LLM Clustering Benchmark

Do LLM clustering methods beat classical baselines?

Paper · ITAT 2026“Do LLM Clustering Methods Outperform Classical Baselines? A Controlled, Cost-Aware Benchmark”

• Python
• LLMs
• Clustering
• Paper

What it does

Five LLM-guided methods against k-means, Ward and spectral clustering on Banking77, CLINC150 and MASSIVE, with one shared LLM and cached embeddings. Against the best classical baseline, only 1 of 45 cells wins beyond seed variance, at 100–1000× the cost.

Highlights

• Methods reimplemented from their original repos on a shared backbone.
• Tracks accuracy (ACC, NMI, ARI) and cost (tokens, USD, time).
• Ships precomputed embeddings for exact reproduction.