Mathematical Reasoning on MGSM 10 non-English languages (2-shot exact-match)
76.4Non-Eng AVG Accuracy (Original)QWEN3-30B-A3B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| QWEN3-30B-A3BTotal layers=482025.10 | 76.4 | — | |
| GPT-OSS-20BTotal layers=242025.10 | 68.9 | — | |
| PHI-3.5-MOETotal layers=322025.10 | 57.5 | — | |
| Routing InterventionBase Model=QWEN3-30B-A3B, Total layers=48, Target layers=(8,35), τ (Δk thresh.)=0.4, Intervention method=soft, λ=0.5, #Experts selected=222025.10 | — | 78 | |
| Routing InterventionBase Model=PHI-3.5-MOE, Total layers=32, Target layers=(8,17), τ (Δk thresh.)=0.3, Intervention method=soft, λ=0.5, #Experts selected=122025.10 | — | 58.9 | |
| Routing InterventionBase Model=GPT-OSS-20B, Total layers=24, Target layers=(4,19), τ (Δk thresh.)=0.3, Intervention method=hard, #Experts selected=92025.10 | — | 71.5 |