NARU - Subtext Interpretation: leaderboard
A NARU (2026) category testing what an utterance means beneath its surface given the context, with four-choice questions on long-form Japanese video. Accuracy in percent; higher is better.
Metric: Accuracy (%; the 148 cultural questions of category C.3; four-choice questions over 155 Japanese long-form videos (146.8 hours), each model with its family's recommended input (Gemini through native video APIs at 0.25 fps, open models at the largest uniform frame count their context allows)). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 64.2 | #135 |
| 2 | Gemini 3 Flash | 57.4 | #159 |
| 3 | Gemini 2.5 Flash | 48.6 | #452 |
| 4 | Qwen 3.5 9B | 34.5 | #803 |
| 5 | Qwen 2.5 VL 7B | 25.7 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=naru-subtext-interpretation · How It Works · Data refreshed daily, snapshot 2026-11-12.