First-error detection on PRM800K
92.9AccuracyStudent
Evaluation results
| Method | Links | |
|---|---|---|
| Studentdeployment=deployable2026.05 | 92.9 | |
| Teacherdeployment=non-deployable2026.05 | 88.4 | |
| Linear Probe2026.05 | 68.8 | |
| LLM-Checkmechanism=attention2026.05 | 49.2 | |
| TL-Entropy2026.05 | 43.8 | |
| TL-Perplexity2026.05 | 39 |