New math framework unifies Transformer attention variants

Original: Kan Extension Transformers: A Categorical Unification of Attention, Diffusion, and Predict-Detach Self-Conditioning

Source: arxiv.org ↗