From Children to Checkpoints: Parallel Shifts from Duplication to Relational Reasoning

When children learn to solve analogies such as ○:●●::□:?, they often begin with associative reasoning, most commonly duplicating the final element (□), before increasingly producing relational responses (■ ■). Eye-tracking studies further show that this behavioural transition is associated with a shift away from disproportionate attention to the final element and towards more balanced attention across relevant problem elements. Large language models (LLMs) likewise show duplication responses, particularly at smaller scales or earlier stages of training. We investigate this relational shift in children and across checkpoints of OLMo. First, we show that OLMo shifts from duplication to relational responding across training in both verbal analogy and abstract sequence-completion tasks. Second, we compare the eye-tracking profiles of children classified by predominantly duplicative versus relational responses with changes in OLMo’s activation-based profiles between pre- and post-shift checkpoints. We find mixed, task-dependent evidence for comparable changes in information allocation: verbal analogies show partial correspondence, whereas abstract reasoning reveals divergent child and model profiles. Finally, exploratory intervention analyses provide preliminary evidence that OLMo’s shift may involve control-like reweighting of competing duplication and relational-retrieval pathways. These findings illustrate how developmental comparisons can generate and constrain mechanistic hypotheses in both cognitive development and AI research.