NarrativeTrack: leaderboard
NarrativeTrack (Ha et al., 2026; Apple) evaluates entity-centric temporal reasoning for video understanding across 1,006 human-verified question-answer pairs over 406 clips drawn from AVA, Video-MME and LVBench, probing an entity's existence, action, outfit and scene changes, and identity ambiguity through binary, multiple-choice and ordering questions. The score is the average exact-match accuracy over the thirteen question cells, as a percentage where higher is better; open-source models are given 20 frames per video and proprietary models 128.
Metric: Accuracy (%; mean exact-match over 13 binary, multiple-choice and ordering question cells across five entity-centric reasoning dimensions). Source: arxiv.org. Saturation forecast: Around February 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 83.8 | #262 |
| 2 | Gemini 3.1 Pro (Preview) | 83.4 | #91 |
| 3 | Gemini 3 Flash | 79.32 | #159 |
| 4 | GPT-4.1 | 74.85 | #492 |
| 5 | Gemini 2.5 Flash | 74.25 | #452 |
| 6 | GPT-4o | 72.27 | #741 |
| 7 | Gemini 2.0 Flash | 60.24 | #702 |
| 8 | Qwen 2.5 VL 32B Instruct | 56.96 | #1055 |
| 9 | InternVL3-38B | 55.47 | #905 |
| 10 | InternVL3-8B | 48.81 | #1695 |
| 11 | Qwen 2.5 VL 7B Instruct | 48.81 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=narrativetrack · How It Works · Data refreshed daily, snapshot 2026-11-12.