Research map / Large language models / Language, safety and interpretability
HUME: Measuring the Human-Model Performance Gap in Text Embedding Tasks
Adnan El Assadi, Isaac Chung, Roman Solomatin, Niklas Muennighoff, Kenneth Enevoldsen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use best embedding model instead of human annotators | MTEB tasks · MTEB human-eval subset (26 task-language pairs, overall) | average score: 77.6 -> 80.1 | Improved |
| use best LLM annotator instead of human annotators | LLM-as-annotator · 19 task-language pairs | average score: 81.2 -> 76.1 | Hurt |
| use best LLM instead of humans on reranking | Mistral-Small-24B · Reranking | MAP: 88.3 -> 78.0 | Hurt |
| use GPT-5 instead of humans on classification | GPT-5 · Classification | accuracy: 79.1 -> 78.9 | No clear effect |
| use best LLM instead of humans on STS | Mistral-Small-24B · STS | Spearman: 76.5 -> 75.0 | Hurt |
| replace human with best model on emotion classification | embedding models · EmotionClassification | accuracy: 45.8 -> 75.4 | Improved |
| replace human with best model on ArXiv clustering | embedding models · ArxivClusteringP2P | V-measure: 49.2 -> 84.6 | Improved |
| replace human with best model on Arabic sentiment | embedding models · MultilingualSentiment (ara) | accuracy: 95.0 -> 77.5 | Hurt |
| replace human with best model on Russian sentiment | embedding models · MultilingualSentiment (rus) | accuracy: 92.5 -> 81.2 | Hurt |
| replace human with best model on Arabic STS | embedding models · STS22 (ara) | Spearman: 67.5 -> 40.9 | Hurt |
| replace human with best model on reranking | embedding models · Reranking (eng) | MAP: 87.2 -> 96.4 | Improved |
| use best model instead of human on multilingual vs English tasks (win rate vs best models) | humans · MTEB human-eval subset | win rate: 29% -> 0% | Mixed |
Closest papers by meaning
- The Scandinavian Embedding Benchmarks: Comprehensive Assessment of Multilingual and Monolingual Text Embedding (NeurIPS 2024 · similarity 0.91)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · similarity 0.90)
- MMTEB: Massive Multilingual Text Embedding Benchmark (ICLR 2025 · similarity 0.89)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.89)
- ACT as Human: Multimodal Large Language Model Data Annotation with Critical Thinking (NeurIPS 2025 · similarity 0.87)
- When is an Embedding Model More Promising than Another? (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
