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

2026-05-25

AI Summary

This work challenges the prevailing assumption that correct in-context examples inherently improve model performance in in-context learning. The authors propose a task-preserving perturbation framework, employing label-update and objective-preserving perturbations to systematically evaluate the practical utility of correct demonstrations across tasks such as sentiment classification, logical reasoning, and mathematical word problems. Their findings reveal a significant gap between example correctness and effectiveness, introducing the concept of “contextual evidence shift” to explain how even accurate examples can degrade performance. Notably, perturbed examples substantially impair in-context learning—especially in smaller models and more challenging tasks—thereby questioning the conventional wisdom that correctness guarantees benefit.

Abstract

In-context learning (ICL) is often motivated by the intuition that demonstrations help because they provide correct input-output examples. However, we reveal a counterintuitive phenomenon: correctness does not guarantee exemplar utility, and some correct demonstrations can even reduce ICL accuracy. To study this correctness-utility gap, we introduce task-preserving perturbations, where only the exemplar input is changed, while the example remains a correct instance of the same task. Concretely, each perturbed exemplar is assigned the target induced by the task mapping. This framework covers both label-updating perturbations, where task-relevant semantics change and targets are recomputed, and stricter target-preserving perturbations, where the original target remains valid. We formalize the resulting failure mode as contextual evidence shift: task-preserving perturbations can change the effective mixture of evidence used by the model for contextual inference, thereby separating exemplar correctness from exemplar utility. Across sentiment classification, logical reasoning, and math word problems, we find that task-preserving perturbed demonstrations can substantially degrade ICL performance, especially for smaller models, harder tasks, and higher perturbation ratios. Our results show that robust ICL requires evaluating not only whether demonstrations are correct, but also how they influence contextual inference. Code is available at https://github.com/Chenghao-Qiu/Task-Preserving-ICL.

Problem

• in-context learning
• correctness-utility gap
• task-preserving perturbations
• contextual evidence shift
• exemplar utility

Innovation

• in-context learning
• task-preserving perturbations
• contextual evidence shift
• demonstration utility
• correctness-utility gap