Experimental procedure.

(A) Illustration of a single trial in Experiment 1 (not to scale). Subjects were briefly presented with 1, 2, 4, 6, or 8 Gabor patches, which they had to keep in memory during the delay period. Thereafter, a randomly oriented Gabor patch would appear at one of the previous stimulus locations. The task was to match the orientation of this stimulus with the remembered orientation of the stimulus that had appeared earlier at this location. After submitting their response, the subject was presented with feedback about the size of their response error. Subjects received points on each trial (larger error = fewer points), which after the experiment were converted into monetary reward. Different groups of subjects received different levels of reward. (B) Predictions of the resource-rational model about the effect of reward level (x axis) on VWM error (y axis). The predictions were obtained by simulating responses of the model presented in Van den Berg & Ma (2018). Simulations were performed at five set sizes (separate lines) and 5 reward levels. Each of the simulations was performed six times, with each run using the maximum-likelihood parameters of one of the six subjects in experiment E4 Van den Berg & Ma (2018). Error bars represent ±1 SEM across the six runs. A two-way Bayesian ANOVA on the simulation data show strong evidence for an effect of both set size (BFincl > 1013) and reward level (BFincl > 108), as well as for an interaction effect (BFincl = 98).