Working memory constructs joint probabilistic task representations for decision-making

The human brain has a remarkable ability in extracting and integrating relevant data for guiding actions and decisions. This capacity in part depends on working memory (WM), which maintains and manipulates task-relevant information in the service of goal-directed behavior. Theories and experimental evidence suggest that the mnemonic mechanisms of WM functions probabilistically, implying its potential to form a joint distribution for integrating multiple working memory representations. Yet, it remains an open question whether this probabilistic operation underpins the WM process in constructing task representation for guiding decisions, especially in the presence of multiple WM inputs. Our study investigates whether WM integrates multifaceted information probabilistically or deterministically. We designed a novel task requiring subjects to make decisions based on multi-dimensional WM content, with four levels of ambiguity associated with each dimension of WM features. We observed that response time and error rates increase with the cumulative ambiguity of WM representations. Through computational modeling, we found that a probabilistic model, which integrates WM uncertainty, outperformed deterministic models. This suggests that WM likely employs a probabilistic operation to integrate multiple representations, guiding decision-making.