Tell, don't show: Declarative facts influence how LLMs generalize

Alexander Meinke, Owain Evans

We examine how large language models (LLMs) generalize from abstract declarative statements in their training data. As an illustration, consider an LLM that is prompted to generate…