It provides a fine-grained interpretation of the model's internal prediction process, as well as a powerful framework for intervening in LM behavior.