We show that autoregressive generation in large language models exhibits a temporal structure: each token is not only conditioned on the past but also reshapes the future continuation space. We call this process proto-interpretation : the probabilistic redistribution across competing continuations through which the model gradually commits to one emerging branch of meaning. Using a minimal ambiguity case, we demonstrate branch competition and sequential commitment during inference. These findings reveal meaning in LLMs as a dynamic, temporally unfolding process, shifting interpretability from static model states to inference-time dynamics.
Mattias Rost (Thu,) studied this question.