VERHallu - Subevent QA: leaderboard
VERHallu (Zhang et al., 2026) probes event-relation hallucination in video LLMs with clips from Mr. Bean films and episodes, whose counterintuitive event chains defeat language and vision-language priors. The 258 subevent questions ask which event occurred during a main event, with seven options including distractors built from language and vision-language priors; accuracy in percent, higher is better.
Metric: Accuracy (%; 258 seven-option questions on which event occurred during a main event, distractors from language and vision-language priors; random 14.3; each model's own frame sampling and default settings, answer given as the option number). Source: arxiv.org. Saturation forecast: Around 2031. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ChatGPT 4o | 42.5 | #610 |
| 2 | Qwen 2.5 VL 7B Instruct | 40.7 | #1842 |
| 3 | Gemini 3 Pro | 40 | #135 |
| 4 | InternVL3-8B | 36.4 | #1695 |
| 5 | Qwen 2.5 VL 72B Instruct | 34.8 | #845 |
| 6 | MiniCPM-V-2.6 | 33.3 | #3095 |
| 7 | Qwen 2.5 VL 32B Instruct | 31.4 | #1055 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=verhallu-subevent-qa · How It Works · Data refreshed daily, snapshot 2026-11-12.