LAMDA: Large Language Model as Decision Analyst

Summary

LAMDA uses large language models to convert unstructured natural-language descriptions of decision problems into draft influence diagrams. The framework combines graph generation and probability elicitation with rule-based verification and iterative regeneration to resolve structural and formatting issues. The study evaluates LAMDA on a dataset of typical decision problems under risk and applies it to expert group discussions about a hypothetical pandemic. The resulting diagrams organize decision-relevant variables, relationships, beliefs, and preferences for human review and refinement, supporting decision analysts in information processing and decision-making.