Vision-Text Arbitration on MC2
83.5Adversarial AccuracyGACL (fixed)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GACL (fixed)Model=Qwen3-VL-2B2026.06 | 83.5 | 40.5 | 100 | 0 | 56.5 | 0.5 | |
| GACL (fixed)Model=Qwen3-VL-8B2026.06 | 82 | 26.5 | 99.5 | 0.5 | 42.5 | 2 | |
| Joint baselineModel=Qwen3-VL-8B2026.06 | 55.5 | — | 100 | 0 | — | — | |
| Joint baselineModel=Qwen3-VL-2B2026.06 | 43 | — | 100 | 0 | — | — |