Visual Question Answering on VQA Average (VQAv2, OK-VQA, GQA)
66.9Accuracy (Clean)UniTok-MLLM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| UniTok-MLLMTokenizer=clean2026.02 | 66.9 | 10 | 6.4 | |
| UniTok-MLLMTokenizer=AT 4/2552026.02 | 62.5 | 45.2 | 35.2 | |
| UniTok-MLLMTokenizer=AT 8/2552026.02 | 59.1 | 47.1 | 39.4 | |
| UniTok-MLLMTokenizer=AT 12/2552026.02 | 57.5 | 47 | 40 | |
| UniTok-MLLMTokenizer=AT 16/2552026.02 | 54.6 | 45.6 | 38.6 | |
| CLIPAttack Type=Clean2026.06 | 54.55 | — | — | |
| AdvXLCLIP-HAttack Type=APGD-Ensemble, Perturbation budget (epsilon)=8/255*2026.06 | 24.51 | — | — |