Text-based Visual Question Answering on TextVQA (Adversarial & Robustness Metrics)
91.35Metric (Clean, No Context)Vanilla Attack
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Vanilla AttackModel=Qwen-VL, Attack Type=Vanilla, Compression=PDrop2026.01 | 91.35 | 63.63 | 32.99 | 33.31 | 36.11 | 47.65 | -0.1624 | |
| Random AttackModel=Qwen-VL, Attack Type=Random, Compression=PDrop2026.01 | 91.35 | 63.63 | 85.34 | 59.39 | 93.42 | 6.66 | 0.0008 | |
| CAAModel=Qwen-VL, Attack Type=Compression-Aware, Compression=PDrop2026.01 | 91.35 | 63.63 | 76.18 | 33.11 | 83.39 | 47.96 | 0.3136 | |
| Vanilla AttackModel=LLaVA, Attack Type=Vanilla, Compression=PDrop2026.01 | 88.17 | 72.22 | 13.06 | 16.56 | 14.81 | 77.07 | -0.0812 | |
| Random AttackModel=LLaVA, Attack Type=Random, Compression=PDrop2026.01 | 88.17 | 72.22 | 74.73 | 60.99 | 84.76 | 15.55 | 0.0031 | |
| CAAModel=LLaVA, Attack Type=Compression-Aware, Compression=PDrop2026.01 | 88.17 | 72.22 | 72.96 | 16.37 | 82.75 | 77.33 | 0.6008 | |
| Vanilla AttackModel=LLaVA-NEXT, Attack Type=Vanilla, Compression=PDrop2026.01 | 63.18 | 55.63 | 15.41 | 15.03 | 24.39 | 72.98 | -0.0263 | |
| Random AttackModel=LLaVA-NEXT, Attack Type=Random, Compression=PDrop2026.01 | 63.18 | 55.63 | 56.95 | 50.08 | 90.14 | 9.98 | 0.0012 | |
| CAAModel=LLaVA-NEXT, Attack Type=Compression-Aware, Compression=PDrop2026.01 | 63.18 | 55.63 | 56.33 | 32.7 | 89.16 | 41.22 | 0.3038 |