Dialogue state tracking (belief tracking) is a key component of task-oriented dialogue systems and aims to estimate the user's goal at each user turn. State of the art approaches for state tracking rely on deep learning methods. These approaches represent dialogue state as a distribution over all possible value s for a slot, for each slot present in the ontology. Such a representation is not scalable for slots for which the set of possible values is unbounded (e.g., date, time or location) or dynamic (e.g., movies, usernames). We introduce a novel framework for state tracking which is independent of the value-set for a slot. The key idea is to obtain a set of values of interest (candidate set) which is bounded in size for each slot and represent the state as a distribution over the candidate set. Such an approach solves the problem of slot-scalability by making the state representation independent of the value set. Furthermore, by leveraging the slot-independent architecture and transfer learning, our model scales well and can be quickly bootstrapped to unseen domains with just a few training examples.