Hallucination Detection on PM-seen vs. hallucination (N=300)
99.4AUROCOracle margin
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Oracle margin2026.05 | 99.4 | 100 | |
| Entropy (output-based)Source=output-based2026.05 | 96.7 | 0.5 | |
| Linear probe (post-hoc)Evaluation Protocol=post-hoc2026.05 | 96.4 | — | |
| MLP head (post-hoc)Evaluation Protocol=post-hoc2026.05 | 95.3 | 18 | |
| MLP head (end-to-end)Evaluation Protocol=end-to-end2026.05 | 91.3 | 0.5 |