Mathematical Reasoning on Minerva Math 7 subjects
39.4NT AccuracyOurs (SP4, α=0.80)
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Ours (SP4, α=0.80)Layer Selection Strategy (SP)=SP ≥ 4.0, Scaling Parameter (α)=0.80, Number of Layers=14/34, SAE-projection=false2026.05 | 39.4 | 41.4 | 67 | 37.2 | 19.6 | 69.6 | 20.5 | 5 | |
| SAE Projection†Scaling Parameter (α)=0.70, SAE rank=3, SAE size=16K, SAE-projection=true2026.05 | 33.9 | 34.6 | 62.1 | 31 | 16.8 | 63.5 | 20.3 | 0 | |
| Full LoRA merge (34L)Number of Layers=342026.05 | 30 | 33.8 | 58.1 | 29.2 | 16.3 | 61.4 | 19.8 | 0 | |
| Base model2026.05 | 29.6 | 33.3 | 61.3 | 30.5 | 16.6 | 62.9 | 20.5 | — |