Large language models (LLMs) appear to solve certain novel tasks with appropriately formatted prompts, which is an ability known as out-of-distribution generalization. However, the underlying ...