Object Hallucination Evaluation on MSCOCO POPE (Popular)
93.87AccuracyQwen2.5-VL-7B-Instr.
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5-VL-7B-Instr.Backbone=Qwen2.5-VL-7B-Instr., Finetuning Objective=NTP + LLL2026.02 | 93.87 | — | |
| LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B, Finetuning Objective=NTP + LLL2026.02 | 92.4 | — | |
| Qwen2.5-VL-7B-Instr.Backbone=Qwen2.5-VL-7B-Instr., Finetuning Objective=NTP2026.02 | 90.5 | — | |
| LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B, Finetuning Objective=NTP2026.02 | 90.03 | — | |
| SPINBackbone=LLaVA-v1.5-13B2026.02 | 88.83 | — | |
| ICTModel=Qwen-VL2024.11 | 88.16 | 87.33 | |
| Qwen2.5-VL-7B-Instr.Backbone=Qwen2.5-VL-7B-Instr., Finetuning Objective=NTP + LLL, Evaluation Protocol=RefCOCO → POPE (zero-shot transfer)2026.02 | 87.9 | — | |
| ICT w/o imageModel=Qwen-VL2024.11 | 87.63 | 86.78 | |
| ICTModel=LLaVA-v1.52024.11 | 87.5 | 87.6 | |
| OPERAModel=Qwen-VL2024.11 | 87.44 | 86.68 | |
| ICT w/o imageModel=LLaVA-v1.52024.11 | 87.43 | 87.09 | |
| ICT w/o objectModel=LLaVA-v1.52024.11 | 87.3 | 86.65 | |
| VCDModel=Qwen-VL2024.11 | 87.12 | 86.4 | |
| ICT w/o objectModel=Qwen-VL2024.11 | 87.1 | 86.28 | |
| Qwen2.5-VL-7B-Instr.Backbone=Qwen2.5-VL-7B-Instr., Finetuning Objective=Base Model2026.02 | 86.77 | — | |
| OPERAModel=LLaVA-v1.52024.11 | 86.64 | 86.62 | |
| Qwen2.5-VL-7B-Instr.Backbone=Qwen2.5-VL-7B-Instr., Finetuning Objective=NTP, Evaluation Protocol=RefCOCO → POPE (zero-shot transfer)2026.02 | 86.3 | — | |
| LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B, Finetuning Objective=Base Model2026.02 | 86.23 | — | |
| VCDBackbone=LLaVA-v1.5-13B2026.02 | 85.74 | — | |
| VCDModel=LLaVA-v1.52024.11 | 85.38 | 85.06 | |
| RegularModel=Qwen-VL2024.11 | 84.13 | 82.06 | |
| RegularModel=LLaVA-v1.52024.11 | 81.88 | 80.06 |