KnowHal - Positive Questions: leaderboard
Accuracy of multimodal LLMs on the positive questions of KnowHal (Shandong University, 2026), which ask about visual evidence or factual knowledge that is actually present for the depicted entity, averaged over the entity, attribute, relation and knowledge dimensions. Answers are scored by standard VQA answer matching; percent, higher is better.
Metric: Accuracy (%; mean over the entity, attribute, relation and knowledge dimensions on questions about valid visual evidence or facts; 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) | 86.6 | #144 |
| 2 | Seed 2.0 Mini | 85 | #508 |
| 3 | Kimi K2.5 | 79.2 | #256 |
| 4 | GLM-4.6V | 77.1 | #701 |
| 5 | Qwen 3 VL 32B Instruct | 76.9 | #585 |
| 6 | Qwen 2.5 VL 72B Instruct | 74.9 | #846 |
| 7 | GPT-5.6 Sol | 73.6 | #30 |
| 8 | GPT-5.2 | 67.9 | #176 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=knowhal-positive-questions · How It Works · Data refreshed daily, snapshot 2026-11-13.