Visual Grounding on RefCOCO+
88.3Accuracy @ 0.5 IoUGiT-H universal
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GiT-H universal#Params=756M2024.03 | 88.3 | — | — | |
| SGPtoken ratio=64%2024.12 | 86.71 | 100.14 | — | |
| InternVL2-26Btoken ratio=100%2024.12 | 86.67 | 100 | — | |
| FastVtoken ratio=64%2024.12 | 86.66 | 99.84 | — | |
| GiT-L universal#Params=387M2024.03 | 86 | — | — | |
| SGPtoken ratio=35%2024.12 | 84.35 | 98.36 | — | |
| GiT-B universal#Params=131M2024.03 | 84.2 | — | — | |
| Qwen2.5-VLParam Size=7B2025.05 | 84.2 | — | — | |
| DeepEyesParam Size=7B2025.05 | 83.6 | — | — | |
| Qwen2.5-VL*Param Size=7B2025.05 | 82.6 | — | — | |
| DeepEyesreproduced=true2025.12 | 82.56 | — | — | |
| SubagentVL2025.12 | 81.97 | — | — | |
| Qwen2.5-VLSize=7B, reproduced=true2025.12 | 81.62 | — | — | |
| ToMetoken ratio=64%2024.12 | 78.91 | 94.24 | — | |
| GiT-HBackbone=Huge, Training=Universal2024.03 | 78.3 | — | — | |
| FastVtoken ratio=35%2024.12 | 77.61 | 88.28 | — | |
| GiT-LBackbone=Large, Training=Universal2024.03 | 73.9 | — | — | |
| InternVL2-2Btoken ratio=100%2024.12 | 73.53 | 87.25 | — | |
| GiT-BBackbone=Base, Training=Universal2024.03 | 72.5 | — | — | |
| SGPtoken ratio=9%2024.12 | 72.22 | 89.58 | — | |
| IAGModel=InternVL-2.5-8B2025.08 | 68.1 | — | — | |
| ToMetoken ratio=35%2024.12 | 64.97 | 85.2 | — | |
| ImperioModel=InternVL-2.5-8B2025.08 | 63.8 | — | — | |
| IAGModel=Llava-v1.5-7B2025.08 | 54.7 | — | — | |
| ImperioModel=Llava-v1.5-7B2025.08 | 51.1 | — | — | |
| IAGModel=Ferret-7B2025.08 | 40.7 | — | — | |
| ImperioModel=Ferret-7B2025.08 | 34.8 | — | — | |
| MarksmanModel=Ferret-7B2025.08 | 30.1 | — | — | |
| ToMetoken ratio=9%2024.12 | 17.74 | 54.28 | — | |
| FastVtoken ratio=9%2024.12 | 16.66 | 46.99 | — | |
| MarksmanModel=InternVL-2.5-8B2025.08 | 15.5 | — | — | |
| RandomModel=InternVL-2.5-8B2025.08 | 9.3 | — | — | |
| MarksmanModel=Llava-v1.5-7B2025.08 | 7.8 | — | — | |
| RandomModel=Ferret-7B2025.08 | 4.9 | — | — | |
| One-to-NModel=InternVL-2.5-8B2025.08 | 4.1 | — | — | |
| One-to-NModel=Llava-v1.5-7B2025.08 | 3.5 | — | — | |
| RandomModel=Llava-v1.5-7B2025.08 | 1.8 | — | — | |
| One-to-NModel=Ferret-7B2025.08 | 1.6 | — | — | |
| DeepEyes2026.04 | — | — | 83.6 | |
| EraX-VL-7B-V1.0Checkpoint Type=Fine-tuned2025.05 | — | — | 40.99 | |
| EVR + CCA2026.04 | — | — | 85.2 | |
| Individual GeometryProtocol=full fine-tuning2025.05 | — | — | 19.61 | |
| Individual GroundingProtocol=full fine-tuning2025.05 | — | — | 71.35 | |
| Individual OCRProtocol=full fine-tuning2025.05 | — | — | 68.01 | |
| InternVL2.5-InstructProtocol=full fine-tuning2025.05 | — | — | 65.41 | |
| Iso-CProtocol=full fine-tuning2025.05 | — | — | 66.56 | |
| Iso-CMethod Class=Merging2025.05 | — | — | 82.1 | |
| Mixture TrainingProtocol=full fine-tuning2025.05 | — | — | 65.93 | |
| olmOCR-7B-0225-previewCheckpoint Type=Fine-tuned2025.05 | — | — | 71.55 | |
| OptMergeMerging Method=Optimization-based Merging2025.05 | — | 65.9 | — | |
| OptMergeProtocol=full fine-tuning2025.05 | — | — | 69.72 | |
| OptMergeMethod Class=Merging2025.05 | — | — | 82.97 | |
| Qwen2-VL-7B-GRPO-8kCheckpoint Type=Fine-tuned2025.05 | — | — | 79.12 | |
| Qwen2-VL-7B-PokemonCheckpoint Type=Fine-tuned2025.05 | — | — | 68.8 | |
| Qwen2-VL-InstructMethod Class=Baseline2025.05 | — | — | 77.87 | |
| Qwen2.5VLParameters=7B2026.04 | — | — | 82.6 | |
| TA w/ DAREProtocol=full fine-tuning2025.05 | — | — | 68.07 | |
| TA w/ DAREMethod Class=Merging2025.05 | — | — | 82.79 | |
| Task ArithmeticProtocol=full fine-tuning2025.05 | — | — | 68.15 | |
| Task ArithmeticMethod Class=Merging2025.05 | — | — | 81.6 | |
| TIES MergingProtocol=full fine-tuning2025.05 | — | — | 68.45 | |
| TIES MergingMethod Class=Merging2025.05 | — | — | 83.52 | |
| TIES w/ DAREProtocol=full fine-tuning2025.05 | — | — | 68.48 | |
| TIES w/ DAREMethod Class=Merging2025.05 | — | — | 82.93 | |
| TSV MergingProtocol=full fine-tuning2025.05 | — | — | 58.51 | |
| TSV MergingMethod Class=Merging2025.05 | — | — | 82.25 | |
| Weight AverageProtocol=full fine-tuning2025.05 | — | — | 45.29 | |
| WUDI MergingProtocol=full fine-tuning2025.05 | — | — | 70.14 | |
| WUDI MergingMethod Class=Merging2025.05 | — | — | 81.5 |