Language Understanding on Global-MMLU Medical 13 non-English languages (0-shot)
68.2Non-Eng AVG Original Accuracy (0-shot)QWEN3-30B-A3B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| QWEN3-30B-A3BTotal layers=482025.10 | 68.2 | — | |
| GPT-OSS-20BTotal layers=242025.10 | 63.8 | — | |
| PHI-3.5-MOETotal layers=322025.10 | 57.8 | — | |
| Routing InterventionBase Model=QWEN3-30B-A3B, Total layers=48, Target layers=(8,35), τ (Δk thresh.)=0.5, Intervention method=hard, #Experts selected=232025.10 | — | 69.1 | |
| Routing InterventionBase Model=PHI-3.5-MOE, Total layers=32, Target layers=(8,17), τ (Δk thresh.)=0.25, Intervention method=soft, λ=0.5, #Experts selected=22025.10 | — | 58.8 | |
| Routing InterventionBase Model=GPT-OSS-20B, Total layers=24, Target layers=(4,19), τ (Δk thresh.)=0.3, Intervention method=soft, λ=0.5, #Experts selected=62025.10 | — | 64.5 |