Support  Vector  Generation: Kernelizing Large Language Models for Efficient Zero‑Shot NLP