LLMs Do Not Have Human-Like Working Memories

This seminar presentation explores the critical limitations of Large Language Models (LLMs) regarding working memory capabilities. Learn how human working memory functions as an active cognitive system that enables temporary information storage, processing, and utilization—a capability current LLMs lack. The speaker, Jen-Tse (Jay) Huang from Johns Hopkins University, demonstrates this limitation through three experimental scenarios: the Number Guessing Game, Yes or No Game, and Math Magic. Discover how these experiments reveal that leading LLM families fail to exhibit human-like cognitive behaviors in tasks requiring working memory, highlighting a significant challenge to achieving artificial general intelligence. Dr. Huang, a postdoctoral researcher at the Center for Language and Speech Processing with publications in top AI venues including ICLR 2024, presents this research to encourage development of LLMs with improved working memory capabilities.