Fig. 4: Dynamics of compositionality in SWM with increased length.

a Illustration of two types of neuronal resource allocation strategies as new items are incorporated into WM. Recruit: extra neurons are recruited. Recycle: neurons initially engaged in earlier items are reused. Recycled neurons include stable (green; encode both earlier and new items) and flexible (blue; switch from earlier to new items) subtypes. b Proportion of recycled neurons and extra neurons for ranks 1-3 in length-3. c An example neuron previously encoding an earlier item was recycled to encode a new item (L3-R3). Shading: SEMs across trials. Line/dot colors correspond to spatial locations. d-g Early condition results (WM not saturated), comparing across lengths (L1-R1 vs. L2-R1) and ranks (L2-R1 vs. L2-R2). d Subspace comparison across lengths, similar to Fig. 3h. e Subspace comparison across ranks. Same format as Fig. 3f. f Correlation between generalization (1-|NSS| across lengths) and orthogonality (|NSS| across ranks) for recycled neurons (n = 363). g Correlation between generalization (180-φ_diff across lengths) and orthogonality (φ_diff across ranks) in stable neurons (n = 253 neurons). h-k Late condition results (near capacity). h-i Comparing across lengths (L2-R2 vs. L3-R2) and ranks (L3-R2 vs. L3-R3). Same format as (d-e). j-k Correlation between generalization and orthogonality for recycled neurons (n = 227) and stable neurons (n = 137). f, g, j, k, The dark line shows the least squares best fit, with the blue region as the 95% confidence interval of the fitted mean, the Spearman correlation coefficient (r) is shown. l-m Saturated condition results, comparing across lengths (L3-R3 vs. L4-R3) and ranks (L4-R3 vs. L4-R4). Same format as (d-e). n Correlation of NSS across monkeys (n = 4 monkeys; both-sided Wilcoxon rank sum test). o Correlation based on φ_diff (n = 4 monkeys; both-sided Wilcoxon rank sum test). p Left: the percentage of overlapping neurons across lengths significantly decreased with increasing items stored in WM (Chi-square test, χ²(2) = 200.87, p = 0; Post hoc pairwise chi-square tests (two-sided) with Bonferroni correction. Early vs. late: χ²(1) = 5.73, p = 0.05; late vs. saturated: χ²(1) = 97.92, p = 0). Right: CDFs of φ_diff shift toward the control level indicated by a dashed line. Group differences were tested using the Kruskal–Wallis test (H(2) = 114.33, all p = 0, two-sided), followed by post hoc pairwise comparisons with Bonferroni correction. q Left: the percentage of disjoint neurons (two sided Chi-square test, χ²(2) = 5.49, p = 0.06, no multiple comparison correction was applied). Right: CDFs of φ_diff shift away from the control level. Same format as (p). Group differences were tested using the two sided Kruskal–Wallis test (H(2) = 1.78, p = 0.41, two-sided). p, q The boxes show the range from first to third quartiles divided by the median line. The whiskers extend to the most extreme data points within 1.5 times the interquartile range (Tukey method), and circles indicate individual data.