We also studied the scaling effect of such high-level test-time computation: more compute leads to better decisions. On out-of-distribution Book Organization, next-subtask accuracy improves from 50%…