Bayesian Network Structure Learning through Large Language Models

Constructing Bayesian Networks in data-scarce scenarios typically relies on costly expert knowledge. While Large Language Models (LLMs) offer a promising data-free alternative, they often suffer from hallucinations and generate structurally invalid networks containing cycles. To address these challenges, we propose a novel multi-agent framework comprising a Decider, Critic, and Arbiter (DCA) for automated BN structure learning. By integrating tripletbased causal reasoning with a confidence-driven network refinement strategy, our approach effectively eliminates redundant edges andensuresthegenerationofvalidDirectedAcyclicGraphs(DAGs). Experimental results on standard benchmarks demonstrate that our method significantly outperforms existing LLM-based baselines in terms of F1-score and Structural Hamming Distance (SHD), while successfully avoiding cycles and isolated nodes.

Authors

• Jiebin Cai ShenZhen University
• Yinghui Pan ShenZhen University
• Yifeng Zeng Northumbria University
• Han Liu ShenZhen University

Keywords

• Large Language Models
• Bayesian Networks
• Causal Discovery