We introduce the task of recommendation set generation for entity-oriented exploratory search. Given an input search query which is open-ended or under-specified, the task is to present the user with an easily-understandable collection of query recommendations, with the goal of facilitating domain exploration or clarifying user intent. Traditional query recommendation systems select recommendations by identifying salient keywords in retrieved documents, or by querying an existing taxonomy or knowledge base for related concepts. In this work, we build a text-to-text model capable of generating a collection of recommendations directly, using the language model as a "soft" knowledge base capable of proposing new concepts not found in an existing taxonomy or set of retrieved documents. We train the model to generate recommendation sets which optimize a cost function designed to encourage comprehensiveness, interestingness, and non-redundancy. In thorough evaluations performed by crowd workers, we confirm the generalizability of our approach and the high quality of the generated recommendations.