Part-aware explanation quality on PartImageNet OOD
9.67TGROurs(Prompt-side)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Ours(Prompt-side)Backbone=Qwen2-VL-7B, Emphasis on target token=false, Attention layers=Last transformer layer2026.06 | 9.67 | 7.26 | 0.65 | |
| PV-TAMBackbone=Qwen2-VL-7B2026.06 | 9.67 | 7.26 | 0.65 | |
| CAMBackbone=Qwen2-VL-7B, Emphasis on target token=true, Attention layers=Last transformer layer2026.06 | 8.47 | 6.2 | 0.72 | |
| TAMBackbone=Qwen2-VL-7B, Emphasis on target token=true, Attention layers=Last transformer layer2026.06 | 8.29 | 6.36 | 1.1 | |
| TAM#Backbone=Qwen2-VL-7B2026.06 | 8.29 | 6.36 | 1.1 | |
| TAM#Backbone=Qwen2-VL-2B2026.06 | 8.13 | 5.21 | 1.24 | |
| PV-TAMBackbone=Qwen2.5-VL-7B2026.06 | 7.6 | 6.08 | 0.92 | |
| CAMBackbone=Qwen2-VL-7B, Emphasis on target token=false, Attention layers=Last transformer layer2026.06 | 7.03 | 5.25 | 0.77 | |
| TAMBackbone=Qwen2-VL-7B, Emphasis on target token=false, Attention layers=Last transformer layer2026.06 | 6.63 | 5.21 | 1.25 | |
| TAM#Backbone=Qwen2.5-VL-7B2026.06 | 5.34 | 5.26 | 1.04 | |
| TAM#Backbone=Qwen2.5-VL-3B2026.06 | 5.02 | 3.23 | 1.26 | |
| PV-TAMBackbone=Qwen2-VL-2B2026.06 | 4.87 | 4.08 | 0.91 | |
| PV-TAMBackbone=Qwen2.5-VL-3B2026.06 | 4.21 | 2.96 | 1.12 | |
| CP-LRPBackbone=Qwen2-VL-7B, Emphasis on target token=false, Attention layers=Last transformer layer2026.06 | 4.12 | 6.02 | 0.71 | |
| CP-LRPBackbone=Qwen2-VL-7B, Emphasis on target token=true, Attention layers=Last transformer layer2026.06 | 3.63 | 6.02 | 0.71 | |
| PV-TAMBackbone=InternVL3-1B2026.06 | 3.42 | 5.7 | 0.81 | |
| Attention RolloutBackbone=Qwen2-VL-7B, Emphasis on target token=false, Attention layers=Last three layers2026.06 | 2.46 | 3.09 | 1.62 | |
| TAM#Backbone=InternVL3-1B2026.06 | 2.36 | 3.34 | 1.47 |