Language Models use Lookbacks to Track Beliefs w/ Nikhil Prakash of Northeastern University