Conference on Cognitive Computational Neuroscience

Advice-giving is an instrumental component of teaching, mentorship, and a common component of human interaction. The effect of receiving a piece of advice can range from inconsequential to life-changing. How would you choose the best advice to give? We hypothesize that good advice combines accurate inference about the correct answer in the given context and accurate inference about the recipient’s knowledge and tendencies. We created a novel task where a learner agent navigates through a graph. A teacher, who does not know what edges are known to the learner, gives a single piece of advice by revealing one edge, to improve the learner's total rewards. We find that participants give a mixture of advice tailored to the other's knowledge and advice the is more generic. We show that tailored advice generated by an Optimal Bayesian Mentor (OBM) model that infers the learner's unobservable knowledge from their previous trajectory outperforms generic advice generated by a Prior Only Mentor (POM). These findings suggest that an estimate of the other's unobservable factors is key to selecting which piece of information will most benefit the other.