Mislabel detection on Weak Reference Labels Mislabels + subjective
76.2AP4PL Δℓi (proposed)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| 4PL Δℓi (proposed)Criterion=Forced-ceiling likelihood ratio2026.05 | 76.2 | 92.7 | 81 | |
| top-10 disagreement2026.05 | 73.8 | 90.8 | 78.6 | |
| XGBoost (4PL params)Model=XGBoost, Features=4PL parameters2026.05 | 70.3 | 88.6 | 74 | |
| plain 2PL (low ai)Model=2PL, Selection=low discrimination2026.05 | 70 | 78.7 | 70.9 | |
| 4PL, low diModel=4PL, Selection=low difficulty2026.05 | 69.8 | 88.3 | 73.8 | |
| plain 4PL, single-stage (low di)Model=4PL, Protocol=single-stage, Selection=low difficulty2026.05 | 69.8 | 91.5 | 74 | |
| GLAD2026.05 | 66.1 | 85.1 | 67.7 | |
| low riSelection=low response probability2026.05 | 62.6 | 87.6 | 67.2 | |
| overall disagreement2026.05 | 57.3 | 81.7 | 66.2 |