Figure 3. Comparison of cross-validated prediction weights for high- and low-improvement groups.

Relational score gains were predicted separately for the eleven highest and eleven lowest improvement participants across 11 leave-one-out fits ( $\text{low}{R}_{cv}^2=.30;\text{high}{R}_{cv}^2=.44$). Each value was predicted by a separate model that had no access to the data for the corresponding individual. Panel (a) shows the average prediction weight matrix for the low-improvement group and panel (b) for the high-improvement group. A comparison of the prediction weight matrices shows markedly more diffuse scanning in the low-improvement group (panel a) and a gain in systematicity in the high-improvement group (panel b).