Global statistical constraint and local partner-sensitive inference support dynamic, open-ended verbal coordination

Abstract

Effective verbal communication requires individuals to align their communicative intentions through the precise use of utterances, yet how this is achieved in dynamic, open-ended communication remains unclear. Combining predictive deep language models (DLMs) with cognitive computational modeling, we tested a computational division of labor between context-sensitive global constraint on the utterance space and partner-sensitive local adjustment of selection priorities. We formalized local adjustment using an active inference inspired updating rule within a statistical-pragmatic hybrid model. Thirty dyads performed a dynamic, open-ended collaborative word-generation task during functional near infrared spectroscopy (fNIRS) hyperscanning. Interlocutors converged in their lexical choices, and lexical alignment was best explained by a hybrid model integrating contextual selection priors with interaction-specific local adjustment among the models tested. This hybrid process formed a dynamic hypothesis landscape comprising a global semantic search region and local selection priorities. Critically, interaction did not simply reduce uncertainty: within individuals, the hypothesis space of global semantic search progressively narrowed while local uncertainty among plausible alternatives increased; between partners, both global hypothesis space and local selection became increasingly aligned, with alignment of local selection priorities most strongly predicting subsequent lexical alignment. The contribution of local adjustment was stronger during precise verbal communication than in the control tasks. Neural analyses further identified correlates of the local adjustment process: temporoparietal and anterior temporal regions supported partner-related inference and semantic prediction error, respectively, and their activity during thinking predicted subsequent inferior frontal activity associated with self-updating during speaking. The participant's own hypothesis landscape was also represented in the left inferior parietal-supramarginal cortex. The main findings generalized across language models and an independent Czech dataset. Together, these findings reveal that alignment in open-ended communication emerges from complementary statistical and pragmatic computations, providing a framework for understanding human and human-AI communication.