HallusionBench (average)
Multimodal · 2023-10-23
HallusionBench probes language hallucination and visual illusion in image-text reasoning with yes/no questions on original and edited images. This row is the average of its question-level, figure-level and all-question accuracies, as Qwen reports it ("HallBench avg").
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen2 VL 72B Instruct | 58.1 |
| InternVL2.5-78B | 57.4 |
| GPT-4o (2024-05-13) | 55.0 |
| Qwen2.5-VL-7B-Instruct | 52.9 |
| Qwen2 VL 7B Instruct | 50.6 |
| MiniCPM-o 2.6 | 48.1 |
| MiniCPM-V 2.6 | 48.1 |
| Qwen2.5-VL-3B-Instruct | 46.3 |
| GPT-4o Mini | 46.1 |
| Qwen2 VL 2B Instruct | 41.7 |