MARETopic uses manifold-aware rank-based prototypes for topic modeling
September 25, 2026
MARETopic is a training-free framework that selects K exemplar documents by projecting embeddings onto a low-dimensional manifold and using ordinal neighborhood structures. It outperforms neural and clustering-based models on Purity and NMI benchmarks by avoiding hubness and anisotropy distortions.
HOW THIS AFFECTS YOU
●
researcherYou can implement this to improve topic discovery accuracy without the computational overhead of training new neural architectures.