NarrativeBench
Models seek, include, exclude, and recover narratives from multilingual news articles. Citation-based F1 or recall scores narrative coverage after checking the response language.
The release includes 1,333 articles, 460 queries, and 4,800 evaluation cases. Five tasks use different metrics, and language or framing gaps are lower-is-better.
- Made by
- Nikhil Sharma, Kelly Marchisio, Kenton Murray, and Ziang Xiao
- Released
- November 2026
- Kind
- Reading in many languages
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Subject
- Media and communications
- Language
- Many languages
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Many languages
- Strongest model tested
-
Gemini 4 Carbon #3 on the capability index
- Last model added
- November 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
Large language models (LLMs) increasingly mediate how people seek and consume information, including accounts that differ across languages and regions, especially on contested topics. Supporting diverse information seeking requires models to help users navigate competing narratives across languages and information streams. We define six capabilities necessary for diverse multilingual information seeking: three foundational capabilities for narrative inclusion, narrative exclusion, and missing narrative identification; and three higher-level capabilities for presenting diverse narratives, preserving narrative coverage under confirmatory queries, and maintaining information parity across languages. To guide the development of those information seeking capabilities, we introduce NarrativeBench, a benchmark comprising 1,333 news articles in 23 publication languages across 23 contested topics and 460 Reddit-grounded queries. Evaluating 14 LLMs across 24 query languages, we find that current models lack the capabilities needed to reliably navigate diverse narratives, limiting users' control over the perspectives they receive and leaving substantial gaps in exposure to competing accounts even when diversity is explicitly requested. Information access also differs across query framings and languages, risking echo chambers, linguistic filter bubbles, and information disparities that weaken the common ground needed for democratic discourse. NarrativeBench provides a foundation for developing information systems that give users across languages and regions agency to navigate diverse narratives.
Gemini 4 Argon leads with 90.7%
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 90.7
Gemini 4 Argon
- 88.0
Gemini 4 Hydrogen
- 78.6
Gemini 4 Carbon
- 52.5
Gemini 3.8 Flash
- 26.0 GPT-6 Luna
- 25.2
Gemini 2.5 Pro
- 21.8
Gemini 2.5 Flash
- 21.4
Gemma 4 26B A4B
- 21.0 GPT-5.4 Mini
- 20.4
DeepSeek-V4-Flash
- 19.8 Kimi K2.6
- 18.8
Qwen3.5-9B
- 16.5 gpt-oss-120b
- 14.3
Cohere Command A
- 14.3 gpt-oss-20b
- 10.7
Qwen3.5-9B-Base
- 2.60
Qwen3.5-0.8B
- 1
Gemini 4 Argon high 90.7%
- 2
Gemini 4 Hydrogen xhigh 88%
- 3
Gemini 4 Carbon high 78.6%
- 4
Gemini 3.8 Flash low 52.5%
- 5 GPT-6 Luna 26%
- 6
Gemini 2.5 Pro Jun 2025 25.2%
- 7
Gemini 2.5 Flash Sep 2025 21.8%
- 8
Gemma 4 26B A4B Reasoning disabled 21.4%
- 9 GPT-5.4 Mini 21%
- 10
DeepSeek-V4-Flash Reasoning disabled 20.4%
- 11 Kimi K2.6 19.8%
- 12
Qwen3.5-9B 18.8%
- 13 gpt-oss-120b low, 3K 16.5%
- 14
Cohere Command A 14.3%
- 15 gpt-oss-20b medium, 16K 14.3%
- 16
Qwen3.5-9B-Base 10.7%
- 17
Qwen3.5-0.8B 2.60%
Source: Official NarrativeBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 90.7% | high | Sep 2026 | Official leaderboard |
| 2 | | 88% | xhigh | Nov 2026 | Official leaderboard |
| 3 | | 78.6% | high | Oct 2026 | Official leaderboard |
| 4 | | 52.5% | low | Sep 2026 | Official leaderboard |
| 5 | GPT-6 Luna | 26% | Sep 2026 | Official leaderboard | |
| 6 | | 25.2% | Jun 2025 | Official leaderboard | |
| 7 | | 21.8% | Sep 2025 | Official leaderboard | |
| 8 | | 21.4% | Reasoning disabled | Apr 2026 | Official leaderboard |
| 9 | GPT-5.4 Mini | 21% | Mar 2026 | Official leaderboard | |
| 10 | | 20.4% | Reasoning disabled | Apr 2026 | Official leaderboard |
| 11 | Kimi K2.6 | 19.8% | Apr 2026 | Official leaderboard | |
| 12 | | 18.8% | Feb 2026 | Official leaderboard | |
| 13 | gpt-oss-120b | 16.5% | low, 3K | Aug 2025 | Official leaderboard |
| 14 | | 14.3% | Mar 2025 | Official leaderboard | |
| 15 | gpt-oss-20b | 14.3% | medium, 16K | Aug 2025 | Official leaderboard |
| 16 | | 10.7% | Feb 2026 | Official leaderboard | |
| 17 | | 2.60% | Feb 2026 | Official leaderboard |
Rankings by part
We read 5 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Narrative inclusion, Narrative exclusion and Include Missing Narratives. Gemini 4 Hydrogen leads 3 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Narrative inclusion
17 models we track have a score here.
