The paper presents an approach to improve causal inference using Large Language Models (LLMs) such as GPT-3.5-turbo and GPT-4. The authors argue that full graph information is not necessary for causal inference, the topological order over graph variables (causal order) suffices. They propose a method to obtain causal order from LLMs, which act as virtual domain experts. The study also explores integrating LLMs with established causal discovery algorithms to enhance their performance. The results show significant improvement in causal ordering accuracy, demonstrating the potential of LLMs in enhancing causal inference across varied fields.
Publication date: 23 Oct 2023
Project Page: https://arxiv.org/abs/2310.15117
Paper: https://arxiv.org/pdf/2310.15117
