VCR-Bench - CoT Score: leaderboard

On the same 1,034 video questions, GPT-4o splits each answer into steps and matches them against human-checked reference steps; the score is the F1 of step recall and step precision. Unlike the VCR-Bench accuracy board, a right answer reached by wrong or missing steps scores low.

Metric: CoT score (F1, %). Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 72B Instruct50.5#849
2Qwen 2.5 VL 7B Instruct38.2#1893
3MiniCPM-o-2.633.8#2417
4InternVL2.5-78B33.7#1036
5LLaVA-Video-72B-Qwen233.6#1344
6Aria33.5#2189
7VideoLLaMA3-7B32#2015
8InternVL2.5-8B28#2401
9InternVideo2.5-Chat-8B21.2#2031

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=vcr-bench-cot-score · How It Works · Data refreshed daily, snapshot 2026-11-17.