CCTVBench: leaderboard

CCTVBench (2026) tests video LLMs on 305 real traffic-accident videos paired with generated counterfactual videos and mutually exclusive yes/no hypotheses; a question pair counts only when all four video-question answers are consistent. Quadruple accuracy (%), one question per request; higher is better.

Metric: Quadruple accuracy (%): share of question pairs in which the model answers Yes to the true hypothesis on the real video and No to the other three video-question combinations, on 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, each question in its own request; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 10 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1InternVL3.5-8B11.38#1604
2Qwen 3 VL 8B10.65#917
3Qwen 3 VL 2B7.62#2134
4InternVL3.5-2B7.05#3284
5VideoLLaMA3-7B6.58#2050
6VideoLLaMA3-2B6.48#2593
7PLLaVA-13B6.19#3377
8Video-ChatGPT4.13#6929
9PLLaVA-7B3.9#5139
10Phi-4 Multimodal Instruct2.92#4154

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cctvbench · How It Works · Data refreshed daily, snapshot 2026-11-21.