From Token Prediction to Interaction Dynamics

Abstract Large language models (LLMs) are commonly described as systems that perform next-token prediction over large corpora. While this characterisation is accurate at the level of the training objective and local generation, it does not fully account for the structured, meaning-making behaviour observed in sustained dialogue. This paper proposes a complementary perspective in which LLM behaviour is understood as a trajectory through a semantic state space, evolving under iterative interaction. We argue that coherence in dialogue emerges through the progressive resolution of interpretive uncertainty, corresponding to a reduction in semantic entropy and the formation of attractor-like dynamics. By formally defining this state as *Resonance*, we show that coherence is not solely a property of the model, but of the coupled human–model system. Situating dialogue within dynamical systems theory, information theory, and basal cognition, this framework offers a more complete account of interactional stability and provides a teleological perspective on alignment and system design.