Dissociating Value and Motor Learning in Human Habit Formation: Insights from Neuromodulation and Computational Modeling

Habit formation in humans emerges from the interaction between reward-guided choice and the progressive stabilization of stimulus-action associations. It remains debated whether these components rely on shared or dissociable mechanisms. Concurrently, recent experiments in rodents demonstrate that phasic frontostriatal dopaminergic signals not only encode reward prediction errors (RPEs), but also action prediction errors (APEs). However, studies in humans requiring complex motor execution are missing. To investigate this, 22 healthy participants performed a reward probabilistic learning task, selecting between two visual stimuli via continuous joystick movements. One stimulus carried a higher reward probability per block, with contingencies changing unannounced. Reward delivery was conditional on successful motor execution. Participants underwent three, single-blind, inhibitory transcranial static magnetic stimulation (tSMS) sessions: real and sham medial prefrontal (PFC) and real parieto-occipital (PO). Behavioural analyses assessed value and motor accuracy. We developed a new dual-controller reinforcement learning model, combining value learning and an action-history component, respectively updated through RPEs and APEs. Model comparisons strongly supported the dual-controller account. Value update was gated by motor success. Moreover, the action-history module included a strong decay, especially between blocks. tSMS conditions did not significantly affect behavior, but PO inhibition increased the action-history’s learning rate. These findings support the idea that human learning in probabilistic tasks emerges depends on dissociable value learning and action stabilization processes. Moreover, this study helps bridge the gap between animal learning models and human paradigms in requiring complex, non-ballistic responses.