Object Hallucination on POPE
91.9F1 ScoreCALRD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CALRDDecoding Strategy=Beam, Backbone=Qwen3-VL2026.06 | 91.9 | — | |
| CALRDDecoding Strategy=Greedy, Backbone=Qwen3-VL2026.06 | 91.6 | — | |
| CALRDDecoding Strategy=Nucleus, Backbone=Qwen3-VL2026.06 | 91.4 | — | |
| DeCoDecoding Strategy=Beam, Backbone=Qwen3-VL2026.06 | 91.2 | — | |
| DeCoDecoding Strategy=Nucleus, Backbone=Qwen3-VL2026.06 | 91.2 | — | |
| DeCoDecoding Strategy=Greedy, Backbone=Qwen3-VL2026.06 | 90.6 | — | |
| ILVADBase Model=Qwen3-VL-8B2026.05 | 89.62 | 89.8 | |
| OPERADecoding Strategy=Beam, Backbone=Qwen3-VL2026.06 | 89.3 | — | |
| ILVADBase Model=Qwen2-VL-7B2026.05 | 89.15 | 89.45 | |
| ILVADBase Model=LLaVA-NeXT-7B2026.05 | 89.08 | 88.92 | |
| GreedyBase Model=Qwen2-VL-7B2026.05 | 88.91 | 89.27 | |
| VCDDecoding Strategy=Nucleus, Backbone=Qwen3-VL2026.06 | 88.9 | — | |
| GreedyBase Model=Qwen3-VL-8B2026.05 | 88.82 | 89.19 | |
| VanillaDecoding Strategy=Nucleus, Backbone=Qwen3-VL2026.06 | 88.7 | — | |
| DoLaDecoding Strategy=Greedy, Backbone=Qwen3-VL2026.06 | 88.6 | — | |
| VanillaDecoding Strategy=Beam, Backbone=Qwen3-VL2026.06 | 88.6 | — | |
| GreedyBase Model=LLaVA-NeXT-7B2026.05 | 88.51 | 88.69 | |
| VanillaDecoding Strategy=Greedy, Backbone=Qwen3-VL2026.06 | 88.5 | — | |
| CALRDDecoding Strategy=Beam, Backbone=LLaVA-1.62026.06 | 88.3 | — | |
| CALRDDecoding Strategy=Greedy, Backbone=LLaVA-1.62026.06 | 88.2 | — | |
| ILVADBase Model=InternVL3-8B2026.05 | 87.95 | 88.31 | |
| DeCoDecoding Strategy=Beam, Backbone=LLaVA-1.62026.06 | 87.9 | — | |
| GreedyBase Model=InternVL3-8B2026.05 | 87.81 | 88.16 | |
| DeCoDecoding Strategy=Greedy, Backbone=LLaVA-1.62026.06 | 87.6 | — | |
| OPERADecoding Strategy=Beam, Backbone=LLaVA-1.62026.06 | 87.5 | — | |
| CALRDDecoding Strategy=Nucleus, Backbone=LLaVA-1.62026.06 | 87.5 | — | |
| DeCoDecoding Strategy=Nucleus, Backbone=LLaVA-1.62026.06 | 87.3 | — | |
| DoLaDecoding Strategy=Greedy, Backbone=LLaVA-1.62026.06 | 87.1 | — | |
| VanillaDecoding Strategy=Beam, Backbone=LLaVA-1.62026.06 | 87.1 | — | |
| CALRDDecoding Strategy=Beam, Backbone=LLaVA-1.52026.06 | 86.9 | — | |
| DeCoDecoding Strategy=Beam, Backbone=LLaVA-1.52026.06 | 86.7 | — | |
| AGLABase Model=LLaVA-1.5-7B2026.05 | 86.38 | 85.63 | |
| ILVADBase Model=LLaVA-1.5-7B2026.05 | 86.12 | 85.76 | |
| CALRDDecoding Strategy=Nucleus, Backbone=LLaVA-1.52026.06 | 85.9 | — | |
| VARBase Model=LLaVA-1.5-7B2026.05 | 85.87 | 84.82 | |
| VCDDecoding Strategy=Nucleus, Backbone=LLaVA-1.62026.06 | 85.6 | — | |
| ONLYBase Model=LLaVA-1.5-7B2026.05 | 85.54 | 85.13 | |
| VanillaDecoding Strategy=Nucleus, Backbone=LLaVA-1.62026.06 | 85.5 | — | |
| VHRBase Model=LLaVA-1.5-7B2026.05 | 85.45 | 84.74 | |
| BeamBase Model=LLaVA-1.5-7B2026.05 | 85.4 | 85.41 | |
| VanillaDecoding Strategy=Greedy, Backbone=LLaVA-1.62026.06 | 85.4 | — | |
| CALRDDecoding Strategy=Greedy, Backbone=InstructBLIP2026.06 | 85.4 | — | |
| OPERADecoding Strategy=Beam, Backbone=LLaVA-1.52026.06 | 85.4 | — | |
| VCDBase Model=LLaVA-1.5-7B2026.05 | 85.31 | 84.74 | |
| SPARCBase Model=LLaVA-1.5-7B2026.05 | 85.28 | 84.6 | |
| GreedyBase Model=LLaVA-1.5-7B2026.05 | 85.22 | 84.5 | |
| CODEBase Model=LLaVA-1.5-7B2026.05 | 85.06 | 84.19 | |
| DeCoDecoding Strategy=Greedy, Backbone=InstructBLIP2026.06 | 84.9 | — | |
| VanillaDecoding Strategy=Beam, Backbone=LLaVA-1.52026.06 | 84.9 | — | |
| DeCoDecoding Strategy=Beam, Backbone=InstructBLIP2026.06 | 84.9 | — | |
