A Unified Framework for Critical Scaling of Inverse Temperature in Self-Attention

Provides unified theory for length-dependent logit rescaling in self-attention, resolving conflicting inverse-temperature laws by deriving critical scale from gap-counting function N_n of attention rows, determining how to scale inverse temperature with context length n.