VURB: leaderboard
Video Understanding Reward Benchmark (VURB, Wei et al., South China University of Technology, Peking University, HKU and Tencent, 2026): 2,100 pairs of long chain-of-thought answers to questions from eight video benchmarks, one correct and one incorrect answer of similar length per pair, both checked by human reviewers for reasoning that supports the final answer. The model acts as a pairwise judge eight times with the answer order shuffled and is credited when it prefers the correct answer in more than four judgments; the score is accuracy in percent over pairs with a valid verdict, and higher is better.
Metric: Majority-vote pairwise accuracy (%; eight judgments with the response order shuffled, a pair counts when the chosen response wins more than four, over pairs with a valid verdict; 2,100 pairs). Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 62.9 | #172 |
| 2 | Qwen 3 VL 32B (Thinking) | 54.5 | #557 |
| 3 | Qwen 3 VL 8B (Thinking) | 53 | #887 |
| 4 | Qwen 3 VL 8B Instruct | 52.4 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vurb · How It Works · Data refreshed daily, snapshot 2026-11-12.