Assuming you’re in control of the test data set, you do know if a task is unsolvable. At that point you can reward the model based on how quickly they give up.