LLMs learn different forms of metacognition when trained to predict their own accuracy
This work investigates how large language models acquire metacognitive monitoring, the ability to know what one knows, by training 10 open-weight LLMs to predict their own accuracy on factual multiple-choice questions before answering them and finds that trained confidence reflects two distinct signals.
Nicolas Yax, Stefano Palminteri, P. Oudeyer
· 0 citations