Despite large inter-individual differences in experience and conceptual structures, humans converge on referents largely underspecified by the signals exchanged during communication. Many neural networks have become extremely sensitive to context-dependent relationships between signals, but remain relatively blind to their referents outside the signal space. Here, we study how human interlocutors dynamically coordinate both their signal and referential spaces over extended communicative interactions. We identify a latent control parameter, representational complexity, that may regulate referential coordination in human communication. Using a custom hierarchical Transformer model, we generate movement- and interaction-level embeddings from neurotypical (NT) and autistic (ASC) dyads engaged in an experimental semiotic task. By leveraging ASC-related communicative variance, we identify changes in parameters that track the representational dimensionality of signals and referents as communication unfolds. Movement-level embeddings (within-trial dependencies) could not differentiate the two groups, indicating comparable communicative behaviors. In contrast, interactionlevel embeddings (across-trial dependencies) distinguished ASC from NT dyads with high accuracy. Crucially, representational complexity, i.e. dyadic alignment in the interaction-level intrinsic dimensionality used to encode communicative histories, tracked referential coordination demands, with greater misalignment in ASC dyads under referential volatility. These findings suggest that referential alignment is an interaction-level process, driven by the dynamic adaptation of representational complexity, rather than statistical relationships between signals alone.