Why does every decision come with a feeling of confidence? Common roles ascribed to decision confidence include its use in post-decision wagering, credit assignment in sequential decision-making, and the computation of expected reward. In this talk I will outline another decision components for which such a sense of confidence is essential: learning to make good decisions. To show this, I will focus on diffusion models – one of the most successful models in decision neuroscience. They unrealistically assume that decision-related evidence is encoded in a readily interpretable neuron/anti-neuron pair. I will instead assume it to be encoded by a larger neural population, and will ask how the brain could learn to interpret the population’s activity to improve its choices. Bayes-optimal learning turns out to yield a learning rule whose rate of adjustment to the decision strategy is strongly modulated by decision confidence. This learning rule significantly outperforms other, common heuristics, and results in sequential choice dependencies when applied in volatile environments. We find that such learning was required to explain the behavior of rats performing an odor categorization and identification reaction time task. Standard, non-learning diffusion models and other heuristics were neither able to fit the behavior in both conditions simultaneously, nor featured the sequential choice dependencies observed in the data. Overall, this provides a theoretical justification for, and experimental validation of, the use of confidence in learning to make better decisions.