Visual Hallucination Evaluation on MSCOCO
18.2CHAIR_iM3ID
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| M3IDBackbone=InstructBLIP, Max Token=1282025.12 | 18.2 | 62.3 | — | |
| VCDBackbone=InstructBLIP, Max Token=1282025.12 | 18 | 60.7 | — | |
| VanillaBackbone=InstructBLIP, Max Token=1282025.12 | 17.6 | 57.4 | — | |
| WoodpeckerBackbone=InstructBLIP, Max Token=1282025.12 | 17.6 | 60.8 | — | |
| M3IDBackbone=Qwen-VL, Max Token=1282025.12 | 17.4 | 49.8 | — | |
| VCDBackbone=Qwen-VL, Max Token=1282025.12 | 17.2 | 50.4 | — | |
| FastVBackbone=MiniGPT-4, method_type=image token sparsity2025.01 | 16.72 | 41.32 | 38.29 | |
| WoodpeckerBackbone=LLaVA-1.5, Max Token=1282025.12 | 16.7 | 57.6 | — | |
| VanillaBackbone=Qwen-VL, Max Token=1282025.12 | 16.7 | 52.1 | — | |
| VCDBackbone=LLaVA-1.5, Max Token=1282025.12 | 16.6 | 54.4 | — | |
| VanillaBackbone=LLaVA-1.5, Max Token=1282025.12 | 16.4 | 55.1 | — | |
| SparseVLMBackbone=MiniGPT-4, method_type=image token sparsity2025.01 | 16.38 | 40.93 | 37.81 | |
| WoodpeckerBackbone=Qwen-VL, Max Token=1282025.12 | 16.3 | 51.8 | — | |
| DOLABackbone=InstructBLIP, Max Token=1282025.12 | 15.9 | 48.4 | — | |
| M3IDBackbone=LLaVA-1.5, Max Token=1282025.12 | 15.8 | 56.6 | — | |
| ONLYBackbone=InstructBLIP, Max Token=1282025.12 | 15.7 | 52.5 | — | |
| HALCBackbone=InstructBLIP, Max Token=1282025.12 | 15.7 | 53.8 | — | |
| HALCBackbone=Qwen-VL, Max Token=1282025.12 | 15.4 | 49.6 | — | |
| ONLYBackbone=LLaVA-1.5, Max Token=1282025.12 | 14.9 | 51.2 | — | |
| HALCBackbone=LLaVA-1.5, Max Token=1282025.12 | 14.8 | 51 | — | |
| ONLYBackbone=Qwen-VL, Max Token=1282025.12 | 14.4 | 48.1 | — | |
| OPERABackbone=InstructBLIP, Max Token=1282025.12 | 14.2 | 46.4 | — | |
| DOLABackbone=LLaVA-1.5, Max Token=1282025.12 | 13.8 | 47.8 | — | |
| VCDBackbone=Qwen-VL, Max Token=642025.12 | 13.1 | 33.3 | — | |
| LLama3-8BBackbone=LLama3-8B2026.06 | 13 | 44.8 | — | |
| VanillaBackbone=Qwen-VL, Max Token=642025.12 | 12.9 | 33.8 | — | |
| DOLABackbone=Qwen-VL, Max Token=1282025.12 | 12.9 | 46.8 | — | |
| HeRABackbone=Vicuna-7B2026.06 | 12.9 | 47.6 | — | |
| OPERABackbone=LLaVA-1.5, Max Token=1282025.12 | 12.8 | 44.6 | — | |
| Qwen2.5-3BBackbone=Qwen2.5-3B2026.06 | 12.6 | 44.2 | — | |
| Vicuna-7BBackbone=Vicuna-7B2026.06 | 12.6 | 46.6 | — | |
| DeCoBackbone=InstructBLIP, Max Token=1282025.12 | 12.4 | 41.2 | — | |
