CCTVBench - Balanced Accuracy: leaderboard
CCTVBench (2026) plain binary classification quality: balanced accuracy (%) over all 7,104 yes/no video-question instances of real and counterfactual traffic videos; proprietary API models answer each video's questions in one batched request; higher is better.
Metric: Balanced accuracy (%) over all binary video-question instances of CCTVBench's 305 contrastive traffic scenes (a real accident video paired with a world-model-generated counterfactual video) and 1,776 pairs of mutually exclusive yes/no hypothesis questions (7,104 binary instances), single-word Yes/No answers at temperature 0, proprietary models answer per-video batches, parsed as independent binary responses; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 14 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3.5-8B | 64.85 | #1604 |
| 2 | Qwen 3 VL 2B | 62.96 | #2134 |
| 3 | Qwen 3 VL 8B | 60.26 | #917 |
| 4 | PLLaVA-13B | 59.81 | #3377 |
| 5 | VideoLLaMA3-7B | 58.47 | #2050 |
| 6 | InternVL3.5-2B | 58.46 | #3284 |
| 7 | PLLaVA-7B | 57.38 | #5139 |
| 8 | Phi-4 Multimodal Instruct | 57.07 | #4154 |
| 9 | VideoLLaMA3-2B | 55.38 | #2593 |
| 10 | Video-ChatGPT | 49.85 | #6929 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=cctvbench-balanced-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.