Object Hallucination Evaluation on POPE (Score, #F)
89.1POPE ScoreAVIS
Evaluation results
| Method | Links | ||
|---|---|---|---|
| AVISCheckpoint=Vision-R1 [5], Pruning rate (ρ)=Adaptive, Rollouts (K)=Adaptive2026.06 | 89.1 | 40 | |
| FixedCheckpoint=Vision-R1 [5], Pruning rate (ρ)=75%, Rollouts (K)=52026.06 | 88.9 | 57 | |
| BaselineCheckpoint=Vision-R1 [5], Pruning rate (ρ)=0%, Rollouts (K)=12026.06 | 88.5 | 100 | |
| VanillaToken Reduction Ratio=Full Tokens (100%), Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 87.5 | — | |
| CLSEToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 85.6 | — | |
| PDropToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 84.9 | — | |
| FixedCheckpoint=Qwen 2.5 VL [1], Pruning rate (ρ)=75%, Rollouts (K)=52026.06 | 84.7 | 50 | |
| AVISCheckpoint=VL-Rethinker [18], Pruning rate (ρ)=Adaptive, Rollouts (K)=Adaptive2026.06 | 84.7 | 40 | |
| CLSEToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 84.7 | — | |
| AVISCheckpoint=Qwen 2.5 VL [1], Pruning rate (ρ)=Adaptive, Rollouts (K)=Adaptive2026.06 | 84.5 | 38 | |
| AVISCheckpoint=OpenVLThinker [75], Pruning rate (ρ)=Adaptive, Rollouts (K)=Adaptive2026.06 | 84.5 | 34 | |
| BaselineCheckpoint=Qwen 2.5 VL [1], Pruning rate (ρ)=0%, Rollouts (K)=12026.06 | 84.3 | 100 | |
| FixedCheckpoint=OpenVLThinker [75], Pruning rate (ρ)=75%, Rollouts (K)=52026.06 | 83.2 | 48 | |
| BaselineCheckpoint=VL-Rethinker [18], Pruning rate (ρ)=0%, Rollouts (K)=12026.06 | 82.9 | 100 | |
| FastVToken Reduction Ratio=↓ 66.7%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 82.8 | — | |
| BaselineCheckpoint=OpenVLThinker [75], Pruning rate (ρ)=0%, Rollouts (K)=12026.06 | 82.6 | 100 | |
| FixedCheckpoint=VL-Rethinker [18], Pruning rate (ρ)=75%, Rollouts (K)=52026.06 | 82 | 52 | |
| PDropToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 82 | — | |
| CLSEToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 81.4 | — | |
| FastVToken Reduction Ratio=↓ 77.8%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 78 | — | |
| PDropToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 75.6 | — | |
| FastVToken Reduction Ratio=↓ 88.9%, Model=Qwen2-VL-72B, Quantization=4-bit2026.06 | 71.3 | — |