Multi-modal preference alignment on WildVision
40.2Winning RateLLaVANext
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LLaVANextData=OmniAlign-Vmix, LLM=Qwen2.5-32B2025.02 | 40.2 | -14.9 | |
| LLaVANextData=OmniAlign-Vmix, LLM=InternLM2.5-7B2025.02 | 29.6 | -31.3 | |
| LLaVAData=OmniAlign-Vmix, LLM=InternLM2.5-7B2025.02 | 28.2 | -34.6 | |
| LLaVANextData=LLaVANext-778k, LLM=Qwen2.5-32B2025.02 | 25.2 | -41.3 | |
| LLaVANextData=LLaVANext-778k, LLM=InternLM2.5-7B2025.02 | 23.4 | -45 | |
| LLaVAData=LLaVANext-778k, LLM=InternLM2.5-7B2025.02 | 18.4 | -55.1 |