How Models Learn

Confidence target

An instruction that wrong answers cost points, so a model should answer only when it is confident enough.

Their proposed fix is to change how tests are marked: state a confidence target in the instructions, saying that wrong answers cost points.
From Why models hallucinate, step “Why training rewards a good guess”

See it in action

Try confidence target hands-on in the lesson Why models hallucinate, part of the How AI assistants work course.

Open the lesson