Decoupling Is Not Identification: Supervised Evidential Learning in Next-Token Prediction

Ge Wang

A next-token probability says what a model predicts, not how much training support lies behind it. A Dirichlet head can represent this distinction by separating mean $m$ from conce…