`Generalization is hallucination' through the lens of tensor completions

Abstract

In this short position paper, we introduce tensor completions and artifacts and make the case that they are a useful theoretical framework for understanding certain types of hallucinations and generalizations in language models.