Influence of intrinsic rewards on working memory allocation

Visual working memory (VWM) has been characterized as a limited resource shared between visual stimuli. Previous studies demonstrated that this resource can be flexibly allocated depending on task demands, e.g. to prioritize an item that is more likely to be tested for recall. In these cases, the observed resource allocation has been interpreted as evidence for an optimal strategy that minimizes expected errors. In the present study, observers memorized two motion stimuli of different colours and later reproduced the motion direction of one cued item. In the first experiment, we manipulated the feedback of the true motion given at the end of each trial, artificially increasing the error for items of one colour and decreasing it for the other. Modelling resource sharing in VWM showed that the error-minimizing strategy was to allocate more resources to the items for which feedback error was magnified. Surprisingly, we found strong evidence that observers instead allocated more resources to the stimuli for which the feedback error was reduced, i.e. those perceived as easier to remember. This preference for ‘easier’ items was confirmed in a second experiment where we manipulated the true difficulty by using different motion coherences for the two colours. Two further experiments tested optimal allocation in contexts with previously documented unequal allocation, specifically, variable cueing probability and variable rewards. Despite replicating previously observed patterns, modelling again revealed suboptimal resource allocation. Altogether, these results challenge previous assumptions about how resources are shared in VWM, and suggest an implicit bias in allocation towards stimuli associated with lower estimation uncertainty, which observers may experience as intrinsically rewarding. We show that such a bias provides an alternative account of seemingly error-minimizing strategies observed in previous studies, while also being consistent with the broader literature on human preferences for rewarding stimuli.