51 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Understanding The Structure of Reasoning In Language Models
University of Maryland, College Park
Recording
Abstract
Large language models are becoming increasingly capable reasoners, but reliability requires more than solving individual problems correctly: models must reason under uncertainty, allocate computation effectively, recognize their limitations, and decide when and how to use external information or tools. This talk discusses recent work on the structure and dynamics of reasoning in language models, their ability to reason under uncertainty, and mechanisms for controlling inference-time computation. Across these settings, a recurring theme is that strong underlying capabilities do not always translate into reliable reasoning behavior, and understanding this gap, together with the internal signals and mechanisms that govern reasoning, will be central to building more reliable and autonomous AI systems.