The authors report higher question-answering accuracy when a shared simulator and value estimator guide a frozen agent, but ...