NarrativeBench - NE F1 (%): leaderboard
NarrativeBench v1,4800cases perconfiguration,1333articles,460queries,23publication/24query languages. Three articles perselected narrative. Document-ID citations measure instruction-following coverage: NI target-only F1,NE target-exclusion F1,MNI omitted-narrative recall,BP/DP all-narrative recall. Wrong-language responses zero;low-confidence language detection passes. Source parentheses are SE,not scores. ExactAPI/checkpoint,budget andthinking settings retained;9BBase distinct. Five native0–100 metrics,not agreement or signed bias/gap diagnostics.
Metric: NE F1 (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gemini-2.5-pro (NarrativeBench Run: thinking-budget=128;output=3000) | 74.2 | |
| 2 | gemini-3.8-flash (NarrativeBench Run: low;output=3000) | 65.2 | |
| 3 | gemini-2.5-flash (NarrativeBench Run: thinking-budget=0;output=3000) | 64.8 | |
| 4 | gemma-4-26B-A4B-it (NarrativeBench Run: disabled;output=3000) | 61.5 | |
| 5 | kimi-k2.6 (NarrativeBench Run: disabled;output=3000) | 60.4 | |
| 6 | Qwen3.5-9B (NarrativeBench Run: disabled;output=3000) | 55.8 | |
| 7 | gpt-5.4-mini-2026-03-17 (NarrativeBench Run: reasoning=none;output=3000) | 54.8 | |
| 8 | gpt-oss-120b (NarrativeBench Run: low;output=3000) | 52.3 | |
| 9 | DeepSeek-V4-Flash (NarrativeBench Run: disabled;output=3000) | 51.2 | |
| 10 | gpt-oss-20b (NarrativeBench Run: medium;output=16384) | 49 |
Interactive version: theaggregate.ai/benchmark?slug=narrativebench-ne-f1 · How It Works · Data refreshed daily, snapshot 2026-11-12.