- 1
Gemini 4 Argon 90.7%
- 2
Gemini 4 Hydrogen 88%
- 3
Gemini 4 Carbon 78.6%
- 4
Gemini 3.8 Flash 52.5%
- 5 GPT-6 Luna 26%
- 6
Gemini 2.5 Pro Jun 2025 25.2%
- 7
Gemini 2.5 Flash Sep 2025 21.8%
- 8
Gemma 4 26B A4B 21.4%
- 9 GPT-5.4 Mini 21%
- 10
DeepSeek-V4-Flash 20.4%
Narrative exclusion
17 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 88.4%
- 2
Gemini 4 Hydrogen 88%
- 3
Gemini 4 Argon ↓2 84.2%
- 4
Gemini 2.5 Pro Jun 2025 ↑2 74.2%
- 5
Gemini 3.8 Flash ↓1 65.2%
- 6
Gemini 2.5 Flash Sep 2025 ↑1 64.8%
- 7
Gemma 4 26B A4B ↑1 61.5%
- 8 Kimi K2.6 ↑3 60.4%
- 9
Qwen3.5-9B ↑3 55.8%
- 10 GPT-5.4 Mini ↓1 54.8%
Include Missing Narratives
17 models we track have a score here.
- 1
Gemini 4 Hydrogen ↑1 96.3%
- 2
Gemini 4 Carbon ↑1 90.8%
- 3
Gemini 4 Argon ↓2 89.1%
- 4
Gemini 3.8 Flash 82.8%
- 5
Gemini 2.5 Pro Jun 2025 ↑1 79.4%
- 6
Gemini 2.5 Flash Sep 2025 ↑1 74.3%
- 7 gpt-oss-120b ↑6 73.7%
- 8 Kimi K2.6 ↑3 73%
- 9
Gemma 4 26B A4B ↓1 70%
- 10
DeepSeek-V4-Flash 64.8%
Baseline prompt
17 models we track have a score here.
- 1
Gemini 4 Hydrogen ↑1 80.3%
- 2
Gemini 4 Carbon ↑1 77.8%
- 3
Gemini 4 Argon ↓2 75.5%
- 4
Gemini 2.5 Pro Jun 2025 ↑2 70.6%
- 5
Gemini 2.5 Flash Sep 2025 ↑2 60.5%
- 6
Gemini 3.8 Flash ↓2 57.8%
- 7
Gemma 4 26B A4B ↑1 55.8%
- 8 gpt-oss-120b ↑5 51.4%
- 9
Qwen3.5-9B ↑3 48.9%
- 10 Kimi K2.6 ↑1 48.3%
Diverse prompt
17 models we track have a score here.
- 1
Gemini 4 Hydrogen ↑1 91%
- 2
Gemini 4 Carbon ↑1 86%
- 3
Gemini 4 Argon ↓2 82.4%
- 4
Gemini 2.5 Pro Jun 2025 ↑2 74.9%
- 5
Gemini 3.8 Flash ↓1 68.7%
- 6
Gemini 2.5 Flash Sep 2025 ↑1 65.9%
- 7
Gemma 4 26B A4B ↑1 61.2%
- 8 gpt-oss-120b ↑5 56.6%
- 9 GPT-5.4 Mini 55.4%
- 10
Qwen3.5-9B ↑2 52.5%
More effort adds a median 5 points
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official NarrativeBench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| high | 90.7% | ||||
| xhigh | 88% | ||||
| low | 85.3% | ||||
| high | 78.6% | ||||
| low | 78.1% | ||||
| low | 72.4% | ||||
| low | 52.5% | ||||
| GPT-6 Luna | 26% | ||||
| 25.2% | |||||
| 21.8% | |||||
| Reasoning disabled | 21.4% | ||||
| GPT-5.4 Mini | 21% | ||||
| Reasoning disabled | 20.4% | ||||
| Kimi K2.6 | 19.8% | ||||
| 18.8% | |||||
| gpt-oss-120b | low, 3K | 16.5% | |||
| 14.3% | |||||
| gpt-oss-20b | medium, 16K | 14.3% | |||
| gpt-oss-20b | low, 3K | 12.9% | |||
| 10.7% | |||||
| 2.60% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official NarrativeBench leaderboard
Gemini 4 Argon on the official leaderboard passed 85% in September 2026
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official NarrativeBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Multilingual
- MMMLU In use MMLU professionally translated into 14 languages. Multiple-choice accuracy. Claude Mythos Preview 92.7%
- WMT24++ In use English-to-55-language translation of WMT24 test data. Scored with automatic metrics such as MetricX or xCOMET.
K-EXAONE 90.5%
- MMLU-ProX In use MMLU-Pro translated into up to 29 languages with expert review. Multiple-choice accuracy per language.
Qwen3.7-Max 87%
- CLIcK In use Models answer 1,995 multiple-choice questions on Korean language and culture. Questions come from official Korean exams and textbooks in eleven categories.
Solar Open 2 90.7%
- Global MMLU Lite 3.0 In use Models answer 400 human-translated MMLU questions per language in 23 languages. Version 3.0 adds Czech, Hungarian, Oriya, Slovak and Tajik and updates Italian. Accuracy. Claude Opus 5 95.7%
- MMMB In use Models answer multiple-choice questions about images in six languages across 15 categories. The set has about 12,000 questions. LFM2.5-VL-3B 83%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (17 entries).
- Pass Models scored 17 models we track.
- Pass Newest score The newest is from November 2026, for Gemini 4 Hydrogen.
- Note Frontier models 2 of the top 10 have a score, which is fine for a retired test. The strongest tested is Gemini 4 Carbon (#3).
- Note Headroom Saturated. The best is 90.7%, past 85% since September 2026.
Leaderboard import 3 pass
- Pass Importer We read the board on isle-dev.github.io: 21 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.