Long-form Generation Hallucination Evaluation on CHAIR
58.8CS ScoreVCD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VCDBackbone=LLaVA-1.5-7B, max new token=128, Decoding strategy=Greedy2026.05 | 58.8 | 17.5 | 70.1 | |
| VCDBackbone=LLaVA-1.5-7B, max new token=256, Decoding strategy=Greedy2026.05 | 54.6 | 17.3 | 71.3 | |
| BeamBackbone=LLaVA-1.5-7B, max new token=256, Decoding strategy=Beam2026.05 | 52.4 | 14.5 | 73.8 | |
| BeamBackbone=LLaVA-1.5-7B, max new token=128, Decoding strategy=Beam2026.05 | 51.6 | 13.8 | 75 | |
| GreedyBackbone=LLaVA-1.5-7B, max new token=256, Decoding strategy=Greedy2026.05 | 50 | 13.1 | 76.2 | |
| GreedyBackbone=LLaVA-1.5-7B, max new token=128, Decoding strategy=Greedy2026.05 | 49.4 | 13.1 | 76.4 | |
| MemVRBackbone=LLaVA-1.5-7B, max new token=2562026.05 | 43.4 | 14.3 | 77.2 | |
| MemVRBackbone=LLaVA-1.5-7B, max new token=1282026.05 | 42.6 | 14.2 | 74.7 | |
| VCDBackbone=Qwen-VL, max new token=256, Decoding strategy=Greedy2026.05 | 41.4 | 15.6 | 73.8 | |
| VCDBackbone=Qwen-VL, max new token=128, Decoding strategy=Greedy2026.05 | 39 | 15.8 | 76.5 | |
| MemVRBackbone=Qwen-VL, max new token=2562026.05 | 32.3 | 11.6 | 75.7 | |
| VCDBackbone=LLaVA-1.5-7B, max new token=64, Decoding strategy=Greedy2026.05 | 29.2 | 9.7 | 84.5 | |
| GreedyBackbone=Qwen-VL, max new token=256, Decoding strategy=Greedy2026.05 | 28.2 | 14.7 | 79.4 | |
| MemVRBackbone=Qwen-VL, max new token=1282026.05 | 27.4 | 13.2 | 75.5 | |
| BeamBackbone=Qwen-VL, max new token=256, Decoding strategy=Beam2026.05 | 27.2 | 8.5 | 82.5 | |
| HGAIBackbone=LLaVA-1.5-7B, max new token=256, Decoding strategy=Greedy2026.05 | 26.6 | 7.4 | 86.2 | |
| HGAIBackbone=LLaVA-1.5-7B, max new token=128, Decoding strategy=Greedy2026.05 | 25.2 | 6.9 | 70.1 | |
| MemVRBackbone=Qwen-VL, max new token=642026.05 | 25 | 7.7 | 87.1 | |
| MemVRBackbone=LLaVA-1.5-7B, max new token=642026.05 | 24 | 7.3 | 87.4 | |
| BeamBackbone=Qwen-VL, max new token=128, Decoding strategy=Beam2026.05 | 23.2 | 7.5 | 85.5 | |
| GreedyBackbone=LLaVA-1.5-7B, max new token=64, Decoding strategy=Greedy2026.05 | 22.4 | 6.4 | 88.9 | |
| GreedyBackbone=Qwen-VL, max new token=128, Decoding strategy=Greedy2026.05 | 21.8 | 11.7 | 84.1 | |
| AFIPBackbone=LLaVA-1.5-7B, max new token=128, Decoding strategy=Greedy2026.05 | 21.6 | 4.9 | 89.3 | |
| HGAIBackbone=Qwen-VL, max new token=64, Decoding strategy=Greedy2026.05 | 20.8 | 10.7 | 85.8 | |
| BeamBackbone=Qwen-VL, max new token=64, Decoding strategy=Beam2026.05 | 20.2 | 7.2 | 88.9 | |
| GreedyBackbone=Qwen-VL, max new token=64, Decoding strategy=Greedy2026.05 | 19.6 | 10.6 | 86 | |
| BeamBackbone=LLaVA-1.5-7B, max new token=64, Decoding strategy=Beam2026.05 | 19.2 | 5.9 | 90.4 | |
| VCDBackbone=Qwen-VL, max new token=64, Decoding strategy=Greedy2026.05 | 19.2 | 9.8 | 83.6 | |
| HGAIBackbone=LLaVA-1.5-7B, max new token=64, Decoding strategy=Greedy2026.05 | 18.6 | 5.3 | 90.5 | |
| HGAIBackbone=Qwen-VL, max new token=128, Decoding strategy=Greedy2026.05 | 16.4 | 7.4 | 90.2 | |
| AFIPBackbone=Qwen-VL, max new token=128, Decoding strategy=Greedy2026.05 | 15.6 | 7.2 | 89.7 | |
| HGAIBackbone=Qwen-VL, max new token=256, Decoding strategy=Greedy2026.05 | 15.4 | 6.7 | 90.6 | |
| AFIPBackbone=LLaVA-1.5-7B, max new token=256, Decoding strategy=Greedy2026.05 | 15 | 3.6 | 92 | |
| AFIPBackbone=Qwen-VL, max new token=64, Decoding strategy=Greedy2026.05 | 14.4 | 6.6 | 90 | |
| AFIPBackbone=LLaVA-1.5-7B, max new token=64, Decoding strategy=Greedy2026.05 | 13 | 4 | 92.6 | |
| AFIPBackbone=Qwen-VL, max new token=256, Decoding strategy=Greedy2026.05 | 12.2 | 6 | 91.7 |