Proceedings of the Annual Meeting of the Cognitive Science Society

Cognitive theories propose that multiplication learning progresses from slow calculation-like responding toward faster responding often attributed to retrieval and automaticity. We test whether a data-driven graph-based skill dependency model, GraafTel, fit with k latent components (k=3 to 6), recovers an item representation that includes a robust RT-linked "slow" component across learning stages. Using 315,690 practice trials from 540 children ages 6 to 10 in an adaptive fact-learning system, we relate item-level component requirements to response time and to a system-estimated forgetting parameter _ that is partly RT-mediated in Level 3. Across k, one component shows a strong positive association with RT across levels and encounter positions, whereas the remaining components show weaker or more stage-dependent time associations. These results suggest that graph-based latent structure can serve as a computational marker of strategy-sensitive dissociations in large-scale learning traces.