Most research on question answering implicitly assumes that the answer and its evidence appear close together in a single document. How can we teach computers to answer questions that require reasoning over multiple pieces of evidence?