Robustness on R-Bench
61.01R-Bench Dis MetricSymMPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SymMPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 61.01 | — | 65.18 | |
| IC-VCOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 60.71 | — | 64.67 | |
| mDPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 60.4 | — | 63.77 | |
| SymMPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 60.2 | — | 63.77 | |
| mDPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 60 | — | 65.79 | |
| S-VCOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 59.8 | — | 63.36 | |
| IC-VCOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 59.7 | — | 63.87 | |
| DPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 58.99 | — | 63.16 | |
| S-VCOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 58.99 | — | 63.36 | |
| V-DPOContrastive Sample Source=Synthetic, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 58.59 | — | 62.15 | |
| V-DPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 57.98 | — | 61.34 | |
| DPOContrastive Sample Source=Contrastive Sample Editing, Backbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 57.58 | — | 61.34 | |
| LLaVA-NeXT-Interleave-Qwen-7BBackbone=LLaVA-NeXT-Interleave-Qwen-7B2026.05 | 55.96 | — | 59.11 | |
| BAGEL# Params=7B, Model Type=Unified model2026.05 | — | 70.1 | — | |
| BAGEL+Ours# Params=7B, Model Type=Unified model2026.05 | — | 71.7 | — | |
| InternVL2.5# Params=4B, Model Type=Understanding Only2026.05 | — | 66.1 | — |