How does GPT-2 Predict Acronyms? Extracting and Understanding a Circuit via Mechanistic Interpretability | Publicación