3/3/26 (in-person) Dr. Weishun Zhong, Institute for Advanced Study at PrincetonÂ
Title: Statistical mechanics of real memories in natural language
Abstract: There has long been a divide in the study of memory: on one side, memories are complex and subjective, shaped by our individual experiences; on the other, traditional theories rely on oversimplified assumptions in artificial settings, rendering them largely irrelevant to real memories. We attempt to bridge this gap by introducing a statistical mechanical model that captures real memories acquired from and recalled in natural language. Analytical solutions of the model align with observations from large-scale narrative recall experiments, accounting for key features of memory such as summarization and abstraction. Grounded in the hierarchical organization of language itself, our model enables a unified framework for memory and language. Surprisingly, when applied directly to language, the model yields a prediction for the entropy of natural language, which we confirm through experiments with modern large language models.Â