arXiv:2604.02339v1 Announce Type: new Abstract: Natural language context-such as instructions, knowledge, or feedback-contains rich signal for adapting language models. While in-context learning provides adaptation via the prompt, parametric learning persists into model weights and can improve perfo
