In-Context Fixation: When Demonstrated Labels Override Semantics in Few-Shot Classification

arXiv:2605.08295v1 Announce Type: cross Abstract: While random demonstration labels barely hurt in-context learning (Min et al., 2022), we show that homogeneous labels--even semantically valid ones--collapse accuracy to <=12% across six models (Pythia, Llama, Qwen; 0.8B--8B) and four tasks. The...