Multi-image Understanding on NLVR2 (test)
87.3Accuracyattention-masking strategy
Evaluation results
| Method | Links | |
|---|---|---|
| attention-masking strategybackbone=LLaVA-OV-7B, attention masking=true2026.01 | 87.3 | |
| LLaVA-OV-7B2026.01 | 84.2 | |
| procedural data-generation strategybackbone=LLaVA-OV-0.5B, attention masking=false2026.01 | 68 | |
| attention-masking strategybackbone=LLaVA-OV-0.5B, attention masking=true2026.01 | 65.1 | |
| LLaVA-OV-0.5B2026.01 | 61.2 | |
| Qwen2VL-2B2026.01 | 41.6 | |
| Qwen2VL-7B2026.01 | 41.5 | |
| InternVL2-2B2026.01 | 18.9 | |
| InternVL2-8B2026.01 | 8.7 |