“In an experiment on the TriviaQA dataset, using reinforcement learning with an oracle reward function taught a model the optimal thresholding behavior for when to answer versus refuse.”