled framework for answering observational, interventional, and counterfactual queries from observational data. However, many deep causal models rely on expressive architectures with opaque mechanisms, limiting auditability in high-stakes domains. This paper introduces the Kolmogorov-Arnold causal generative model, a mixed-type tabular causal generative model that parameterizes each structural equation with a Kolmogorov-Arnold Network. The resulting mechanisms can be inspected through symbolic approximations and parent-child visualizations while preserving query-agnostic causal generative semantics. The model includes a validation pipeline based on distributional matching and independence diagnostics of inferred exogenous variables, allowing adequacy to be assessed using observational data alone. Experimentally, on eleven synthetic mixed-type benchmarks, the Kolmogorov-Arnold causal generative model is not statistically distinguishable from the best state-of-the-art model: aggregate post-hoc $p$p-value of 0.69. On the semi-synthetic Sachs' additive dataset, with 11 variables, the model obtains a $p$p-value of 0.18 against the best model. A real- world cardiovascular case study demonstrates the interpretability workflow, showing how learned mechanisms can be simplified into structural equations and visualized to audit parent-child effects across the causal system. This suggest that expressive causal generative modeling and functional transparency can be achieved jointly, supporting trustworthy deployment in tabular decision-making settings.