Neural and computational underpinnings of biased confidence in human reinforcement learning

This study investigated the neural and computational underpinnings of biased confidence in human reinforcement learning. Using fMRI and a reinforcement learning paradigm, the researchers found a functional dissociation between the dorsal prefrontal network (negatively correlated with condition-specific confidence) and the VMPFC network (positively encoding task-wide confidence incorporating valence-induced bias). VMPFC activity correlated better with confidence than with option values from reinforcement learning models. These findings suggest VMPFC plays a key role in building global feeling-of-confidence signals from latent decision variables and contextual biases.

Chih-Chung Ting, Nahuel Salem-Garcia, Stefano Palminteri, Jan B. Engelmann, Maël Lebreton