CauTion preprint reports reliable causal discovery through selective LLM use

Researchers report that CauTion, a framework for inferring causal relationships from observational data, resolves up to 96% of relationships through agreement among statistical algorithms, with near-perfect accuracy on that agreed subset. The preprint reports better results than statistical and LLM-assisted baselines across six datasets. The framework estimates the relative reliability of a language model and the statistical algorithms without requiring annotations. It uses that estimate to weight votes, restricting language-model arbitration to relationships where algorithmic evidence is unreliable. A final repair step removes cycles from the resulting causal graph. The near-perfect accuracy finding applies to the consensus subset, not the entire graph.