Logical Reasoning on LogiQA (Acc)
60.22AccuracyInFamilySteer
Evaluation Results
| Method | Links | |
|---|---|---|
| InFamilySteerBest Layer=L2, Best λ=-1.52026.05 | 60.22 | |
| DenseSteerBest Layer=L16, Best λ=-4.02026.05 | 58.22 | |
| BaselineBest Layer=–, Best λ=–2026.05 | 44.22 | |
| Thoughts-as-PlanningAvg Edits=472026.04 | 21.6 | |
| CoTGenAvg Edits=1802026.04 | 20.1 | |
| RLCoTAvg Edits=1502026.04 | 19.8 | |
| SoftCoT2026.04 | 19.5 | |
| Manual CoT2026.04 | 18.7 |