Attention Head Interpretability
Activation-level analysis of transformer attention heads and their role in model behavior.
This project studies how individual attention heads encode, route, and influence information inside transformer models. The work focuses on identifying which heads matter for specific behaviors, how their effects propagate through the model, and how activation-level mechanisms can inform broader questions in AI interpretability and safety.