Causality Elicitation from Large Language Models

Researchers propose a new pipeline to extract causal relationships from large language models by sampling documents, identifying events, and using causal discovery methods. The approach aims to reveal the causal hypotheses that LLMs assume rather than establishing real-world causality.

• →A five-step pipeline is proposed to elicit causal relationships from LLM-generated content.
• →The method involves sampling documents, extracting events, grouping canonical events, and constructing binary indicator vectors.
• →Causal discovery methods are used to estimate candidate causal graphs from the processed data.
• →The approach focuses on revealing LLM assumptions rather than establishing real-world causality.
• →The framework provides an inspectable way to understand causal hypotheses embedded in language models.