Step-level error discrimination on MATH and GSM8k (test)
0.762AUROC (Step-level Error Discrimination)Fine-tuned
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Fine-tunedBackbone=Qwen3-1.7B2026.05 | 0.762 | 0.382 | 73.1 | |
| Role PromptedBackbone=gpt-4o-mini2026.05 | 0.554 | 0.161 | 16.4 | |
| Naive LLMBackbone=gpt-4o-mini2026.05 | 0.519 | 0.128 | 11 | |
| Random Guessingbaseline=true2026.05 | 0.5 | 0.118 | 11.8 |