First-error detection on ProcessBench
68.7AccuracyTeacher
Evaluation Results
| Method | Links | |
|---|---|---|
| Teacherdeployment=non-deployable2026.05 | 68.7 | |
| TL-Entropy2026.05 | 46.3 | |
| LLM-Checkmechanism=attention2026.05 | 43.8 | |
| TL-Perplexity2026.05 | 43.2 | |
| Studentdeployment=deployable2026.05 | 34.4 | |
| Linear Probe2026.05 | 24.4 |