Understanding LLM Embeddings for Regression