Strategic Control of Working Memory Operations in Dynamic Reward-based Learning

Rapid adaptation in dynamic and shifting environments is crucial, and working memory’s (WM) capacity to swiftly integrate information is widely considered fundamental to humans’ ability to flexibly respond to unpredictable change. The distinct contribution of WM to learn- ing under fluctuating reward contingencies extends beyond mere information maintenance, to encompass strategic control mechanisms. The specific control processes humans utilize in this context and how those mechanisms impact memory contents, and thereby learning, remain unclear. Although WM is recognized to play a role in reward-based learning problems, the implementation-level details of these processes are often left ambiguous, typically character- ized by qualitative rather than information processing models of WM influence on learning and choice. This study implements a mechanistic process model of WM involving discrete information storage to identify behavioral signatures of complex ’write’ and ’clearance’ op- erations. We use this model to disentangle the relationship between these mechanisms and behavior, as well as to predict internal WM representations, in a novel task designed to high- light WM contributions to reward-based learning. Our results indicate that human behavior is best captured by models integrating strategic WM control processes, revealing complex in- teractions between individual mechanisms and learning outcomes. Furthermore, we find that the influence of these mechanisms on task performance is mediated by their impact on WM contents, which affect accuracy in a context-specific manner. These findings clarify the role of WM during reward learning tasks and demonstrate the capacity of specific WM processes to actively mediate complex learning strategies.