A new preprint on arXiv argues that language models are missing a fundamental part of the input that human children receive: prosodic structure. Children learn from continuous, prosodically structured speech long before they can read, yet most language models are trained on orthographic text. The authors position their work as building on earlier phonemic BabyLM training, specifically citing Goriely et al., and propose that prosody should be a critical component in this setting.
The paper's core claim is not just that phonemes matter, but that the rhythmic and intonational structure of speech is essential for modelling language acquisition at a BabyLM scale. By shifting away from orthographic input and toward phonemic representations that preserve prosody, the authors suggest a more faithful training signal for models intended to learn like children.
As this is a single preprint abstract, the full experimental details and results are not yet visible in the source. The significance lies in the direction it points: making language models more sensitive to the acoustic and structural cues that are available to human learners from the very start.