Generative Hallucination Evaluation on AMBER-g
2.2CHAIR ScoreUE-DPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| UE-DPOBackbone=LLaVA-v1.5-13B, Datasize=5.7k2026.05 | 2.2 | — | 0.9 | 14.4 | |
| UE-DPOBackbone=LLaVA-v1.5-7B, Datasize=16k, Training Data Source=RLAIF-V Data2026.05 | 2.5 | — | 0.9 | 12.5 | |
| UE-DPOBackbone=LLaVA-v1.5-7B, Datasize=5.7k, Training Data Source=RLHF-V Data2026.05 | 2.9 | — | 1 | 17.4 | |
| RLAIF-VBackbone=LLaVA-v1.5-7B, Datasize=16k2026.05 | 3 | — | 1 | 16.2 | |
| UE-DPOBackbone=Qwen2.5-VL-3B, Datasize=5.7k2026.05 | 3 | — | 0.6 | 14.7 | |
| mDPOBackbone=LLaVA-v1.5-7B, Datasize=10k2026.05 | 4.4 | — | 2.4 | 24.5 | |
| Qwen2.5VL-7B + ACTArchitecture=Qwen2.5VL-7B, ACT=Enabled2026.04 | 4.5 | 21.8 | 1.3 | — | |
| LLaVA-1.5-13B + ACTArchitecture=LLaVA-1.5-13B, ACT=Enabled2026.04 | 4.6 | 25.3 | 2.3 | — | |
| Qwen2VL-7B + ACTArchitecture=Qwen2VL-7B, ACT=Enabled2026.04 | 4.9 | 20.4 | 1.2 | — | |
| LLaVA-1.5-7B + ACTArchitecture=LLaVA-1.5-7B, ACT=Enabled2026.04 | 5.1 | 27.4 | 2.7 | — | |
| Qwen2.5VL-7BArchitecture=Qwen2.5VL-7B, ACT=Disabled2026.04 | 5.4 | 22.6 | 1 | — | |
| V-DPOBackbone=LLaVA-v1.5-7B, Datasize=5.7k2026.05 | 5.6 | — | 2.7 | 27.3 | |
| DPOBackbone=Qwen2.5-VL-3B, Datasize=5.7k2026.05 | 5.7 | — | 1.4 | 22.6 | |
| RLHF-V(HD)Backbone=LLaVA-v1.5-13B, Datasize=1.4k2026.05 | 6.3 | — | 2.1 | 25.1 | |
| HALVABackbone=LLaVA-v1.5-13B, Datasize=22.5k2026.05 | 6.4 | — | 3.2 | 30.4 | |
| Qwen2VL-7BArchitecture=Qwen2VL-7B, ACT=Disabled2026.04 | 6.5 | 42.8 | 4 | — | |
| HALVABackbone=LLaVA-v1.5-7B, Datasize=21.5k2026.05 | 6.6 | — | 3.4 | 32.2 | |
| LLaVA-v1.5-13BBackbone=LLaVA-v1.5-13B2026.05 | 6.8 | — | 3.3 | 31.8 | |
| InternVL2.5-8B + ACTArchitecture=InternVL2.5-8B, ACT=Enabled2026.04 | 7.1 | 63.8 | 7.6 | — | |
| LLaVA-1.5-13BArchitecture=LLaVA-1.5-13B, ACT=Disabled2026.04 | 7.2 | 34.2 | 3.8 | — | |
| POVIDBackbone=LLaVA-v1.5-7B, Datasize=17k2026.05 | 7.4 | — | 3.9 | 34.3 | |
| LLaVA-1.5-7BArchitecture=LLaVA-1.5-7B, ACT=Disabled2026.04 | 7.5 | 35.9 | 4.3 | — | |
| InternVL2-8B + ACTArchitecture=InternVL2-8B, ACT=Enabled2026.04 | 7.6 | 56.2 | 6.6 | — | |
| InternVL2.5-8BArchitecture=InternVL2.5-8B, ACT=Disabled2026.04 | 7.6 | 63.5 | 7.3 | — | |
| LLaVA-v1.5-7BBackbone=LLaVA-v1.5-7B2026.05 | 7.7 | — | 4.2 | 34.7 | |
| LLaVA-RLHFBackbone=LLaVA-v1.5-13B, Datasize=122k2026.05 | 7.7 | — | 4 | 38.6 | |
| Qwen2.5-VL-3BBackbone=Qwen2.5-VL-3B2026.05 | 8 | — | 4.3 | 44.4 | |
| InternVL2-8BArchitecture=InternVL2-8B, ACT=Disabled2026.04 | 8.3 | 68.2 | 8.1 | — | |
| LLaVA-RLHFBackbone=LLaVA-v1.5-7B, Datasize=122k2026.05 | 9.7 | — | 5.3 | 46.6 |