Conference Paper Published

Abstract

How similar are the learning trajectories of language models and children? Recent work has narrowed the data-efficiency gap by training language models on child-scale input—roughly 108 tokens by early adolescence. However, evaluation remains largely based on adult-oriented English benchmarks and rarely involves direct comparison with human developmental data. We introduce ChineseDevBench, a Mandarin developmental benchmark comprising eight tasks that probe word meaning comprehension, association structure, acquisition dynamics, and language production. Crucially, the benchmark includes behavioral data from both children and adults, enabling direct model–human comparison. We train Chinese GPT-2 models on child-scale Mandarin input using an age-based curriculum and evaluate alignment between model and human response patterns across training checkpoints. Across most tasks, alignment improves with training but developmental age does not predict performance. ChineseDevBench provides a framework for systematically characterizing where model learning converges with—and diverges from—human language development.

Keywords