Proceedings of the Annual Meeting of the Cognitive Science Society

The cortical temporal hierarchy—sensory regions process rapidly while association regions operate slowly—is well-documented, yet whether slow processing enables deeper integration remains debated. We analyzed fMRI data from four participants viewing ~65 hours of naturalistic movies, computing Temporal Dynamics Speed (TDS) and encoding performance across seven functional networks. Networks differed significantly in TDS (F(6, 993) = 18.42, p < .001): Frontoparietal fastest (1.20), Ventral Attention slowest (0.64), Default Mode intermediate (0.78). We hypothesized slower networks would benefit more from extended temporal integration; however, this was not supported: temporal averaging impaired encoding across all networks (negative TWG, p < .001), with no relationship between TDS and integration benefit. Critically, LSTM models outperformed ridge regression—even with concatenated temporal features—by 47–57% uniformly across all networks (p < .001). These findings resolve the paradox: all networks benefit from temporal integration, but only when sequential dependencies are explicitly modeled. Code: https://github.com/xiaoyanLi629/temporal-efficiency-paradox