HUME Paper: Humans Outperform LLMs as Annotators in Text Embedding Tasks

Muennighoff · x · 2026-08-20

The new HUME paper introduces a Human Evaluation Framework for Text Embeddings. Measuring human performance across 16 MTEB datasets, researchers found humans averaged 77.6%, slightly below the best embedding model's 80.1%, though models struggled with low-resource languages. The study also benchmarked 9 LLMs as annotators, finding that despite scalability advantages, LLMs (76.1%) fell short of human performance (81.2%).

Original post →

More from Research

Research channel →