LLMs face the challenge of learning new tasks with minimal data, as exemplified by the MTOB benchmark, where models translate between English and Kalamang using only a book of grammar explanations, ...