Seonglae ChoFor black-box text-embedding-3-small embedding model
SAE Probing for label prediction.
While using only embeddings shows higher performance, obtaining interpretable NLP features is a notable advantage.
Token embedding → text embedding by pooling