VidPair-Halluc: leaderboard
VidPair-Halluc (2026) probes hallucination in video-language models with adversarial pairs of generated videos that share a near-identical background but differ in foreground actions, attributes or event order. This board is binary-question weighted accuracy in percent, which credits a question pair or video only when every answer in it is correct; higher is better.
Metric: Binary-QA weighted accuracy wAcc (%): question-pair and video-pair accuracy (all answers in an adversarial pair correct) weighted by their sample sizes on VidPair-Halluc adversarial video pairs (generated clips with near-identical backgrounds and different foreground semantics). Source: arxiv.org. Saturation forecast: Around 2029. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 49.15 | #266 |
| 2 | Qwen 2.5 VL 7B Instruct | 41.66 | #1833 |
| 3 | Gemini 2.5 Flash | 29.83 | #453 |
| 4 | GPT-5 Mini | 29.33 | #348 |
| 5 | GPT-4o | 26.97 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vidpair-halluc · How It Works · Data refreshed daily, snapshot 2026-11-13.