I think alignment in this context refers to distribution alignment: “aligning” it so its outputs more efficiently steer the target model with less tokens