Tag
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.
This paper investigates how lexical diversity affects intrinsic dimensionality estimates in language model representations, revealing a scale-dependent transition between two regimes and deriving a formula for the reversal point.
This paper presents a spectral phase diagram for binary few-shot classification, analyzing intrinsic dimensionality and geometric saturation for representational diagnosis.