Use of Metamorphic Testing Techniques for Improving Causal Discovery using LLMs

Abstract

This thesis explores the use of Large Language Models (LLMs) in causal discovery, a critical task for understanding complex relationships in data. Traditional methods often fall short when dealing with high-dimensional datasets and confounders. LLMs, leveraging metadata and contextual information, offer a novel approach by emulating expert domain reasoning. To address the inherent biases and limitations of LLMs, this research integrates metamorphic testing and prompt engineering. Metamorphic testing ensures the reliability of causal inferences by systematically validating model outputs, while prompt engineering optimizes input queries for improved accuracy. The study employs datasets such as Asia and Child to evaluate these methods. Results indicate that combining metamorphic testing and prompt engineering significantly enhances the robustness and precision of LLM-based causal discovery, offering a promising direction for future AI applications. This framework not only advances the theoretical understanding but also provides practical tools for researchers and practitioners in the field.