Multimodal Perception and Reasoning on MME-Realworld ID
60.5Overall ScoreMGPO
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MGPOPost-training paradigm=MGPO, Training samples=21K2025.07 | 60.5 | 86.4 | 54 | 78 | 46.7 | 44 | 74 | 69 | 52.7 | 39.3 | |
| SFTPost-training paradigm=SFT, Training samples=21K2025.07 | 58.7 | 85.6 | 55.3 | 78 | 43.6 | 43.7 | 70 | 60 | 50.7 | 41 | |
| GRPOPost-training paradigm=GRPO, Training samples=21K2025.07 | 55.1 | 81.6 | 51.3 | 75 | 42.9 | 43.7 | 67 | 53 | 43.3 | 38.3 | |
| Qwen2.5-VL-7B2025.07 | 46.1 | 75.6 | 33.3 | 65 | 30.1 | 29.4 | 61 | 56 | 30.7 | 34 | |
| OpenAI GPT-4o2025.07 | 45.2 | 77.7 | 28.9 | 46.7 | 33.9 | 22.4 | 61.4 | 44.8 | 36.5 | 26.4 |