CapRL++ Prism - MVBench: leaderboard
CapRL++ (arXiv 2606.09393v1) Table 6 measures video captioners by caption utility: each model captions the video, then "a text-only LLM answers task-specific questions based solely on the generated caption", with the answering LLM fixed "to a fine-tuned Qwen2.5-3B-Instruct". The score shows how much answerable information a caption carries, not native video QA; all Qwen3-VL rows are Instruct variants and the frame count is not stated. Questions from MVBench. Metric: Accuracy (%).
Metric: MVBench accuracy (%) answered from the model's caption (Prism caption-utility protocol: the listed model captions each video, then a fixed text-only answering LLM (a fine-tuned Qwen2.5-3B-Instruct) answers the benchmark's questions from the caption alone; all Qwen3-VL models are Instruct variants; frame count not stated). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 60.5 | #291 |
| 2 | Qwen 3 VL 235B A22B Instruct | 59.5 | #281 |
| 3 | Qwen 3 VL 8B Instruct | 58.3 | #420 |
| 4 | Qwen 3 VL 4B Instruct | 56.8 | #502 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=caprl-plus-plus-prism-mvbench · How It Works · Data refreshed daily, snapshot 2026-11-06.