Large Language Models as Tools to Improve Bayesian Causal Discovery

Bruna Bazaluk, Benjie Wang, Denis Mauá, Flavio S Correa da Silva

Keywords: Bayesian Structure Learning, Large Language Models, Causal Discovery

TL;DR: We propose using LLMs beside Bayesian Structure Learning (BSL) methods to do Causal Discovery.

Abstract: Causal discovery is the task of automatically infer- ring causal structures, typically from observational data. Recently, there has been much interest in utilizing domain knowledge from large language models (LLM) in causal discovery. However, existing LLM-based approaches only output a single directed acyclic graph (DAG) without uncertainty, which can be unreliable. In this work, we investigate using LLMs alongside Bayesian structure learning (BSL) methods for causal discovery, which output a distribution of possible graphs. In particular, we propose to harness the domain knowledge from the LLMs in the prior distribu- tion over graphs, in place of uninformed priors or human expertise. Our experiments show that LLM- informed priors can improve the performance of Bayesian structure learning methods.

Submission Number: 14