Differential Attention (V2)

Softmax attention spreads a small amount of probability over every non-matching token. Over 100k tokens that noise adds up and drowns the signal. Differential Transformer (Ye et al., ICLR 2025) fixes it by computing attention as the differ...