As machine learning datasets grow, they often do so by adding samples that are semantically similar to what is already present, rather than introducing genuinely new information. This is especially common in pipelines that blend human-authored content with AI-generated data, where the sheer volume of generated samples can mask a lack of true diversity. The result is a phenomenon sometimes called semantic collapse, where a dataset appears large but offers little variety to a model.
A new arXiv preprint proposes PLDiv, a method that uses persistence landscapes—a tool from topological data analysis—to quantify dataset diversity and detect semantic collapse. By representing the geometric structure of a dataset's semantic content, PLDiv aims to give researchers a concrete way to measure whether additional samples are expanding the dataset's coverage or merely repeating existing patterns.
The work is described only in the abstract at this stage, so details on the exact algorithm and experimental validation are not yet available. Still, the motivation is clear: as AI-generated data becomes more common in training and evaluation, tools like PLDiv could help ensure that bigger datasets also mean more diverse ones.