KnowHal: leaderboard
KnowHal (Shandong University, 2026) evaluates hallucination in multimodal LLMs with paired questions over 1,800 entity images in 10 domains: a positive question about evidence that is present and a negative question carrying a false premise, across entity, attribute, relation and knowledge dimensions. The overall score is the mean of positive and negative accuracy in percent; higher is better.
Metric: Overall score (%; mean of the positive-question and negative-question average accuracies; 1,800 entity-image samples in 10 domains with paired positive and false-premise negative questions; positive answers scored by standard VQA answer normalization and soft matching, negative answers by a GPT-4o-mini judge (option matching for relation questions); zero-shot). Source: arxiv.org. Saturation forecast: Not forecast. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 67.9 | #144 |
| 2 | Qwen 3 VL 32B Instruct | 66.6 | #585 |
| 3 | Qwen 2.5 VL 72B Instruct | 66 | #846 |
| 4 | Seed 2.0 Mini | 61.1 | #508 |
| 5 | Kimi K2.5 | 60.5 | #256 |
| 6 | GPT-5.2 | 54.2 | #176 |
| 7 | GLM-4.6V | 54.1 | #701 |
| 8 | GPT-5.6 Sol | 52.1 | #30 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=knowhal · How It Works · Data refreshed daily, snapshot 2026-11-13.