Visual Question Answering on GQA (ECE)
6.09ECEMixUp
Evaluation Results
| Method | Links | |
|---|---|---|
| MixUpVision Token Count=Full, Training=ranking loss2025.10 | 6.09 | |
| +MixUp^rlVision Token Budget=Full Vision Tokens2025.10 | 6.09 | |
| MergeMixVision Token Count=Full, Training=ranking loss2025.10 | 6.5 | |
| +MergeMix^rlVision Token Budget=Full Vision Tokens2025.10 | 6.5 | |
| CutMixVision Token Count=Full, Training=ranking loss2025.10 | 6.74 | |
| +CutMix^rlVision Token Budget=Full Vision Tokens2025.10 | 6.74 | |
| SFT VisionVision Token Count=Full2025.10 | 8.52 | |
| SFT VisionVision Token Budget=Full Vision Tokens2025.10 | 8.52 | |
| MergeMixVision Token Count=50%, Training=ranking loss2025.10 | 10.34 | |
| +MergeMix^rlVision Token Budget=50% Vision Tokens2025.10 | 10.34 | |
| CutMixVision Token Count=50%, Training=ranking loss2025.10 | 10.48 | |
| +CutMix^rlVision Token Budget=50% Vision Tokens2025.10 | 10.48 | |
| +ResizeMix^rlVision Token Budget=25% Vision Tokens2025.10 | 10.66 | |
| +CutMix^rlVision Token Budget=25% Vision Tokens2025.10 | 12.23 | |
| ResizeMixVision Token Count=Full, Training=ranking loss2025.10 | 12.53 | |
| +ResizeMix^rlVision Token Budget=Full Vision Tokens2025.10 | 12.53 | |
| +MergeMix^rlVision Token Budget=25% Vision Tokens2025.10 | 12.55 | |
| ResizeMixVision Token Count=50%, Training=ranking loss2025.10 | 12.6 | |
| +ResizeMix^rlVision Token Budget=50% Vision Tokens2025.10 | 12.6 | |
| +MixUp^rlVision Token Budget=25% Vision Tokens2025.10 | 12.97 | |
| SFT VisionVision Token Budget=25% Vision Tokens2025.10 | 13.32 | |
| MixUpVision Token Count=50%, Training=ranking loss2025.10 | 13.4 | |
| +MixUp^rlVision Token Budget=50% Vision Tokens2025.10 | 13.4 | |
| Baseline2025.10 | 14.57 | |
| BaselineVision Token Budget=Full (Implied)2025.10 | 14.57 | |
| SFT VisionVision Token Count=50%2025.10 | 18.13 | |
| SFT VisionVision Token Budget=50% Vision Tokens2025.10 | 18.13 |