Researchers have trained a model on readers' highlights that can point a newcomer to where earlier readers converged on a page, and it works even when a page has few readers. The approach turns collective reading behaviour into a navigational signal.

A year of stronger language models suggests two upgrades, according to the paper's abstract: use the language models directly, or use them as a teacher. The title frames the underlying question as whether more reader data or a better teacher improves the model.

The abstract is truncated, so the details of the second upgrade — the teacher route — are not fully specified in the source. With only one source available, there are no conflicting findings to compare; the paper's own framing is the only evidence presented.