A new topic-modeling method finds representative documents for a corpus without training anything.
Researchers built MARETopic, a framework that picks real documents from a corpus to represent each topic, rather than generating abstract cluster centers the way most topic models do. It works by projecting document embeddings onto a lower-dimensional space and ranking neighbors instead of measuring raw distances, which sidesteps known distortions (called hubness and anisotropy) that skew distance measurements in high-dimensional embedding spaces. A greedy algorithm then picks exactly K documents whose neighborhoods collectively cover the corpus. One variant, MARETopic-Corr, scores candidates with a rank-correlation measure and beats existing neural and clustering-based topic models on two accuracy metrics, Purity and NMI, on the benchmarks with the most categories. A second variant, MARETopic-Diff, drops that scoring step for a diffusion-based one and runs 1.7 to 1.9 times faster.
The appeal here is grounding: instead of a black-box cluster center, you get an actual document you can read to understand what a topic means, generated with zero gradient updates. That matters for anyone using topic models to sort or audit large document sets, where interpretability and speed to deploy count as much as raw accuracy. A separate step borrowed from search-result ranking, Maximal Marginal Relevance, also pads out vocabulary diversity across topics without hurting coherence much.
Worth tempering the enthusiasm: MARETopic leads on topic coherence in two of three tested datasets, not all three, and this is an unreviewed arXiv preprint. Training-free methods have promised to leapfrog neural pipelines before; whether this one holds up outside its own benchmarks is still an open question.