FORWARD AND BACKWARD HUMAN REPLAY DURING SEQUENCE LEARNING

Core cognitive functions such as memory, and learning emerge when neural activity is decoupled from external input, yet the mechanisms underlying this spontaneous activity remain debated. One influential framework proposes that these dynamics implement statistical inference. Neural replay, the sequential reactivation of neural states within fast oscillations, has been proposed as a candidate mechanism. Competing accounts propose that replay either supports goal-directed learning or forms offline, behaviour-independent representations. To distinguish between these views, we recorded high-density electrophysiology while 40 human participants (mean age 24.55 ± 3.18; 20F/20M) performed an eight-block sequential category-learning task (Fig.1A). Participants categorized images from five classes, with blockwise sequence probabilities decreasing from 100% to 0% in counterbalanced order. Resting-state data were collected before the task, between blocks, and after task completion. Behavioural learning slopes revealed two subgroups: high learners, who improved across the sequence, and low learners, who did not. Neural replay was assessed using temporally delayed modelling to detect forward or backward replay during rest (Fig.1B). Replay emerged only during inter-block rest periods, not offline after task completion. High learners showed early forward replay compressed to ~40 ms, which diminished over time, whereas low learners exhibited later-emerging backward replay at ~50 ms. Replay effects were localized to occipital sensors. Forward replay in high learners was associated with high- frequency (>90 Hz) activity (Fig.1C), while backward replay in low learners involved lower frequencies (20–60 Hz; Fig.1D). These findings indicate that human neural replay primarily supports online optimisation during learning, rather than offline consolidation.