When Correct Demonstrations Hurt: Rethinking the Role of Exemplars in In-Context Learning

In three linesAn arXiv study reveals that correct demonstrations can degrade in-context learning (ICL) performance. Researchers introduce task-preserving perturbations to show that correctness does not guarantee utility: changing an exemplar's input while keeping a correct output can reduce accuracy, especially for smaller models and harder tasks.

Summary generated by Claude — human-verified