Abstract: Multi-document machine reading comprehension requires the machine to read and comprehend multi documents simultaneously, so as to answer subsequent questions. Compared with the traditional single-document machine reading comprehension, it is more concerned with the semantic logic among the documents within the constraints of questions. The paper firstly explores the main datasets for multi-document machine reading comprehension to analyze its development trends, including multi-source datasets, linear inference datasets, and fine-grained semi-structured datasets. Secondly, it investigates the mainstream evaluation metrics, then a prototype of four-tier reading structure is discussed in detail. This paper also reviews the methods based on document selection, Transformers, Graph Convolution Networks, and external knowledge fusion. Finally, it gives an outlook on the research of multi-document reading comprehension, as well as the problems and challenges in this study.