FineBench: leaderboard
Fine-grained human activity understanding in video, FineBench (Faure et al., 2026): four-option questions generated from AVA's human-annotated action labels ask what a designated person in a 15-minute movie clip is doing, covering body movement, interactions with other people and object manipulation. Accuracy in percent over all 199,420 questions (chance 25%); higher is better.
Metric: Accuracy (%; four-option multiple-choice questions on fine-grained human activity (movement, interaction and object manipulation; question-weighted mean); all 64 AVA movie clips, 199,420 questions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B Instruct | 68.8 | #1842 |
| 2 | MiniCPM-V-2.6 | 59.2 | #3095 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finebench · How It Works · Data refreshed daily, snapshot 2026-11-12.