FitAQA - Perception: leaderboard
Metric: Question-macro accuracy (%; recognizing the quality-relevant visual manifestation, equal weight per unique question; over FitAQA's fitness action-quality videos (30 bodyweight exercises, 38-error taxonomy); perception task). Source: arxiv.org. Saturation forecast: Around 2030. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 54.3 |
| 2 | GPT-5.5 | 54.1 |
| 3 | GPT-5.4 | 51.3 |
| 4 | Gemma 4 31B | 46.3 |
| 5 | Qwen 3.5 35B A3B (Thinking) | 45.9 |
| 6 | Qwen 3.5 27B (Thinking) | 45.7 |
| 7 | Qwen 3.5 35B A3B (Non-reasoning) | 45.6 |
| 8 | Qwen 3.5 27B (Non-reasoning) | 45.2 |
| 9 | Qwen 3.5 9B (Non-reasoning) | 44.6 |
| 10 | Qwen 3 VL 8B | 44.2 |
| 11 | Qwen 3.5 9B (Thinking) | 43.9 |
Interactive version: theaggregate.ai/benchmark?slug=fitaqa-perception · How It Works · Data refreshed daily, snapshot 2026-09-26.