AI4 Society Seminar: Emre Kiciman

"Advances in Causal Reasoning with LLMs"

Correct causal reasoning requires domain knowledge beyond observed data. Framing cause-and-effect questions in medicine, science, law, and engineering depends on eliciting expert assumptions about system dynamics and mechanisms—a labor-intensive process that is a bottleneck to the broader use of causal methods.
Large language models (LLMs) can provide general-purpose assistance for constructing causal arguments: extracting assumptions from text, proposing and critiquing causal graphical models, and identifying missing mechanisms. While not replacing experts, they reduce the effort and error of end-to-end causal reasoning, extending the reach of causal inference in observational studies and experimental design. More importantly, they make rigorous causal reasoning feasible in real-world settings where formal modeling was previously impractical. This includes identifying actual causes in complex narratives, decomposing multifaceted questions into tractable subproblems, conducting counterfactual analyses, and integrating disparate causal evidence into coherent system-level explanations. Together, these advances allow us to expand causal methods to a broader variety of real-world causal tasks previously outside of our field’s focus. 

This talk will present recent studies, empirical results, and emerging methodologies that illustrate both the progress made and the opportunities ahead at this intersection of causal reasoning and large language models.