Caption Hallucination Analysis on COCO
47.8CHAIRsVCD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VCDBackbone=LLaVA-1.5-7B2026.06 | 47.8 | 14.1 | 82.7 | |
| LLaVA-1.5-7BBackbone=LLaVA-1.5-7B2026.06 | 46 | 12.8 | 78.9 | |
| VISTABackbone=LLaVA-1.5-7B2026.06 | 44.6 | 13.4 | 77 | |
| SPARCBackbone=LLaVA-1.5-7B2026.06 | 43.4 | 16.8 | 61.1 | |
| DPO (RLHF-V)Backbone=LLaVA-1.5-7B2026.06 | 43.2 | 11.9 | 76.4 | |
| ASDBackbone=LLaVA-1.5-7B2026.06 | 40 | 11.3 | 82 | |
| VTIBackbone=LLaVA-1.5-7B2026.06 | 35.8 | 11.1 | 76.8 | |
| SHARPBackbone=LLaVA-1.5-7B2026.06 | 34.8 | 10.6 | 59.7 | |
| TLVSBackbone=LLaVA-1.5-7B2026.06 | 31.2 | 9.7 | 82.2 | |
| VCDBackbone=Qwen2.5-VL-7B2026.06 | 29.8 | 9.7 | 58 | |
| FV2026.06 | 28.26 | 25.44 | 27.69 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.06 | 24 | 7 | 86.7 | |
| LoRA2026.06 | 17.42 | 11.55 | 42.93 | |
| 32-ICLmode=32-shot In-Context Learning2026.06 | 16.78 | 9.77 | 42.59 | |
| DPO (RLHF-V)Backbone=Qwen2.5-VL-7B2026.06 | 13.4 | 3.5 | 62.7 | |
| TLVSBackbone=Qwen2.5-VL-7B2026.06 | 12 | 5.2 | 83.9 | |
| TV2026.06 | 8.88 | 7.5 | 36.22 | |
| LIVE2026.06 | 8.65 | 6.05 | 42.84 | |
| MimIC2026.06 | 8.51 | 5.74 | 43.3 | |
| Hyper-ICL2026.06 | 8.29 | 5.63 | 44.04 | |
| Z-Shotmode=Zero-shot2026.06 | 5.93 | 5.58 | 30.72 |