Multi-type Hallucination Evaluation on MHumanEval
21.9Object Hallucination RateRLHF-V
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RLHF-VBackbone=Vicuna v1.0 13B, Visual Module=BEiT-3, Resolution=4482023.12 | 21.9 | 7.5 | 14.4 | 55.5 | |
| GPT-4V2023.12 | 22.6 | 12.3 | 11 | 45.9 | |
| InstructBLIP2023.12 | 30.8 | 15.1 | 17.1 | 63.7 | |
| LLaVA 1.52023.12 | 30.8 | 17.8 | 17.1 | 61 | |
| Muffin2023.12 | 33.6 | 16.4 | 26 | 74.7 | |
| Qwen-VL-Chat2023.12 | 34.9 | 16.4 | 15.8 | 61 | |
| LLaVA-RLHF2023.12 | 37.7 | 17.8 | 18.5 | 72.6 | |
| LRV2023.12 | 43.2 | 11.6 | 19.2 | 82.9 | |
| LLaVA2023.12 | 46.6 | 21.2 | 19.9 | 80.8 |