Causal AI Scientist: Towards End-to-End Causal Inference with Large Language Models

Causal inference is a cornerstone of empirical research, yet conducting rigorous analysis on tabular datasets requires substantial methodological expertise to navigate complex identification assumptions. While current agentic frameworks aim to lower this barrier, they often require users to manually specify methods or variables, or they support only a narrow set of techniques with limited validation on real-world data. To address these limitations, we present Causal AI Scientist (CAIS), an autonomous agent designed for end-to-end causal inference. Given a dataset and its description, CAIS leverages Large Language Models (LLMs) to interpret metadata and navigate a structured decision tree guided by causal inference principles to select methods. To ensure reliability, CAIS implements a validation feedback loop that performs post-selection empirical diagnostics, allowing the agent to verify assumptions and self-correct method choices before implementation. We evaluate CAIS on CauSciBench, a rigorous benchmark comprising synthetic and real-world empirical studies, designed to assess LLMs' causal estimation capabilities in the real world. Our results demonstrate that while specific frontier models can excel with alternative prompting, CAIS achieves the best average performance across model families, including GPT-4o and GPT-4o mini. More broadly, CAIS showcases how agentic frameworks can democratize expert-level causal analysis by automating the path from natural language queries to rigorous causal estimates.