Visual Question Answering on TextVQA 1k (test)
96.46ASR (%)MIE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MIETarget Model=InternVL3.5-4B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 96.46 | 3.01 | 13.02 | |
| COATarget Model=LLaVA-1.5-7B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 96.04 | 2.29 | 1.37 | |
| MIETarget Model=Qwen2.5-VL-7B-Instruct, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 95.58 | 3.73 | 12.61 | |
| EGATarget Model=InternVL3.5-4B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 95.17 | 4.1 | 23.42 | |
| EGATarget Model=Qwen2.5-VL-7B-Instruct, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 93.64 | 5.37 | 24.74 | |
| COATarget Model=Qwen2.5-VL-7B-Instruct, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 91.88 | 6.85 | 0 | |
| COATarget Model=InternVL3.5-4B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 90.74 | 7.86 | 0 | |
| PGDTarget Model=Qwen2.5-VL-7B-Instruct, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 90.27 | 16.65 | 0 | |
| MIETarget Model=LLaVA-1.5-7B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 83.85 | 9.38 | 11.77 | |
| VLATarget Model=InternVL3.5-4B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 81.99 | 15.29 | 0 | |
| VLATarget Model=Qwen2.5-VL-7B-Instruct, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 81.34 | 15.75 | 0 | |
| EGATarget Model=LLaVA-1.5-7B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 80.75 | 11.13 | 28.62 | |
| PGDTarget Model=InternVL3.5-4B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 79.44 | 17.46 | 0 | |
| PGDTarget Model=LLaVA-1.5-7B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 77.5 | 13.01 | 1.76 | |
| VLATarget Model=LLaVA-1.5-7B, Epsilon (ϵimg)=8/255, Decoding Strategy=greedy decoding2025.12 | 69.15 | 17.83 | 2.21 |