New AI Tool Figures Out What Really Causes What

A new approach combines two complementary AI methods to build more accurate causal graphs. Statistical Bayesian network learning is good at finding which variables are likely connected, while large language models are better at determining the direction of cause and effect — but each source alone has weaknesses. By representing both as probabilistic uncertainty over whether an edge exists and how it is oriented, then fusing them through weighted averaging, researchers improved causal graph accuracy. Across 26 benchmark networks, a simple 50/50 fusion beat the better of either source in 22 cases, with a statistically significant average F1 gain of 0.056. The key insight: statistical methods provided a high-recall skeleton (80% recall vs 60% for LLMs), while LLMs provided far more accurate edge orientation (96% vs 77%), making their strengths complementary.