| HeRABackbone=Qwen2.5-7B2026.06 | 12.4 | 44.8 | — | |
| OPERABackbone=MiniGPT-42025.01 | 12.34 | 32.63 | 5.57 | |
| WoodpeckerBackbone=Qwen-VL, Max Token=642025.12 | 12.3 | 31.1 | — | |
| Qwen3-8BBackbone=Qwen3-8B2026.06 | 12.3 | 42 | — | |
| HeRABackbone=Vicuna-13B2026.06 | 12.2 | 47.4 | — | |
| Qwen2.5-14BBackbone=Qwen2.5-14B2026.06 | 12.2 | 42.6 | — | |
| GreedyBackbone=MiniGPT-42025.01 | 12.17 | 31.47 | 36.64 | |
| WoodpeckerBackbone=MiniGPT-4, method_type=post-hoc2025.01 | 12.09 | 31.69 | — | |
| VCDBackbone=MiniGPT-42025.01 | 11.9 | 30.6 | 17.69 | |
| M3IDBackbone=Qwen-VL, Max Token=642025.12 | 11.9 | 32.3 | — | |
| HeRABackbone=Qwen3-4B2026.06 | 11.9 | 43.8 | — | |
| Qwen2.5-7BBackbone=Qwen2.5-7B2026.06 | 11.9 | 45.2 | — | |
| Vicuna-13BBackbone=Vicuna-13B2026.06 | 11.9 | 43 | — | |
| SIDBackbone=MiniGPT-42025.01 | 11.85 | 31.73 | 22.95 | |
| LUREBackbone=MiniGPT-4, method_type=post-hoc2025.01 | 11.8 | 31.67 | — | |
| HeRABackbone=Qwen2.5-3B2026.06 | 11.8 | 44 | — | |
| Qwen3-4BBackbone=Qwen3-4B2026.06 | 11.7 | 42 | — | |
| HeRABackbone=LLama3-8B2026.06 | 11.7 | 40.2 | — | |
| Qwen3-14BBackbone=Qwen3-14B2026.06 | 11.7 | 43.2 | — | |
| HALCBackbone=MiniGPT-42025.01 | 11.69 | 31.76 | 3.86 | |
| DoLaBackbone=MiniGPT-42025.01 | 11.62 | 30.58 | 25.01 | |
| Beam SearchBackbone=MiniGPT-42025.01 | 11.57 | 31.8 | 32.27 | |
| FastVBackbone=mPLUG-Owl2, method_type=image token sparsity2025.01 | 11.4 | 38.49 | 24.6 | |
| VanillaBackbone=InstructBLIP, Max Token=642025.12 | 11.4 | 31.5 | — | |
| HeRABackbone=Qwen3-14B2026.06 | 11.4 | 41.2 | — | |
| VASparseBackbone=MiniGPT-42025.01 | 11.35 | 30.19 | 30.87 | |
| SparseVLMBackbone=mPLUG-Owl2, method_type=image token sparsity2025.01 | 11.35 | 38.99 | 23.73 | |
| DeCoBackbone=LLaVA-1.5, Max Token=1282025.12 | 11.1 | 37.8 | — | |
| HeRABackbone=Qwen3-8B2026.06 | 11 | 39.2 | — | |
| WoodpeckerBackbone=InstructBLIP, Max Token=642025.12 | 10.8 | 31.2 | — | |
| HeRABackbone=Qwen2.5-14B2026.06 | 10.6 | 39 | — | |
| M3IDBackbone=InstructBLIP, Max Token=642025.12 | 10.5 | 31.1 | — | |
| DeCoBackbone=Qwen-VL, Max Token=1282025.12 | 10.4 | 42.2 | — | |
| VCDBackbone=InstructBLIP, Max Token=642025.12 | 10.1 | 30 | — | |
| Vanilla + VDCBackbone=LLaVA-1.5, Max Token=1282025.12 | 9.8 | 33.6 | — | |
