Multi-image Reasoning on MIRB
63.57AccuracyQwen2.5-VL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5-VLModel Size=7B, Method Variant=Baseline2026.02 | 63.57 | — | |
| DelimScalingBackbone=Qwen2.5-VL, Model Size=7B2026.02 | 63.05 | — | |
| Qwen2VL-7BModel Size=7B2026.01 | 60.8 | — | |
| CcDPOModel Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=7B2026.01 | 60.7 | — | |
| LLaVA-NeXTSOFA=true2025.03 | 60.67 | — | |
| VISCModel Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=7B2026.01 | 60.2 | — | |
| GLM4.1VBase + CAPLParams=9B2026.03 | 60.06 | — | |
| GLM4.1VBaseParams=9B2026.03 | 59.96 | — | |
| InternVL2-Llama3-76BModel Size=76B, Backbone=Llama32026.01 | 58.2 | — | |
| DelimScalingBackbone=InternVL3, Model Size=14B2026.02 | 57.59 | — | |
| DelimScalingBackbone=Qwen2.5-VL, Model Size=3B2026.02 | 57.38 | — | |
| LLaVA-NeXTSOFA=false2025.03 | 57.15 | — | |
| Qwen2.5-VL + CAPLParams=7B2026.03 | 56.55 | — | |
| InternVL2.5 + CAPLParams=8B2026.03 | 56.55 | — | |
| Qwen2.5-VLModel Size=3B, Method Variant=Baseline2026.02 | 56.45 | — | |
| InternVL3Model Size=14B, Method Variant=Baseline2026.02 | 56.45 | — | |
| Two-stage RL (Ours)Model Category=Multi-image/Video Enhancing MLLMs, Training Strategy=Two-stage RL (DPS and annealing)2026.01 | 55.7 | — | |
| MantisSOFA=true2025.03 | 55.68 | — | |
| DelimScalingBackbone=Qwen2.5-VL, Model Size=32B2026.02 | 55.21 | — | |
| DAPO (Ours)Model Category=Multi-image/Video Enhancing MLLMs, Training Strategy=DAPO2026.01 | 55.2 | — | |
| Qwen2.5-VLModel Size=32B, Method Variant=Baseline2026.02 | 54.9 | — | |
| MantisSOFA=false2025.03 | 54.59 | — | |
| DPS (Ours)Model Category=Multi-image/Video Enhancing MLLMs, Training Strategy=DPS2026.01 | 54.4 | — | |
| InternVL2.5Params=8B2026.03 | 54.18 | — | |
| InternVL2Params=7B2026.03 | 53.15 | — | |
| GPT-4VModel Size=Proprietary2026.01 | 53.1 | — | |
| Qwen2.5-VLParams=7B2026.03 | 52.73 | — | |
| DelimScalingBackbone=InternVL3, Model Size=8B2026.02 | 52.63 | — | |
| InternVL2.5Model Category=Open-Source General MLLMs, Parameter Scale=8B2026.01 | 52.5 | — | |
| InternVL3Model Size=8B, Method Variant=Baseline2026.02 | 52.32 | — | |
| Qwen2.5-VL + FOCUSModel Size=≥7B2025.08 | 52.3 | — | |
| LLaVA-OneVisionModel Category=Open-Source General MLLMs, Parameter Scale=7B2026.01 | 51.2 | — | |
| VILASOFA=true2025.03 | 51.15 | — | |
| Ours (masked) (LLaVA-OV-7B)Model Size=7B, Backbone=LLaVA-OV, Attention Masking=True2026.01 | 51 | — | |
| Qwen2.5-VLModel Size=≥7B2025.08 | 50.53 | — | |
| VILASOFA=false2025.03 | 50.36 | — | |
| Qwen2.5-VL + FOCUSModel Size=<7B2025.08 | 50 | — | |
| Qwen2.5-VLModel Size=<7B2025.08 | 49.27 | — | |
| LLaVA-OV + FOCUSModel Size=≥7B2025.08 | 48.85 | — | |
| Idefics2SOFA=true2025.03 | 48.83 | — | |
| InternVL2-8BModel Size=8B2026.01 | 48.6 | — | |
| Qwen2.5-VLModel Category=Open-Source General MLLMs, Parameter Scale=7B2026.01 | 48.3 | — | |
| DelimScalingBackbone=LLaVA-OV, Model Size=7B2026.02 | 48.19 | — | |
| LLaVA-OVModel Size=≥7B2025.08 | 48.02 | — | |
| LLaVA-OVModel Size=7B, Method Variant=Baseline2026.02 | 47.88 | — | |
| Avg + FOCUSModel Size=≥7B2025.08 | 47.88 | — | |
| LLaVA-OVParams=7B2026.03 | 47.3 | — | |
| LLaVA-OV-7BModel Size=7B2026.01 | 47.2 | — | |
| DelimScalingBackbone=InternVL3, Model Size=2B2026.02 | 46.96 | — | |
| AvgModel Size=≥7B2025.08 | 46.8 | — | |
