NovBench - Coverage: leaderboard

NovBench (2026) tests whether LLMs can write the novelty assessment of a research paper from the novelty claims in its introduction, on 1,684 EMNLP 2023 papers with expert peer reviews. Coverage is the share of the expert reviewers novelty comments that a generated statement matches by embedding similarity, from 0 to 1, higher is better.

Metric: Coverage (0-1): the share of the consolidated human novelty comments matched by a generated statement with embedding cosine similarity above 0.7; zero-shot prompt; 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: Not forecast. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1SEA-E0.261
2SEA-S0.2576
3CycleReviewer-70B0.2198#1880
4Gemini 2.5 Flash0.212#463
5DeepSeek R1 Distill Llama 70B0.2074#1669
6Qwen 3 32B0.2065#972
7DeepSeek R1 Distill Qwen 14B0.2053#2771
8Qwen 3 8B0.1996#1873
9Qwen 3 14B0.1974#1269
10GPT-OSS-20B0.1843#1183
11GPT-OSS-120B0.1787#704
12DeepSeek-R1-8B (NovBench checkpoint unspecified)0.1603
13GPT-50.1591#178
14CycleReviewer-8B0.1577
15DeepReviewer-14B0.1173#844

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=novbench-coverage · How It Works · Data refreshed daily, snapshot 2026-11-21.