in-context learning (ICL), the remarkable ability to solve a task from only input exemplars, has commonly been assumed to be a unique hallmark of transformer models. In this study, we demonstrate that →