Causal AI Scientist: Facilitating Causal Data Science with Large Language Models

Vishal Verma, Sawal Acharya, Devansh Bhardwaj, Samuel Simko, Yongjin Yang, Anahita Haghighat, Dominik Janzing, Mrinmaya Sachan, Bernhard Schölkopf, Zhijing Jin

Keywords: LLMs, Causal AI

Abstract:

Large Language Models~(LLMs) have recently been used to automate the causal inference process by overcoming the expertise barrier. However, existing LLM-powered approaches for causal effect estimation often require human users to manually specify variables and methods, and those that do not require manual specification support only a limited set of causal effect measures. To address these limitations, we present Causal AI Scientist (CAIS), an LLM-augmented causal tool with self-correction capabilities.
Specifically, given a natural language query and a dataset along with its description, CAIS uses LLMs to understand the user query and dataset, and then selects a method based on a decision tree approach.
It then executes the selected method, applies a validation feedback loop for self-correction, and uses the results to answer the input question, enabling fully autonomous causal analysis. Extensive experiments across diverse queries curated from textbooks, synthetic data, and real-world datasets demonstrate CAIS's ability to produce precise causal effect estimates through improved method selection and self-corrections, while reducing runtime errors. We believe CAIS will serve as a strong foundation for enabling fully automated causal inference with LLMs.

Submission Number: 51