Extended Data Fig. 3. Pilot experiment with partial feedback (n=11).

The design was identical to Exp. 2, except that all item pairs (neighbours and non-neighbours) were presented equally frequently (like in Exp. 1). Left, Performance simulation shows a similar benefit of asymmetric updating as we observed in simulation of Exp. 2-4 (where neighbouring pairs were presented more frequently, cf. Figure 2e and S2, right). Right, Mean proportions of correct choices in the pilot experiment. The overall learning level was relatively low, with n=9 (of 20) pilot participants not meeting our inclusion threshold for above-chance performance (cf. Methods: Participants). The descriptive choice data of the remaining 11 pilot participants (shown in right) indicate a similar learning asymmetry as we observed in our main experiments.