Children, but not language models, show accelerating returns in word learning
- Source
- arxiv.org
- Author
- Michael C. Frank
- Date

Context
Children's vocabulary growth fits a model in which later language experience contributes more to learning than earlier experience. The study analyzed repeated parental vocabulary reports for 1,841 children across five datasets. A statistical model allowing learning efficiency to accelerate fit their development better than versions without that acceleration.
The tested small language models, including models trained on speech directed at children, showed no comparable acceleration. This identifies a possible clue to children's efficient learning: what changes may be their ability to learn from experience. The evidence comes from statistical comparisons, so it does not establish which developmental process causes the difference.
articleRecursive Language Models Generalize Out of DomainChenxiao Yang, Zhiyuan Li, David McAllester, Nathan Srebro
articleModular Cognitive Architecture Emerges in Large Language ModelsPengrui Han, Jacob Andreas, Evelina Fedorenko, Andrea Gregor de Varda
articleAccelerating LLM Inference via Vector Index Based Output EmbeddingsMartin Loretz, Sepp Hochreiter
Checking sign-in…
Loading comments…