Visual Question Answering on ScienceQA (ASR, Coh, Rat, Acc, ASR-C)
4.1CoherenceReShift
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ReShiftBase model=Qwen2.5-VL-7B2026.07 | 4.1 | 100 | 4.17 | 89 | 4 | |
| ReShiftBase model=InternVL3.5-8B2026.07 | 4.1 | 99 | 4.31 | 90 | 3 | |
| BaseBase model=Qwen2.5-VL-7B2026.07 | 3.95 | — | 4.43 | 89 | — | |
| BaseBase model=InternVL3.5-8B2026.07 | 3.93 | — | 4.39 | 92 | — | |
| RewriteBase model=InternVL3.5-8B2026.07 | 3.39 | 90 | 3.47 | 79 | 13 | |
| RewriteBase model=Qwen2.5-VL-7B2026.07 | 3.37 | 77 | 3.35 | 79 | 13 | |
| BadVisionBase model=InternVL3.5-8B2026.07 | 3.23 | 39 | 3.34 | 69 | 7 | |
| BadTokenBase model=InternVL3.5-8B2026.07 | 3.07 | 92 | 2.58 | 84 | 13 | |
| BadVisionBase model=Qwen2.5-VL-7B2026.07 | 2.72 | 45 | 3.59 | 81 | 17 | |
| BadTokenBase model=Qwen2.5-VL-7B2026.07 | 2.64 | 84 | 3.28 | 83 | 11 |