Don't let the LLM speak, just probe it

A researcher describes a technique for using a large language model to classify text based on a given criterion without generating text, by training a small MLP to read the model's hidden state at a specific layer.