Visual Grounding on RefCOCOg
88.44AccuracyInternVL2-26B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| InternVL2-26Btoken ratio=100%2024.12 | 88.44 | 100 | — | — | |
| FastVtoken ratio=64%2024.12 | 88.3 | 99.84 | — | — | |
| SGPtoken ratio=64%2024.12 | 88.05 | 100.14 | — | — | |
| EVR + CCA2026.04 | 87.4 | — | — | — | |
| Qwen2.5-VLParam Size=7B2025.05 | 87.2 | — | — | — | |
| DeepEyesParam Size=7B2025.05 | 86.7 | — | — | — | |
| DeepEyes2026.04 | 86.7 | — | — | — | |
| Qwen2.5-VL*Param Size=7B2025.05 | 86.1 | — | — | — | |
| Qwen2.5VLParameters=7B2026.04 | 86.1 | — | — | — | |
| SGPtoken ratio=35%2024.12 | 86.07 | 98.36 | — | — | |
| DeepEyesreproduced=true2025.12 | 83.5 | — | — | — | |
| SubagentVL2025.12 | 82.96 | — | — | — | |
| Qwen2.5-VLSize=7B, reproduced=true2025.12 | 82.86 | — | — | — | |
| FastVtoken ratio=35%2024.12 | 81.39 | 88.28 | — | — | |
| ToMetoken ratio=64%2024.12 | 80.35 | 94.24 | — | — | |
| InternVL2-2Btoken ratio=100%2024.12 | 77.55 | 87.25 | — | — | |
| SGPtoken ratio=9%2024.12 | 77.45 | 89.58 | — | — | |
| WUDI MergingProtocol=full fine-tuning2025.05 | 74.48 | — | — | — | |
| OptMergeProtocol=full fine-tuning2025.05 | 73.94 | — | — | — | |
| TIES MergingProtocol=full fine-tuning2025.05 | 73.65 | — | — | — | |
| TIES w/ DAREProtocol=full fine-tuning2025.05 | 73.55 | — | — | — | |
| TA w/ DAREProtocol=full fine-tuning2025.05 | 73.02 | — | — | — | |
| Task ArithmeticProtocol=full fine-tuning2025.05 | 72.75 | — | — | — | |
| Individual GroundingProtocol=full fine-tuning2025.05 | 70.09 | — | — | — | |
| Individual OCRProtocol=full fine-tuning2025.05 | 69.1 | — | — | — | |
| Iso-CProtocol=full fine-tuning2025.05 | 68.5 | — | — | — | |
| ToMetoken ratio=35%2024.12 | 68.08 | 85.2 | — | — | |
| Mixture TrainingProtocol=full fine-tuning2025.05 | 67.46 | — | — | — | |
| InternVL2.5-InstructProtocol=full fine-tuning2025.05 | 67.4 | — | — | — | |
| InternVL3Type=AR2026.05 | 67.1 | — | — | — | |
| LLaDA-V + OursType=Diff.2026.05 | 65 | — | — | — | |
| LLaDA-VType=Diff.2026.05 | 64.8 | — | — | — | |
| TSV MergingProtocol=full fine-tuning2025.05 | 59.17 | — | — | — | |
| LLaVA-OVType=AR2026.05 | 56.9 | — | — | — | |
| Weight AverageProtocol=full fine-tuning2025.05 | 52.39 | — | — | — | |
| ADAPTBackbone=LLaVA-v1.5-7B2026.06 | 50.3 | — | — | 57 | |
| CLIP2026.06 | 45.9 | — | — | 51.3 | |
| LaViDa + OursType=Diff.2026.05 | 44 | — | — | — | |
| SpectralBackbone=LLaVA-v1.5-7B2026.06 | 42.3 | — | — | 49.8 | |
| APIBackbone=LLaVA-v1.5-7B2026.06 | 39.8 | — | — | 56 | |
| SpatialBackbone=LLaVA-v1.5-7B2026.06 | 37 | — | — | 45.1 | |
| LaViDaType=Diff.2026.05 | 36.9 | — | — | — | |
| LLaVA-v1.5-7B2026.06 | 32.6 | — | — | 41.8 | |
| LLaVA-1.6Type=AR2026.05 | 28.1 | — | — | — | |
| FocusBackbone=LLaVA-v1.5-7B2026.06 | 27.9 | — | — | 37 | |
| Individual GeometryProtocol=full fine-tuning2025.05 | 23.84 | — | — | — | |
| ToMetoken ratio=9%2024.12 | 19.36 | 54.28 | — | — | |
| FastVtoken ratio=9%2024.12 | 17.22 | 46.99 | — | — | |
| Qwen2.5Type=AR2026.05 | 16.6 | — | — | — | |
| GiT-HBackbone=Huge, Training=Universal2024.03 | 0.807 | — | — | — | |
| GiT-LBackbone=Large, Training=Universal2024.03 | 0.789 | — | — | — | |
| GiT-BBackbone=Base, Training=Universal2024.03 | 0.769 | — | — | — | |
| Qwen2.5-VL-7BModel=Qwen2.5-VL-7B2026.06 | — | — | 81.1 | — | |
| Qwen2.5-VL-7B + LOCUSModel=Qwen2.5-VL-7B, Method Extension=LOCUS2026.06 | — | — | 83 | — |