KnowHal - Negative Questions: leaderboard
Accuracy on the negative questions of KnowHal (Shandong University, 2026), which embed a misleading premise, such as a substituted attribute or object, a false spatial relation or a fabricated action, that the model must reject. Averaged over the four dimensions and scored by a GPT-4o-mini judge or option matching; percent, higher is better.
Metric: Accuracy (%; mean over the entity, attribute, relation and knowledge dimensions on questions carrying a misleading premise the model must reject; 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 | Qwen 2.5 VL 72B Instruct | 57 | #846 |
| 2 | Qwen 3 VL 32B Instruct | 56.3 | #585 |
| 3 | Gemini 3 Flash (Preview) | 49.3 | #144 |
| 4 | Kimi K2.5 | 41.8 | #256 |
| 5 | GPT-5.2 | 40.4 | #176 |
| 6 | Seed 2.0 Mini | 37.1 | #508 |
| 7 | GLM-4.6V | 31.2 | #701 |
| 8 | GPT-5.6 Sol | 30.6 | #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-negative-questions · How It Works · Data refreshed daily, snapshot 2026-11-13.