Reviso Captioning - CaReBench Action F1: leaderboard
Reviso Table 1: common maximum 128 input frames, uniform sampling for direct model baselines and adaptive revisits for trained Reviso. Original native metric units retained. Proprietary rows are the authors’ reference runs and remain evidence, despite their highlighting convention; no site exclusion is inferred. Reviso is a trained Qwen 3.5-9B derivative with tools and remains distinct. Reasoning effort otherwise not reported. Metric: F 1 (%).
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Around 2031. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 44.3 | #6 |
| 2 | Gemini 3.8 Flash | 39.7 | #30 |
| 3 | Qwen 3 VL 32B (Thinking) | 33.3 | #291 |
| 4 | Qwen 3 VL 32B Instruct | 32.2 | #290 |
| 5 | Qwen 3 VL 8B Instruct | 29.5 | #419 |
| 6 | Qwen 3.8 27B | 29.2 | #123 |
| 7 | Molmo2-8B | 29 | |
| 8 | Qwen 3.6 27B | 25.5 | #182 |
| 9 | Qwen 3.5 9B | 25.2 | #391 |
| 10 | Qwen 3.6 35B A3B | 24 | #238 |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=reviso-captioning-carebench-action-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.