Probabilistic Fusion of LLMs and Bayesian Network Structure Learning
August 31, 2026
A novel Probabilistic Dependency Graph (PDG) representation fuses LLM causal knowledge with Bayesian network structure learning (BNSL) algorithms. A 50/50 weighted fusion of BNSL ensembles and LLMs improved F1 scores in 22 of 26 benchmarks, leveraging BNSL for edge recall and LLMs for orientation.
HOW THIS AFFECTS YOU
●
researcherYou can mitigate LLM unreliability in causal discovery by fusing them with algorithmic edge skeletons.