LLMs can now spot valid causal relationships in messy real-world data — if you ask them the right way

Researchers built a multi-agent system where language models propose, critique, and refine instrumental variables — statistical tools economists use to isolate cause from correlation when confounding factors make the relationship murky. The system recovered known instruments from published research and avoided discredited ones, suggesting LLMs can help economists identify which real-world variables actually cause changes in outcomes.

Econometricians spend months or years identifying valid instrumental variables because it requires deep domain knowledge, creativity, and contextual understanding — you need to know the literature, understand what's theoretically plausible, and spot what's been tried and failed. If LLMs can reliably do parts of this work, researchers can move faster from question to causal claim. The catch is that validity lives in the details: a variable looks like it could work until someone empirically tests it and finds it doesn't, which means LLM proposals still need human expertise to validate.