Dividing activations into blocks and scaling each independently to handle outliers without affecting overall precision.