“The performance of in-context learning (ICL) is not monotonically improving with an increasing number of samples; it fluctuates and eventually hits a performance cliff at the model's context length limit.”