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

arXiv:2405.04156v1 Announce Type: new
Abstract: Transformer-based language models are treated as black-boxes because of their large number of parameters and complex internal interactions, which is a serious safety concern. Mechanistic Interpretability (MI) intends to reverse-engineer neural network behaviors in terms of human-understandable components. In this work, we focus on understanding how GPT-2 Small performs the task of predicting three-letter acronyms. Previous works in the MI field have focused so far on tasks that predict a single token. To the best …

abstract arxiv components cs.lg engineer gpt gpt-2 human interactions interpretability language language models network neural network parameters safety terms transformer type understanding via work