A new arXiv paper reports evidence that language models with fewer than one billion parameters can track entities across natural narratives. Entity tracking means following where people or objects are, how they change, and what remains implied even when a text does not repeat every detail.
The finding matters because this skill is central to reading comprehension. If smaller models can learn it, some narrative understanding may not require only the largest frontier systems. That could help make certain language tasks cheaper to run.
The claim is limited to the study’s benchmarks and setup, including a comparison in which the models exceeded human performance. It does not mean small models understand stories as people do, but it suggests useful structure can emerge at smaller scales than many users assume.