Answers are grounded in this paper and cite their sources.
As the interest in language models (LMs) gains traction in academia, there has been an impulse to compare LMs with children— the idea seems to be that if these models perform well, perhaps this reveals something about the acquisition process. The problem with this line of reasoning is that it is often at odds with the nature of the input, as LMs are typically trained on billions of words extracted from written sources. In order to transcend the difficulties with making such a comparison, there have been attempts to scale down these models to better approximate the input that children are exposed to in real life. In this research, we focus on word learning and compare human and model performance in the adjectival domain. Here we compare humans and models directly by first running an experimental study with participants using a modified version of the Human Simulation Paradigm, in which we analyze human learners’ abilities to generate frame-compliant responses to syntactic frames. We then run the same task with BabyBERTa, an LM that is trained specifically on Child Language Data Exchange System (CHILDES) data. For human learners, we find that frames trigger adjective responses, and responses shift as a result of the frame, and that additional semantic information enhances subcategorization patterns. Conversely, BabyBERTa diverges from our participants—at times failing to predict adjectives, and not improving when semantic information is included. More troubling, similar mismatches continue at different scales of LMs, suggesting a flaw with the claim that models are representative of acquisition. While LMs can generate grammatical results, they fail to capture human performance. Although the Human Simulation Paradigm (HSP) and BabyBERTa have both been argued to replicate the acquisition process, divergent results suggest that models are not learning in the same way as humans.
No takes yet. Share an insight, caveat, or question.
Gotowski et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: