Language Understanding on OLMoE Downstream Evaluation Suite (MMLU, ARC-E, BoolQ, HellaSwag, PIQA, SciQ)
27.6MMLUL2R (SIPS)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| L2R (SIPS)Training Tokens=10B, Scoring Geometry=SIPS2026.01 | 27.6 | 43.8 | 61.7 | 40.6 | 62.3 | 71.5 | 43.4 | |
| L2R (Cosine)Training Tokens=10B, Scoring Geometry=Cosine2026.01 | 24.9 | 42.2 | 61.1 | 42.2 | 60 | 72.7 | 42 | |
| X-MoETraining Tokens=10B, Router=Cosine scoring2026.01 | 24.6 | 46.9 | 60 | 40.6 | 64.5 | 68.8 | 42.1 | |
| L2R (Dot)Training Tokens=10B, Scoring Geometry=Dot2026.01 | 24.4 | 43.4 | 59.2 | 39.5 | 60.5 | 64.8 | 40.6 | |
| LinearTraining Tokens=10B, Router=Linear2026.01 | 22.7 | 47.7 | 60 | 38.7 | 61.1 | 66.8 | 40.5 |