Trajectory quality correlation analysis on Teacher-generated Reasoning Datasets Qwen-2.5-7B student
0.92Spearman Correlation (Abs)Rank-Surprisal Ratio
Evaluation results
| Method | Links | |
|---|---|---|
| Rank-Surprisal RatioMetric Category=Student-Specific2026.01 | 0.92 | |
| Rule-based QualityMetric Category=Student-Agnostic2026.01 | 0.75 | |
| Avg-RankMetric Category=Student-Specific2026.01 | 0.68 | |
| GRACEMetric Category=Student-Specific2026.01 | 0.66 | |
| Avg-SurprisalMetric Category=Student-Specific2026.01 | 0.55 | |
| Avg-Surpr_localMetric Category=Student-Specific2026.01 | 0.54 | |
| G-NormMetric Category=Student-Specific2026.01 | 0.51 | |
| LLM-judged QualityMetric Category=Student-Agnostic2026.01 | 0.46 | |
| Avg-Token LengthMetric Category=Student-Agnostic2026.01 | 0.45 | |
| Influence ScoreMetric Category=Student-Specific2026.01 | 0.32 | |
| Verified AccuracyMetric Category=Student-Agnostic2026.01 | 0.25 | |
| Teacher ParamsMetric Category=Student-Agnostic2026.01 | 0.2 | |
| Teacher PerformanceMetric Category=Student-Agnostic2026.01 | 0.13 |