54 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Iterated Models: Expressive Power, Learning, and Chain of Thought
Toyota Technological Institute at Chicago
Recording
Abstract
Sequence models often apply the same learned function repeatedly to generate successive tokens. The talk examines the computations possible even with simple underlying function classes and compares the samples and computation needed to learn these models from final answers or from complete intermediate reasoning sequences.