Figure 5. Individual differences in belief-updating complexity were reflected in across-hazard modulations of baseline, but not evoked, pupil diameter.

(a) Strength of modulation of baseline pupil diameter by absolute belief strength from the complex learning model plotted as a function of the subject’s belief-updating complexity. Subjects using simple strategies tended to have baseline pupil dynamics that were uncorrelated with the complex model’s belief-updating process, whereas subjects using complex strategies tended to have baseline pupil dynamics that were anti-correlated (i.e., pupils were less dilated when the model predicted higher belief certainty) with the complex model’s belief-updating process. The solid line is a linear fit. (b) Strength of modulation of peak evoked changes from baseline by surprise from the complex learning model plotted as a function of the subject’s belief-updating complexity. Evoked changes in pupil diameter tended to be modulated positively by surprise in the complex model, regardless of the subject’s belief-updating complexity. Points correspond to data from individual subjects. In both panels, the ordinates show the strength (computed via linear regression) of pupil modulations with respect to a particular, complex belief-updating strategy: the Bayesian model with a wide prior on hazard (Figure 3—figure supplement 1).