General VQA on MMMU (val)
73.4ScoreInternVL3.5
Evaluation results
| Method | Links | |
|---|---|---|
| InternVL3.5Openness=semi-open, Model Size=8B2025.10 | 73.4 | |
| Keye-VLOpenness=semi-open, Model Size=8B2025.10 | 71.4 | |
| Bee-8BOpenness=fully open, Model Size=8B, Training Strategy=SFT2025.10 | 66.8 | |
| Bee-8BOpenness=fully open, Model Size=8B, Training Strategy=RL2025.10 | 66.1 | |
| DreamPRMFramework Type=Training-Based, Base Model=InternVL-2.5-8B2026.03 | 61.4 | |
| Ours (CDRL + CA-TTS)Framework Type=Training-Based, Base Model=Qwen2.5-VL-7B2026.03 | 59.9 | |
| Qwen2.5-VLOpenness=semi-open, Model Size=7B2025.10 | 58.6 | |
| Majority VotingFramework Type=Training-Free, Base Model=Qwen2.5-VL-7B2026.03 | 53.2 | |
| VL-RethinkerFramework Type=Training-Based, Base Model=Qwen2.5-VL-7B2026.03 | 51.6 | |
| R1-OnevisionFramework Type=Training-Based, Base Model=Qwen2.5-VL-7B2026.03 | 51.6 | |
| DeepconfFramework Type=Training-Free, Base Model=Qwen2.5-VL-7B2026.03 | 51.5 | |
| We-ThinkFramework Type=Training-Based, Base Model=Qwen2.5-VL-7B2026.03 | 50.8 | |
| LLaVA OneVisionOpenness=fully open, Model Size=7B2025.10 | 48.8 | |
| MolmoOpenness=fully open, Model Size=7B, Training Strategy=Distilled (-D)2025.10 | 45.3 | |
| Pass@1Framework Type=Training-Free, Base Model=Qwen2.5-VL-7B2026.03 | 40.1 |