New methods target attention, distillation, and lightweight adaptation costs

A structure-aware exact relative-position attention kernel reports 1.8–6.4× forward and 2–21× backward speedups, while P^2-MOPD reports higher recovery than its baseline in a 1.7B setting. Weight-space programming claims a large MATH-500 gain from editing only 1,023 parameters, but transfer is uneven; K-prop and FewStepRL add leads on online-learning efficiency and few-step diffusion post-training.