This paper studies a generalization of the classic binary search problem of locating a desired value within a sorted list. The classic problem can be viewed as determining the correct one-dimensional, binary-valued threshold function from a finite class of such functions based on queries taking the form of point samples of the function. The classic problem is also equivalent to a simple binary encoding of the threshold location. This paper extends binary search to learning more general binary-valued functions. Specifically, if the set of target functions and queries satisfy certain geometrical relationships, then an algorithm, based on selecting a query that is maximally discriminating at each step, will determine the correct function in a number of steps that is logarithmic in the number of functions under consideration. Examples of classes satisfying the geometrical relationships include linear separators in multiple dimensions. Extensions to handle noise are also discussed. Possible applications include machine learning, channel coding, and sequential experimental design.