Visual Reasoning on CoBSAT 4-shot
92.7Color (I) AccuracyTBDN
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| TBDNshot=4-shot, Backbone Model=InternVL3-8B (I3)2026.01 | 92.7 | 95.1 | 49 | 87.5 | 62.7 | 78.3 | 98.3 | 54.7 | 86.8 | 63.6 | 76.9 | |
| TBDNshot=4-shot, Backbone Model=Qwen2-VL-7B-Instruct (Q2)2026.01 | 92 | 94.8 | 58.2 | 82.2 | 63.3 | 85.1 | 98 | 55 | 78.8 | 59.3 | 76.7 | |
| Baseshot=4-shot, Backbone Model=InternVL3-8B (I3)2026.01 | 88.1 | 90.8 | 40.1 | 82.4 | 57.3 | 65.9 | 93.6 | 52.2 | 83.7 | 58.8 | 71.3 | |
| TBDNshot=4-shot, Backbone Model=Qwen2.5-VL-7B-Instruct (Q2.5)2026.01 | 86.8 | 83.2 | 39.2 | 75.2 | 53.1 | 67.5 | 96 | 45.7 | 74 | 51.7 | 67.2 | |
| Baseshot=4-shot, Backbone Model=Qwen2-VL-7B-Instruct (Q2)2026.01 | 81.5 | 73 | 34.6 | 46.6 | 50.4 | 65.2 | 92 | 44.3 | 81.2 | 44.9 | 61.4 | |
| ThinkDiffshot=4-shot2026.01 | 63.8 | 36.2 | 25.4 | 43.4 | 31.7 | 61 | 59 | 43.2 | 66.4 | 33.2 | 46.3 | |
| Baseshot=4-shot, Backbone Model=Qwen2.5-VL-7B-Instruct (Q2.5)2026.01 | 49.8 | 55.8 | 14.9 | 44.6 | 25.5 | 36.9 | 49.3 | 26.5 | 62 | 29.6 | 39.5 | |
| SLshot=4-shot2026.01 | 48.2 | 21.1 | 14.1 | 5.3 | 12.2 | 25.2 | 7.6 | 26.8 | 20.7 | 10.5 | 19.2 | |
| GILLshot=4-shot2026.01 | 10.6 | 4.4 | 4.1 | 7.3 | 8.7 | 2.2 | 5.9 | 4.4 | 3.2 | 6.7 | 5.8 | |
| Emushot=4-shot2026.01 | 6.3 | 1.8 | 4.5 | 4.8 | 9.7 | 3.7 | 12.2 | 10.9 | 7.7 | 8.8 | 7 |