Fig. 3. Empirical results and model fits.

a,b, Mean proportions of correct choices observed over all trials in each experiment (with full feedback in a and partial feedback in b). c,d, Mean choice probabilities predicted by the best-fitting model in each experiment. e,f, Model comparison (for Exp. 1 in e and Exps. 2–4 in f). The markers show model fits using a pseudo-R² (left y axis; the diamonds and error bars show mean ± s.e.m., and the dots show individual participants). R² is inversely related to BIC, with larger values indicating better fit. Intuitively, R² = 0 is equivalent to random chance, while R² = 1 corresponds to a theoretically perfect model. The overlaid red bar graphs indicate each model’s probability of describing the majority of participants best (right y axis; Methods). The model space is described with the following nomenclature: Q, item-level learning; 1/2, symmetric/asymmetric; *, difference-weighted updating; P, pair-level learning; i, pair-relation-based inference.