Further analysis substantiates our hypothesis that our improvement can be attributed to reduced overfitting to instruction tuning datasets. It is worth noting that we are not proposing IM as a replacement for the current instruction tuning process. Instead, our work aims to provide practical guidance for instruction tuning LMs, especially in low-resource scenarios.

Forward-mode AD that are closer estimations of the true gradients.