| OPERADecoding Strategy=Beam, Backbone=InstructBLIP2026.06 | 84.8 | — | |
| DeCoDecoding Strategy=Nucleus, Backbone=LLaVA-1.52026.06 | 84.8 | — | |
| CALRDDecoding Strategy=Beam, Backbone=InstructBLIP2026.06 | 84.7 | — | |
| CALRDDecoding Strategy=Greedy, Backbone=LLaVA-1.52026.06 | 84.5 | — | |
| VAFBase Model=LLaVA-1.5-7B2026.05 | 84.43 | 83.03 | |
| VanillaDecoding Strategy=Beam, Backbone=InstructBLIP2026.06 | 84.4 | — | |
| DeCoDecoding Strategy=Greedy, Backbone=LLaVA-1.52026.06 | 83.8 | — | |
| DoLaDecoding Strategy=Greedy, Backbone=InstructBLIP2026.06 | 83.4 | — | |
| CALRDDecoding Strategy=Nucleus, Backbone=InstructBLIP2026.06 | 83.4 | — | |
| DoLaDecoding Strategy=Greedy, Backbone=LLaVA-1.52026.06 | 83.2 | — | |
| VanillaDecoding Strategy=Nucleus, Backbone=LLaVA-1.52026.06 | 83.1 | — | |
| VCDDecoding Strategy=Nucleus, Backbone=LLaVA-1.52026.06 | 83.1 | — | |
| CALRDDecoding Strategy=Beam, Backbone=Qwen2.5-VL2026.06 | 82.6 | — | |
| CALRDDecoding Strategy=Greedy, Backbone=Qwen2.5-VL2026.06 | 82.3 | — | |
| VanillaDecoding Strategy=Greedy, Backbone=LLaVA-1.52026.06 | 82.2 | — | |
| DoLaDecoding Strategy=Greedy, Backbone=Qwen2.5-VL2026.06 | 82.1 | — | |
| CALRDDecoding Strategy=Nucleus, Backbone=Qwen2.5-VL2026.06 | 81.9 | — | |
| DeCoDecoding Strategy=Nucleus, Backbone=InstructBLIP2026.06 | 81.8 | — | |
| DeCoDecoding Strategy=Greedy, Backbone=Qwen2.5-VL2026.06 | 81.4 | — | |
| DeCoDecoding Strategy=Beam, Backbone=Qwen2.5-VL2026.06 | 81.4 | — | |
| DeCoDecoding Strategy=Nucleus, Backbone=Qwen2.5-VL2026.06 | 81.3 | — | |
| VCDDecoding Strategy=Nucleus, Backbone=Qwen2.5-VL2026.06 | 80.9 | — | |
| VanillaDecoding Strategy=Greedy, Backbone=Qwen2.5-VL2026.06 | 80.8 | — | |
| OPERADecoding Strategy=Beam, Backbone=Qwen2.5-VL2026.06 | 80.8 | — | |
| VanillaDecoding Strategy=Nucleus, Backbone=Qwen2.5-VL2026.06 | 80.8 | — | |
| VanillaDecoding Strategy=Beam, Backbone=Qwen2.5-VL2026.06 | 80.7 | — | |
| VanillaDecoding Strategy=Greedy, Backbone=InstructBLIP2026.06 | 80 | — | |
| VCDDecoding Strategy=Nucleus, Backbone=InstructBLIP2026.06 | 79.9 | — | |
| VanillaDecoding Strategy=Nucleus, Backbone=InstructBLIP2026.06 | 79.8 | — | |
| BaseModel=Qwen3-VL-4B2026.07 | — | 87.16 | |
| CDPrunerRatio=60%, Backbone=InternVL3.5-8B2026.05 | — | 87 | |
| CDPrunerRatio=40%, Backbone=InternVL3.5-8B2026.05 | — | 87.68 | |
| CDPrunerRatio=20%, Backbone=InternVL3.5-8B2026.05 | — | 86.5 | |
| CDPrunerToken Retention Ratio=60%2026.05 | — | 89.68 | |
| CDPrunerToken Retention Ratio=40%2026.05 | — | 88.74 | |
| CDPrunerToken Retention Ratio=20%2026.05 | — | 87.31 | |
| CPOModel=Qwen3-VL-4B2026.07 | — | 86.93 | |
| DivPruneRatio=60%, Backbone=InternVL3.5-8B2026.05 | — | 86.54 | |
| DivPruneRatio=40%, Backbone=InternVL3.5-8B2026.05 | — | 86.83 | |
| DivPruneRatio=20%, Backbone=InternVL3.5-8B2026.05 | — | 84.74 | |
| DivPruneToken Retention Ratio=60%2026.05 | — | 89.77 | |
| DivPruneToken Retention Ratio=40%2026.05 | — | 88.92 | |
| DivPruneToken Retention Ratio=20%2026.05 | — | 86.84 | |
| DPVR-KVTrainable=202M2026.06 | — | 84.5 | |
| DPVR-LFTrainable=202M, evaluation=3-seed average2026.06 | — | 85.5 | |
| DPVR-PCTrainable=202M, evaluation=3-seed average2026.06 | — | 85 | |
| F3ARatio=60%, Backbone=InternVL3.5-8B2026.05 | — | 87.1 | |
| F3ARatio=40%, Backbone=InternVL3.5-8B2026.05 | — | 87.54 | |
| F3ARatio=20%, Backbone=InternVL3.5-8B2026.05 | — | 87.12 | |
| F3AToken Retention Ratio=60%2026.05 | — | 90.1 |