Language models struggle with compartmentalization

Thomas Vincent Howe, David Wingate

In the training data used by large language models (LLMs), the same latent concept is often presented in multiple distinct ways: the same facts appear in English and Swahili; many…