Multimodal Captioning on MAPLE-Caption (test)
66.78V ScoreMAPO
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| MAPOBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score2026.02 | 66.78 | 83.46 | 87.5 | 66.44 | 67.99 | 78.59 | 66.4 | 73.88 | |
| MAPO + adpw + curBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Adaptive weighting (adpw)=true, Static curriculum (cur)=true2026.02 | 66.43 | 82.91 | 85.57 | 66.3 | 68.48 | 78.23 | 66.38 | 73.47 | |
| MAPO + adpw + adpcurBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Adaptive weighting (adpw)=true, Dynamic curriculum (adpcur)=true2026.02 | 66.2 | 81.42 | 84.97 | 66.98 | 69.1 | 81.71 | 67.58 | 74 | |
| MAPO + adpwBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Adaptive weighting (adpw)=true2026.02 | 65.65 | 82.39 | 85.82 | 65.89 | 67.72 | 77.57 | 65.18 | 72.89 | |
| MAPO + adpcurBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Dynamic curriculum (adpcur)=true2026.02 | 65.11 | 79.89 | 86.14 | 65.12 | 66.94 | 76.26 | 64.17 | 71.95 | |
| MAPO + curBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Static curriculum (cur)=true2026.02 | 65.03 | 81.65 | 86.39 | 64.71 | 66.21 | 76.83 | 63.67 | 72.07 | |
| MUPOBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score2026.02 | 63.78 | 73.69 | 68.2 | 64.46 | 65.69 | 71.71 | 65.18 | 67.67 | |
| Zero-shotBackbone=Qwen2.5-Omni-3B, Evaluation Protocol=LLM-as-Judge score, Mode=Zero-shot2026.02 | 56.41 | 63.25 | 61.98 | 59.51 | 59.1 | 63.63 | 58.38 | 60.32 |