Midbrain Dopaminergic State Influences Reward Learning Maintenance but Not Acquisition
Abstract
Reinforcement learning is known to depend on dopaminergic activity, with reward prediction errors signalling expectation violations. Whilst this formalism provides a canonical understanding of this framework, humans appear to be subject to various sources of noise, often captured using a variety of different learning biases (Dubois et al., 2021; Wagner et al., 2025; Wyart & Koechlin, 2016). Here, we speculated that this learning variability might be driven by fluctuations in dopaminergic midbrain activity. Such spontaneous fluctuations have been observed widely in the brain, and we previously showed that they are behaviourally meaningful (Chew et al., 2019). To probe this, we deployed a real-time fMRI approach whilst participants learned a stationary probabilistic bandit task. Using activity-dependent trial presentation, we found that low dopaminergic activity was associated with a stronger preference for the most rewarding stimuli. This effect was most prominent during later learning stages, especially in the final block of the task (F(3,880) = 2.86, p = .036). Our findings extend the behavioural impact of endogenous dopaminergic fluctuations to reward learning.