Methods, systems, and devices for a decision directed multi-modulus searching algorithm are described. A receiver may receive a signal including a set of data symbols. The receiver may iteratively determine a set of centroids for demodulating the set of data symbols (e.g., as part of a training procedure). The centroids may be used to demodulate the set of data symbols according to a modulation constellation associated with the set of data symbols. The training procedure may include, for each data symbol of a subset of data symbols, assigning a centroid of the set of centroids to each data symbol and updating the set of centroids based on assigning the centroid to each data symbol. The receiver may demodulate the set of data symbols based on the updated set of centroids.