Can LLMs Propose Instrumental Variables for Causal Reasoning?

In the presence of confounding between an endogenous variable and the outcome, instrumental variables (IVs) are used to isolate causal effects. Identifying valid instruments requires interdisciplinary knowledge and contextual understanding, making it a difficult task. In this paper, we examine whether large language models (LLMs) can assist. We adopt a two-stage evaluation: first, testing whether LLMs recover established instruments from the literature, and second, assessing whether they avoid empirically or theoretically discredited ones. Building on these results, we introduce \textbf{IV Co-Scientist}, a multi-agent system that proposes, critiques, and refines IVs, along with a statistical test to contextualize consistency without ground truth. Our results show the potential of LLMs to identify valid IVs from large observational data.