Researchers find that AI models learn more robustly from expert-novice interactions than from expert demonstrations alone

Cognitive science study examined which features of multi-person interactions improve AI learning performance

Transformer models trained on pedagogical interactions between experts and novices showed greater robustness across diverse scenarios

Controlled experiments used synthetic datasets of spatial navigation tasks to precisely measure differences between interactive learning and solo expert demonstrations

Ability to represent expertise emerged as a key factor supporting the effectiveness of interaction-based training data