neural-interpretability

This paper investigates how the grammatical role of tokens shapes the geometry of transformer representations across layers, finding distinct evolution patterns in encoder versus decoder models.