The somatic marker hypothesis (SMH) holds that bodily arousal signals shape decisions before their basis can be stated. This study tested whether pre-choice pupil dilation during reinforcement learning behaves as the SMH predicts, or is better explained by reward expectation (arousal tracking the gain an option is expected to bring) or decision uncertainty (arousal reflecting conflict between closely matched options). A sample of 32 students learned by feedback which of four abstract stimuli, with reward probabilities of 80%, 70%, 30% and 20%, were more likely to pay out, while pupil diameter was recorded. Stimuli appeared one at a time and the response was cued after both had been removed.

Learned values (Q-values) came from a hierarchical Bayesian Q-learning model. The SMH predicts a negative linear relation between a stimulus’s Q-value and the pupil response, and reward expectation a positive one. Neither was found, although an exploratory analysis revealed an inverted-U shape relationship, with dilation peaking at intermediate values, which are the most difficult choices, supporting the decision uncertainty account. At the decision, dilation grew larger as the two options approached equal value, b = −0.54, p = .002, consistent with decision uncertainty. Pre-choice dilation in reinforcement learning may therefore reflect decision uncertainty more than a marker attached to individual options.