Visual Question Answering on OKVQA (AUROC, TPRs, Correlation, ECE, AURAC)
0.761AUROCFUSE
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| FUSEBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.761 | 37.2 | 21.8 | 9.4 | 0.945 | 0.944 | 3.7 | 0.77 | |
| Mean-only u*description=Uses only the posterior mean2026.06 | 0.757 | 36.9 | 21.2 | 9.2 | 0.943 | 0.94 | 3.8 | 0.769 | |
| Quantile θ*description=Global threshold via coverage constraint2026.06 | 0.757 | 37 | 21.3 | 9 | 0.941 | 0.935 | 4 | 0.766 | |
| Global Tuningdescription=Jointly tuned parameters2026.06 | 0.752 | 36.9 | 20.9 | 9 | 0.943 | 0.934 | 4.4 | 0.763 | |
| Logistic u*description=Fitted logistic function on calibration dataset2026.06 | 0.751 | 36.6 | 21.4 | 9.1 | 0.941 | 0.939 | 3.9 | 0.767 | |
| OLS V~description=Variance fitting using least squares2026.06 | 0.751 | 37.1 | 20.8 | 9.1 | 0.942 | 0.937 | 4.2 | 0.765 | |
| UMPIREBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.75 | 36.8 | 21 | 9.1 | 0.945 | 0.934 | 4.1 | 0.761 | |
| Unweighteddescription=Unweighted semantic evidence2026.06 | 0.75 | 36.8 | 21 | 9.1 | 0.944 | 0.94 | 4 | 0.764 | |
| EigenScoreBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.741 | 34 | 19.5 | 7.5 | 0.79 | 0.842 | 19 | 0.753 | |
| Semantic EntropyBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.712 | 32.1 | 14.5 | 6.8 | 0.444 | 0.532 | 18.9 | 0.734 | |
| LN-EntropyBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.698 | 24.4 | 11.3 | 3 | 0.856 | 0.699 | 5.7 | 0.741 | |
| LoFreeCPBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.67 | 32.3 | 19 | 9.2 | 0.484 | 0.621 | 21 | 0.693 | |
| CLMBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.668 | 34 | 17.8 | 9.2 | 0.592 | 0.612 | 18.6 | 0.693 | |
| NeighborhoodBackbone=CLIP-ViT-L-336px, stochastic responses (n)=502026.06 | 0.512 | 9.5 | 3.1 | 0.8 | 0.577 | 0.58 | 50.2 | 0.629 |