Abstract
LLMs are ideal for decision-making thanks to their ability to reason overlong contexts. However, challenges arise when processing speech transcriptsthat describe complex scenarios, as they are verbose and include repetition,hedging, and vagueness. E.g., during a company's earnings call, an executivemight project a positive revenue outlook to reassure investors, despiteuncertainty regarding future earnings. It is crucial for LLMs to incorporatethis uncertainty systematically when making decisions. In this paper, weintroduce \textsc{DeFine}, a modular framework that constructs probabilisticfactor profiles from complex scenarios. It then integrates these profiles withanalogical reasoning, leveraging insights from similar past experiences toguide LLMs in making critical decisions in new situations. Our frameworkseparates the tasks of quantifying uncertainty and incorporating it into LLMdecision-making. This approach is particularly useful in areas such asconsulting and financial deliberation, where making decisions under uncertaintyis vital.