Step-level hallucination detection on PRM800K
99.8AUROCStudent
Evaluation Results
| Method | Links | |
|---|---|---|
| StudentDeployability=deployable2026.05 | 99.8 | |
| TeacherDeployability=non-deployable2026.05 | 98.5 | |
| Linear Probe2026.05 | 91.3 | |
| TL-Entropy2026.05 | 54.5 | |
| LLM-CheckMechanism=attention2026.05 | 48 | |
| TL-Perplexity2026.05 | 45.8 |