Hallucination Evaluation on HallBench
73.6AccuracyToR-DAPO
Evaluation results
| Method | Links | |
|---|---|---|
| ToR-DAPOData Size=39K (ViRL-39K), Base Model=Qwen-2.5-VL-7B2026.03 | 73.6 | |
| ToR-GRPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-7B2026.03 | 72.4 | |
| NoisyRollout-7BData Size=2.1K, Base Model=Qwen-2.5-VL-7B2026.03 | 72.1 | |
| DR-MMSearchAgent (w/ SPAI)Model Category=Open-source Models2026.04 | 72 | |
| ToR-DAPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-7B2026.03 | 71.8 | |
| ThinkLite-7B-VLData Size=11K, Base Model=Qwen-2.5-VL-7B2026.03 | 71 | |
| OpenVLThinker-7BData Size=35K+15K, Base Model=Qwen-2.5-VL-7B2026.03 | 70.8 | |
| ThinkLite-VL-7BModel Category=Open-source Models2026.04 | 70.2 | |
| VL-Rethinker-7BModel Category=Open-source Models2026.04 | 69.9 | |
| GRPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-7B2026.03 | 69.8 | |
| VLAA-Thinker-7BData Size=25K, Base Model=Qwen-2.5-VL-7B2026.03 | 68.6 | |
| MM-Eureka-Qwen-7BData Size=15K, Base Model=Qwen-2.5-VL-7B2026.03 | 68.3 | |
| DAPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-7B2026.03 | 67.9 | |
| InternVL-2.5-8BData Size=-, Base Model=Open-source Models2026.03 | 67.3 | |
| R1-OneVision-7BData Size=155K+10K, Base Model=Qwen-2.5-VL-7B2026.03 | 65.6 | |
| Qwen2.5-VL-7BModel Category=Open-source Models2026.04 | 65.2 | |
| Qwen2.5-VL-7B-InstructData Size=-, Base Model=Open-source Models2026.03 | 64.6 | |
| Ovis2Number of Parameters=2.5B2025.12 | 60.8 | |
| ToR-DAPOData Size=39K (ViRL-39K), Base Model=Qwen-2.5-VL-3B2026.03 | 60.8 | |
| OpenVLThinker-7BModel Category=Open-source Models2026.04 | 59.1 | |
| ToR-DAPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-3B2026.03 | 56.3 | |
| Claude-3.7-SonnetModel Category=Close-source Models2026.04 | 55.4 | |
| WeThink-7BModel Category=Open-source Models2026.04 | 55.1 | |
| GPT-4oModel Category=Close-source Models2026.04 | 55 | |
| DAPOData Size=2.1K (Geo3K), Base Model=Qwen-2.5-VL-3B2026.03 | 54.7 | |
| Latent DenoisingArchitecture=Qwen-2.5-VL2026.04 | 54.7 | |
| BaselineArchitecture=Qwen-2.5-VL2026.04 | 52.9 | |
| AndesVLNumber of Parameters=2.4B2025.12 | 51.8 | |
| SAIL-VL2Number of Parameters=2.7B2025.12 | 51.7 | |
| HyperVL ViTLNumber of Parameters=2.0B2025.12 | 51.5 | |
| Qwen3-VLNumber of Parameters=2.1B2025.12 | 51.4 | |
| Qwen3-VL-8BRetained Tokens=1024 Tokens (100%)2026.07 | 51.2 | |
| SAIL-VL1.5Number of Parameters=2.5B2025.12 | 49.8 | |
| Vision-R1-LlamaV-CI-11BParameters=11B2025.03 | 49.5 | |
| HyperVLNumber of Parameters=1.8B2025.12 | 49.2 | |
| Mulberry-Llama-11BParameters=11B2025.03 | 48.9 | |
| InternVL3.5Number of Parameters=2.3B2025.12 | 48.6 | |
| LLaVA-OneVision-7BData Size=-, Base Model=Open-source Models2026.03 | 48.4 | |
| LLaVA-Cot-11BParameters=11B2025.03 | 47.8 | |
| UnicornText-only=true, Pretrain Scale=1.2M2025.03 | 43.01 | |
| EADPRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 42.8 | |
| InternVL3Number of Parameters=2.1B2025.12 | 42.5 | |
| EADPRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 42.5 | |
| Qwen2-VLNumber of Parameters=2.2B2025.12 | 41.7 | |
| CDPrunerRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 41.6 | |
| CDPrunerRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 41.3 | |
| BunnyText-only=false, Pretrain Scale=1.2M2025.03 | 41.26 | |
| DivPruneRetained Tokens=512 Tokens (↓ 50.0%)2026.07 | 41 | |
| Unicorn*Text-only=true, Pretrain Scale=558K2025.03 | 40.38 | |
| Llama-3.2-11B-VParameters=11B2025.03 | 40.3 | |
| EADPRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 40.2 | |
| LLaVA-1.5Text-only=false, Pretrain Scale=558K2025.03 | 38.84 | |
| DivPruneRetained Tokens=256 Tokens (↓ 75.0%)2026.07 | 38.7 | |
| DivPruneRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 36.7 | |
| CDPrunerRetained Tokens=128 Tokens (↓ 87.5%)2026.07 | 34.7 | |
| Latent DenoisingArchitecture=LLaVA+SigLIP2026.04 | 30.7 | |
| Latent DenoisingArchitecture=LLaVA+CLIP2026.04 | 29.6 | |
| BaselineArchitecture=LLaVA+SigLIP2026.04 | 28.3 | |
| BaselineArchitecture=LLaVA+CLIP2026.04 | 27.6 |