In light of diversifying societies, how do we learn social group boundaries and categorize others into "us" and "them"? Previous research relied on explicitly labelled, static groups (e.g., team membership, race, etc.) and suggested that group memberships can be inferred via dyadic similarity; similarity on a salient feature drives our understanding of "us" and "them". In contrast, this line of research suggests that the context-dependent and flexible natures of our abilities to categorize others can be more comprehensively captured in a model of latent structure learning.