Presenter: Zahra Jamshidi, University of Calgary

Field: Business Intelligence

Paper: Large Language Models' Statistical Reasoning in Heavy-Tailed Contexts

Abstract: "The emergence of generative artificial intelligence represents what Lévesque et al., (2022) describe as a hot spot for entrepreneurship research - a technological revolution with the potential to fundamentally reshape both the theory and practice of entrepreneurship. Increasingly, entrepreneurs may rely on Large Language Models (LLMs) to make autonomous decisions or to augment human judgment throughout the entrepreneurial process and the extent to which LLMs are competent entrepreneurs has become an important industry benchmark for testing their capabilities. Researchers have begun to study how LLMs perform in entrepreneurial decision making and how they can support human entrepreneurial decision making. In particular a line of research applies the reasoning capabilities of LLMs to evaluate opportunities, pitches, and make startup investment decisions. But this research is inhibited by a gap in our understanding of how LLMs think and reason about uncertainty in entrepreneurial contexts and outcomes. Given the centrality of uncertainty to the theory and practice of entrepreneurship, this is a critical gap to address. In our research, we aim to make a contribution to our understanding of how LLMs reason about uncertainty by asking if they are capable of thinking in heavy tails when they try to predict entrepreneurial outcomes. Additionally, we ask: how can we get LLMs to be better at reasoning about heavy-tailed outcomes through prompt engineering? By investigating these questions, we aim to inform all LLM-based research in decision making that involves asking LLMs for evaluations of opportunities or predictions of outcomes."

Host: Allan Würtz