Researchers introduce Cluster-R1, a reasoning-based AI system that outperforms standard embedding models by autonomously interpreting user instructions to determine optimal text clustering without manual intervention.

Large reasoning models (LRMs) are trained as clustering agents to overcome limitations of general embedding models that fail to follow specific user instructions

The approach reframes clustering as a generative task, enabling LRMs to both interpret high-level instructions and autonomously infer latent corpus structures like optimal cluster numbers

ReasonCluster benchmark introduced with 28 diverse clustering tasks spanning domains including daily dialogue, legal cases, and financial reports

Reasoning-driven training pipeline demonstrates consistent performance improvements over competing approaches across diverse datasets and clustering scenarios