VidMsg QA: leaderboard
The multiple-choice probe of VidMsg (2026), a benchmark of implicit messages in short YouTube clips: for each clip a model picks the intended message from five options drawn from the same topic, with only sampled video frames as input. Accuracy in percent (chance 20); higher is better.
Metric: Overall accuracy (%), five-option multiple choice: pick the intended implicit message of a short YouTube clip from distractors sampled within the same topic (one question per clip over 9 topics and 52 target messages; the per-topic cells imply 395 scored clips); chance 20; visual frames only (native frame protocol of each model, at most 1 frame per second); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B Instruct | 71.9 | #1833 |
| 2 | GPT-5.4 Mini (2026-03-17) | 70.6 | #287 |
| 3 | GPT-5.4 (2026-03-05) | 70.4 | #87 |
| 4 | Qwen 2.5 VL 32B Instruct | 67.6 | #1053 |
| 5 | Molmo2-8B | 57.7 | #1483 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vidmsg-qa · How It Works · Data refreshed daily, snapshot 2026-11-13.