The most significant achievement, so far, of the MindSharing project has been the identification of a latent control parameter used by human interlocutors during referential communication. The project aims at understanding how interlocutors regulate the referential process to effectively coordinate novel, context-dependent mappings in real-time interactions. To address this issue, we have combined two complementary approaches to characterize interaction-specific mappings between signals and referents across communicative turns. First, we employ an experimental semiotic task (the Tacit Communication Game, TCG) that amplifies natural generative demands by requiring participants to communicate without preexisting shared signal-referent mappings. In the TCG, dyads collaborate to arrange geometric shapes into designated configurations across multiple turns, minimizing reliance on conventional linguistic or gestural cues while enabling precise quantification of communicative behaviors across diverse referential challenges. Second, we have developed a hierarchical Transformer model (HTM) to generate movement- and interaction-level embeddings of communicative behaviors (113 adult dyads). Unlike standard large language models, this approach captures dependencies not only between tokens, but also between the actual referents of those tokens over the broader communicative exchange. By leveraging full access to both signal trajectories and referential spaces, we move beyond surface-level signal analysis to uncover how discrete behavioral sequences evolve into structured patterns of referential coordination over time. This approach allowed us to generate movement- and interaction-level embeddings. We experimentally generated communicative variance by sampling from neurotypical (NT) and autistic (ASC) dyads engaged in the TCG. 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, interaction-level 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. We are currently in the process of testing the reproducibility and generalizability of these findings, applying the same general methodology to data obtained from linguistic and multimodal referential communication. We are also testing whether and how interlocutors track representational complexity during a dialogue.
