Perturbation-based evaluation on VQA v2
1.13Positive Perturbation AUCDEX-AR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DEX-ARModel=BakLlaVA-v12026.03 | 1.13 | 0.81 | |
| RISEModel=BakLlaVA-v12026.03 | 1.09 | 0.86 | |
| Attn×GradModel=BakLlaVA-v12026.03 | 1.06 | 0.86 | |
| CheferCAMModel=BakLlaVA-v12026.03 | 1.05 | 0.84 | |
| Int.GradModel=BakLlaVA-v12026.03 | 1.03 | 0.84 | |
| IIAModel=BakLlaVA-v12026.03 | 1.02 | 0.86 | |
| GradCAMModel=BakLlaVA-v12026.03 | 1.01 | 0.86 | |
| AttentionModel=BakLlaVA-v12026.03 | 0.98 | 0.86 | |
| CheferCAMModel=Florence22026.03 | 0.97 | 0.97 | |
| DEX-ARModel=Florence22026.03 | 0.97 | 0.95 | |
| Attn×GradModel=PaliGemma2026.03 | 0.96 | 0.75 | |
| RolloutModel=BakLlaVA-v12026.03 | 0.95 | 0.9 | |
| Attn×GradModel=Florence22026.03 | 0.95 | 0.99 | |
| IIAModel=Florence22026.03 | 0.95 | 0.99 | |
| RISEModel=PaliGemma2026.03 | 0.95 | 0.79 | |
| DEX-ARModel=PaliGemma2026.03 | 0.95 | 0.75 | |
| AttentionModel=Florence22026.03 | 0.94 | 1 | |
| RolloutModel=Florence22026.03 | 0.94 | 0.98 | |
| AttentionModel=PaliGemma2026.03 | 0.94 | 0.75 | |
| CheferCAMModel=PaliGemma2026.03 | 0.94 | 0.74 | |
| CheferCAMModel=LlaVA-1.52026.03 | 0.93 | 0.78 | |
| DEX-ARModel=LlaVA-1.52026.03 | 0.93 | 0.77 | |
| IIAModel=PaliGemma2026.03 | 0.93 | 0.76 | |
| RISEModel=LlaVA-1.52026.03 | 0.92 | 0.78 | |
| IIAModel=LlaVA-1.52026.03 | 0.92 | 0.77 | |
| GradCAMModel=LlaVA-1.52026.03 | 0.91 | 0.77 | |
| Int.GradModel=LlaVA-1.52026.03 | 0.91 | 0.8 | |
| Attn×GradModel=LlaVA-1.52026.03 | 0.9 | 0.8 | |
| RolloutModel=PaliGemma2026.03 | 0.9 | 0.8 | |
| RISEModel=Florence22026.03 | 0.89 | 0.89 | |
| AttentionModel=LlaVA-1.52026.03 | 0.88 | 0.89 | |
| Int.GradModel=PaliGemma2026.03 | 0.86 | 0.86 | |
| GradCAMModel=PaliGemma2026.03 | 0.85 | 0.87 | |
| Int.GradModel=Florence22026.03 | 0.76 | 0.74 |