Correct Demonstrations Can Reduce In-Context Learning Accuracy

A new arXiv preprint (2605.26350) reveals a counterintuitive phenomenon in in-context learning (ICL): correct demonstrations do not guarantee utility, and some can even reduce accuracy. The authors introduce task-preserving perturbations, where only the exemplar input changes while remaining a correct instance of the same task. This includes label-updating perturbations (task-relevant semantics change, targets recomputed) and target-preserving perturbations (original target remains valid). They formalize the failure mode as contextual evidence shift, where perturbations alter the effective mixture of evidence used by the model. The study challenges the common intuition that correct examples always help ICL.