Hallucination Detection on RAGTruth 1.0 (test)
0Detection Delay (ARL0=50)Oracle
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OracleInput=observes labels2026.06 | 0 | — | 0 | — | 0 | — | |
| Lorden boundInput=features2026.06 | 0.9 | — | 1.3 | — | 1.2 | — | |
| ForwardGRUMethod=CUSUM2026.06 | 9 | 19 | 11.5 | 24 | 13.3 | 32 | |
| ForwardGRUMethod=threshold2026.06 | 11.2 | 26 | 13.4 | 30 | 14 | 27 | |
| HistGBMModel Type=nonlinear per-token2026.06 | 14.4 | 38 | 17.9 | 40 | 21.4 | 33 | |
| ForwardGRU-shuffledMethod=CUSUM, Input=token-shuffled2026.06 | 15.3 | 17 | 15.6 | 21 | 15.2 | 29 | |
| LogRegModel Type=linear per-token2026.06 | 30.7 | 35 | 30.8 | 31 | 35.9 | 20 | |
| Naive Gaussian CUSUMstatus=misspecified, off-scale2026.06 | 41 | — | — | — | — | — |