Fig. 7: Relating representational change to recall accuracy on the final test.

A/B. When averaging across 80 participants, word pairs that showed greater within-pair similarity after learning (A; r₍₅₈₎ = 0.46, p < 0.001, 95% CI = [0.232 0.638]), and those that showed more representational change (difference in within-pair similarity between final and initial assessment) (B; r₍₅₈₎ = 0.39, p = 0.002, 95% CI = [0.154 0.583]) were recalled with greater accuracy at the final test. N = 60 word pairs. Closed circles reflect average similarity (A) or change in similarity (B) and average final test accuracy across 80 participants. Shaded area reflects 95% CI. C More change in cue representation after learning (i.e. lower correlation between initial and final) is associated with a higher probability of recall at Day 2 (z = 2.453, p = 0.014, η_p² = 0.0015, 95% CI = [−0.772 −0.086]). D The impact of the change in target representation across learning on subsequent recall depended on the semantic relatedness of the pair and whether the pair was tested or restudied at Day 1. For unrelated pairs that were tested at Day 1, more change in the representation of the target across learning (i.e. lower values on x-axis) was associated with higher probability of subsequent recall (z = 2.691, p = 0.007, d = 0.16, 95% CI = [−1.678 −0.048]). No other relationships between target word representational change and behavior were significant. Green indicates semantically related pairs, yellow indicates semantically unrelated pairs. Dashed lines indicate tested pairs of words, non-dashed lines indicate restudied pairs of words. For both C and D, N = 3902 pairs of words from 80 participants. Values on x-axis reflect Fisher’s z-transformed correlation of cue representations across learning (C) and of targets across learning (D); values on y-axis reflect average marginal predicted probability (probability of subsequent recall across all participants across change of the predictor of interest, holding other predictors constant). Shaded areas reflect 25th–75th percentile of the average marginal predicted probability. All tests were two-tailed and no corrections were made for multiple comparisons.