NovBench (RAG) - Relevance: leaderboard
Metric: Relevance (AvgIMS, 0-5; the paper states a 1-5 scale but prints values below 1): for each novelty-description sentence the best information-matching score against any sentence of the generated evaluation, averaged; retrieval-augmented prompt with the five most similar ACL, EMNLP and NAACL 2019-2022 titles and abstracts; 1,684 EMNLP 2023 papers: from the novelty-description sentences of the paper introduction (selected by GPT-5) the model writes a novelty evaluation as positive, neutral and negative statements, compared with the expert reviewers novelty comments (extracted by GPT-4o-mini, then deduplicated, consolidated and sorted by sentiment by GPT-4o); greedy decoding, at most 4,096 tokens; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 3.51 |
| 2 | GPT-4o | 3.45 |
| 3 | Qwen 3 8B | 3.3 |
| 4 | Qwen 3 32B | 3.26 |
| 5 | GPT-5 | 3.23 |
| 6 | Qwen 3 14B | 3.2 |
| 7 | GPT-OSS-120B | 3.14 |
| 8 | GPT-OSS-20B | 3.14 |
| 9 | DeepSeek R1 Distill Qwen 14B | 3.07 |
| 10 | DeepSeek R1 Distill Llama 70B | 3.04 |
Interactive version: theaggregate.ai/benchmark?slug=novbench-rag-relevance · How It Works · Data refreshed daily, snapshot 2026-10-07.