A flexible coding scheme underlying working memory generalization in human parietal and frontal cortices

Abstract

Humans can rapidly extract abstract, common knowledge from distinct experiences, enabling generalization of this knowledge across tasks within working memory (WM) to guide behavior. Although task-specific representations of WM have been observed in a distributed network encompassing sensory, parietal, and frontal cortices, whether these regions construct task-generalizable representations and whether such representations can be dynamically shaped by task demands remain unclear. Across two functional MRI experiments, we investigated WM generalization across distinct WM tasks (location and object tasks) based on shared underlying task structures. Trial-specific task demands varied between passive maintenance and active, rule-guided manipulation of mnemonic stimuli. The posterior parietal cortex (PPC) demonstrated task-generalizable stimulus representations during both maintenance and manipulation. In contrast, the lateral (LFC) and medial (MFC) frontal cortices supported manipulation- and maintenance-based generalization, respectively. Critically, manipulation-based generalization in the PPC and LFC emerged even when participants did not explicitly learn the mapping between task spaces, indicating that active exploration of task structure can spontaneously facilitate generalization across tasks. Together, these findings reveal that flexible generalization for goal-directed behavior is achieved via a distributed WM network, with distinct regions serving active versus passive task demands.