Recent topic modeling research has diverged into two broad paths, according to a new paper on arXiv. The first path uses efficient embedding-based methods that scale well but often underperform on standard clustering metrics. The second path leverages large language models (LLMs), which tend to produce better results but are not described in the same scalability terms.

The paper, titled "TopicMerge: A Divide and Conquer Approach to Topic Modeling," introduces a method that appears to bridge these two directions. By adopting a divide-and-conquer strategy, TopicMerge aims to break the topic modeling problem into smaller parts and then merge the results, potentially capturing the strengths of both approaches.

The abstract is brief and does not include experimental details or specific performance numbers. Still, the framing suggests that TopicMerge is designed as a practical compromise between speed and quality, a central concern in current topic modeling work.