Researchers build a cheaper way to organize text by topic using AI labels and lightweight models

A team combined large language models with simpler clustering methods to sort text into precise topics using far fewer AI queries than existing approaches. This means researchers and text analysts can now build interpretable topic models that run locally and cost less, without needing the scale of commercial AI systems.

Text analysis at scale — sorting through corpora to find what people actually claim and argue about — has been trapped between two bad choices: use massive language models and pay for it, or use cheap methods that miss nuance. This paper shows you can get precision without the cost. The practical effect is that smaller organizations, academic labs, and individual researchers can now build systems to track what's being said about a narrow topic (a policy dispute, an election, a scientific claim) in ways that are fast to debug and deployable on their own servers, not dependent on expensive external APIs.