student.uva.nl
What is your study programme?
What is your study programme?
Colloquium credits

Presentation Master's thesis - Eva Wardenaar - Psychological Methods

Colloquium credits

Presentation Master's thesis - Eva Wardenaar - Psychological Methods

Last modified on 19-08-2026 11:19
Attentional Divergence in Abstract Reasoning across Developing Minds and Machines
Show information for your study programme
What is your study programme?
or
Start date
21-08-2026 13:00
End date
21-08-2026 14:00
Location

Children and large language models can undergo similar shifts toward relational reasoning while their attention changes in opposite directions.

The canonical developmental relational shift involves moving from surface-based or duplication-like responses toward responses based on the underlying relational structure. For example, in the sequence ABCAABC?, a child showing duplication-like responding may repeat the final element (C), whereas relational reasoning requires identifying the underlying rule and selecting the correct answer (A).

This study compares this developmental shift in 4–6-year-old children with an analogous behavioral transition across training in OLMo 3 32B. Children and OLMo completed structurally matched abstract sequence-completion tasks, with attentional focus assessed through eye-tracking in children and self-attention weights in OLMo.

Both children and OLMo moved from duplication-like responding toward correct relational responses, but their attentional changes diverged. Children shifted gaze toward earlier sequence positions, whereas OLMo attentional profiles in causally relevant heads shifted toward later positions. Human–OLMo comparisons showed that these changes were systematically opposed. Causal interventions further showed that OLMo’s post-shift attention patterns promoted correct over duplication-like responding, whereas the human post-shift attentional profile did not functionally substitute for OLMo’s own post-shift pattern.

These findings distinguish behavioral correspondence from process-level and mechanistic alignment, showing that similar reasoning outcomes can arise through divergent attentional reorganization.