Statistical learning involves extracting regularities of token appearances, and inferring the underlying abstract structure governing pattern formation. In graph-learning tasks, this structure corresponds to the rules defining transitions within the graph. The acquired structural knowledge should enable transfer across environments sharing the same organization but differing in observable tokens. A key question is whether individuals who learn faster within graphs also show stronger transfer across graphs, or learning rate and transfer reflect independent components of behavior.
We reanalyzed the graph-learning dataset of Mark et al. (eLife, 2024), focusing on the graphs that shared hexagonal grid structure. Within days, graphs had identical structure but different tokens labeling the states, whereas across days they were variants of the grid structure. Accuracy across blocks within each graph was analyzed using a hierarchical Bayesian model separating baseline performance, within-graph learning, and graph-specific shifts.
Transfer appeared primarily as performance shifts across graphs rather than changes in learning rate. Shifts were largest across days, whereas within-day transitions were smaller and positively correlated across participants, suggesting stable individual differences in benefiting from prior experience with the exact same structure. Transfer effects overall showed little relation to learning rate, consistent with learning and transfer reflecting distinct behavioral components.