The behaviour bends, the code holds: working-memory category information in the multiple-demand network across graded interruption

Abstract

Protecting working memory (WM) from interference is essential for goal-directed behaviour, yet its neural mechanisms remain unclear. The multiple-demand (MD) network - a set of frontoparietal regions recruited by diverse cognitive demands - is a candidate system for maintaining WM content under interference. We examined how the MD network represents a memorandum when an unrelated task interrupts maintenance. Participants memorised an image of a tool or animal and, during the delay, performed a digit task whose difficulty was parametrically varied (passive, easy, hard), allowing us to test how task engagement and associated cognitive demands affected concurrently maintained information. Using fMRI and multivariate pattern analysis, we decoded the maintained category (tool vs. animal) from 13 MD regions and early visual cortex. Behaviourally, interruption reduced recall accuracy, while the most demanding interrupter slowed responses. Neurally, category information was present during the delay, most strongly in visual cortex and bilateral intraparietal sulcus. This information was unaffected by task engagement (active vs. passive), consistently across decoding accuracy and cross-validated Mahalanobis (crossnobis) distance, with Bayesian evidence supporting the null effect. Increasing the demand of the active interrupter, however, produced a modest, network-wide reduction in crossnobis distance that was not detectable in decoding accuracy. Together, these findings reveal a dissociation between the behavioural costs of interruption and the neural representation of maintained category information. Although interruption impaired recall, engaging in an interrupter did not measurably alter category-level information in the MD network or visual cortex. Increasing the demand of the interrupter produced only a modest, network-wide reduction in representational distance, without affecting decoding accuracy, suggesting that the behavioural costs of interruption may not be fully explained by changes in the categorical representation of the memorandum.