[christianqchung]

The problem is also that the model may have a very high confidence in token probability and is still wrong, but I'm sure it could help in some cases.