Introduction A central question in the discussion of large language models (LLMs) concerns the extent to which they memorize their training data versus how they generalize to new tasks and settings. Most practitioners seem to (at least informally) believe that LLMs do some degree of both: they cl