Visual Question Answering on GQA (Epsilon Robustness Accuracy)
68Accuracy (clean)UniTok-MLLM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| UniTok-MLLMTokenizer=clean2026.02 | 68 | 13.4 | 10 | |
| UniTok-MLLMTokenizer=AT 4/2552026.02 | 67 | 46.2 | 31.4 | |
| UniTok-MLLMTokenizer=AT 8/2552026.02 | 65.4 | 48.4 | 38.8 | |
| UniTok-MLLMTokenizer=AT 12/2552026.02 | 64.2 | 49.4 | 40.2 | |
| UniTok-MLLMTokenizer=AT 16/2552026.02 | 61.2 | 48.4 | 40 |