LLMs Improve Causal Discovery Through Probabilistic Graph Fusion

Summary

Researchers introduce Probabilistic Dependency Graphs to combine uncertainty from Bayesian network structure learning and large language models. Across 26 benchmark networks, a 50/50 fusion of three structure-learning algorithms and Gemini, Claude, and GPT improved F1 over the better standalone source on 22 networks, with a significant mean gain of 0.056. The experiments show that Bayesian methods recover more candidate edges, while LLMs orient edges more accurately, suggesting that the two sources provide complementary evidence for causal graph discovery.