In this work, I tried making progress towards identifying the circuit for choosing the right pronouns (e.g. he vs she vs it vs they) to complete a rhetorical question (so it doesn’t spoil the answer!) like “{name} is a great friend, isn’t” implemented by GPT-2 Small, a 12 layer and 80M parameter transformer model.
Here is the google doc summarizing the results.