A new framework utilizes Large Language Models (LLMs) for full causal graph discovery, addressing limitations of prior methods. Previous LLM-based approaches used a pairwise query strategy, resulting in a quadratic number of queries, making them impractical for larger causal graphs.
The proposed framework, in contrast, implements a breadth-first search (BFS) approach. This innovative method significantly reduces the query count to a linear number, drastically improving efficiency. Furthermore, it can easily incorporate available observational data to enhance performance in causal discovery.
Beyond its enhanced time and data efficiency, this framework has achieved state-of-the-art results on real-world causal graphs of varying complexities. These findings demonstrate the method's effectiveness and efficiency in uncovering causal relationships, indicating its broad applicability across diverse causal graph discovery domains.