Fig. 3: Overview of the main analyses for experiment 1.

In this experiment, participants learned about the personalities of four strangers. Results indicate that participants used fine-grained correlation structures during learning. a Model comparison results using fixed-effects analysis (losing model as reference) indicate Model 5 [Fine Granularity (FG) & Population Reference Point (RP)] as the best fitting model (n = 35). b Simulated data (n = 35), the best performing model indicates which of the models performs the task most optimally (see Supplementary Fig. 7). The best performing model (Model 5) demonstrates that participants used the best strategy. c A decrease of the prediction errors (PEs) over time can be interpreted as learning. Both plots display the average absolute PEs over time ± SEM. We calculated a pairwise Pearson correlation between trial number and the mean absolute PEs to determine if the PEs decrease over time. Top) Participants’ data shows a decrease in the PEs over time (ρ:−0.523, least squares line (red)). Bottom) Simulated data using Model 5 shows a similar decrease in PEs over time (ρ: −0.563). d General linear model (GLM) on three core model features: (1) Rescorla–Wagner RL, 2) coarse models, and 3) fine models that predict the accuracy per trial per participant. Only the third regressors was significant (n = 35), indicating participants’ use of fine granularity: (one-sided t-test) regressor 1: t(34) = 0.927, p = 0.8198, regressor 2: t(34) = 1.2915, p = 0.8974, regressor 3: t(34) = −1.7109, p = 0.0481. Individual data points are participants’ parameter estimates which are summarized by boxplots (median (middle line), 25th, and 75th percentile (box), the whiskers extend to most extreme data points not considered outliers (1.5 times interquartile range), outliers are indicated with + signs). Conclusions based on this GLM should take into account that all three regressors are highly correlated (ρ between 0.76 and 0.92). [One-sided t-test; * indicates p < 0.05, ** indicates p < 0.001, no correction for multiple comparisons]. CG coarse granularity, FG fine granularity, RP reference point, # number of, PEs prediction errors, SEM standard error of the mean, LSLine least squares line.