![[2511.00574] Bayesian Network Structure Discovery Using Large Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2511.00574] Bayesian Network Structure Discovery Using Large Language Models
Summary
This article presents a novel framework for Bayesian network structure discovery using large language models (LLMs), highlighting its effectiveness in both data-free and data-aware settings.
Why It Matters
The integration of LLMs into Bayesian network structure discovery represents a significant advancement in machine learning, particularly for scenarios with limited data. This research addresses the limitations of traditional methods, offering a more efficient and accurate approach to understanding complex systems.
Key Takeaways
- Introduces PromptBN and ReActBN for Bayesian network discovery.
- Demonstrates improved performance in low-data and out-of-distribution scenarios.
- Offers a unified framework that centralizes LLMs in the learning process.
Computer Science > Machine Learning arXiv:2511.00574 (cs) [Submitted on 1 Nov 2025 (v1), last revised 23 Feb 2026 (this version, v2)] Title:Bayesian Network Structure Discovery Using Large Language Models Authors:Yinghuan Zhang, Yufei Zhang, Parisa Kordjamshidi, Zijun Cui View a PDF of the paper titled Bayesian Network Structure Discovery Using Large Language Models, by Yinghuan Zhang and 3 other authors View PDF HTML (experimental) Abstract:Understanding probabilistic dependencies among variables is central to analyzing complex systems. Traditional structure learning methods often require extensive observational data or are limited by manual, error-prone incorporation of expert knowledge. Recent studies have explored using large language models (LLMs) for structure learning, but most treat LLMs as auxiliary tools for pre-processing or post-processing, leaving the core learning process data-driven. In this work, we introduce a unified framework for Bayesian network structure discovery that places LLMs at the center, supporting both data-free and data-aware settings. In the data-free regime, we introduce \textbf{PromptBN}, which leverages LLM reasoning over variable metadata to generate a complete directed acyclic graph (DAG) in a single call. PromptBN effectively enforces global consistency and acyclicity through dual validation, achieving constant $\mathcal{O}(1)$ query complexity. When observational data are available, we introduce \textbf{ReActBN} to further refine the ...