Visual Question Answering on VQA v2 (Epsilon Robustness)
73.2Accuracy (clean)UniTok-MLLM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| UniTok-MLLMTokenizer=clean2026.02 | 73.2 | 14.4 | 8.5 | |
| UniTok-MLLMTokenizer=AT 4/2552026.02 | 67.4 | 50.5 | 44.4 | |
| UniTok-MLLMTokenizer=AT 8/2552026.02 | 62.7 | 52.4 | 46.6 | |
| UniTok-MLLMTokenizer=AT 12/2552026.02 | 61.6 | 52.1 | 45.7 | |
| UniTok-MLLMTokenizer=AT 16/2552026.02 | 57.5 | 49.7 | 43.7 |