Ben Prystawski, a PhD candidate in the psychology department at Stanford University who studies the nexus between machine learning and cognitive science, joins us today. The topic of discussion is Ben’s most recent paper, “Why Think Step by Step? “The locality of experience is where reasoning emerges,” he said in a recent presentation at NeurIPS 2023. We begin by discussing some fundamental issues regarding LLM reasoning, such as whether it exists, how to characterize it, and how methods such as chain-of-thought reasoning seem to support it. We next delve into Ben’s article, which tries to explain the effectiveness of step-by-step thinking and shows that the essential feature of LLM training data that makes it possible is local structure.
