Analysis of Internal Representations of Knowledge with Expressions of Familiarity

Abstract

Research on the ability of large language models (LLMs) to judge the familiarity of knowledge is progressing. However, little attention has been given to whether LLMs can assess the familiarity of knowledge when linguistic expressions such as "It is known that..." are included during training. This study investigates how familiarity is internally represented in LLMs. To achieve this, we trained the models on descriptions of knowledge accompanied by linguistic expressions indicating familiarity. The internal representations of the models were then analyzed. The findings reveal that (1) familiarity information is separately retained in the internal representations of knowledge, for the linguistic expressions provided during training, and (2) familiarity information is separately maintained for each position of the linguistic expressions. This study provides a foundation for understanding the mechanisms underlying LLMs' ability to judge familiarity.