| VideoRFTModel Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=7B2026.01 | 46.7 | — | |
| Idefics2SOFA=false2025.03 | 46.58 | — | |
| InternVL2SOFA=true2025.03 | 46.19 | — | |
| Qwen2VL-2BModel Size=2B2026.01 | 45.9 | — | |
| InternVL3Model Size=2B, Method Variant=Baseline2026.02 | 44.38 | — | |
| InternVL2SOFA=false2025.03 | 42.66 | — | |
| InternVL3 + FOCUSModel Size=≥7B2025.08 | 42.48 | — | |
| InternVL3Model Size=≥7B2025.08 | 41.86 | — | |
| mPLUG-Owl3Model Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=8B2026.01 | 41.2 | — | |
| DelimScalingBackbone=InternVL3, Model Size=1B2026.02 | 40.25 | — | |
| LLaVA-NeXT-InterleaveModel Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=7B2026.01 | 39.3 | — | |
| InternVL3Model Size=1B, Method Variant=Baseline2026.02 | 38.49 | — | |
| LLaVA-OV-1.5BModel Size=1.5B2026.01 | 37.7 | — | |
| Idefics2Params=8B2026.03 | 37.56 | — | |
| VILA1.5Model Category=Open-Source General MLLMs, Parameter Scale=8B2026.01 | 36.5 | — | |
| Ours (LLaVA-OV-1.5B)Model Size=1.5B, Backbone=LLaVA-OV, Attention Masking=False2026.01 | 36 | — | |
| Avg + FOCUSModel Size=<7B2025.08 | 34.83 | — | |
| Mantis-Idefics2Model Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=8B2026.01 | 34.8 | — | |
| AvgModel Size=<7B2025.08 | 34.38 | — | |
| Ours (LLaVA-OV-0.5B)Model Size=0.5B, Backbone=LLaVA-OV, Attention Masking=False2026.01 | 32.8 | — | |
| InternVL3 + FOCUSModel Size=<7B2025.08 | 32.36 | — | |
| DelimScalingBackbone=LLaVA-OV, Model Size=0.5B2026.02 | 32.3 | — | |
| LLaVA-OV-0.5BModel Size=0.5B2026.01 | 31.8 | — | |
| LLaVA-OVModel Size=0.5B, Method Variant=Baseline2026.02 | 31.79 | — | |
| Qwen2VLParams=7B2026.03 | 31.68 | — | |
| InternVL3Model Size=<7B2025.08 | 31.32 | — | |
| Ours (masked) (LLaVA-OV-1.5B)Model Size=1.5B, Backbone=LLaVA-OV, Attention Masking=True2026.01 | 30.6 | — | |
| LLaVA 1.6Model Category=Open-Source General MLLMs, Parameter Scale=7B2026.01 | 29.8 | — | |
| OpenFlamingo-v2Model Category=Open-Source General MLLMs, Parameter Scale=9B2026.01 | 28.8 | — | |
| LLaVA-v1.5-7BModel Size=7B2026.01 | 28.5 | — | |
| Ours (masked) (LLaVA-OV-0.5B)Model Size=0.5B, Backbone=LLaVA-OV, Attention Masking=True2026.01 | 28.5 | — | |
| InternVL2-2BModel Size=2B2026.01 | 25 | — | |
| TW-GRPOModel Category=Multi-image/Video Enhancing MLLMs, Parameter Scale=7B2026.01 | 24.3 | — | |
| LLaVA-OVModel Size=<7B2025.08 | 22.55 | — | |
| LLaVA-OV + FOCUSModel Size=<7B2025.08 | 22.13 | — | |
| Random Guessing2025.03 | 19.13 | — | |
| GPT-4VModel Name=GPT-4V2024.12 | — | 53.1 | |
| InternVL-Chat-V1.5Model Name=InternVL-Chat-V1.52024.12 | — | 50.3 | |
| InternVL2-1BModel Name=InternVL2-1B2024.12 | — | 31.5 | |
| InternVL2-26BModel Name=InternVL2-26B2024.12 | — | 53.7 | |
| InternVL2-2BModel Name=InternVL2-2B2024.12 | — | 32.1 | |
| InternVL2-40BModel Name=InternVL2-40B2024.12 | — | 55.2 | |
| InternVL2-4BModel Name=InternVL2-4B2024.12 | — | 39.9 | |
| InternVL2-8BModel Name=InternVL2-8B2024.12 | — | 50 | |
| InternVL2-Llama3-76BModel Name=InternVL2-Llama3-76B2024.12 | — | 58.2 | |
| InternVL2.5-1BModel Name=InternVL2.5-1B2024.12 | — | 35.6 | |
| InternVL2.5-26BModel Name=InternVL2.5-26B2024.12 | — | 55.7 | |
| InternVL2.5-2BModel Name=InternVL2.5-2B2024.12 | — | 36.4 | |
| InternVL2.5-38BModel Name=InternVL2.5-38B2024.12 | — | 61.2 | |
| InternVL2.5-4BModel Name=InternVL2.5-4B2024.12 | — | 51.7 |