NextMotionQA - Error Correction: leaderboard
NextMotionQA (2026) tests how well vision-language models understand human motion shown as rendered 3D motion clips, with expert-verified items stratified by body part, direction, action and difficulty. This board is correction of 400 captions with a planted error, the mean of span identification, token recall and judged correctness on a 0 to 100 scale; higher is better.
Metric: Task 3 caption error correction score (0-100): mean of error-span identification, token recall against the gold span and Gemini-judged correctness of the rewrite, on 400 captions of rendered AMASS human-motion clips given to the model as RGB video; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 Plus | 65.25 | #222 |
| 2 | Qwen 3.5 27B | 62.85 | #359 |
| 3 | Qwen 3.5 4B | 56.35 | #1175 |
| 4 | Qwen 3.5 9B | 52.34 | #803 |
| 5 | GPT-5.4 Mini | 43.57 | #389 |
| 6 | Qwen 3.5 0.8B | 11.41 | #8081 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=nextmotionqa-error-correction · How It Works · Data refreshed daily, snapshot 2026-11-13.