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

Jorge Garca-Carrasco, Alejandro Mat, Juan C. Trujillo

Browse the full AISTATS paper archive.