| VCD + VDCBackbone=InstructBLIP, Max Token=1282025.12 | 9.7 | 38.6 | — | |
| M3ID + VDCBackbone=Qwen-VL, Max Token=1282025.12 | 9.7 | 34.2 | — | |
| M3ID + VDCBackbone=LLaVA-1.5, Max Token=1282025.12 | 9.6 | 33.8 | — | |
| VCD + VDCBackbone=Qwen-VL, Max Token=1282025.12 | 9.6 | 34.6 | — | |
| VanillaBackbone=LLaVA-1.5, Max Token=642025.12 | 9.4 | 26.5 | — | |
| VCD + VDCBackbone=LLaVA-1.5, Max Token=1282025.12 | 9.2 | 32.9 | — | |
| M3ID + VDCBackbone=InstructBLIP, Max Token=642025.12 | 9.2 | 21.6 | — | |
| VCDBackbone=mPLUG-Owl22025.01 | 9.13 | 24.89 | 9.89 | |
| Vanilla + VDCBackbone=InstructBLIP, Max Token=1282025.12 | 9.1 | 38.2 | — | |
| Vanilla + VDCBackbone=Qwen-VL, Max Token=1282025.12 | 9.1 | 34.2 | — | |
| HALCBackbone=Qwen-VL, Max Token=642025.12 | 9.1 | 28.2 | — | |
| OPERABackbone=mPLUG-Owl22025.01 | 9.07 | 24.48 | 3.56 | |
| WoodpeckerBackbone=mPLUG-Owl2, method_type=post-hoc2025.01 | 8.99 | 25.05 | — | |
| GreedyBackbone=mPLUG-Owl22025.01 | 8.94 | 24.42 | 20.36 | |
| ONLY + VDCBackbone=LLaVA-1.5, Max Token=1282025.12 | 8.9 | 31.4 | — | |
| M3ID + VDCBackbone=InstructBLIP, Max Token=1282025.12 | 8.9 | 33.9 | — | |
| DoLaBackbone=mPLUG-Owl22025.01 | 8.88 | 24.67 | 14.74 | |
| VCD + VDCBackbone=Qwen-VL, Max Token=642025.12 | 8.8 | 19.1 | — | |
| Beam SearchBackbone=mPLUG-Owl22025.01 | 8.72 | 23.87 | 19.62 | |
| ONLY + VDCBackbone=Qwen-VL, Max Token=1282025.12 | 8.7 | 32.6 | — | |
| VCD + VDCBackbone=InstructBLIP, Max Token=642025.12 | 8.6 | 24.2 | — | |
| ONLYBackbone=Qwen-VL, Max Token=642025.12 | 8.6 | 27.7 | — | |
| SIDBackbone=mPLUG-Owl22025.01 | 8.54 | 23.55 | 12.95 | |
| FastVBackbone=LLaVA-1.5, method_type=image token sparsity2025.01 | 8.53 | 26.76 | 33.21 | |
| SparseVLMBackbone=LLaVA-1.5, method_type=image token sparsity2025.01 | 8.44 | 26.11 | 32.47 | |
| Vanilla + VDCBackbone=Qwen-VL, Max Token=642025.12 | 8.4 | 19.4 | — | |
| ONLYBackbone=InstructBLIP, Max Token=642025.12 | 8.3 | 23.9 | — | |
| Vanilla + VDCBackbone=InstructBLIP, Max Token=642025.12 | 8.1 | 21.4 | — | |
| M3ID + VDCBackbone=Qwen-VL, Max Token=642025.12 | 8.1 | 19.2 | — | |
| VCDBackbone=LLaVA-1.5, Max Token=642025.12 | 8 | 24.8 | — | |
| HALCBackbone=InstructBLIP, Max Token=642025.12 | 8 | 24.5 | — | |
| LUREBackbone=mPLUG-Owl2, method_type=post-hoc2025.01 | 7.78 | 22.53 | — | |
| HALCBackbone=mPLUG-Owl22025.01 | 7.71 | 23.48 | 1.52 |