Teaching an AI model to say it does not know
A new method nudges a few internal switches so a model declines to guess.



Picture made for ABUREAU.
The hardest sentence for many people is also hard for machines: "I don't know." A new paper, Dr. No, shows a way to make a vision-and-language model hold back when the evidence is not enough.
The method finds a small set of internal features linked to holding back and to committing, then turns them up or down while the model reads the question. The base model is not retrained. The authors tested it on two models and several question sets, and it balanced right answers and fair refusals better than the other methods they compared.
It worked best where the picture and the question did not fit each other, and on maths-style reasoning. It was weaker where the answer was simply absent, and on questions that need several facts joined. The authors conclude that knowing when to refuse depends on why the model is unsure.
For anyone who owns or uses data, the test sets here are the point. Questions that cannot be answered from the evidence, labelled as such, are scarce, and they are what a model needs to learn when to stay quiet. A well-made set of them is worth building.
