Overview
Explore the concept of weak-to-strong generalization in reasoning models through this 47-minute lecture by Tom Goldstein from the University of Maryland. Delve into the importance of recurrent architectures in enabling models to dynamically scale computation for solving increasingly complex problems. Examine examples demonstrating weak-to-strong generalization in recurrent networks across various reasoning tasks. Discover how transformer-based Large Language Models (LLMs) can benefit from recurrence, enhancing their performance on weak-to-strong arithmetic challenges. Gain insights into the potential of recurrent architectures to push the boundaries of AI problem-solving capabilities beyond their initial training parameters.
Syllabus
Using recurrence to achieve weak to strong generalization
Taught by
Simons Institute