Vision-Language Reasoning on Winoground
59.88Simple AccLLaVA-1.5 13B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| LLaVA-1.5 13BIntervention Strategy=No-intervention baseline2026.01 | 59.88 | 24.63 | 74.38 | |
| LLaVA-1.5 13BIntervention Strategy=Q4 system redistribution (prop)2026.01 | 58.56 | 21.5 | 79.31 | |
| Q4 system redistr (prop)Intervention=Q4 system redistribution, strategy=proportional2026.01 | 57.88 | 19.5 | 75.38 | |
| Q4 text redistr (prop)Intervention=Q4 text redistribution, strategy=proportional2026.01 | 55 | 12 | 92 | |
| Image×2.0Intervention=Image×2.02026.01 | 53.69 | 8.85 | 94.56 | |
| Q4 system ablIntervention=Q4 system ablation2026.01 | 53.56 | 8.13 | 94.81 | |
| PAIIntervention=PAI2026.01 | 53.5 | 7.63 | 95.25 | |
| No intervention baselineIntervention=None2026.01 | 53.38 | 7.38 | 95.38 | |
| AD-HHIntervention=AD-HH2026.01 | 53.06 | 6.63 | 95.94 |