> I am generally skeptical whether such a highly formal approach can help us answer the practical question of whether we can get LLMs to answer 'I don't know' m

> I am generally skeptical whether such a highly formal approach can help us answer the practical question of whether we can get LLMs to answer 'I don't know' more often I feel like writing an entire paper about the practical approach to the problems posed in this paper, but you'll probably have to first formally define the language used in the training data before you can try to map it (through training and sampling algos, which this paper conveniently skipped) to the target form. This sounds really fun at first, but then we're once again talking about the strict formalisation of natural language (which you could still do - the training data is limited and fixed!)