Predicting model responses to cognitive scaffolding from architecture and training details

Identify architectural and training-based predictors that explain and allow a priori prediction of cross-model differences in response to cognitive scaffolding interventions.

Background

Across 16 models, the authors observe dramatic variability in how models respond to cognitive structure guidance, with larger and better-aligned models benefiting while smaller models often degrade.

They explicitly state that these differences cannot currently be predicted from architectural or training details, underscoring the need for predictive factors that relate model design to intervention efficacy.