When Context Misleads: In-context Learning with Jurisdiction in Large Language Models.

arXiv:2609.27603v2 Announce Type: replace-cross Abstract: In-Context Learning (ICL) has become a cornerstone of modern LLM deployment. However, existing ICL post-training methods have a critical blind spot: they excel at extracting patterns from demonstrations