Stanford's Frank shows children beat LMs with accelerating word learning

In a new arXiv paper, Stanford professor Michael C. Frank challenges the assumption that language models learn like children. Analyzing vocabulary growth trajectories, Frank shows children's word learning is best modeled as accelerating accumulation: each additional unit of linguistic experience produces more learning than the one before, a virtuous cycle that kicks in during early development. This contrasts sharply with the scaling-law behavior of neural language models, which exhibit constant proportional returns on new data—each token contributes a fixed relative improvement, no matter how much training data is added.

The gap persists even when language models are trained exclusively on child-directed speech, the exact input a human toddler might receive. While children master hundreds of words with a fraction of the data, LMs need billions of tokens to approximate even basic vocabulary knowledge. Frank suggests the key difference lies in children's increasingly efficient use of their learning input—a mechanism absent in current architectures. The findings, posted as arXiv:2608.17120 on August 17, 2026, add to a growing body of research questioning whether scaling alone can replicate human language acquisition, and highlight the need for models that optimize not just for data volume, but for learning efficiency.