Finding IVs is a heuristic and creative process, and justifying exclusion restrictions is largely rhetorical. We propose using LLMs to search for new IVs systematically through narratives and counterfactual reasoning. We develop a discovery pipeline (multi-step, role-playing prompts) and an evaluation pipeline (expert surveys and an LLM judge trained on the survey data). 

*This project is featured in the causal inference course at Stanford︎︎︎ and the Lindau Nobel Laureate Meetings 2025︎︎︎.