We assess the ability of large language models (LLMs) to answer causal questions by analyzing their strengths and weaknesses against three types of causal question. We believe that current LLMs can answer causal questions with existing causal knowledge as combined domain experts. However, they are not yet able to provide satisfactory answers for discovering new knowledge or for high-stakes decision-making tasks with high precision. We discuss possible future directions and opportunities, such as enabling explicit and implicit causal modules as well as deep causal-aware LLMs. These will not only enable LLMs to answer many different types of causal questions for greater impact but also enable LLMs to be more trustworthy and efficient in general.
@article{arxiv.2304.05524,
title = {Understanding Causality with Large Language Models: Feasibility and Opportunities},
author = {Cheng Zhang and Stefan Bauer and Paul Bennett and Jiangfeng Gao and Wenbo Gong and Agrin Hilmkil and Joel Jennings and Chao Ma and Tom Minka and Nick Pawlowski and James Vaughan},
journal= {arXiv preprint arXiv:2304.05524},
year = {2023}
}