Can LLMs Grow Like Children? A Developmental Analysis of Capability Trajectories in Language Model Training

Within two years, an infant who can barely lift her head learns to sit, reach, and walk. Developmental science found that this growth is both continuous and phased: skills improve smoothly, yet the pace of improvement changes at identifiable ages, and the field built statistical tools to locate those changes. Language models also transform across training, and whether their capabilities grow smoothly or in jumps is actively debated. The present study argues that the developmental answer applies to artificial learners as well, and that the developmental tools can show it. We apply an analysis pipeline built for infant motor development to the training trajectories of 16 open-source language models, each tested at 27 points during training on five benchmarks. The results mirror the infant findings. Growth is continuous but organized into phases, and transitions cluster at a small set of shared points that are deployed differently across domains. Growth proceeds slowly, then rapidly, then slowly, and models differ reliably in developmental rate. Unlike infants, those rate differences are almost entirely explained by model size. The transitions also survive changes of measurement scale, and on a continuous measure they appear earlier and more sharply than on accuracy. The paper contributes a formal test for phase structure in capability growth, a construct mapping between infant development and language model development, and evidence that phase structure may be a shared property of learning systems.