Inferring the input grammar accepted by a program is central for a variety of software engineering problems, including parsers verification, grammar-based fuzzing, communication protocol inference, and documentation. Sound and complete active learning techniques have been developed for several classes of languages and the corresponding automaton representation, however there are outstanding challenges that are limiting their effective application to the inference of input grammars. We focus on active learning techniques based on L∗ and propose two extensions of the Minimally Adequate Teacher framework that allow the efficient learning of the input language of a program in the form of symbolic automata, leveraging the additional information that can extracted from concolic execution. Upon these extensions we develop two learning algorithms that reduce significantly the number of queries required to converge to the correct hypothesis.